Content uploaded by Ben DeVries
Author content
All content in this area was uploaded by Ben DeVries on Oct 25, 2015
Content may be subject to copyright.
Ben DeVries
Monitoring tropical forest dynamics
using Landsat time series and
community-based data
Monitoring tropical forest dynamics
using Landsat time series and
community-based data
Ben DeVries
Thesis committee
Promotor:
Prof. Dr M. Herold
Professor of Geo-information Science and Remote Sensing
Wageningen University
Co-promotors:
Dr L. Kooistra
Assistant Professor, Laboratory of Geo-information Science and Remote Sensing
Wageningen University
Dr J. Verbesselt
Assistant Professor, Laboratory of Geo-information Science and Remote Sensing
Wageningen University
Other members:
Prof. Dr L. Poorter, Wageningen University
Dr M. Wulder, Canadian Forest Service, Victoria, Canada
Dr R. Fensholt, University of Copenhagen, Denmark
Dr I. Jonckheere, Food and Agriculture Organization of the United Nations, Rome, Italy
This research was conducted under the auspices of the C.T. de Wit Graduate School
of Production Ecology & Resource Conservation (PE&RC)
Monitoring tropical forest dynamics
using Landsat time series and
community-based data
Ben DeVries
Thesis
submitted in fulfilment of the requirements for the degree of doctor at
Wageningen University
by the authority of the Rector Magnificus
Prof. Dr A.P.J. Mol,
in the presence of the
Thesis Committee appointed by the Academic Board
to be defended in public
on Friday 16 October 2015
at 4 p.m. in the Aula.
Ben DeVries
Monitoring tropical forest dynamics using Landsat time series and community-based data
170 pages.
PhD thesis, Wageningen University, Wageningen, NL (2015)
With references, with summary in English
ISBN 978-94-6257-476-2
Acknowledgements
This PhD thesis represents four years of research, learning, discovery and collaboration.
While the contents herein are the final product of many hours of hard work, they cannot
completely reflect the input I have had from so many people along the way.
First of all, I would like to thank my promoter, Martin Herold, under whose guidance I
have developed tremendously in the past four years. Your commitment to your students
and department is clearly evident and very much appreciated.
Second, I would like to thank my co-promoters, Jan Verbesselt and Lammert Kooistra.
After countless discussions over coffee, hashing through details, debating methods and
results, I am honoured to have been able to work together with you both on this thesis.
Thank you also to Valerio Avitabile for your valuable assistance and guidance along the
way, both while in the field and in the office.
I consider myself very fortunate to have had the chance to carry out much of my research
in the field in Ethiopia. This work would not have been possible without the continued
support of NABU. From the German office, I would like to thank Svane for her strong
leadership, Dani for the very dependable support and company while in Bonga, and Bianca
for keeping this wonderful project running. I would like to give a big thanks to Bekele,
Sisay and Bety in Addis Ababa and Mesfin, Wondu, Fiseha and Wolde for the leadership
and support on the ground in Ethiopia. Finally, my work was also made possible with
the kind assistance from partners, especially EWNHS in Addis Ababa: thank you to
Ato Mengistu, Alemayeu, Tewabe and especially Tesfaye for your assistance in these past
years.
While in Ethiopia, I benefited from a very fruitful partnership between NABU and the
Kafa Zone Bureau of Agriculture. Thank you to Terefe for your continued leadership in
Kafa. I would like to give a huge thank you to the forest rangers who continue to show
an incredible dedication to their work, and have made a concrete contribution to this
research: Serkalem, Mitiku, Atnafu, Admasu H., Zintalem, Admasu A., Matiwos, Kasim,
Tadalech, Gizachew, Zinabu, Asegadech, Abdu, Bethelihem, Girma, Nasir, Mekonnen,
Israel, Bizuayehu, Mamo, Getinet, Siraj, Abera, Tesfaye, Tadesse, Mohammed, Abebe,
Wodajo, Teka, and Adinew - Yeerimba! Finally, I would like to thank Muluken for his
vi
constant support and facilitation during my five field campaigns in Kafa.
A big thanks goes out to the very vibrant Laboratory of Geo-Information Science and
Remote Sensing group at Wageningen University. Arun, Johannes R., Roberto, Niki,
Tsoefiet, Mathieu, Richard, Simon, Michael, Manos, Maria, Arend, Philip, Eskender,
Lo¨ıc, Peter, Kim, Johannes E., Sytze, Eliakim, Harm, John, Ron, Jan C., Arnold, Angela,
Brice, Nandika, Hans, Jose, Sarah, Astrid, Alvaro, Rosa, Erika, Jessica, Kathleen, Juha,
Aldo, Willy, Frans, Roland, Lucasz, Corne, Ben, Giulia, Yang, Marston, Konstantin,
Qijun, Marcio, Jalal, Laure, Valerie, Titia, Rogier, Daniel, Daniela - thanks for four great
years and keep up the excellent work! A special thanks goes to Truus and Antoinette for
your endless support and problem solving on the administrative side. Finally, I would like
to thank the MSc thesis students whom I have had the pleasure to supervise: Frehiwot,
Kalkidan, Stephan, Elias and Dereje.
I did not spend all of my time behind the desk and in the field, of course. My other great
passion - music - helped to keep me on track and well-balanced in these past years in
Wageningen. For this, I would like to thank the entire music community in Wageningen
and surroundings. An especially big thanks goes out to Caperune: Johan, Jimi, Sam,
Erikwim, Erik, Jelmer, Daniel, and the countless guest musicians who have helped to
make this project the success it was. A big thanks goes to Guus Tangelder and the Big
Band Sound of Science, and all of the other musicians and friends with whom I have had
the pleasure to meet and get to know in this small but very charismatic town.
Finally, I would like to give a big thanks and much love to my wonderful family, both
immediate and extended, near and far. Ik wil de Dykstra familie hier in Nederland
bedanken voor alles tijdens de afgelopen zeven jaar in Nederland. Ik ben zo blij dat
ik jullie beter heb leren kennen. Bovendien wil ik Max en Tineke bedanken voor hun
steun vanaf mijn allereerste dag in Wageningen. To my family in Canada: I’ve missed
you and am grateful for everything you’ve given me over all these years. First, to my
parents - Andy and Siska DeVries - thanks for your unfailing love and support in all of
my endeavours, without which I would never had made it this far. A huge thanks goes
out to my brothers and their families as well: Chris, Amanda, Joe, Elle, Aaron, Sarah,
Maya, Julian, Raine, Helena, Nolan and baby number three (who I can’t wait to meet!).
I’m honoured to call you all family, and couldn’t have asked for a better one!
Contents
Page
Acknowledgements v
Contents 1
Summary 3
Chapter 1 Introduction 5
Chapter 2 Monitoring small-scale disturbances using Landsat time series 17
Chapter 3 Tracking post-disturbance regrowth using Landsat time series 45
Chapter 4 Combining satellite data and community-based observations 73
Chapter 5 Integrating Landsat time series and community-based moni-
toring data to characterize forest changes 97
Chapter 6 Synthesis 123
References 139
List of publications 163
Short biography 167
PE&RC Training and Education Statement 169
Summary
Tropical forests cover a significant portion of the earth’s surface and provide a range of
ecosystem services, but are under increasing threat due to human activities. Deforestation
and forest degradation in the tropics are responsible for a large share of global CO2
emissions. As a result, there has been increased attention and effort invested in the
reduction of emission from deforestation and degradation and the protection of remaining
tropical forests in recent years. Methods for tropical forest monitoring are therefore vital
to track progress on these goals. Two data streams in particular have the potential to
play an important role in forest monitoring systems. First, satellite remote sensing is
recognized as a vital technology in supporting the monitoring of tropical forests, of which
the Landsat family of satellite sensors has emerged as one of the most important. Owing
to its open data policy, a large range of methods using dense Landsat time series have
been developed recently which have the potential to greatly enhance forest monitoring
in the tropics. Second, community-based monitoring is supported in many developing
countries as a way to engage forest communities and lower costs of monitoring activities.
The development of operational monitoring systems will need to consider how these data
streams can be integrated for the effective monitoring of forest dynamics.
This thesis is concerned with the monitoring of tropical forest dynamics using a combi-
nation of dense Landsat time series and community-based monitoring data. The added
value conferred by these data streams in monitoring deforestation, degradation and re-
growth in tropical forests is assessed. This goal is approached from two directions. First,
the application of econometric structural change monitoring methods to Landsat time
series is explored and the efficacy and accuracy of these methods over several tropical
forest sites is tested. Second, the integration of community-based monitoring data with
Landsat time series is explored in an operational setting. Using local expert monitoring
data, the reliability and consistency of these data against very high resolution optical
imagery are assessed. A bottom-up approach to characterize forest change in high the-
matic detail using a priori community-based observations is then developed based on
these findings.
Chapter 2 presents a robust data-driven approach to detect small-scale forest disturbances
driven by small-holder agriculture in a montane forest in southwestern Ethiopia. The
4 Summary
Breaks For Additive Season and Trend Monitoring (BFAST Monitor) method is applied
to Landsat NDVI time series using sequentially defined one-year monitoring periods. In
addition to time series breakpoints, the median magnitude of residuals (expected versus
observed observations) is used to characterize change. Overall disturbances are mapped
with producer’s and user’s accuracies of 73%. Using ordinal logistic regression (OLR)
models, the extent to which degradation and deforestation can be separately mapped is
explored. The OLR models fail to distinguish between deforestation and degradation,
however, owing to the subtle and diffuse nature of forest degradation processes.
Chapter 3 expands upon the approach presented in Chapter 2 by tracking post-disturbance
forest regrowth in a lowland tropical forest in southeastern Peru using Landsat Normalized
Difference Moisture Index (NDMI) time series. Disturbance between 1999 and 2013 are
mapped using the same sequential monitoring method as in Chapter 2. Pixels where
disturbances are detected are then monitored for follow-up regrowth using the reverse of
the method employed in Chapter 2. The time of regrowth onset is recorded based on a
comparison to defined stable history period. Disturbances are mapped with 91% accuracy,
while post-disturbance regrowth is mapped with a total accuracy of 61% for disturbances
before 2006.
Chapter 4 and 5 explore the integration of community-based forest monitoring data and
remote sensing data streams. Major advantages conferred by community-based forest dis-
turbance observations include the ability to report on drivers and other thematic details
of forest change and the ability to detect low-level forest degradation before these changes
are visible above the forest canopy. Chapter 5 builds on these findings and presents a
novel bottom-up approach to characterize forest changes using local expert disturbance
reports to calibrate and validate forest change models based on Landsat time series. Using
random forests and a selection of Landsat spectral and temporal metrics, models describ-
ing forest state variables (deforested, degraded or stable) at a given time are produced.
As local expert data are continually acquired, the ability of these models to predict forest
degradation are shown to improve.
Chapter 6 summarizes the main findings of the thesis and provides a future outlook, given
the prospect of increasing availability of satellite and in situ data for tropical forest mon-
itoring. This chapter argues that forest change methods should strive to utilize satellite
time series and ground data to their maximum potential. As “big data” emerges in the
field of earth observation, new data streams need to be accommodated in monitoring
methods. Operational forest monitoring systems that are able to integrate such diverse
data streams can support broader forest monitoring goals such as quantitative monitoring
of forest dynamics. Even with a wealth of time series based forest disturbance methods
developed recently, forest monitoring systems require locally calibrated forest change esti-
mates with higher spatial, temporal and thematic resolution to support a variety of forest
monitoring objectives.
Chapter 1
Introduction
Parts of this chapter are adapted with permission from:
DeVries, B. & Herold, M. 2013. The Science of Measuring, Reporting and Verifi-
cation (MRV). In R. Lyster, C. MacKenzie, & C. McDermott (Eds.), Law, Tropical
Forests and Carbon: The Case of REDD+, pp. 151-183. Cambridge: Cambridge Univ
Press.
6 Introduction
1.1 Monitoring Tropical Forest Change
The tropical forest has long held a place of wonder and mystery in the collective imagi-
nation of the Western world. Thomas Belt, a 19th century English naturalist, said of the
cloud forests of Nicaragua:
“Though I had dived into the recesses of these mountains again and again,
and knew that they were covered with beautiful vegetation and full of animal
life, yet the sight of that leaden-coloured barrier of cloud resting on the forest
tops, whilst the savannahs were bathed in sunshine, ever raised in my mind
vague sensations of the unknown and the unfathomable.” (Belt, 1874)
To many, the tropical forest was wild, pristine and untouchable. The fragility of these
environments was not lost on Belt, however. What he observed was a highly dynamic
system left vulnerable by human activities:
“I have been led to the conclusion that the forest formerly extended much
further towards the Pacific, and has been beaten back principally by the agency
of man.” (Belt, 1874)
Deforestation in the tropics came to the attention of the broader scientific community
nearly one century following Belt’s observations (Myers, 1979). Several studies revealed
tropical forest removal using aerial photography and field observations (Bernstein et al.,
1976), linked conversion of tropical forests to other land uses to climate change and air
pollution (Bach, 1976) and issued calls for the establishment of protected forest areas for
primates and other wildlife (Veblen, 1976; Bernstein et al., 1976).
1.1.1 Tropical forests and carbon emissions
Today, tropical forests cover 15% of the earth’s surface (FAO, 2014). In the absence of
disturbances, tropical forests have been shown to act as carbon sinks as CO2is actively
removed from the atmosphere and assimilated into biomass (Lewis et al., 2004; Phillips
et al., 1998). This carbon-absorbing function is not limited to pristine primary forests,
as secondary forests have been shown recently to be key players in the mitigation against
global climate change (Bongers et al., 2015). Due to these changes, tropical forests are
estimated to be a net source of CO2emissions since 1990, with gross emissions outweighing
atmospheric removals by 1.3 billion tonnes per year (Pan et al., 2011).
Curbing deforestation in the tropics has the potential to contribute significantly to car-
bon emission reductions globally (Gullison et al., 2007). The role of tropical forests in
climate change mitigation has been recognized through negotiated mechanisms such as the
Reducing Emissions from Deforestation and Forest Degradation (REDD+) programme,
which aims to credit developing countries who enact measures to stem emission due to
1.1 Monitoring Tropical Forest Change 7
Figure 1.1: Schematic outlining the concept of forest change employed in this thesis. Defor-
estation (DEF) is a transition from a forested (F) to non-forested (NF) state, while regrowth
(REG) is the reverse process. Degradation (DEG) is a decrease in percent tree cover (TC)
within a preserved forested state.
forest loss and conserve remaining forests (Schmidt & Scholz, 2008). The successful im-
plementation of REDD+ depends on the Measuring, Reporting and Verification (MRV)
of emissions reductions (DeVries & Herold, 2013; UNFCCC, 2009b), where activity data
(forest area change) and emission factors (carbon stock per unit forest area) are linked to
derive estimated CO2emissions (Penman et al., 2003).
MRV is concerned with three broad forest change processes: deforestation, forest degrada-
tion and forest regrowth. Making a distinction between these processes requires a robust
definition of forests (DeVries & Herold, 2013). Thresholds defined at the UNFCCC 7th
Conference of Parties (COP7), commonly known as the Marrakech Accords, are often
used to define forests as follows: “a minimum area of land of 0.05-1.0 hectares with tree
crown cover (or equivalent stocking level) of more than 10-30 percent with trees with the
potential to reach a minimum height of 2-5 metres” (UNFCCC, 2001). Deforestation thus
refers the process by which forested land is converted to non-forest lands (Figure 1.1).
Degradation, on the other hand, is defined by the IPCC as “reductions in carbon stocks
within forests remaining forests” (Penman et al., 2003). Forest regrowth can thus be
defined as the reverse of either of these two processes, where forest carbon stocks recover,
removing atmospheric CO2in the process. The definitions of forest change employed
in this thesis are shown in Figure 1.1. These change classes are defined below and are
explained in more detail in each chapter.
Deforestation is defined as a transition from a forested stated to a non-forested
state, where “forest” is defined based on a tree cover threshold within one Landsat
pixel.
Degradation is defined as a reduction in tree cover where the end result is still
defined as a forest. This definition is analogous to that recommended by the
IPCC (Penman et al., 2003), except that the definition used here only considers
the forest canopy, rather than height and area.
Regrowth is defined as a transition from a non-forest state to a forested state.
8 Introduction
In reality, the regrowth forest canopy and structure differ greatly from that of the
original intact forest in terms of carbon stocks and other variables. The definition
used here is simplified for the purpose of detecting changes at the Landsat pixel
scale.
1.1.2 Understanding tropical forest dynamics
Forest monitoring extends beyond the measurement of forest cover loss and carbon emis-
sions and removals. Changes in forest cover and structure have been shown to have an
effect on the resilience of forest systems, bringing them closer to critical forest/non-forest
transitions as tree cover is reduced (Hirota et al., 2011). Such non-linear transitions be-
tween forest and non-forest states are in line with modern thinking around ecosystem
dynamics (Holling, 1973; van de Leemput et al., 2015). A detailed understanding of for-
est dynamics is necessary to unravel the complex interactions between forests, climate
and human systems (Bonan, 2008). Such an understanding requires information on forest
status with high spatial, temporal and thematic detail. While conventional plot-based
forest research has the most potential for addressing the thematic dimension, key barriers
are faced in the spatial and temporal domains (Salk et al., 2013).
1.2 Monitoring Forests from Space
Remote sensing technologies, including aerial and space-borne sensors, have become
prominent in operational forest research and monitoring in the past few decades (De
Sy et al., 2012) and are poised to fill the spatio-temporal gaps present in plot-based forest
research. At present there are numerous satellite constellations monitoring the earth’s
surface using a variety of wavelength ranges, including optical and radar sensors. Optical
sensors detect reflected radiation from the earth’s surface at wavelengths in the visible,
Near-Infrared (NIR), Short-Wave Infrared (SWIR) and Thermal Infrared (TIR) ranges.
At these ranges, spectral characteristics of the detected radiation can reveal a great deal
about the physical characteristics of the earth’s surface, including land cover and vege-
tation cover. Optical sensors suffer several drawbacks, however. First, most space-borne
optical sensors are passive, meaning they can only collect data during the day when
the sun is illuminating the earth. Second, optical radiation is not able to penetrate cloud
cover, a problem which plagues many areas within the humid tropics (Asner, 2001). Radar
sensors can overcome some of the shortcomings of optical sensors. For example, as active
sensors (meaning they act as the source and receiver of electromagnetic radiation at the
same time), they do not rely on reflected solar energy and can make observations at day
or night (Thiel et al., 2006). Additionally, radiation at wavelengths in the radar range
is able to penetrate clouds (De Sy et al., 2012), thus providing an alternate data source
1.2 Monitoring Forests from Space 9
for areas with high cloud cover. Future applications of remote-sensing science to forest
monitoring will likely include the integration of radar and optical remote-sensing based
methods (Reiche et al., 2013, 2015a,b). Although each of these types of sensors provides
unique advantages to the field of forest and land cover change monitoring, this thesis
focuses on the use of optical sensors, as they currently provide the most widely accessible
data and associated methods.
1.2.1 Spectral Response to Forest Changes
As the bulk of the sun’s optical radiation does not penetrate the forest canopy, monitoring
forest changes using passive optical sensors is mostly concerned with changes to the forest
canopy. A simplified illustration of some of the forest change processes studied in this the-
sis is shown in Figure 1.2. Here, reflectance over a single pixel (bottom row) is shown for
forest stands under different disturbance conditions (top row). In this illustration, a hy-
pothetical index is shown, with darker values related to forest canopies and lighter values
related to soil reflectance. Reflected radiation from intact forests (Figure 1.2A) describes
vegetation characteristics with no background influence of soil or understorey vegetation.
Forest degradation, due to selective logging for example, often results in canopy gaps,
whereby a fraction of the pixel contains background soil reflectance from the forest floor,
with the remaining fraction containing forest canopy reflectance (Figure 1.2B). Spectral
un-mixing algorithms are commonly used to disentangle these contributing reflectance
profiles towards degradation mapping (Souza et al., 2005). Complete removal of a forest
stand within a pixel results in a completely altered reflectance profile, the result be-
ing dominated by bare soil or other vegetation reflectance characteristics (Figure 1.2C).
Finally, regrowing forests following a stand-replacing disturbance can result in canopy
closure by woody successional species within several years (Finegan, 1996; Howorth &
Pendry, 2006). Research using optical Landsat data has suggested that reflectance in the
SWIR domain may be used to distinguish young successional forest canopies from intact
forest canopies, since primary and secondary forest canopies have different moisture con-
tents, a physical parameter to which these wavelength ranges are sensitive (Fiorella &
Ripple, 1993; Gao, 1996; Wilson & Sader, 2002; Jin & Sader, 2005).
1.2.2 The Landsat legacy
The launch of the first Landsat sensor in 1972 marked the beginning of the longest ever
continuous earth observation record (Roy et al., 2014; Williams et al., 2006). Since that
time, a suite of Landsat-bound sensors have facilitated the collection of terrestrial imagery
over the globe. The Multi-Spectral Scanner (MSS) sensors capture reflected radiation from
the visible and near infra-red (NIR) wavelengths. The Thematic Mapper (TM) on board
Landsat-4 and Landsat-5 and the Enhanced Thematic Mapper Plus (ETM+) on board
10 Introduction
Figure 1.2: Illustration of canopy reflectance as detected by passive optical sensors. Intact
forests (A), degraded forests with canopy gaps (B), cleared forests (C) and regrowing forests
(D) differ in canopy refectance due to varying degrees of background soil reflectance, under-
storey vegetation, canopy structure and moisture content. The bottom row illustrates the
impact of these reflectance profiles on a hypothetical vegetation index over a 30m Landsat
pixel.
Landsat-7 have been responsible for a spectrally richer dataset, with the capabilities of
recording radiation at a much wider range of wavelengths. The Landsat programme has
not been without its disappointments, however. First, the Enhanced Thematic Mapper
(ETM) on board Landsat 6 failed to reach orbit, a loss which was partially compensated
by the unanticipated lifespan of the Landsat-5 TM sensor. Second, the scan-line corrector
(SLC) on board the ETM+ sensor failed in 2003, resulting in a loss of approximately 22%
of data from each scene acquired thereafter (Maxwell et al., 2007). ETM+ data have
nevertheless remained an important component of the Landsat archive, especially in light
of the end of the TM sensor’s life in 2011. The eighth Landsat mission, featuring the
Operational Land Imager (OLI), was launched in 2013 (Irons et al., 2012), once again
filling the gaps left by the ETM+ SLC-off errors and the decommissioned TM sensor.
OLI expands on the spectrum of wavelengths offered by TM and ETM+, including extra
bands for enhanced atmospheric correction and cloud masking (Irons et al., 2012).
In recognition of the importance of the Landsat archive, recent efforts have been placed
in consolidating (Loveland & Dwyer, 2012), pre-processing (Masek et al., 2006b; Vermote
et al., 1997) and publicly releasing (Wulder et al., 2012) Landsat data across the globe.
This effort has led to a surge in high temporal resolution change detection methods, not
least of which have been designed expressly for forest monitoring purposes (Banskota
1.3 Structural Change Monitoring 11
et al., 2014). Landsat-based multi-temporal methods have taken their cues from time
series approaches developed with coarse resolution datasets with frequent return times
and higher-level products available (Cihlar et al., 1997; Roerink et al., 2000; Cihlar et al.,
2004; Roerink et al., 2003). More recently, these types of approaches have been developed
using Landsat Time Series (LTS), either on annual composites (Kennedy et al., 2010;
Huang et al., 2010), or by ingesting all available observations in the archive to derive
higher level thematic products (Broich et al., 2011; Potapov et al., 2012; Zhu et al.,
2012b; Zhu & Woodcock, 2014). A more detailed background on methods developed for
LTS is given in Chapters 2 and 3.
1.2.3 Operational forest monitoring systems
A number of examples of operational forest monitoring systems designed to support
REDD+ MRV or other forest monitoring objectives have emerged as a result in advances
in time series based change detection methods. The Brazilian Space Agency (INPE),
for example, has developed systems for reporting forest cover changes at annual time
scales (INPE, 2014b) and on a near real-time basis (INPE, 2014a). The Global For-
est Watch initiative (World Resources Institute, 2014) provides an open access platform
for the visualization and analysis of LTS-based forest cover change estimated at unprece-
dented spatial scale and resolution (Hansen et al., 2013). These forest monitoring systems
represent significant and important steps towards bringing remote sensing based forest
monitoring methods into public use. The interpretations, conclusions and applications
that can be drawn from such large area datasets remain controversial, however (Hansen
et al., 2014; Tropek et al., 2013). Describing forest change in greater thematic detail using
the full spectral and temporal arsenal of Landsat and similar datasets therefore represents
a key research gap.
1.3 Structural Change Monitoring
Many of the methods applied to satellite time series data for terrestrial monitoring have
their roots in different disciplines. For example, methods such as dynamic time warp-
ing, originally developed for voice recognition algorithms, and statistical control charts,
developed as a multivariate process monitoring method (Reynolds & Cho, 2011), have
found their place in the satellite time series literature to map land cover changes (Pe-
titjean et al., 2012; Brooks et al., 2014). Structural change monitoring encompasses a
series of methods used to track breaks in a time series that were originally developed for
application to econometric models (Bai, 1997; Bai & Perron, 2003; Zeileis et al., 2005).
Structural change monitoring is based upon hypothesis tests regarding stability of a time
series compared to a previously established stable period. Under the null hypothesis,
12 Introduction
data being monitored do not significantly deviate from expected patters established in a
history period. Rejection of this null hypothesis is thus interpreted as a significant break
(e.g. based on a 95% probability threshold) in the time series (Zeileis et al., 2005).
Rather than testing the null hypothesis on singular observations within the time series,
structural change monitoring employs cumulative or temporal neighbourhood statistics,
thus preserving the structural characteristics of the time series (Chu et al., 1992; Zeileis
et al., 2005). One such measure is based on moving sums of residuals (MOSUM), where
residuals are calculated as the difference between expected values and actual observations
in a monitoring period. The MOSUM (MOt) corresponding to an observation yat time
tis computed as follows:
MOt=
t
X
i=t−h+1
yi−ˆyi
ˆσ√n(1.1)
In this form, a bandwidth (h) is typically set as a fraction of the number of observations
(n) in the time series such that MOtis the “normalized” sum of all residuals within
the window preceding (and including) observation yt. This sum is normalized by the
estimated deviance from the stable history period (ˆσ), implying that with a noisy time
series (i.e. with high variance throughout), there is less chance that the null hypothesis
will be rejected due to noise alone. In other words, a statistically significant and persistent
break is required to trigger a breakpoint (Zeileis et al., 2005; Verbesselt et al., 2010, 2012).
Structural change monitoring thus provides a data-driven method for detecting change,
whereby the threshold above which change is detected is not determined externally (e.g.
by a user), but depends on the local structure of the time series residuals (the numerator
of Equation 1.1) and the level of noise expected in the time series (the denominator of
Equation 1.1).
1.3.1 Applying the theory: BFAST
Structural change monitoring based on the MOSUM has been previously applied to coarse
resolution satellite time series in the form of the Breaks For Additive Season and Trend
(BFAST) family of algorithms (Verbesselt et al., 2010, 2012). These algorithms have been
designed either to divide a time series into a number of stable segments based on a number
of breaks (Verbesselt et al., 2010), to select the single most important break in a time
series, if any (de Jong et al., 2013), or to monitor incoming observations in a time series
against a defined stable history period (Verbesselt et al., 2012). These approaches rely on
the additive fitting of a season-trend curve to the data as follows:
yt=Tt+St+t(1.2)
where Ttis the trend component, Stis the seasonal component and tis the remainder
(related to ˆσin Equation 1.1). By combining Equations 1.2 and 1.1, the BFAST-family
1.4 Community-Based Forest Monitoring 13
of algorithm allow for robust monitoring of time series changes above and beyond what
might be expected due to seasonality, monotonic trend or random noise. Few studies have
applied structural change monitoring methods to irregular time series such as LTS (Reiche
et al., 2015a; Dutrieux et al., 2015), and whether such a data-driven approach can work
in an operational forest monitoring setting is unknown.
1.4 Community-Based Forest Monitoring
Ground-based forest monitoring approaches have developed in parallel with, and in many
cases in synergy with, satellite-based technologies. One such data source which is es-
pecially relevant for tropical countries engaging in REDD+ originates from Community-
Based Monitoring (CBM) of forest. CBM data are a product of community forest manage-
ment systems, which have been promoted and supported throughout the tropics (Pratihast
et al., 2013; Bowler et al., 2012). CBM has been promoted in REDD+ MRV systems for a
number of reasons. First, involvement of local people in REDD+ monitoring activities is
seen as a way to enhance the acceptance, sustainability and equity of REDD+ projects on
the ground (Palmer Fry, 2011). Second, where CBM is effectively implemented, monitor-
ing costs have been shown to dramatically decrease, giving added incentives to countries
to invest in REDD+ MRV systems (Pratihast et al., 2012).
Aside from the obvious social and practical benefits to including communities in forest
monitoring activities, CBM has the potential to complement existing forest monitoring
technologies, such as satellite-based monitoring, towards enhanced monitoring systems.
In this sense, communities can be viewed as “human sensors” who measure and monitor
important forest-related variables in situ. This concept is analogous to volunteered geo-
information (VGI) or citizen science data, where volunteers or lay people provide geo-
localized qualitative or quantitative data which are used as reference data for various
land monitoring applications (Goodchild, 2007). CBM has been shown to be useful and
cost-effective in measuring forest carbon stocks at the plot level (Pratihast et al., 2012;
Brofeldt et al., 2014) and has been put forward as a key component in national-level forest
monitoring systems in support of REDD+ MRV (Pratihast et al., 2013).
CBM data have the potential to support satellite-based forest change analysis in two key
ways. First, these data could play a vital role in validation of change estimates. Validating
change maps is notoriously difficult due to the simple fact that we cannot travel back in
time to verify if and when a change occurred. Validation thus often relies on other imagery
(such as very high resolution optical imagery), which is difficult to access in many cases.
Other studies have used human interpreters and the same LTS source data that were used
to derive the change maps to validate changes (Cohen et al., 2010), which has been shown
to be effective in the absence of other reference data (Kennedy et al., 2010). CBM data
could therefore help to fill this gap by tapping into the knowledge of local people, thereby
14 Introduction
supplying a bottom-up data stream for forest change monitoring.
The second potential value added of CBM data to satellite-based change estimates is
the addition of thematic details to change estimates. Human sensors have the distinct
advantage of being able to document forest changes underneath the canopy independently
of satellite-based alerts. This capability is a key advantage in forest areas where low-level
degradation is driven by fuelwood collection in the understorey, for example (Herold et al.,
2011; Dresen et al., 2014), and could serve as an early warning for degradation before
changes to the forest canopy become visible. Additionally, this “bottom-up” monitoring
of forest changes could result in additional training data to gage the extent to which subtle
forest canopy changes can be detected from space.
While the idea of CBM is gaining considerable traction in the frame of REDD+ and other
similar forest-related schemes, the potential of CBM to support forest change monitoring
from the ground remains largely unexplored. Several research questions need to be ad-
dressed before CBM can be integrated with satellite data in an operational setting. First,
the question of quality and reliability of CBM-based forest change observations needs to
be addressed. Differences in forest change interpretations between local communities and
professionals, or inadequate technical support (e.g. GPS devices for geo-location) may
preclude the effective integration of CBM data with satellite-based observations. Second,
the question regarding the level at which integration of CBM and satellite data should
be integrated remains. As noted above, this integration could occur at the calibration or
validation stages of forest change monitoring. Addressing these research gaps represents
a significant step towards operational integration of CBM and satellite-data in a forest
monitoring system.
1.5 Problem Statement and Research Objectives
With the rich Landsat archive stretching back to the 1970’s and the ever-expanding
prospects of continuous satellite-borne data streams in the future, time series based meth-
ods such as those developed for Landsat data will become increasingly important. While
forest disturbance monitoring using LTS has been operationalized in many cases, most
of the methods that have been developed still under-utilize the time series or rely on
user-defined or otherwise arbitrary thresholds. There is a need for more robust, data-
driven methods that can detect forest changes in high temporal, spatial and thematic
resolution.
As described above, CBM has seen similarly significant advances in recent years. Evolving
technologies are opening the doors to enhanced in situ and community-based monitoring
of forest resources. CBM has the potential to help address some of the gap in LTS-based
forest monitoring identified above. Despite a growing body of research into the topic of
1.6 Thesis Overview and Study Sites 15
Figure 1.3: Flowchart for the chapters in of this thesis in relation to each of the research
questions (RQ).
CBM for forest monitoring, however, very little research has been done to address the
question as to how such a data stream can complement satellite-based time series for
operational forest monitoring.
The research described in this thesis seeks to address these two research gaps by exploring
two potential bridges: (1) that between robust structural change monitoring methods and
satellite-based time series; and (2) that between satellite-based time series and CBM data
streams. To this end, I pose the following research questions:
1. To what extent can we track small-scale forest disturbances in complex forest land-
scapes using dense LTS?
2. Can structural change monitoring be used with LTS to monitor post-disturbance
regrowth in tropical forests?
3. How can community-based monitoring data and Landsat time series be integrated
to enhance forest monitoring?
1.6 Thesis Overview and Study Sites
This thesis includes six chapters, including this Introduction Chapter. The outline of
subsequent chapters is presented in Figure 1.3.
Chapter 2 describes a robust approach for monitoring small-scale forest disturbances in a
forest-agriculture matrix landscape in southern Ethiopia. Structural change monitoring
methods are applied iteratively on Landsat Time Series data and a change magnitude
16 Introduction
method is derived statistically to map small-scale agriculture-driven deforestation and
degradation.
Chapter 3 extends on the methodology developed in Chapter 2 and presents data-driven
method for detecting post-disturbance forest regrowth using Landsat Time Series. By
applying the method over a lowland tropical forest system in southern Peru, advantages
and limitations to using Landsat Time Series for post-disturbance regrowth monitoring are
shown, and a framework for continuous monitoring of forest dynamics is proposed.
Chapter 4 presents an assessment of consistency between community-based monitoring
data, professional forest monitoring data and very high resolution remote sensing data
for forest change monitoring. A dataset of local expert observations is compared with
professional forest observations and very high resolution satellite imagery over a site in
southern Ethiopia.
Chapter 5 presents an integrated forest monitoring approach using Landsat time series and
community-based forest monitoring data in southern Ethiopia. Using a suite of spectral-
temporal metrics derived from dense Landsat time series combined with a continuous
stream of local expert monitoring data in random forest models, local expert data are
shown to improve estimates of deforestation and forest degradation over time.
Chapter 6 discusses the results of the thesis and addresses the research questions presented
above. The implications of the results on integrated forest monitoring are discussed and an
outlook regarding the future of satellite-based and in situ monitoring of forest dynamics
is presented.
I conducted my research through two case studies in the tropics. First, the UNESCO Kafa
Biosphere Reserve in southwest Ethiopia is featured in Chapters 2, 4 and 5. The moist
forests in this area represent some of the last remaining Afro-montane forests (Schmitt
et al., 2010a), and harbour valuable Coffea arabica genetic resources (Aerts et al., 2015;
Hein & Gatzweiler, 2006). The second study area, on which Chapter 3 is based, is found
in Madre de Dios, southeastern Peru. This lowland forest has recently been opened up
to rapid developments, including for pasture, croplands and gold mining (Nepstad &
Carvalho, 2001; Shepard et al., 2010; Asner et al., 2013; Alvarez-Berr´ıos & Mitchell Aide,
2015). Specific characteristics of the study areas are explained further in their respective
chapters.
Chapter 2
Monitoring small-scale disturbances
using Landsat time series
This chapter is based on:
DeVries, B., Verbesselt, J., Kooistra L. & Herold, M. 2015. Robust monitoring of
small-scale forest disturbances in a tropical montane forest using Landsat time series.
Remote Sensing of Environment, 161:107-121.
18 Monitoring small-scale disturbances using LTS
Abstract
Remote sensing data play an important role in the monitoring of forest changes. Methods
are needed to provide objective estimates of forest loss to support monitoring efforts at
various scales, and with increasing public availability of remote sensing data, accurate de-
forestation measurements at high temporal resolution are becoming more realistic. While
several time series based methods have recently been described in the literature, there
are few studies focusing on tropical forest areas, where low data availability and complex
change processes present challenges to forest disturbance monitoring. Here, we present
a robust data-driven method to track tropical deforestation and degradation based on
Landsat time series data. Based on the previously reported Breaks For Additive Season
and Trend Monitor (BFAST Monitor) method (Verbesselt et al., 2012), we show that
BFAST Monitor, when applied to Landsat NDVI time series data using sequentially de-
fined monitoring periods, can be used to track small-scale forest disturbances annually in
an Afromontane forest system in southern Ethiopia. Using an ordinal logistic regression
(OLR) approach, change magnitude, calculated based on differences between observed and
expected values in a monitoring period, was found to be an essential predictor variable
for disturbances. After applying a NDVI change magnitude threshold of -0.065, overall
accuracy was estimated to be 78%, and both producer’s and user’s accuracy of the distur-
bance class were estimated to be 73%. The method and results presented here are relevant
to tropical countries engaged in REDD+ for whom data availability and complex forest
change dynamics limit the ability to reliably track forest disturbances over time.
2.1 Introduction 19
2.1 Introduction
With deforestation in the tropics accounting for upwards of 20% of global CO2emis-
sions (Gullison et al., 2007), mitigation efforts against global climate change must include
considerations to reduce tropical deforestation and forest degradation. To this end, in-
ternational climate negotiations include the development of a mechanism aimed at the
“Reduction of Emissions from deforestation and degradation and considerations for con-
servation, enhancement of carbon stocks”, commonly known as REDD+. For a results-
based mechanism such as REDD+ to be successful, countries are required to establish
robust Measuring, Reporting and Verification (MRV) systems with which to report forest
changes and impact of REDD+ activities.
A key component of a REDD+ MRV is the assessment of activity data - the area of for-
est undergoing change processes, including deforestation, forest degradation, and forest
regrowth (Penman et al., 2003). To support REDD+ MRV and other efforts to conserve
tropical forest resources, participating countries need to establish robust forest monitor-
ing systems to track activity data at regular time-frames (Holmgren & Marklund, 2007).
Remote sensing based approaches play a key role in forest monitoring, as they provide
the best opportunity for mapping forest area change over large areas (Herold & Johns,
2007; De Sy et al., 2012; DeVries & Herold, 2013; Sanz-Sanchez et al., 2013). To date,
only few remote sensing based forest monitoring systems exist in tropical countries, the
most advanced of which are the PRODES and DETER systems of the Brazilian Space
Agency (INPE), used for annual deforestation mapping and near real-time deforesta-
tion monitoring, respectively (INPE, 2014a,b). Considerable advancements in monitoring
capacities are needed for other tropical countries to establish similar forest monitoring
systems (Romijn et al., 2012).
To track forest change over time, most change detection methods rely on the selection of
imagery from key points in time, which necessitates the selection of appropriate imagery
from the archive from which to derive change information. These bi-temporal change
detection methods range from simple image differencing methods (Coppin et al., 2004) to
statistically-based methods such as the Multivariate Alteration Detection method (Nielsen
et al., 1998). An important constraint in the selection of imagery for such change detection
methods is the loss of data due to a number of contaminations or errors. First, where
bi-temporal change detection methods require that the source imagery be cloud-free for a
gap-free change product, cloud cover presents a key constraint (Ju & Roy, 2008), especially
in the tropics where cloud cover is frequently high (Mitchard et al., 2011). Second, other
sensor-specific sources of data loss can present significant constraints to the selection
of imagery for detecting change. Notably, the scan-line corrector (SLC) on board the
Landsat 7 Enhanced Thematic Mapper (ETM+) failed in March 2003, resulting in the
loss of approximately 22% of data from each scene (Zhang et al., 2007). While methods
20 Monitoring small-scale disturbances using LTS
exist to fill these gaps with data derived from other scenes or even other sensors (Zhu
et al., 2012a; Chen et al., 2011), introduction of extraneous data (e.g. from other images)
into the data processing chain can introduce additional errors into the processing chain
(B´edard et al., 2008; Alexandridis et al., 2013). Introducing gap-filling or other data
fusion methods into the preprocessing chain for bi-temporal change detection approaches
can also introduce uncertainties related to the actual acquisition date of the source data,
which can have implications on quantitative estimates of forest change (Pelletier et al.,
2011).
Another potential drawback of using a bi-temporal change detection approach relates to
the dynamic behaviour of vegetation over time. Basing change estimates on differencing
between images at only two points in time risks interpreting natural phenological change
as actual land cover change (Verbesselt et al., 2010; Zhu et al., 2012b). This problem is
especially pronounced in tropical regions, where frequent cloud cover can severely limit the
choice of imagery available per year, sometimes necessitating the use of non-anniversary
imagery in change detection studies. Confusion between forest and non-forest spectral
signatures can arise as a result of imagery from different seasons, which can lead to
increased errors in the change classification result (Coppin et al., 2004).
With the opening of the U.S. Geological Service (USGS) Landsat data archive, large
amounts of medium-resolution optical earth observation data have been made freely avail-
able to the public, which combined with continued advances in the field of cloud computing
for geo-spatial data (Evangelidis et al., 2014; Lee & Kang, 2013) has allowed for high tem-
poral resolution forest change monitoring at unprecedented spatial scales (Hansen et al.,
2013). Similar developments in multi-temporal satellite image analysis have been previ-
ously realized in the case of coarse resolution datasets, including AVHRR (Cihlar et al.,
1997, 2004; Pinzon & Tucker, 2014; Tucker et al., 2005) and MODIS (Roerink et al., 2000,
2003; Verbesselt et al., 2010; de Jong et al., 2013) time series data, based on their high re-
turn rates and rich historical archives. A number of temporal trajectory methods based on
Landsat time series data have been developed in recent years to make more extensive use
of the temporal domain. Some methods construct regular (e.g. annual) image composites
to understand disturbance-recovery dynamics (Kennedy et al., 2010; Huang et al., 2010),
while others use all available data to allow for considerations of more complex change dy-
namics, such as phenology (Zhu et al., 2012b) and transient forest changes (Broich et al.,
2011). Despite the number of temporal trajectory change detection approaches recently
published in the literature, many of these methods have been developed in temperate
forests with relatively high data availability (Zhu et al., 2012b; Zhu & Woodcock, 2014;
Kennedy et al., 2010; Huang et al., 2010). There are relatively few studies demonstrating
these methods in tropical areas with lower data availability due to persistent cloud cover
(Mitchard et al., 2011; Duveiller et al., 2008; Ernst et al., 2013) or excessive gaps in the
Landsat archive (Broich et al., 2011).
2.2 Study Area 21
While the monitoring of tropical deforestation at large spatial scales has been well docu-
mented and is largely operational (Achard et al., 2010), methods able to track small-scale
deforestation at high temporal resolution are currently lacking. Small-scale deforesta-
tion driven by small-holder subsistence agriculture is a prominent forest change process
found in sub-Saharan African countries (Fisher, 2010; Potapov et al., 2012; Joseph et al.,
2013). Monitoring these small changes is essential for such countries to play a role in
climate change mitigation and to implement forest protection measures. Much of the
research done on deforestation in tropical Africa has been undertaken in central Africa,
where small-scale forest changes have been mapped using multi-temporal segmentation
(Duveiller et al., 2008; Ernst et al., 2013), classification of annual Landsat time series
(Hirschmugl et al., 2014), or analysis of all available Landsat ETM+ data (Potapov et al.,
2012). Persistent small-scale changes in these landscapes (usually related to small-holder
agricultural expansion) are a major constraint to the accurate mapping and accounting
of deforestation (Tyukavina et al., 2013).
In this paper, we describe a robust and novel approach to monitoring forest disturbance
in the tropics using Landsat time series data using the Breaks For Additive Season and
Trend (BFAST) Monitor method (Verbesselt et al., 2012). Recent work has been done
to demonstrate this algorithm on Landsat times series for tropical forest monitoring us-
ing fused Landsat-SAR time series (Reiche et al., 2015a) or Landsat-MODIS time series
(Dutrieux et al., 2015). The goal of this study was to investigate the suitability of the
BFAST Monitor method to detect forest disturbances using Landsat time series data over
a tropical montane forest system in southwestern Ethiopia. To this end, we addressed two
objectives: (i) to test the method in an area with lower data density typical of tropical
montane forest systems experiencing regular cloud cover; and (ii) to develop an approach
to track small-scale forest disturbances characteristic of changes driven by small-holder
agriculture expansion in the tropics. The monitoring approach demonstrated in this study
and described in this paper can serve a number of purposes, including acting as a key
component in REDD+ monitoring systems (Sanz-Sanchez et al., 2013).
2.2 Study Area
2.2.1 Geographic and Biophysical Characteristics
This study was carried out in the UNESCO Kafa Biosphere Reserve (http://www.kafa-
biosphere.org), located in the Afromontane forests in Southern Nations Nationalities and
People’s Region (SNNPR) state of southern Ethiopia. Due to availability of very high
resolution (VHR) reference imagery, we focused our research on a subset of the Biosphere
Reserve, bound by 7.22◦E to 7.84◦E and 35.59◦N to 37.17◦N (Figure 2.1). The Biosphere is
comprised of three zones related to forest management: core (protected forest area), buffer
22 Monitoring small-scale disturbances using LTS
Figure 2.1: Overview of the study area, located in the UNESCO Kafa Biosphere Reserve in
southwestern Ethiopia. Core, buffer and transition forest areas are indicated in the bottom
inset. The footprint of the SPOT5 time series from K-J coordinates 133-335 is shown as a
black rectangle. White outlines indicate core forest protection areas. The base image is a
band 5-4-3 composite of an ETM+ image (WRS-2 p170r55) from 2001-02-05.
(forest with mixed land use) and transition (agriculture landscape with forest patches).
The area is characterized by vegetated mountains ranging from an altitude of 1400m
to 3000m. Approximately half of the area is covered with fragmented moist broadleaf
evergreen forests, and the rest of the area is characterized by patchy cropland-forest
matrix landscapes. Small-holder agriculture is the major driver of forest loss with coffee
being a major crop for both small-holder farmers and investors alike (Schmitt et al.,
2010b).
The Kafa Biosphere Reserve experiences a humid tropical climate, with a mean annual
temperature in the region of approximately 19◦C (Schmitt et al., 2010a). According
to daily rainfall data acquired from the Tropical Rainfall Monitoring Mission (TRMM),
average annual rainfall between 1998 and 2013 is approximately 1700mm, which is com-
parable to estimates based on local climate stations (Schmitt et al., 2010a). Rainfall is
distributed in a unimodal pattern throughout the year, with one dry and one wet season
per year.
2.3 Data and Methods 23
2.2.2 Drivers of Deforestation in Southern Ethiopia
Despite the relatively low forest cover, deforestation and forest degradation in Ethiopia
is a pressing issue that has implications that reach beyond climate change, threatening
endemic biodiversity and valuable genetic resources (Teketay, 1997; Hein & Gatzweiler,
2006). Land use change studies in several sites of Ethiopia have revealed high rates of
natural forest cover loss since the 1950’s in Southern Ethiopia (Assefa & Bork, 2014;
Getahun et al., 2013; Tadesse et al., 2014b), the Blue Nile watershed (Gebrehiwot et al.,
2013) and the Central Rift Valley (Garedew et al., 2009). These studies attributed high
rates of deforestation and forest degradation to agricultural expansion, decrease in crop
yields and rising fuelwood demand stemming from an increasing population. Tadesse
et al. (2014b) partially attributed forest loss in the Kafa and Sheka zones of Southwestern
Ethiopia to land redistribution and resettlement programmes. In addition to agricultural
expansion and fuelwood harvesting, coffee cultivation is an important driver of forest
change in Southern Ethiopian forest ecosystems (Aerts et al., 2011), and the increasing
intensity of coffee cultivation systems represents a major threat to the remaining native
Afromontane forests (Schmitt et al., 2010b; Tadesse et al., 2014a).
2.3 Data and Methods
2.3.1 Data acquisition and preprocessing
An overview of the methods used in this study is shown in Figure 2.2. We downloaded all
available Landsat ETM+ with WRS-2 coordinates p170r55 at processing level L1T and
cloud cover below 70% from the USGS Glovis repository (http://glovis.usgs.gov). On
average, 9 ETM+ scenes were available per year until 2011, with fewer scenes in 1999,
2000 and 2008 in particular (Table 2.1). To convert raw imagery from digital number
(DN) to Top of Atmosphere (ToA) Reflectance and Surface Reflectance (SR), we used
the LEDAPS method (Masek et al., 2006b), which is based on the 6S radiative transfer
method (Vermote et al., 1997). We used the object-oriented FMASK algorithm (Zhu
& Woodcock, 2012) to detect clouds and cloud shadows subsequently masked these out
of the images. After removing cloud-contaminated and SLC-gap pixels, we estimated
the number of clear-sky observations throughout the period 1999 to 2012 for each pixel
(Figures 2.3 and 2.A1). An average of 55 clear-sky observations were available per pixel
during this time period. Areas in which the least amount of data were available were
associated with high-altitude areas with frequent cloud cover.
24 Monitoring small-scale disturbances using LTS
Figure 2.2: Flowchart used in this study. Rectangles represent image layers resulting from
processing steps.
2.3.2 Benchmark forest mask
To avoid confusion between forest disturbances and other land cover dynamics (e.g. crop-
ping cycles), we produced a benchmark forest mask for 2005 representing all forested
pixels at the beginning of the study period. We selected a base image from the 4th of
March, 2005 and another image from the 1st of December, 2005 to fill gaps left by the
failure of the SLC-corrector (following Zhu et al. (2012a)). We classified the resulting
gap-filled image using a supervised maximum likelihood classifier in the ArcGIS 9.2 soft-
ware package (ESRI, inc.). Our initial land cover map included forest, cropland, wetland,
shrubland, bare soil, plantations and urban classes. We aggregated all non-forest classes
to produce a binary forest/non-forest mask. To avoid high commission errors arising
from patchy mosaic transitional areas (including, for example, cropland demarcated with
planted trees), we further refined the forest mask by masking out all pixels with a tree
2.3 Data and Methods 25
Table 2.1: Number of scenes and mean and standard deviation clear sky pixel observations
across the study area for each year included in this study. Only pixels included in the 2005
benchmark forest mask were included in the means and standard deviations.
year # scenes mean obs. s.d. obs.
1999 2 0.49 0.60
2000 3 1.44 0.77
2001 5 3.68 1.33
2002 9 4.33 1.55
2003 10 4.43 1.70
2004 9 3.89 1.64
2005 9 4.10 1.57
2006 10 3.83 1.82
2007 16 7.39 2.95
2008 6 3.44 1.45
2009 9 3.82 1.76
2010 12 5.56 2.27
2011 12 3.87 2.02
2012 11 4.87 1.72
Figure 2.3: The number of clear-sky ETM+ observations between 1999-2012 for each pixel
overlaid onto a hill-shade layer derived from an SRTM DEM.
cover of less than 30% according to the corresponding MODIS VCF product for 2005
(MOD44B; Hansen et al. (2003)). Finally, we masked all non-forest pixels according to
our 2005 forest mask out of all images in the time series, and produced an NDVI image
stack from the resulting imagery.
26 Monitoring small-scale disturbances using LTS
2.3.3 Breakpoint detection in Landsat time series
To detect forest disturbances, we applied a pixel-wise time series method based on the
BFAST monitoring approach described in Verbesselt et al. (2012). Our approach follows
three main steps to classifying changes in each of the pixels included in the benchmark
forest mask: (1) fitting a harmonic model based on observations within a defined history
period; (2) testing observations in the monitoring period directly following the history
period for structural breaks from the fitted harmonic model; and (3) calculating the
median of the residuals for all expected and actual observations within the monitoring
period. These steps are described in more detail below.
1) Fitting the harmonic model. We divided each pixel time series defined by ti
[t1, tN] into a history period and a monitoring period. For a monitoring period beginning
at time tn, we defined the history period as the period where t1≤ti< tnand the
monitoring period as the remainder of the time series, in which tn≤ti≤tN(see Figure
2.4 for an illustration of these time variables). We assumed that forested pixels were
generally stable in the period leading up to the beginning of the monitoring period (tn).
As such, we fit a harmonic model (Verbesselt et al., 2012) to all observations in the history
period:
yt=α+γsin(2πt
f+δ) + εt(2.1)
where ytand tare the response (dependent variable) and time (independent variable), f
is the temporal frequency, αis the intercept, γ, and δare the amplitude and phase of the
harmonic component, and εtis the residual (noise component). Our time series model
differed from that of Verbesselt et al. (2010) and Verbesselt et al. (2012) in two important
aspects. First, we employed a simple single-order model (Equation 2.1) given the fact
that forest phenology in the study area followed a similar first-order harmonic curve
(Figure 2.4). Moreover, complex seasonal patterns were not expected to be detectable
with Landsat data alone in our study area, due to the irregularity of clear-sky observations
within a given year, and any attempt to model these processes resulted in model over-
fitting. Second, we omitted the trend term from the fitted model. Even in cases where
forest regrowth in the history period might justify the inclusion of a trend, projecting this
trend into the monitoring period resulted in unrealistically high expected NDVI values,
generating false breakpoints and inflated change magnitude values.
We fit the harmonic model independently for each pixel time series using Ordinary Least
Squares (OLS) to determine the model of best fit, as described in more detail in Verbesselt
et al. (2012). Given the short length of the history period, a result of a lack of Landsat
data before the launch of ETM+ in 1999, we chose to include all observations in the history
period in the model fitting process, in contrast to the stable history model employed by
Verbesselt et al. (2012).
2.3 Data and Methods 27
2) Detecting change. To detect breakpoints in pixel time series, we used the approach
described in Verbesselt et al. (2012) for monitoring structural changes, in which moving
sums (MOSUM) are computed using all observations in the monitoring period. Thus for
all tn≤ti≤tN, the MOSUM (MOt) was determined by:
MOt=1
ˆσ√n
t
X
s=t−h+1
(ys−ˆys) (2.2)
where yand ˆyare actual and expected observations, respectively, and ˆσis an estimator
of the variance (Zeileis et al., 2005). The MOSUM bandwidth, h, is defined as a fraction
of the number of observations in the history period (n) (Verbesselt et al., 2012). A
breakpoint is signaled when |MOt|deviates from zero beyond a 95% significance boundary
as described in Verbesselt et al. (2012) and Leisch et al. (2000). We assigned a value of
0.25nto hthroughout this study.
3) Computing change magnitude within the monitoring period. We computed
change magnitude (M) by taking the median of residuals within the monitoring, in which
tn≤ti≤tN:
M=median{yt−ˆyt}(2.3)
where ytand ˆytare actual and expected observations, respectively. While this measure is
closely related to the criterion used for the MOSUM test (Equation 2.2), the median of the
residuals is expected to be less sensitive to noise in the monitoring period than the sum (as
in the MOSUM test), and thus provides an added insurance against spurious breakpoints.
As the length of the monitoring period is increased, this measure of change magnitude
could be affected by an increased number of observations before and after the change
event. For this reason, we chose to limit the monitoring period to one year and applied
the monitoring method described here in an iterative fashion, using sequentially defined
monitoring periods (2005-2006, 2006-2007, and so on). For each monitoring period, we
trimmed the time series such that tNwas equal the date of the final observation within
the 1-year monitoring period. The sequential monitoring approach we described here is
demonstrated in Figure 2.4.
2.3.4 Sample-based reference dataset
We centred our validation approach on changes in 2009 to ensure availability of enough
high-resolution observations before and after to assess changes (Table 2.2). We con-
structed a reference dataset using a stratified random sampling approach from two pop-
ulations of pixels: (1) those having breakpoints assigned in 2009 (and not in previous
monitoring periods, to prevent assessing redundant change pixels); and (2) those having
no breakpoints assigned in any monitoring periods up to and including 2009. Both popu-
lations were stratified into three magnitude quantiles, resulting in a total of six strata from
28 Monitoring small-scale disturbances using LTS
Figure 2.4: Demonstration of the sequential monitoring approach used in this study. In this
illustration, results of a series of sequential 1-year monitoring periods from 2006 to 2009 (from
top to bottom). The presence or absence of a breakpoint (dotted red line) and magnitude
(M) are shown on each plot. The residuals, defined as the difference between observed values
(black dots) and expected values (blue line) are shown as broken grey lines in each monitoring
period. The solid black line denotes both the beginning of the monitoring period and the end
of the history period.
which roughly equal numbers of pixels were selected. After removing 19 pixels lacking
enough data to confidently classify disturbances (e.g. due to excessive cloud cover in the
high resolution data), the reference dataset consisted of a total of 112 breakpoint pixels
and 109 non-breakpoint pixels.
To classify changes in the randomly sampled pixels, we used a combination of Landsat
time series and very high resolution (VHR) imagery. First, we viewed time series of
Landsat RGB composites (using both bands 3-2-1 and 7-4-5 composites) in 35 x 35 pixel
windows following Cohen et al. (2010) to identify potential disturbance events. Second,
we complimented the Landsat time series with anniversary 2.5m resolution SPOT5 im-
agery from 2007 until 2011 inclusive (Table 2.2) to confirm change classes. We chose
a relatively cloud-free image from 2009 from the SPOT5 dataset to be co-registered to
an orthorectified Landsat image, and registered all other SPOT5 images to that image.
Where available, we complemented the SPOT5 time series data with RapidEye imagery
from 2012 and 2013, and GoogleEarth imagery from 2012 and 2014 (based on SPOT5 and
DigitalGlobe imagery). Cases of deforestation were possible to identify with the Landsat
time series and were confirmed with the VHR datasets. In cases of degradation, we flagged
potential degradation in the Landsat time series and assigned a visual estimate of canopy
cover (a value between 0 and 1) to each of the available SPOT5 images for that location.
According to the definition of forest employed in this study, we assigned a degradation
label to the pixel when a change was evident but estimated canopy cover from the VHR
2.3 Data and Methods 29
Table 2.2: High resolution imagery used for validation of change results.
sensor scene ID date
SPOT5
133-335
2007-11-02
2009-01-12
2010-01-01
2011-03-24
132-334 2011-02-05
132-335 2011-02-05
RapidEye
3742502 2012-01-05
2012-10-17
3642428 2012-12-12
3642528 2012-12-12
3642627 2012-12-12
images did not drop below 0.2 (e.g. Figure 2.9).
We calculated total accuracy (TA), user’s accuracy (UA; inversely related to commission
errors) and producer’s accuracy (PA; inversely related to omission errors) using these
reference data. Our accuracy estimates were limited to disturbance and non-disturbance
classes, where disturbance included both the deforestation and degradation classes de-
scribed above.
2.3.5 Modeling the effect of breakpoints and magnitude on disturbance prob-
abilities
Since breakpoints are sometimes associated with near-zero or positive magnitudes which
are not related to forest disturbances, commission errors are high when the presence or
absence of breakpoints is used as the sole criterion for change labeling. Detected break-
points are therefore not sufficient criteria for classifying change and the magnitude must
also be factored into the classification. We used an ordinal logistic regression (OLR) mod-
eling approach (Walker & Duncan, 1967) to further understand the relationship between
magnitude and varying degrees of disturbances found in the study area. We used three
classes as ordered response variables in the OLR models: deforestation, degradation and
no change.
The OLR model takes the general form:
P(Y≥j|X) = 1
1 + e−αj+Xβ (2.4)
where P(Y≥j|X) is the cumulative probability of class jgiven a measured covariate X,
30 Monitoring small-scale disturbances using LTS
αjis the intercept for each class j, and
Xβ =β1X1+. . . +βkXk(2.5)
for kcovariates X. By solving for model coefficients βusing the Maximum Likelihood
(ML) method, a model for likelihood of each class jcan be constructed. Rearranging
equations 2.4 and 2.5, the model takes on the linear form:
logit(Y≥j|X) = αj+β1X1+. . . +βkXk(2.6)
where:
logit(Y≥j|X) = log(P
1−P); P=P(Y≥j|X) (2.7)
We tested two models describing the response of the three ordered classes to several
predictor variables. The first model related P(as defined in Equation 2.7) as a function
of magnitude (M) and the second model related Pas a function of Mand breakpoints
(C). We compared these models using chi-squared probabilities (compared to a null
intercept-only model) and the Akaike Information Criterion (AIC) (Akaike, 1973), which
is a measure of goodness of fit and includes a penalization for additional parameters to
prevent model over-fitting.
2.4 Results
2.4.1 Robustness of the MOSUM breakpoints test
The MOSUM breakpoint test proved effective in discriminating deforestation events from
stable forest trajectories in the presence of noise and data gaps in the time series data.
Discrete deforestation events (Figure 2.5B) resulted in a sudden and persistent decrease
in NDVI, whereas stable forests were largely free of breakpoints (Figure 2.5A). Forests
undergoing degradation intense enough to cause canopy openings also led to breakpoint
detection (Figure 2.5C), even though some degree of seasonality due to the remaining
forest canopy was still present in the follow-up time series. In regions where low availability
of clear-sky observations caused the pixel time series to be irregular, our method was still
able to capture forest changes in most cases (Figure 2.6).
2.4.2 Relationship between magnitude and forest change processes
The examples shown in Figure 2.5 demonstrate the differences in magnitude (∆NDV I)
commonly observed between varying degrees of forest change intensity. Here, a stable
forest pixel with no breakpoint detected in the monitoring period (A) had a magnitude of
2.4 Results 31
Figure 2.5: Results from the 2009-2010 monitoring period (left panel), chosen to illustrate
different change processes. Change magnitude for all pixels with breakpoints are shown over-
laid on a greyscale SPOT5 image (band 2) from February 2011. Time series plots for intact
forest (A), deforestation and subsequent cropping (B), and progressive canopy clearing (C).
Photos taken in April 2013 from areas in the neighbourhoods of each corresponding pixel are
also shown.
-0.026, a pixel where a discrete forest clearance event occurred (B) had a breakpoint and
a magnitude of -0.18, and a pixel where progressive reduction of canopy cover (C) had a
breakpoint and a magnitude of -0.084.
Figure 2.7 shows the magnitudes for sequential monitoring periods (top panel) compared
to a SPOT5 time series (bottom panel) corresponding to each monitoring period over an
area experiencing progressive forest loss over this time period. In 2008, very few changes
were detected in this area and the SPOT5 imagery shows mostly full canopy cover. By
2009, several discrete forest perforations appeared which correspond with high-magnitude
pixel clusters in the BFM results. By 2011, most of this entire region had been cleared
and converted to cropland, but due to cloud cover in the SPOT5 image from 2010 (not
shown), it was not possible to ascertain visually when this change exactly happened. The
magnitude shown in the top panel of Figure 2.7 reveals that most of the area surrounding
this perforation was cleared within the 2009-2010 monitoring period.
32 Monitoring small-scale disturbances using LTS
Figure 2.6: Example of a small change correctly identified in the 2009-2010 monitoring
period, despite an irregular time series with frequent gaps. Landsat 3-2-1 composites for four
key dates are shown below the time series plots. The Landsat pixel for which time series plots
are shown is indicated by a red arrow. Pixels masked out using the FMASK product are
shown in black.
Figure 2.7: Magnitude values for annual monitoring periods between 2008 and 2012 (labeled
from A to D, respectively) with SPOT5 images acquired between 2007 and 2011 and one
RapidEye image acquired in 2012. The change magnitude in the top panel is only shown for
pixels for which a breakpoint with negative magnitude in that monitoring period was detected.
Land cover types in the initial SPOT5 image (bottom left) are labeled for forest (f ), wetland
(w) and cropland (c). An example of discrete deforestation is shown as a solid arrow and an
example of gradual clearing is shown as a broken arrow.
2.4.3 Predicting forest disturbances using breakpoints and magnitude
The ordinal logistic models (OLR) relating the cumulative probability (P) of each of the
change classes (deforestation, degradation or no-change) as a function of magnitude (M)
2.4 Results 33
Table 2.3: Stepwise comparison of ordinal logistic regression models including magnitude
only or breakpoints and magnitude to the null model (intercept only).
regressors residual d.f. residual deviance P(χ2) AIC
(null) 201 383.2 – 387.2
magnitude 200 304.7 <0.0001 310.7
magnitude, breakpoint 199 299.7 0.025 307.7
Figure 2.8: Ordinal logistic regression (OLR) model including magnitude and breakpoints for
three disturbance classes: “deforestation”, “degradation” and “no change”. The probability
of each class (P) is shown on the y-axis as a function of magnitude (M) on the x-axis. The
distribution of reference observations are shown as small segments along the x-axis.
showed that Mwas a significant predictor of P, both in the presence or absence of a
breakpoint (C) as a second binary regressor. Introducing Cto the model resulted in a
slight decrease in the AIC (from 326.6 to 305.9; Table 2.3), but had otherwise little effect
on the fitted model. The probability distribution for each of the disturbance classes (de-
forested, degraded, no-change) is shown in Figure 2.8. The probability of the deforestation
class increased with decreasing magnitude, while the probability of the no-change class
increased with decreasing magnitude. The probability of the degradation class showed a
weak relationship with magnitude, with a slight peak at approximately M=−0.1, but
did not exceed probability of the other classes for any values of M.
The total accuracy (TA) and user’s and producer’s accuracy of the disturbance class (UA
and PA, respectively) are shown for a range of magnitude thresholds in Figure 2.10 for
disturbances classified using magnitude only (left panel) and magnitude and breakpoints
34 Monitoring small-scale disturbances using LTS
Figure 2.9: Example of degradation, where canopy cover is visibly reduced without complete
conversion to non-forest. Landsat images are shown as band 3-2-1 composites in the top panel.
SPOT5 images for the same extent are shown in the 2nd row, with a GoogleEarthTM image
(DigitalGlobe, 2012) in the last panel. The time series plot (top) is shown for monitoring
period 2009-2010.
(right panel). Here, disturbances were defined to include both deforestation and degrada-
tion. A maximum TA of 79% was achieved in both cases with a magnitude threshold of
-0.080. The UA/PA cross-over point for the magnitude-only case occurred at a threshold
between -0.075 and -0.080 and at -0.065 for the magnitude+breakpoint case. In both
cases, the UA and PA were 73% at their cross-over points. In the latter case (which was
selected as a final disturbance threshold), TA was estimated at 78%. Including break-
points in the disturbance classification had a noticeable effect for magnitude thresholds
approaching zero. Here, the decrease in UA was less severe when breakpoints were in-
cluded in the disturbance classification and the PA leveled off at 81% instead of rising to
92% as in the magnitude-only case.
The UA and PA for the no-change class (not shown in Figure 2.10) were generally higher
than all other accuracies. At the -0.080 threshold, UA and PA for the no-change class
were 82% and 83% respectively when only magnitude was considered, and 81% and 85%
respectively when breakpoints were included.
2.5 Discussion 35
Figure 2.10: Total accuracy (TA) and producer’s and user’s accuracies for the disturbance
class (PA and UA respectively) as a function of magnitude. Disturbance is defined as defor-
estation and degradation and modeled using magnitude only (left panel) or magnitude and
breakpoints (right panel).
2.4.4 Area and spatial distribution of forest disturbances
We chose a magnitude threshold of -0.065 (corresponding to the cross-over point between
UA and PA shown in Figure 2.10, right panel) to generate a final disturbance map (Figure
2.12). From this map, the total disturbed forest area was estimated to be over 11,000
hectares throughout the entire 2005-2012 period, representing roughly 3% of the total
forest area in the Biosphere Reserve. In total, 0.84% of core zone forest, 2.1% of buffer
zone forest, and 1.9% of transition zone forests were disturbed over this time period.
Small incremental disturbances permeated the forest areas throughout this period. The
mean size of disturbance pixel clusters was 0.6 hectares, with a maximum change cluster
of 285 hectares.
2.5 Discussion
2.5.1 Forest change in the Ethiopian Afromontane forests
In this paper we present the first detailed study on forest disturbances in the Ethiopia
Afromontane forests, in which we estimate total forest loss of roughly 11,000 hectares be-
tween 2005 to 2012 in the UNESCO Kafa Biosphere Reserve, representing approximately
3% of the total forest area. This deforestation rate corresponds to a loss of roughly 0.4%
of forest land per year, which is comparable to change rates estimated by Getahun et al.
(2013) for a neighbouring Afromontane forest area (0.19% between 1975 and 2007). Sim-
36 Monitoring small-scale disturbances using LTS
Figure 2.11: Example of a commission error arising from several unmasked clouds within one
monitoring period (2009-2010). A Landsat 3-2-1 composite for four key dates are shown below
the time series plots. The Landsat pixel for which time series plots are shown is indicated by
a red arrow. Pixels masked out using the FMASK product are shown in black.
ilarly to Getahun et al. (2013), we found that forest change was largely small-scale and
driven by small-holder agriculture and occurred at higher rates in more remote locations.
We additionally note that in-migration of resettlers from other areas in Ethiopia (Figure
2.12A) was largely responsible for abrupt larger-scale changes in these remote areas.
The ability to describe these change processes with high temporal detail highlights the ad-
vantage of the time series curve-fitting approach used in this study over more conventional
forest change detection studies conducted in Ethiopia (Getahun et al., 2013; Gebrehiwot
et al., 2013; Garedew et al., 2009; Assefa & Bork, 2014). By making use of the magnitude
parameter, we have shown that forest change does not typically occur as large discrete
clearing events, but rather in an incremental manner (Figure 2.7). Discussions with local
farmers in the study area revealed that clearing of a forest demarcated for agriculture is
a process that can last up to several years, as forest is left standing in the first years and
used for cattle grazing and harvest of forest resources (wild coffee and spices). During
this time fuel-wood and timber is gradually harvested by hand, and local crops such as
sorghum, tef are eventually cultivated in place of the forest.
2.5.2 Forest change tracking with sequential monitoring periods
We employed a sequential monitoring approach based on the BFAST monitoring method
Verbesselt et al. (2012), where a harmonic curve was fit to a historical time series and a
MOSUM test for monitoring change was applied to the Landsat time series stack using
non-overlapping 1-year monitoring periods (Figure 2.4). The reason for defining equal
monitoring period length was related to the output magnitude parameter. Magnitude is
defined as the median of the residuals within a monitoring period between observations
and expected values (based on the history period model; Equation 2.3). It was therefore
2.5 Discussion 37
Figure 2.12: Forest change map produced using a magnitude threshold of -0.065. The predicted time of deforestation is shown both
in the main panel and insets A to C, and areas of possible degradation are also shown in the insets. The original forest mask from
2005 (green) and core forest areas defined under the UNESCO Kafa Biosphere Reserve (hatched areas) are shown in the main panel.
Three examples of different change processes occurring in the study area are shown in the insets (using a SPOT5 image from 2011 as
a basemap), including in-migration from other parts of the country (A), gradual reduction of forest cover in forest-agriculture matrix
landscapes (B) and construction of a rural road (C).
38 Monitoring small-scale disturbances using LTS
important to keep a consistent definition of the monitoring period, as change magnitude in
longer monitoring periods would be affected by follow-up land use (e.g. seasonal changes
due to cropping cycles after deforestation), and the metric would no longer be comparable
across change processes. In the case of this study, we assumed that no significant follow-
up changes occurred within the 1-year monitoring period, and that the change magnitude
could therefore be explained by the specific change being observed in that year (rather
than a combination of change and follow-up change processes).
In this study, we found that small-scale forest changes could be tracked using NDVI time
series. It has been previously shown that NDVI is affected by saturation effects when
forest cover is dense (Carlson & Ripley, 1997) and is less sensitive to forest changes than
other metrics based on the short-wave infrared (SWIR) region of the electromagnetic
spectrum (Zhu et al., 2012b; Kennedy et al., 2010; Jin & Sader, 2005). Despite these
limitations, our approach relies not only on absolute difference between observations, but
on structural breaks in the time series (Leisch et al., 2000; Verbesselt et al., 2012). As
such, even a saturated NDVI signal followed by an altered seasonal NDVI curve (due to
regenerating vegetation or follow-up cropping cycles, for example) would be sufficient for
a breakpoint to be detected. The types of forest change encountered in this study resulted
in a general decline in NDVI in cases of incomplete forest clearing (Figures 2.5C, 2.7 and
2.9) and a significantly altered harmonic curve in cases of complete conversion from forest
to cropland (Figure 2.5B).
2.5.3 Modeling disturbance processes with breakpoints and magnitude
The ordinal logistic regression (OLR) model including breakpoint and magnitude as ex-
planatory variables performed only slightly better than that of the model including only
magnitude (Table 2.3). Both the AIC and the accuracies (Figure 2.10) show that addi-
tion of breakpoints as an additional predictor variable resulted in a slight improvement
in accuracy. Adding breakpoints to the disturbance definition shifted the cross-over point
of the UA and PA (Figure 2.10) to a higher magnitude threshold and reduced the com-
mission errors that arise when the magnitude threshold approaches zero. In practice, all
breakpoints with negative magnitudes could be defined as disturbances for the forest types
featured in this study, but would only achieve accuracies on the order of 60%. We thus
conclude that while magnitude alone is sufficient to predict disturbances for the montane
forests in Southern Ethiopia, the best prediction is achieved using both magnitude and
breakpoints.
The reliance on magnitude to detect disturbances has implications on the reproduction of
our method in similar study areas. Rather than applying the curve-fitting and breakpoint
detection algorithm in a purely data-driven manner, calibration of a magnitude threshold
against reference data are needed to derive a final disturbance label. Similar work in
2.5 Discussion 39
southern Peru has shown that with relatively large, discrete forest clearances, breakpoints
are sufficient to detect disturbances when only negative magnitudes are taken (DeVries
et al., 2015). The importance of magnitude in our study is thus related to the scale
of disturbances experienced in the study. In areas with small-scale, incremental forest
changes (e.g. due to small-holder agricultural expansion), the magnitude of the change is
an essential variable for mapping disturbances.
2.5.4 Comparison with other time series curve-fitting methods
The approach we used in this study, based on previous work on MODIS time series
(Verbesselt et al., 2010, 2012), joins the ranks of other recent methods using all available
Landsat data to detect changes in forest cover (Zhu et al., 2012a; Broich et al., 2011;
Ernst et al., 2013; Hansen et al., 2013). Of all the time series methods using all available
Landsat observations for forest change detection, our approach is most similar to that of
Zhu et al. (2012a). Both approaches use similar model-fitting approaches using historic
data from each pixel, and subsequent time series data are checked against the fitted model
for deviation from the fitted model. Our approach differs from that of Zhu et al. (2012a)
in two aspects. First, our approach flagged potential disturbances using a MOSUM test
for monitoring change in the monitoring period while accounting for seasonal variability
(Verbesselt et al., 2012; Leisch et al., 2000), whereas Zhu et al. (2012a) flagged distur-
bances when residuals exceeded a multiple of the RMSE on several consecutive instances.
Second, our method also features the calculation of a consistent magnitude metric in-
dependent of the presence or absence of breakpoints. In addition to being an essential
predictor variable of disturbances in our study area (Table 2.3), the magnitude allowed
for mapping of incremental changes over time (Figure 2.7).
2.5.5 Sources of errors
The two major sources of error arising from this study were unmasked clouds or cloud
shadows and errors arising from the benchmark forest mask.
Unmasked clouds or cloud shadows
Occasional unmasked clouds or cloud shadows resulted in greatly reduced NDVI values for
those pixels. The method used in this study featured two safeguards against these types
of noise in the monitoring period. First, a single outlier observation was unlikely to trigger
a breakpoint, as the MOSUM value is a sum of all observations within a moving window
(Equation 2.2). Second, the magnitude was calculated as the median of all residuals
within the monitoring period (Equation 2.3), which is robust against single outliers. These
safeguards became less effective with multiple occurrences of unmasked clouds or cloud
40 Monitoring small-scale disturbances using LTS
shadows within one monitoring period, as demonstrated in Figure 2.11. Persistent cloud
cover (e.g. over mountain peaks in our study area, shown in Figures 2.3 and 2.A1) can
thus present a significant challenge to the application time series based change detection
methods in tropical montane forest systems (Broich et al., 2011).
Benchmark forest mask
Classification errors in the benchmark forest mask propagated through to the curve-fitting
and MOSUM breakpoints test, leading to breakpoints resulting from non-forest related
dynamics (e.g. cropping cycles). While our revised mask, modified using the the MODIS
Vegetation Continuous Field (VCF) product (Hansen et al., 2003), was more conservative
than the original Landsat-based classified mask, some of the disturbances shown in the
first monitoring period (2005; Figure 2.12) may have arisen from mis-classified non-forest
pixels in the benchmark year, causing them to be mis-labeled as deforestation in the
initial monitoring period. On the other hand, changes occurring in small forest patches
may have been omitted as a result of the application of the MODIS VCF mask, which
would contribute to omission errors in the overall change result.
2.5.6 Other limitations
In addition to the error sources noted above, several other limitations encountered in this
study warrant attention here.
Historical data availability
A common requirement of curve-fitting time series methods is the need for historical data
from which to establish a “stable” reference model (Zhu et al., 2012b; Verbesselt et al.,
2012). In many areas in the tropics, including southern Ethiopia, there is a near-complete
lack of Landsat 5 observations in the 1990’s (Broich et al., 2011). For this reason, the time
series data used in this study were limited to Landsat 7 observations acquired after 1999
as in other similar studies in the tropics (Broich et al., 2011; Potapov et al., 2012). For
many parts of the study area, there were not enough observations in the 1999-2005 period
to monitor for changes in those periods. For example, the 2002-2003 monitoring period
would require enough observations from 1999-2001 (inclusive) to fit a history model. The
significant data gaps throughout the 1990’s in the Landsat archive over our study area,
combined with the need for a history period from which to fit the season model, presents a
constraint to tracking historic changes, a prerequisite for establishing monitoring baselines
for REDD+ (Olander et al., 2008; Kim et al., 2014).
2.5 Discussion 41
Ability to map degradation
Even though we were able to identify several cases of degradation (e.g. Figure 2.9) and
include them in the OLR model, there was a large spread in the relationship between
probability of degradation and magnitude (Figure 2.8), both with and without break-
points. For this reason, we could not conclusively determine a relationship between mag-
nitude and probability of degradation. Several limitations to our approach contributed to
uncertainties in degradation detection. First, as noted above, NDVI is known to be less
sensitive to forest structure than other spectral indices. In addition to other ratio-based or
transformation-based indices used for disturbance monitoring, sub-pixel un-mixing met-
rics, such as the Normalized Difference Fraction Index (NDFI) have shown promise in
mapping degradation due to selective logging in tropical forest systems (Souza et al.,
2005). However, the sensitivity of these indices to the types of degradation encountered
in this study remains largely unexplored.
Second, the fact that our reference dataset was limited to visual interpretation of optical
satellite imagery represents an additional imitation to our ability to assess the ability of
our method to detect degradation. Combining the time series approach described in this
paper with advances in forest monitoring methods, such as LiDAR (Zhuravleva et al.,
2013; Thompson et al., 2013; Ahmed et al., 2014) and field-based monitoring methods
(Gonsamo et al., 2013; Pratihast et al., 2014), can provide new opportunities for mapping
and validation of low-intensity degradation processes. The integration of these monitoring
technologies with Landsat time series for monitoring forest degradation is a topic for future
research.
Finally, our inability to conclusively predict degradation was related to the low intensity
of degradation processes in our study area, a constraint which has been reported in other
areas in sub-Saharan Africa (Zhuravleva et al., 2013). Recent field studies have shown that
fuelwood harvesting at the community scale is the dominant driver of forest degradation
in Kafa (Pratihast et al., 2014; Dresen et al., 2014). The impact of degradation due to
fuelwood harvesting remains invisible to optical satellites until its cumulative effects result
in changes to the forest canopy (Pratihast et al., 2014; Herold et al., 2011).
Repeat disturbances
The method we used in this study does not allow for detection of follow-up regrowth
and repeat disturbances, a similar limitation to that reported by Zhu et al. (2012b). We
instead assumed that the earliest encountered breakpoint with magnitudes lower than the
calibrated magnitude threshold represented a permanent change. Other trajectory-based
approaches, on the other hand, monitor disturbance and follow-up regrowth in a holistic
manner (Huang et al., 2010; Kennedy et al., 2010). To further develop our approach
to include disturbance-regrowth dynamics, continued monitoring of forest regrowth and
42 Monitoring small-scale disturbances using LTS
iterative disturbance monitoring needs to be built into the algorithm (DeVries et al.,
2015).
2.5.7 Implications for national monitoring activities
For the method presented in this paper to be applicable to a national-scale MRV and
National Forest Monitoring System (NFMS) in Ethiopia, several key questions need to
be addressed. The first question relates to the applicability of the method over all forest
types in the country. Despite the fact that we have focused our research on the moist
Afromontane forests of the south, we expect the method presented here to be applicable
over all forest types, as it is robust to noise and data gaps. The second question relates to
the impact of errors on overall CO2emissions estimates, a key requirement for REDD+
MRV. The accuracies estimated in this study indicate that despite the robustness of our
method to noise and data gaps, the nature of forest change in our study area presents key
challenges to reliably tracking changes over time. Similar challenges were encountered
by Tyukavina et al. (2013), who noted considerable scale-related uncertainties in above-
ground carbon stock changes in the Democratic Republic of the Congo due to the small-
scale of forest changes there. Similarly, uncertainties arising from small-scale changes in
our study area are expected to propagate through to forest area loss and carbon stock
change estimates.
2.6 Conclusions
In this paper, we show that the BFAST Monitor method (Verbesselt et al., 2012) for
breakpoint detection in time series is applicable to Landsat ETM+ time series data for
forest disturbance monitoring for an area in Southwestern Ethiopia characterized by highly
fragmented Afromontane forests. Using high resolution time series imagery as reference
data, we estimated an overall accuracy of 78%, with associated user’s and producer’s
accuracy of 73% for disturbances. Our results show that magnitude of residuals during
the monitoring period is essential for mapping disturbances in the landscape featured in
the study area, while MOSUM-based breakpoints slightly improved disturbance results.
Applying the algorithm to Landsat time series allows for regular monitoring of the small-
scale forest change processes characteristic of Ethiopia and other tropical countries where
small-holder agriculture is the main driver of disturbance. Between 2005 and 2012, the
mean disturbance cluster size was found to be 0.6 hectares, which demonstrates the scale
at which this disturbance driver operates. A major constraint faced in this study was
the fact that we could not quantitatively describe degradation processes, due to the small
scale of degradation in the study area and limited data availability in the Landsat NDVI
time series. The method was shown to be useful in monitoring of forest disturbances,
2.6 Conclusions 43
and given the opening and expansion of the Landsat data archive, could be a key asset to
regions and countries developing REDD+ MRV and National Forest Monitoring Systems
(NFMS).
Acknowledgements
We would like to thank Dr. Nikolaus Umlauf (University of Innsbruck) and Dr. Achim
Zeileis (University of Innsbruck) for assistance with the ordinal logistic regression models,
and Michael Schultz (Wageningen University) for assistance in pre-processing the SPOT5
data.
44 Monitoring small-scale disturbances using LTS
2.A Appendix: Clear-sky Observations
Figure 2.A1: Clear sky observations per year over a subset of the study area. Non-forest
pixels according to the 2005 benchmark forest mask are shown in black.
Chapter 3
Tracking post-disturbance regrowth
using Landsat time series
This chapter is based on:
DeVries, B., Decuyper, M., Verbesselt, J., Zeileis, A., Herold, M. & Joseph, S.
2015. Tracking disturbance-regrowth dynamics in tropical forests using structural
change detection and Landsat time series. Remote Sensing of Environment, 169:320-334.
doi:10.1016/j.rse.2015.08.020.
46 Tracking post-disturbance regrowth using LTS
Abstract
Increasing attention on tropical deforestation and forest degradation has necessitated more
detailed knowledge of forest change dynamics in the tropics. With an increasing amount
of satellite data being released to the public free of charge, understanding forest change
dynamics in the tropics is gradually becoming a reality. Methods to track forest changes
using dense satellite time series allow for description of forest changes at unprecedented
spatial, temporal and thematic resolution. We developed a data-driven approach based
on structural change monitoring methods to track disturbance-regrowth dynamics using
dense Landsat Time Series (LTS) in a tropical forest landscape in Madre de Dios, southern
Peru. Whereas most existing post-disturbance regrowth monitoring methods rely on
annual or near-annual time series, our method uses all available Landsat data. Using
our disturbance-regrowth method, we detected annual disturbance from 1999 to 2013
with a total area-weighted accuracy of 91 ±2.3%. Accuracies of the regrowth results
were strongly dependent on the timing of the original disturbance. We estimated a total
area-weighted regrowth accuracy of 61 ±3.9% for pixels where original disturbances were
predicted earlier than 2006. While the user’s accuracy of the regrowth class for these
pixels was high (84 ±8.1%), the producer’s accuracy was low (56 ±9.4%), with markedly
lower producer’s accuracies when later disturbances were also included. These accuracies
indicate that a significant amount of regrowth identified in the reference data was not
captured with our method. Most of these omission errors arose from disturbances late in
the time series or a lack of sensitivity to long-term regrowth due to lower data densities
near the end of the time series. Omission errors notwithstanding, our study represents the
first demonstration of a purely data-driven algorithm designed to detect disturbances and
post-disturbance regrowth together using all available LTS data. With this method, we
propose a continuous disturbance-regrowth monitoring framework, where LTS data are
continually monitored for disturbances, post-disturbance regrowth, repeat disturbances,
and so on.
3.1 Introduction 47
3.1 Introduction
Rapid changes in tropical forest ecosystems worldwide in recent decades have had tremen-
dous environmental impacts globally, contributing significantly to climate change (Gulli-
son et al., 2007) and biodiversity loss (Laurance et al., 2012). In response to these threats,
international level discussions, frameworks and initiatives have been set up to combat an-
thropogenic forest loss. One such initiative, the “Reducing Emissions from Deforestation
and forest Degradation” (REDD+) programme, features results-based payments to mainly
tropical countries who implement activities to stem CO2emissions arising from deforesta-
tion and forest degradation (Corbera et al., 2010). A key requirement for the successful
implementation of REDD+ is the Measuring, Reporting and Verification (MRV) of forest-
related emissions and emission reductions (Joseph et al., 2013; DeVries & Herold, 2013;
Herold & Skutsch, 2011). The importance of including remote sensing data in MRV for
REDD+ has been widely recognized among the scientific community (Goetz & Dubayah,
2011; De Sy et al., 2012).
A range of tropical forest monitoring systems and initiatives have been developed in re-
cent years to support national and international efforts to stem tropical forest loss. With
an increase in the availability of free satellite imagery, large-area mapping of forest distur-
bances has been operationalised in several key instances in the tropics. First, the Brazil-
ian Space Agency (INPE) launched the PRODES and DETER monitoring databases to
provide data on annual forest change and near real-time disturbance detection, respec-
tively (INPE, 2014a,b). Second, a global map of annual forest change made at 30m
resolution (Hansen et al., 2013) was recently made public under the banner of the Global
Forest Watch (World Resources Institute, 2014). These systems represent an important
development towards the operational monitoring of forest change, not only in support of
REDD+ MRV, but also as a tool to raise public awareness of the scale and rate of tropical
forest change.
Change detection methods increasingly make use of Landsat Time Series (LTS) data,
signaling a shift away from conventional bi-temporal change detection methods (Coppin
et al., 2004). This shift is due in part to the opening of the Landsat archive to the public
in 2008, which was followed by the development of methods which make maximal use of
the data contained in the Landsat archive (Wulder et al., 2012). Following the opening
of the Landsat archive, the pre-processing of imagery to derive surface reflectance and
mask clouds became operationalised (Masek et al., 2006b; Zhu & Woodcock, 2012), facil-
itating the use of these data for a wide range of applications, including change detection.
Table 3.1 outlines a selection of some change detection methods based on dense LTS ei-
ther by creating annual composite time series (e.g. Kennedy et al., 2010; Griffiths et al.,
2013; Huang et al., 2010) or by exploiting all data available in the archive (e.g. Broich
et al., 2011; Zhu et al., 2012b; DeVries et al., 2015; Dutrieux et al., 2015; Reiche et al.,
48 Tracking post-disturbance regrowth using LTS
2015a).
While a wealth of forest disturbance methods and products based on LTS have been devel-
oped in recent years, disturbance-recovery dynamics are less well understood, especially
in tropical forest systems. An understanding of the fate of forests after a disturbance is
important in order to estimate net changes (Brown & Zarin, 2013) or to elucidate the
drivers of forest change (Kissinger et al., 2012). As in the case of disturbance monitor-
ing, LTS data present an opportunity to describe forest dynamics with much more detail
and certainty than is possible with bi-temporal comparison methods (Kennedy et al.,
2014). A number of studies in temperate and tropical forests have used LTS to describe
disturbance-regrowth dynamics, ranging from classification to temporal trajectory based
methods. Carreiras et al. (2014) monitored disturbance-regrowth dynamics at several
sites in the Brazilian Amazon by classifying near-annual Landsat time series data into
forest, secondary forest or non-forest classes, thereby shedding light on age classes and re-
clearance rates of the forests. Schmidt et al. (2015) measured regrowth in a forest-savanna
landscape in Queensland, Australia by measuring trends in annual minimum NDVI. Cz-
erwinski et al. (2014) monitored sudden and gradual positive and negative trends in a
protected forest in Canada by applying the Theil-Sen slope estimator (Fernandes & G.
Leblanc, 2005; Sen, 1968) paired with the Contextual Mann-Kendall test (Neeti & East-
man, 2011; Neeti et al., 2012) on annual LTS data. The LandTrendR method (Kennedy
et al., 2010), which segments LTS into temporal trajectories, has been demonstrated in a
number of contexts to be useful in describing historical forest disturbance-regrowth pat-
terns (Main-Knorn et al., 2013; Powell et al., 2013; Neigh et al., 2014), an approach which
has also proven valuable in predicting above-ground biomass using LTS (Pflugmacher
et al., 2012; Frazier et al., 2014). The most spatially comprehensive analysis of forest re-
growth was undertaken by Hansen et al. (2013), who mined LTS data globally to produce
a map of global forest loss and gain over the period 2000 to 2012 using a thresholding and
bagged decision tree approach.
The spectral band or index used in the disturbance-regrowth method is an important
determinant of the sensitivity of the method to forest change dynamics. The Normalized
Difference Vegetation Index (NDVI) is one of the most commonly used indices in vegeta-
tion monitoring. While NDVI has been shown to be sensitive to forest change when used
in a time series context (DeVries et al., 2015; Reiche et al., 2015a; Dutrieux et al., 2015),
it performs poorly as a measure of forest cover and structure (Freitas et al., 2005), and
tends to saturate over dense forest (Gamon et al., 1995; Huete et al., 2002). A number
of alternative metrics have been proposed in the forest disturbance monitoring literature,
including the Normalized Burn Ratio (NBR) (Key & Benson, 2006), the Normalized Dif-
ference Moisture Index (NDMI; also known as the Normalized Difference Water Index,
NDWI) (a.J. McDonald et al., 1998; Wilson & Sader, 2002; Gao, 1996), and a range of
metrics derived from the Tasseled Cap transformation (Crist & Kauth, 1986; Crist & Ci-
cone, 1984; Healey et al., 2005; Ahmed et al., 2014; Kennedy et al., 2010, 2012). Indices
3.1 Introduction 49
Table 3.1: Selection of forest change detection methods using LTS data.
Method Study Area Reference(s)
1 LandTrendR - temporal segmentation on
annual LTS
temperate forests (U.S., Europe) Kennedy et al. (2010); Griffiths et al.
(2013)
2 CMFDA - temporal trajectory-based
method for all available LTS data based
on modeled historical time series
Eastern U.S. Zhu et al. (2012b); Zhu & Woodcock
(2014)
3 VCT - change detection on annual Inte-
grated Forest Z-scores (IFZ) derived from
LTS data
U.S. Huang et al. (2010)
4 BFAST Monitor - temporal trajectory
based method for all available LTS data
based on monitoring structural changes in
a monitoring period
tropical forests DeVries et al. (2015); Dutrieux et al.
(2015); Reiche et al. (2015a); Verbesselt
et al. (2012)
5 Global forest change mapping using
thresholding and bagged decision tree
classifiers
Global Hansen et al. (2013)
6 Time series of forest probabilities for for-
est change monitoring
Indonesia Broich et al. (2011)
50 Tracking post-disturbance regrowth using LTS
exploiting difference in reflectance between the SWIR and near infra-red (NIR) regions
of the electromagnetic spectrum (e.g. NDMI or Tasseled Cap Wetness) have been found
to be particularly useful in discriminating forest age classes (Fiorella & Ripple, 1993) due
to their sensitivity to canopy moisture content (Hardisky et al., 1983; Hunt Jr & Rock,
1989; Jin & Sader, 2005).
Most regrowth monitoring algorithms rely on annual or near-annual time series con-
structed either by selecting a representative image or composite of images for each time
period (Carreiras et al., 2014; Kennedy et al., 2012; Czerwinski et al., 2014). As with
disturbance monitoring, the reduction of the temporal resolution of the data can lead
to losses of information as off-season images are excluded from the analysis (Zhu et al.,
2012b). The inclusion of all available data in a time series, on the other hand, gives a
clear advantage for specific monitoring objectives, including near real-time disturbance
monitoring, for example (Verbesselt et al., 2012; Reiche et al., 2015a; Zhu et al., 2012b).
Other monitoring objectives such as post-disturbance regrowth can similarly benefit from
the inclusion of all available data. Temporally dense time series with multiple observa-
tions per season can shed light on phenological dynamics in forests (Schmidt et al., 2015;
Verbesselt et al., 2010), potentially reducing the need for training data in separating
permanent land use change (e.g. forest to cropland) from transient changes (e.g. forest
harvest cycles).
Structural change monitoring methods rooted in the econometrics discipline have been
shown to be useful in describing time series data in rich detail (Zeileis et al., 2005; Leisch
et al., 2000; Chu et al., 1992). Structural change monitoring is based on the presumption
that current observations in a stable time series should approximately follow historically
defined behaviours (e.g. existing seasonal patterns or linear trends). A structural break-
point is defined as the moment at which time series observations deviate significantly
from a previously established model, similar to the change detection methods based on
statistical quality control charts (Brooks et al., 2014). Structural change monitoring has
been demonstrated more recently on remote sensing time series in the form of the Breaks
for Additive Season and Trend (BFAST) family of methods (Verbesselt et al., 2010, 2012).
These algorithms are based on the decomposition of pixel time series into trend, season
and noise components, and have been shown to be robust against sensor noise in 16-
day MODIS composite time series (Verbesselt et al., 2010, 2012) as well as in irregular
LTS (DeVries et al., 2015). However, there are few comparably robust methods that can
monitor post-disturbance forest regrowth. Structural change monitoring can additionally
allow for a more data-driven approach to regrowth monitoring, whereby post-disturbance
time series profiles are continually compared to historical stable forest profiles. By apply-
ing the reverse logic of the breakpoint monitoring functionality of the structural change
monitoring methods, the stability of time series data after a disturbance could conceiv-
ably be monitored by observing the moving sum (MOSUM) of the residuals against a
statistically defined stability boundary (Leisch et al., 2000).
3.2 Study Area 51
The overall aim of this study is to develop and test a robust data-driven method for
describing disturbance-regrowth dynamics in a tropical forest system. Here, we inter-
pret “data-driven” to imply that minimal user input is needed to derive an output. A
robust data-driven method is thus able to ingest all available observations and is reason-
ably insensitive to noisy observations. To this end, we addressed the following specific
objectives:
1. Map disturbances over the period 2000-2013 using all available LTS data.
2. Test the ability of structural change monitoring methods to identify post-
disturbance regrowth using all available LTS data.
We addressed these objectives using a two-staged disturbance-regrowth monitoring ap-
proach. First, we monitored forest disturbances at an annual timescale using BFAST
Monitor (Verbesselt et al., 2012) applied using sequential 1-year monitoring periods (De-
Vries et al., 2015). We then developed a data-driven method based on moving sums
(MOSUM) and structural change monitoring methods (Zeileis et al., 2005; Chu et al.,
1992) to track post-disturbance regrowth processes. We applied these methods Land-
sat NDMI time series to test how close the canopy reflectance trajectory matches that
of the historical forest signature. We tested this approach over a tropical forest system
in Madre de Dios, southeastern Peru, and assessed the accuracy of the results using a
human-interpreter approach (Cohen et al., 2010).
3.2 Study Area
Our study was conducted in Madre de Dios (MDD), a province in southeastern Peru. The
study area is defined by the Landsat scenes found in WRS-II path/row 2/68 and 2/69
and is bound approximately by 11.480◦S, 69.568◦W and 12.795◦S, 68.889◦W (Figure 3.1).
Mean annual precipitation from 2000 to 2014 for the area estimated from data from Trop-
ical Rainfall Monitoring Mission (TRMM) is approximately 2039 mm/y. Rainfall has a
unimodal distribution through the year, with the peak of the wet season occurring around
January, and the dry season in July, with average minimum and maximum monthly rain-
fall estimated at 28mm/month and 371mm/month, respectively.
MDD holds a significant portion of Peru’s share of the Amazon forest system. The
Amazon forest is the largest tropical forest system in the world and boasts some of the
worlds largest and most important stores of terrestrial carbon (Baccini et al., 2012; Saatchi
et al., 2011) and biodiversity (Malhi et al., 2008). For its share, MDD is part of the
Tropical Andes biodiversity hotspot (Myers et al., 2000) and is home to some of the last
remaining un-contacted human groups (Shepard et al., 2010).
Few studies have been undertaken in MDD to quantify forest change. Joshi et al. (2015)
52 Tracking post-disturbance regrowth using LTS
estimated a disturbance rate of approximately 2.3% between 2007 and 2010 over a similar
study area to that described in this study. Post-disturbance regrowth, on the other hand,
has only been qualitatively described (Joshi et al., 2015), and besides national and global
maps of forest gains and losses (Potapov et al., 2014; Hansen et al., 2013), no quantitative
estimates of net forest changes over MDD exist, to the best of our knowledge.
Forest loss in the Amazon has been mainly driven by cattle ranching, conversion to crop-
lands, logging and road construction (Nepstad & Carvalho, 2001; Walker et al., 2009;
Arima et al., 2008; Chavez, 2013). Crop- or pasture-driven deforestation experienced in
the Amazon basin is sometimes followed by regrowth, resulting in high forest change dy-
namics and an abundance of secondary forests (Caviglia-Harris et al., 2014; Feldpausch
et al., 2007; Joshi et al., 2015). The construction of roads in Madre de Dios after the
1960’s has led to rapid changes in the once pristine forest landscape (Scullion et al.,
2014). Specifically, the implementation of the Inter-Oceanic Highway between Brazil and
Peru, designed to facilitate transport and trade between Brazil and the Pacific coast of
Peru, has been of particular concern to the status of forests in the area (Dourojeanni,
2006). In addition to these drivers, a significant rise in gold mining activities as a re-
sult of the global financial crisis in 2008 and a rise in gold prices has fueled an increase
in deforestation close the Madre de Dios River and its tributaries (Asner et al., 2013;
Alvarez-Berr´ıos & Mitchell Aide, 2015). A series of national-scale REDD+ activities have
been initiated in Peru (Peru: Ministerio del Ambiente, 2011; Potapov et al., 2014), in-
cluding several local REDD+ initiatives in MDD (Hajek et al., 2011) aimed at addressing
these drivers of forest change.
3.3 Methods
A general overview of the methods used in this study is shown in Figure 3.2. Each of the
steps taken in the study are described in detail below.
3.3.1 Data acquisition and preprocessing
We downloaded all available level-1 terrain-corrected (L1T) Landsat scenes at WRS-II
path/row 2/68 and 2/69 (Figure 3.1), for which surface reflectance and cloud masks had
already been produced using the LEDAPS (Vermote et al., 1997; Masek et al., 2006b) and
FMASK (Zhu & Woodcock, 2012) algorithms, respectively. After applying the available
cloud masks to each individual scene to mask out clouds and cloud shadows, we computed
the proportion of no-data pixels per scene and removed all scenes with greater than 80%
no-data pixels. As a result of the failure of the Scan Line Corrector (SLC) on board
Landsat 7 ETM+ in 2003, we found that cloud pixels adjacent to masked SLC-off gaps
were frequently missed by the cloud mask. For this reason, we applied a 5-pixel sieve
3.3 Methods 53
Figure 3.1: Study area located in Madre de Dios province in south-eastern Peru. Forest and
non-forest classes from the 1999 benchmark mask are shown as green and grey, respectively.
Landsat scene footprints at WRS-II path/row 2/68 and 2/69 are also indicated in the inset.
on each remaining scene, which succeeded in removing these outliers. Finally, 15 scenes
having serious mis-registration issues were removed from the dataset.
Statistics for the number of clear-sky observations available per year are shown in the
Supplementary Data. After screening of all images, the available data were still reasonably
dense compared to LTS datasets from other areas of the tropics (DeVries et al., 2015;
Reiche et al., 2015a), forming a continuous time series from 1990 to present. After 2003,
the failure of the SLC on board Landsat 7 was largely compensated by a large number
of TM scenes following 2003. The decommissioning of Landsat 5 at the end of 2011,
however, caused a marked decrease in data availability in 2012 and 2013. An average
of 72% and 69% of the total observations per pixel were valid clear-sky observations (no
cloud, cloud shadows or SLC-off gaps) for path/row 2/68 and 2/69, respectively.
54 Tracking post-disturbance regrowth using LTS
Figure 3.2: Flowchart outlining the main steps implemented in this study.
With the remaining scenes, we computed the Normalized Difference Moisture Index
(NDMI), from the short-wave infrared (SWIR) and near infrared (NIR) bands as fol-
lows (Gao, 1996):
NDM I =NIR −S W IR
NIR +S W IR (3.1)
where the NIR and SWIR bands are the fourth and fifth bands, respectively, of the Landsat
TM and ETM+ sensors. We compared disturbance and regrowth results for NDVI, NDMI
and Band 5 and found that NDMI and band 5 gave comparable results for disturbance
detection and NDMI gave the best results for monitoring regrowth (Figure 3.A1). We
thus chose NDMI for both disturbance and regrowth monitoring for the remainder of the
study.
3.3.2 Benchmark non-forest mask
Since our approach to monitoring forest disturbance requires a history period from which
a stable forest time series can be modeled, it is important to begin with a benchmark
non-forest mask to limit the algorithm to pixels which are known to have been forested at
the beginning of the first monitoring period. For this reason, we produced a benchmark
non-forest mask for 1999 using two cloud-free images for path/row 2/68 and 2/69, both
acquired on the 10th of August 1999. We classified the images using the interactive
supervised classification included in the ArcGIS 10.1 (ESRI, inc.) software package. Initial
3.3 Methods 55
land cover classes included bare soil, pasture, urban and water, which were fused into the
non-forest classes. Primary and secondary forest classes were fused into a single forest
class.
3.3.3 Monitoring for disturbances
In this study, we define “disturbance” to include negative changes to the forest canopy
induced by human activity (deforestation or degradation) or natural disturbances. To
monitor disturbance, we followed the approach of DeVries et al. (2015), who applied the
BFAST Monitor algorithm (Verbesselt et al., 2012) on sequential monitoring periods.
Given the relatively high data density from 1990 to present, we began disturbance moni-
toring in the year 1999 and defined 1-year monitoring periods continuing to 2014. Since
the BFAST Monitor algorithm is described in detail in DeVries et al. (2015) and Verbesselt
et al. (2012), only a brief overview is given here. For each sequential monitoring period,
a first-order harmonic seasonal model is derived from all available data prior to the onset
of the monitoring period and projected into the monitoring period. All data in the moni-
toring period are then checked for consistency with this model by way of a Moving Sums
(MOSUM) of the residuals (Zeileis et al., 2005; Verbesselt et al., 2012). The MOSUM is
defined for each time point (t) in a monitoring period as follows:
MOt=1
ˆσ√n
t
X
s=t−h+1
(ys−ˆys) (3.2)
where ˆσis the predicted variance estimated on the history period, nis the number of
observations in the sample, y−ˆyare the residuals in the MOSUM window, and his the
MOSUM bandwidth, which defined as a fraction of the number of observations in the sam-
ple. Following previous studies using the BFAST Monitor method (Verbesselt et al., 2012;
DeVries et al., 2015), we assigned a value of 0.25nto hfor the disturbance monitoring
component of this study. A structural breakpoint is declared when the null hypothesis of
structural stability (i.e. stability of the seasonality pattern) is rejected (Verbesselt et al.,
2012; Zeileis et al., 2005). The decision to reject this null hypothesis is based on a bound-
ary condition which is set according to a 5% probability level following the Functional
Central Limit Theorem (see Leisch et al. (2000) for more information on how this bound-
ary function is computed). In other words, if the time series pattern is stable compared
to the historical data, all residuals and hence also their MOSUM values should fluctu-
ate closely around zero. However, if there is a breakpoint in that particular monitoring
period, the MOSUM values will deviate significantly from zero.
56 Tracking post-disturbance regrowth using LTS
3.3.4 Calibrating a disturbance magnitude threshold
In addition to breakpoints, the disturbance magnitude was computed by taking the me-
dian of all the residuals within each monitoring period. Even though the breakpoints
detected by the BFAST Monitor method are considered robust against noise and other
variability in the time series (Verbesselt et al., 2012), a magnitude threshold can help to
distinguish between positive and negative breakpoints, and between actual forest change
and spurious breakpoints due to outliers in the monitoring period (DeVries et al., 2015).
These outliers can trigger a breakpoint if the residual (numerator in Equation 3.2) over-
comes the ˆσterm in the denominator of Equation 3.2.
From a reference dataset consisting of randomly sampled pixels with breakpoints in years
2005, 2009 and 2012, we found a strong effect of magnitude on the probability of detecting
disturbance (described further in the Supplementary Data). We derived a magnitude
threshold using a Binomial Logistic Regression (BLR) approach. We chose a magnitude
threshold where the probability of detected actual disturbances was approximately 50%.
A magnitude threshold of -0.006 achieved a 50% probability of actual disturbance. Using a
qualitative assessment of various thresholds on the final mapped disturbance, we revised
the final threshold to -0.01, as this threshold was found to further reduce noise in the
disturbance maps. We eliminated all breakpoints with magnitudes exceeding the selected
threshold and took the earliest remaining breakpoint for each pixel as the “final” change
label. We flagged all disturbance pixels for follow-up regrowth monitoring starting at
their final disturbance dates.
3.3.5 Monitoring for post-disturbance regrowth
For all pixels where a disturbance was detected, we continued monitoring for post-
disturbance regrowth. Here, we define “regrowth” to be the establishment of a secondary
forest canopy following a disturbance. We chose NDMI to monitor this process due to
two associated phenomena known to have a visible impact on the short-wave region of the
electro-magnetic spectrum. First, exposed soil reflectance causes high reflectance in this
region, similarly to the red visible band. Re-establishment of a canopy reduces exposure
of the soil and decreases soil reflectance. Second, and most importantly, reflectance in
the short-wave infrared region has been previously shown to be impacted by differences
in canopy moisture contents (Jin & Sader, 2005). The goal of this step was thus to gauge
whether post-disturbance NDMI trajectories reflected moisture and soil reflectance prop-
erties comparable to that of the stable forest model derived in the history period.
Our approach to monitoring post-disturbance regrowth follows a similar logic to the dis-
turbance monitoring method described above. In short, we are interested in knowing
whether values following a disturbance in a time series return to the previously defined
stable state or not. If such a return to stability occurs, we are also interested in knowing
3.3 Methods 57
after how much time this signal recovery occurs. While MOSUM values at a break-
point deviate significantly from zero (crossing a certain critical boundary, as described
above), forest regrowth is expected to result in MOSUM values returning to near-zero
(i.e., crossing back below the same critical boundary). It is important to note that a
return to stability in the time series does not imply that the conditions of the regrowth
forest match that of the previously intact forest. For this reason, our method returns
a regrowth class label but does not quantitatively describe the post-disturbance forest
state.
Figure 3.3 outlines our general approach to regrowth monitoring. We set the beginning of
the regrowth monitoring period to the beginning of the disturbance year and defined the
history period based on the stable portion of all data prior to this date (Bai & Perron,
2003; Verbesselt et al., 2010). The definition of the history period differs from that of
the disturbance monitoring part of our method (which employs all historical observations
to derive a model) because regrowth monitoring was found to be much more sensitive to
noise in the history period (discussed further in Section 3.5.3). Using the model derived
from these data as described in Section 3.3.3, we took the minimum critical value returned
by the boundary function as the critical boundary for regrowth (shown as a green dotted
line in Figure 3.3). This critical boundary is not a sufficient definition of forest regrowth,
however, since na¨ıvely assigning a regrowth label at the moment when the first MOSUM
value crosses below the critical boundary can result in erroneous flagging of regrowth.
To align the statistical regrowth definition with an ecologically sound definition of forest
regrowth, we imposed a number of additional conditions to be met for a regrowth flag to
be assigned. Thus, for a post-disturbance time series at times tB≤t<tN, a regrowth
flag is assigned at time tRif:
1. (tR−tB) is greater than or equal to 3 years (shown as win Figure 3.3);
2. the period after tRin which MOSUM values remained below the critical boundary
is greater than or equal to 1 year (shown as sin Figure 3.3); and
3. at least some of the MOSUM values at times tB≤t < tRexceed the critical
boundary (i.e. not all pre-regrowth values are lower than the critical boundary).
The third condition is especially important in cases where noise in the history period affect
the MOSUM process in the monitoring period. Noise in the history period is represented
by a high ˆσin the denominator of the MOSUM equation (Equation 3.2), resulting in
particularly low MOSUM values. In some cases, these values may never rise above the
critical boundary and would therefore be erroneously interpreted as regrowth.
The size of the monitoring window used in the MOSUM calculation is expressed as hin
Equation 3.2), and the effect of varying his shown in Figure 3.4. We chose a value of
0.5nfor hfor the regrowth monitoring component of this study, as lower values tended
to signal regrowth too early (e.g. h= 0.25nin Figure 3.4), and the maximum value of
58 Tracking post-disturbance regrowth using LTS
Figure 3.3: Demonstration of the MOSUM-based regrowth test. MOSUM values (bottom
plot) after a disturbance (tB) are evaluated against a statistically defined stability boundary
(green line). A regrowth label is assigned at tRif MOSUM values return below the boundary
at least wyears after the initial disturbance and remain below the boundary for an additional
syears.
Figure 3.4: Demonstration of the effect of varying MOSUM window sizes (h) on the response
of MOSUM values during post-disturbance regrowth.
1.0ntended to omit much of the regrowth due to the extended response time.
For pixels where a final regrowth flag was assigned, we monitored for repeat disturbances
by applying BFAST Monitor to the entire time series with the data between the original
breakpoint (tB) and onset of regrowth (tR) removed. We set the monitoring period directly
after the regrowth flag, and if a breakpoint was detected, the pixel was assigned a “re-
clearance” label, indicating a repeat disturbance.
We conducted all analyses using the “bfast”, “bfastSpatial” and “strucchange” packages
in R (R Core Team, 2014; Zeileis et al., 2002; Verbesselt et al., 2012).
3.3 Methods 59
3.3.6 Accuracy Assessment
We assessed the accuracy of the disturbance and regrowth results following a two-stage
sampling design. First, we selected a random sample of pixel clusters with disturbance
labels, with or without follow-up regrowth. We then randomly selected one pixel from
each of the sampled clusters. Additionally, we selected random samples from all non-
disturbance forest pixels with the aim of assessing the omission errors. Given the rela-
tively large size of the non-disturbance class (e.g. with disturbance across 5% of a forest
landscape, 95% of the pixels would belong to the non-disturbance class), a reliable sample-
based estimate of omission errors is very difficult. To aid in the estimation of omission
errors, we assumed that the probability of forest disturbance is highly affected by the
presence of roads (Pfaff, 1999). We thus stratified the non-disturbance class based on a 4
km buffer around the roads in the study area. Half of the non-disturbance samples were
selected from within this buffer, and the other half from the rest of the non-disturbance
class. All samples were selected from a spatial subset spanning both path/rows hav-
ing a sufficiently high density of very high resolution (VHR) RapidEye or GoogleEarth
imagery.
After rejecting sample pixels where interpretation was not possible (e.g. due to lack of
sufficient VHR data), our final reference dataset consisted of 617 pixels derived from the
following strata: 218 pixels from the disturbance class, 188 from the buffered no-change
class and 211 pixels from the no-change class. We removed all disturbance commission
errors and selected the remaining disturbance pixels for follow-up regrowth. After adding
additional similarly sampled disturbance pixels with and without predicted regrowth, the
number of samples assessed for follow-up regrowth was 216. We assessed the sampled
pixels using the human interpreter approach described in the calibration method above
and in Cohen et al. (2010) and DeVries et al. (2015) to derive a reference label for each
pixels. Specifically, we identified disturbances by interpreting the LTS data. While small
disturbances can also be identified using this approach (DeVries et al., 2015), a lack of
VHR or ground data for earlier dates limited our ability to verify some small diffuse
changes. In pixels where we identified disturbances, we used the VHR data to confirm or
reject the presence of a secondary forest canopy following disturbance.
We produced area-weighted confusion matrices and accuracy estimates following Stehman
et al. (2004) and Olofsson et al. (2013) and computed inclusion probabilities for samples
from the disturbance stratum (πs,d) as follows:
πs,d =ncl
NclNp,x
(3.3)
for sample son pixel cluster x, where ncl is the number of sampled pixel clusters, Ncl is
the total number of pixel clusters in output map and Np,x is the total number of pixels
60 Tracking post-disturbance regrowth using LTS
Table 3.2: Count-based confusion matrix for randomly sampled pixels assessed for forest
disturbances. Classes are defined as disturbance (1), no-change within the 4-km roads buffer
(2) and no-change outside the 4-km roads buffer (3).
Reference
123PUA PA TA
Predicted
1 183 19 12 214 0.86 0.85
0.892 25 169 0 194 0.87 0.90
3 10 0 199 209 0.95 0.94
P218 188 211 617 – – –
within x. We computed inclusion probabilities of non-disturbance pixels from both the
buffered and non-buffered strata (πs,nc) as follows:
πs,nc =ns
Ns
(3.4)
for sample s, where nsand Nsare the number of sampled and total pixels, respectively,
in each stratum. We then weighted all samples by the inverse of their respective inclusion
probabilities as described in detail in Stehman et al. (2004). From the resulting area-
weighted confusion matrix, we calculated the total accuracy (TA), producer’s accuracy
(PA; inversely related to omission errors) and user’s accuracy (UA; inversely related to
commission errors) for the disturbance, regrowth and no-change classes.
3.4 Results
3.4.1 Accuracies of the disturbance and regrowth classes
The count-based confusion matrix for the sampled disturbance pixels is shown in Table 3.2
and the area-weighted confusion matrix is shown in Table 3.3. From the area-weighted
confusion matrix, we estimated a total accuracy (TA) for the disturbance class of 91 ±
2.3% (α= 0.05). User’s accuracy (UA) and producer’s accuracy (PA) of the disturbance
class were 84 ±5.5% and 74 ±5.7% respectively. The count-based confusion matrix
for the regrowth class is shown in Table 3.4 and the area-weighted confusion matrix for
the regrowth class is shown in Table 3.5. Here, we estimated a total accuracy of 48 ±
3.9%, with estimated UA and PA for the regrowth class of 83 ±7.0% and 25 ±7.0%
respectively.
To investigate whether these low accuracies were affected by timing of original disturbance,
we post-stratified the samples into “early” and “late” disturbances, using the original
predicted disturbance date as a cut-off. The resulting accuracies for regrowth following
3.4 Results 61
Table 3.3: Area-weighted confusion matrix for samples shown in Table 3.2
Reference
1 2 3 PUA PA TA
Predicted
1 0.16 0.026 0.0047 0.19 0.84 ±0.055 0.74 ±0.057
0.91 ±0.0232 0.032 0.21 0 0.24 0.87 ±0.063 0.89 ±0.064
3 0.027 0 0.54 0.56 0.95 ±0.070 0.99 ±0.070
P0.22 0.24 0.54 1 – – –
Table 3.4: Count-based confusion matrix for randomly sampled pixels assessed for forest
regrowth. Only sample pixels where disturbances were correctly identified are included.
Reference
regrowth no-regrowth PUA PA TA
Predicted
regrowth 102 21 123 0.83 0.64 0.64
no-regrowth 57 36 93 0.39 0.63
P159 57 216 – – –
Table 3.5: Area-weighted confusion matrix for all samples (n= 216) shown in Table 3.4.
Reference
regrowth no-regrowth PUA PA TA
Predicted
regrowth 0.16 0.034 0.20 0.83 ±0.070 0.25 ±0.070 0.48 ±0.069
no-regrowth 0.49 0.31 0.80 0.39 ±0.10 0.90 ±0.040
P0.66 0.34 1 – – –
62 Tracking post-disturbance regrowth using LTS
Table 3.6: User’s accuracy (UA), producer’s accuracy (PA) and total accuracy (TA) of the
regrowth class for samples with original disturbances predicted prior to a given cut-off year.
cut-off UA PA TA n
2002 0.89 ±0.085 0.88 ±0.089 0.82 ±0.092 78
2003 0.86 ±0.091 0.83 ±0.10 0.78 ±0.095 85
2004 0.85 ±0.083 0.70 ±0.10 0.69 ±0.091 111
2005 0.85 ±0.078 0.66 ±0.099 0.67 ±0.085 129
2006 0.84 ±0.074 0.56 ±0.094 0.61 ±0.081 153
2007 0.83 ±0.074 0.47 ±0.090 0.56 ±0.077 173
2008 0.83 ±0.073 0.40 ±0.085 0.54 ±0.074 188
2009 0.83 ±0.071 0.31 ±0.076 0.50 ±0.070 212
Table 3.7: Area-weighted confusion matrix for samples shown in Table 3.4 with original
predicted disturbances prior to 2006 (n= 153).
Reference
regrowth no-regrowth PUA PA TA
Predicted
regrowth 0.40 0.076 0.48 0.84 ±0.074 0.56 ±0.094 0.61 ±0.081
no-regrowth 0.32 0.21 0.52 0.39 ±0.15 0.73 ±0.16
P0.72 0.28 1 – – –
“early” disturbances are shown in Table 3.6. The area-weighted confusion matrix for
regrowth following disturbances predicted before 2006 (n= 153) is shown in Table 3.7.
In this case, we estimated a total area-weighted accuracy of 61 ±8.1%, with user’s and
producer’s accuracies of the regrowth class of 84 ±7.4% and 56 ±9.4%, respectively.
3.4.2 Disturbed area and follow-up regrowth
The final map of disturbances and regrowth from 1999 to 2013 is shown in Figure 3.5.
The right panel shows the onset of regrowth as determined by our regrowth monitoring
algorithm.
The area disturbed per year is shown in Figure 3.6, which also includes the disturbed area
undergoing post-disturbance regrowth. The mean disturbed area per year (including areas
with and without follow-up regrowth) was 4587 hectares, representing approximately 0.5%
of the total forest area in 1999.
3.5 Discussion 63
Figure 3.5: Disturbances (left panel) and follow-up regrowth (right panel) between 1999 and
2014 over the whole study area. The example area shown in Figure 3.7 is shown as a black
box.
3.5 Discussion
3.5.1 Forest disturbance rates in Madre de Dios
The first objective of this study was to map forest disturbances from 2000 to 2013 in the
study area using LTS. Our estimates show that forest disturbances peaked in 2006 and
2008, after which there was a decline in forest disturbances (Figure 3.6). The disturbed
area in which forest regrowth was subsequently detected is shown as a lighter shade in
Figure 3.6. Subtracting the disturbed forest area where regrowth eventually occurred
from the total gross disturbed area, we estimated a net change between 1999 and 2014
of approximately 5.6% of the total initial forest area. While a conclusion of change
dynamics cannot be made by a direct comparison with 2.44% national-scale disturbance
rate estimated by Potapov et al. (2014) over the same time period, we do note that our
rate could be higher due to high omission errors in the regrowth class. These errors result
in a higher proportion of net disturbances when area under regrowth is subtracted from
64 Tracking post-disturbance regrowth using LTS
Figure 3.6: Disturbed forest area per year from 1999 to 2014. The proportion of disturbed
area where regrowth eventually occurred is shown as a lighter shade. The mean disturbed
area per year is shown as a red dotted line.
the disturbance class. Updating these results as more data become available will result in
more of the disturbances detected in these years assigned to the regrowth class and lower
net disturbance rates.
Joshi et al. (2015) reported a disturbance rate of 0.78% per year for the years 2007, 2008
and 2009 over a very similar study area, which is close to our estimated disturbance
rate of 0.65% per year over the same time period. This comparison indicates that our
estimates of disturbed area compare well with independent estimates based on radar
satellite data (Joshi et al., 2015).
3.5.2 Robustness of the disturbance-regrowth monitoring algorithm
A number of studies have demonstrated the ability of the BFAST Monitor algo-
rithm (Verbesselt et al., 2012) to detect disturbances in tropical forest landscapes with
varying accuracies (Reiche et al., 2015a; DeVries et al., 2015; Dutrieux et al., 2015). In
this study, we have shown that given a reasonably dense LTS dataset for MDD, Peru, this
algorithm can detect disturbances with high accuracy. A major advantage of the BFAST
Monitor method is the inclusion of the noise term (ˆσin Equation 3.2) in the MOSUM
computation. Incorporation of this term effectively allows for consideration of remain-
ing contamination, either from unmasked clouds, cloud shadows, or other sensor-derived
noise, and results in a method which is largely robust against commission errors. In other
words, presence of some noise in the history period means that a single noise value in the
monitoring period would not be expected to trigger a breakpoint.
3.5 Discussion 65
The second objective of this study was to extend the principle of “robustness against
noise” to the monitoring of post-disturbance regrowth, whereby the same MOSUM values
used to monitoring for breakpoints are then used to monitor for a return to time series
stability. In following this principle, our study represents the first attempt at monitoring
post-disturbance regrowth using a purely data-driven approach using all available LTS
data. The examples shown in Figures 3.7 and 3.8 demonstrate how a such a data-driven
approach can discriminate between permanent land use change and post-disturbance re-
growth. In the case of permanent land use change (Figures 3.7A and 3.8A), a new seasonal
profile is visible in the post-disturbance time series and MOSUM values do not return to
below the critical boundary. In the case of regrowth (Figures 3.7B and 3.8B), on the other
hand, the post-disturbance time series eventually returns to a profile that somewhat re-
sembles the historical time series, and the MOSUM values accordingly return below the
critical boundary. In this case, it is apparent that even though the time series profile
does not exactly resemble that of the historical time series, most likely due to differences
in the NDMI reflectance profiles between old-growth and young secondary forests (Jin &
Sader, 2005), the MOSUM values are sufficient to signal a general return to “stability”,
interpreted here as forest regrowth.
The demonstrations outlined above highlight some key differences between the
disturbance-regrowth monitoring method described in this paper features and other
methods based on temporal composites (Huang et al., 2010; Kennedy et al., 2010).
While annual composites are useful in detecting long-term changes, including forest re-
growth (Kennedy et al., 2010; Czerwinski et al., 2014), using all observations in an LTS
conveys a number of distinct advantages in disturbance-regrowth monitoring. First, mul-
tiple observations within a year can reveal different temporal patterns, allowing for the
distinction between actual regrowth processes and permanent land use transitions as de-
scribed above. Second, forest disturbances are often associated with discrete events in
time, and the timing of a disturbance event can be captured with greater temporal reso-
lution using multiple observations per year. On the other hand, the reliance on discrete
breakpoints for monitoring disturbances in our study likely results in a loss in sensitivity to
gradual forest decline due to degradation or other long-term disturbance agents, for which
other trajectory-based approaches such as LandTrendR are well suited (Kennedy et al.,
2010). Recent research in a tropical montane forest system has shown that measuring
change magnitude alongside breakpoints can bolster sensitivity to small-scale changes,
although further research is needed to expand this approach to degradation monitor-
ing (DeVries et al., 2015).
3.5.3 Sources of omission and commission errors in regrowth monitoring
For the regrowth class, we estimated a UA of 83%, indicating that commission errors
were reasonably low. However, a PA of 25% for the regrowth class indicates a very high
66 Tracking post-disturbance regrowth using LTS
Figure 3.7: Dates of disturbance (top left), regrowth (top centre) and re-clearing (top right)
for a small area. Images from GoogleEarthTM in 2004 and 2011 are shown for three pixels of
interest (labeled A, B and C). Pixel time series for points A, B and C are shown in Figure 3.8.
incidence of omission errors, which warrants further investigation here. Post-stratification
of the sample revealed that PA has a strong dependence on the timing of the original
disturbance, ranging from 88% for disturbance prior to 2002 to 31% for disturbances
prior to 2009 (Table 3.6). This trend indicates that in most cases, even though regrowth
had been occurring over a number of years, the MOSUM values had simply not reached
the critical boundary yet (Figure 3.9), but may have done so with more observations
following the disturbance. In the example shown in Figure 3.9, a reduction in data
density at the end of the time series, partly due to the lack of TM data after 2011
(also shown in Table S1 in the Supplementary Materials), resulted in a more slowly
descending MOSUM, and hence a delayed regrowth flag. It is debatable whether these can
be considered actual omission errors, since the regrowth algorithm would eventually have
assigned a regrowth label to these pixels, barring any repeat disturbance. An alternative
interpretation of these cases is that the regrowth monitoring algorithm is not sensitive
enough for fast detection of regrowth processes. If earlier detection of regrowth onset is
desired, this can be accomplished by assigning a smaller value for the monitoring window
(hin Equation 3.2 and Figure 3.4) or by relaxing some of the additional criteria (e.g. w
3.5 Discussion 67
Figure 3.8: NDMI time series with fitted history model (top panel) and corresponding post-
disturbance MOSUM time series (bottom panel) for the three pixels shown in Figure 3.7. The
monitoring period is shown as a vertical black line, breakpoints are shown as vertical red lines,
and the MOSUM stability boundary is shown as a green line.
or sin Figure 3.3). Additionally, forest areas could be stratified based on areas where
rapid recovery of the spectral signal is expected, whereby different values are assigned h
based on these strata.
According to our estimates, commission errors in the regrowth class were markedly lower
than omission errors. One source of commission errors observed in the validation process
was related to the effect of a noisy history period on MOSUM values in the monitoring
period. Figure 3.10A shows an example of a commission error, where a disturbance in
the history period (circa 1991) resulted in a high noise term (ˆσ) and accordingly low
MOSUM values. The implication of the low MOSUM values in this case is that while the
68 Tracking post-disturbance regrowth using LTS
Figure 3.9: Example of an omission error due to gradual post-disturbance recovery with
insufficient data in the post-disturbance period to confirm regrowth.
first disturbance in the monitoring period (2000-2001) was correctly identified, the second
disturbance (2004-2005) failed to bring MOSUM values back above the critical boundary.
Removing the all data prior to 1992 in this example resulted in a correctly detected dis-
turbance without regrowth (Figure 3.10B), indicating that this historical disturbance had
a critical impact on the regrowth monitoring process in this example. The history period
used in the regrowth monitoring process was defined based on a statistical estimate of
stability (Bai & Perron, 2003; Verbesselt et al., 2010), rather than using all historical data
as in the disturbance monitoring method, which succeeded in excluding such disturbances
in most cases. With an estimated UA of 83% for the regrowth class, we conclude that
this phenomenon is relatively uncommon, but serves to demonstrate the effect of noise in
the history period on the MOSUM tests.
3.5.4 Validation of forest change products by human interpretation
Validation of high temporal resolution forest change products is a significant challenge,
given the paucity of historical change reference data. Visualization tools using LTS, such
as TimeSync (Cohen et al., 2010), have been developed to address this gap. By using a
similar approach, we were able to derive reference data for our disturbance and regrowth
products. Disturbance samples were generally straightforward to validate, since most
disturbances were a result of clear-cutting (e.g. conversion to cropland or pasture, or
forest clearing due to mining operations), and thus resulted in a discrete change signal in
the RGB images. The post-disturbance regrowth product, on the other hand, absolutely
required high resolution data to confirm the presence of a secondary forest canopy after
disturbance. Without these data, it was impossible to confirm whether an increase in
NDMI in the LTS was a result of forest regrowth or the presence of other vegetation.
One drawback of the reliance on high resolution data was the presence of a large gap in
these datasets: RapidEye data were only available from 2010 to 2013, and GoogleEarth
data were only available for 2004 and from 2011 to 2013. We derived regrowth refer-
3.5 Discussion 69
Figure 3.10: Example of a regrowth commission error, where a disturbance in the history
period affects the MOSUM values used to monitoring regrowth (A). Removing the historical
disturbance from the history period (B) causes an increase in MOSUM values in the monitoring
period, and eliminates the regrowth flag.
ence data by confirming the presence or absence of a forest canopy in the period where
high-resolution were available, and were not able to address the issue of timing of forest
regrowth. Applying our disturbance-regrowth method in an area where reference datasets
including disturbance dates and stand age of regrowth forests are available will allow for
further fine-tuning of the parameters used in our method to detect forest regrowth.
3.5.5 Towards a continuous monitoring framework for net changes
In this study, we monitored and mapped historical disturbance-regrowth dynamics using
a novel data-driven approach. With the ability to track these change dynamics using LTS,
we propose a framework for continuous monitoring of forest disturbance and regrowth.
While we used 1-year monitoring periods to track disturbances in this study, an approach
using single, newly available observations in near real-time (Verbesselt et al., 2012) is easily
envisioned. Within this framework, disturbance or regrowth monitoring algorithms are
employed for each pixel depending on the current forest status of the pixel. For forested
pixels, disturbances are tracked using the BFAST Monitor algorithm, while for previously
disturbed (but not yet recovered) pixels, the regrowth algorithm described in this paper
70 Tracking post-disturbance regrowth using LTS
is evoked. At the moment a breakpoint is detected in the forested pixels, a disturbance
label is assigned. Conversely, when post-disturbance MOSUM values return below the
critical boundary and satisfy the conditions set for regrowth monitoring (Figure 3.3), a
regrowth label is assigned and the pixel is once again considered to be forested. This
framework can be extended to REDD+ MRV and other forest monitoring systems for the
purpose of accounting for net changes in forest landscapes.
Acknowledgements
We would like to thank Christopher Martius and Louis V. Verchot of CIFOR for their
support in carrying out this study, and Nandika Tsendbazar for helpful comments on the
manuscript.
3.A Appendix 71
3.A Appendix
To derive a magnitude threshold, we first randomly sampled unique change pixels from
several monitoring periods. For each pixel, we checked whether deforestation had occurred
and noted the timing following the TimeSync approach (Cohen et al., 2010), adapted
as in DeVries et al. (2015) to include 6.5m resolution RapidEye time series from 2010
to 2013 and GoogleEarthTM imagery from 2004 and 2011 to 2013. Using the reference
change labels, we constructed a Binary Logistic Regression (BLR) model using magnitude
as a regressor and the reference binary change label as the response variable. As such, we
modeled the probability of correctly identifying deforestation according to the form:
P(Y= 1|X) = 1
1 + e−Xβ (3.5)
where Yis the response variable (binary change/no-change label), Xis the continuous
regressor (magnitude) and βis a regression coefficient. The result of the BLR model
is shown in Figure 3.A1. We chose a magnitude threshold such that P(Y= 1|X) was
approximately equal to 0.5 (shown as a red dotted line in Figure 3.A1) and eliminated all
breakpoints with magnitudes exceeding that threshold.
Figure 3.A1: Results of the binomial logistic regression (BLR) for all reference disturbance
pixels sampled from breakpoints in 2005, 2009 and 2012 monitoring periods. Change magni-
tude (M) is shown on the x-axis and the probability of detecting deforestation (P(def|M))
is shown on the y-axis. A threshold at which P(def) = 0.5 is shown as a red hatched line.
The shaded area represents the 95% confidence interval. Change labels assigned by a human
interpreter are shown as hatches where 1 = deforested and 0 = no-change.
Chapter 4
Combining satellite data and
community-based observations
This chapter is based on:
Pratihast, A.K., DeVries, B., Avitabile, V., de Bruin, S., Kooistra, L. & Herold,
M. 2014. Combining satellite data and community-based observations for forest
monitoring. Forests, 5:2464-2489.
74 Combining satellite CBM data
Abstract
Within the Reducing Emissions from Deforestation and Degradation (REDD+) frame-
work, the involvement of local communities in national forest monitoring activities has
the potential to enhance monitoring efficiency at lower costs while simultaneously pro-
moting transparency and better forest management. We assessed the consistency of forest
monitoring data (mostly activity data related to forest change) collected by local experts
in the UNESCO Kafa Biosphere Reserve, Ethiopia. Professional ground measurements
and high resolution satellite images were used as validation data to assess over 700 forest
change observations collected by the local experts. Furthermore, we examined the com-
plementary use of local datasets and remote sensing by assessing spatial, temporal and
thematic data quality factors. Based on this complementarity, we propose a framework
to integrate local expert monitoring data with satellite-based monitoring data into a Na-
tional Forest Monitoring System (NFMS) in support of REDD+ Measuring, Reporting
and Verifying (MRV) and near real-time forest change monitoring.
4.1 Introduction 75
4.1 Introduction
Forests cover approximately 30% of the Earth’s land surface (Hansen et al., 2013) and
are of immense value to humankind, as they provide habitats for a wide variety of species
and play an important role in the global carbon cycle. However, a loss of approximately
2101 square kilometers of tropical forests per year (Hansen et al., 2013) has made a
significant contribution to the increase of greenhouse gases (GHGs) in the atmosphere,
resulting in accelerated global warming (Gullison et al., 2007). To mitigate this effect, the
United Nations Framework Convention on Climate Change (UNFCCC) has proposed an
international mechanism called Reducing Emissions from Deforestation and Degradation
(REDD+) in developing countries (Gullison et al., 2007; UNFCCC, 2009a). The REDD+
mechanism includes reducing deforestation and forest degradation, forest enhancement,
sustainable forest management and conservation (UNFCCC, 2010). Recently, the 19th
Conference of Parties (COP) of the UNFCCC in Warsaw, November, 2013, agreed on a
collection of seven decisions on REDD+ (UNFCCC, 2013). Together with the REDD+
decisions adopted at previous COPs, these decisions provide international policy guidance
(the Rulebook on REDD+) on how countries should deal with REDD+ in the framework
of the UNFCCC (UNFCCC, 2013). Besides reduction of carbon emissions, the REDD+
mechanism also includes establishment of national institutions, ensuring co-benefits and
safeguards and, above all, creating performance-based financing mechanisms (Gullison
et al., 2007; Danielsen et al., 2011).
A country participating in REDD+ requires a reliable, transparent and credible national-
level forest monitoring system (NFMS) for Measuring, Reporting and Verifying (MRV)
activity data and emission factors (GOFC-GOLD, 2010; Sanz-Sanchez et al., 2013; Herold
& Skutsch, 2011). Activity data is defined as the magnitude of human activity resulting
in emissions or removals. In the case of forest-related emissions and removals, activity
data refers to forest area change (generally measured in hectares), whereas the emission
factor is related to the rate of emission of a given GHG from a given source, relative to
units of activity (generally measured in tons of carbon per hectare) (Penman et al., 2003).
Given that forest change is a dynamic process, monitoring needs to be carried out on a
regular basis to support national MRV requirements. Establishing such monitoring sys-
tems is presumed to be expensive for developing countries (GOFC-GOLD, 2010; Romijn
et al., 2012; Visseren-Hamakers et al., 2012; Bucki et al., 2012). An activity monitoring
system should be based on four broad monitoring objectives related to the location, area,
time and drivers of forest change. These objectives should be properly integrated with
monitoring and MRV systems at the national level. Current schemes for monitoring these
activities are based on remote sensing and field measurements mainly from national forest
inventories.
Remote sensing has proven to be very useful for deforestation monitoring at the global,
76 Combining satellite CBM data
national and subnational scale (Hansen et al., 2013; De Sy et al., 2012; Achard et al., 2010;
DeVries & Herold, 2013). However, remote sensing based monitoring of forest degrada-
tion and regrowth still remains problematic (Achard et al., 2006; Defries et al., 2007;
Vargas et al., 2013), due to cloud cover, seasonality and the limited spatial and tem-
poral resolution of remote sensing observations. Enhancing the interpretation of remote
sensing analyses requires substantial ground verification and validation (Strahler et al.,
2006). Accomplishing these tasks through national forest inventory data is expensive,
time-consuming and difficult to implement across large spatial scales (Gibbs & Herold,
2007; Tomppo et al., 2010).
Community-based monitoring (CBM) is an emerging alternative method for forest change
monitoring that promises to be cheaper than conventional monitoring methods (Danielsen
et al., 2011; Skutsch et al., 2014a, 2011; Pratihast et al., 2013). CBM methodologies
can be organized into two main categories: (i) forest carbon stock measurements for
emission factors; and (ii) forest change monitoring for activity data. Results from well-
designed forest carbon measurement studies (Danielsen et al., 2013; Topp-Jørgensen et al.,
2005; Shrestha, 2010; Pratihast et al., 2012) have demonstrated that local datasets are
comparable to professional measurements, while being cheaper to obtain. Furthermore,
CBM can be considered as a tool to empower the local communities and raise awareness
towards better forest management (Palmer Fry, 2011; Lawlor et al., 2013).
While CBM-based forest carbon stock measurement has been shown to be feasi-
ble (Danielsen et al., 2013; Topp-Jørgensen et al., 2005), monitoring of forest change
through CBM has not been thoroughly investigated yet. Forest change monitoring is a
continuous process, which requires continuous data acquisition, and local communities
may act as active in situ sensors (Goodchild, 2007). Their local knowledge could be es-
pecially valuable in signaling forest change activities (deforestation, forest degradation or
reforestation) and providing valuable information, such as location, time, size, type and
proximate drivers of the change events on a near real-time basis (Skutsch et al., 2011). The
impacts of these activities are rarely captured comprehensively in national forest invento-
ries or from remote sensing (Danielsen et al., 2011; GOFC-GOLD, 2010; Pratihast et al.,
2013). The recent development of hand-held technologies continues to improve and has
significantly enhanced the local capacity in data collection procedures (Pratihast et al.,
2012). Data acquired by communities can therefore play an essential role in enhancing
the efficiency and lowering the cost of monitoring activities, while simultaneously pro-
moting transparency and better management of forests. Thus, local participation within
monitoring programmes holds promise for national REDD+ MRV implementation.
Despite the potentials of CBM, the main challenge of using locally collected data lies in
the lack of confidence in data collection procedures (Palmer Fry, 2011). The accuracy
and reliability of such datasets are often questionable due to inconsistencies arising from
the fact that local participants collect data independently of each other. This can further
4.2 Materials and Methods 77
result in incomplete data collection and a biased representation of changes in a study
area (Danielsen et al., 2010). Therefore, data credibility and trustworthiness are major
obstacles to the integration of CBM data in NFMS (Conrad & Hilchey, 2011; Skarlatidou
et al., 2011). This fact has triggered us to rectify the current shortcomings and expand
the current state of knowledge in community-based forest monitoring and its utility in
NFMS. Specifically, we aim to check the consistency of local datasets and investigate their
complementary use to remote sensing. The purpose of this research is to discover new
perspectives and insights into community-based observations. The aims of this paper are
to: (i) present the details of a local expert-based forest monitoring system; (ii) assess
the spatial, temporal and thematic accuracy of local expert data against independent
field-based measurements and high resolution SPOT and RapidEye satellite imagery; and
(iii) explore the complementarity of local expert data with remote sensing data. While
the UNESCO Kafa Biosphere Reserve in Southwestern Ethiopia is shown here as a case
study, the concepts presented in this study are applicable to a broader geographic scope
and can be scaled up to the national level in support of NFMS and REDD+ MRV.
4.2 Materials and Methods
4.2.1 Study Area Description
The study area is situated in the Kafa Zone, Southern Nations Nationalities and People’s
Region (SNNPR), in Southwestern Ethiopia (Figure 4.1). The Kafa Zone is over 700,000
ha in size and was recognized as a Biosphere Reserve by UNESCO’s Man and the Bio-
sphere (MaB) programme in March, 2011. This region is characterized by Afromontane
cloud forest, with approximately 50% of the land cover still forested. Average annual
precipitation in the area is approximately 1700mm, and average annual air temperature
is approximately 19C (Schmitt et al., 2010a). The topography of the Kafa Biosphere
consists of mountains and undulating hills, with elevations ranging between 400 to 3100
m. The forest ecosystem provides an important contribution to the livelihoods of the
people in the area, including wild coffee, valuable spices and honey from wild bees. It
also represents a significant store of forest carbon as above-ground biomass.
4.2.2 Description of the Forest Monitoring System
According to REDD+ monitoring and implementation guidelines, it is important to in-
volve local community groups and indigenous societies to carry out forest monitoring, in
particular if there is any prospect of payment and credits for environmental services (UN-
FCCC, 2013; Lawlor et al., 2013; Stickler et al., 2009). A variety of practical experiences
from developing countries, such as Nepal, Tanzania, Cameroon, India, Mexico, Indonesia,
78 Combining satellite CBM data
Figure 4.1: Study area in the UNESCO Kafa Biosphere Reserve, Southwestern Ethiopia;
local expert observations (black crosses) were compared with a field-based reference dataset
(red circles) and high resolution remote sensing data from the SPOT (footprint shown as a
blue dotted line) and RapidEye (footprint shown as a black dotted line) sensors.
China, Laos, Cambodia and Vietnam, have demonstrated that local communities can play
an essential role in forest monitoring and management programmes (Danielsen et al., 2011,
2013; Topp-Jørgensen et al., 2005; Shrestha, 2010; Pratihast et al., 2012; Ratner & Terry,
2012). However, most of these experiences are limited to carbon stock measurements
in support of REDD+ MRV, with few prescribed field methods for establishing activity
monitoring (forest change) on the ground (Skutsch et al., 2011; Pratihast et al., 2012). In
this study, we present a ground-based system to monitor activity data because of their
increasing importance in the context of REDD+. The following setup was designed to
contribute an efficient and continuous forest monitoring system for the Kafa Biosphere
Reserve.
Selection of local experts. Selection and recruitment of local experts acts as the
backbone for a forest monitoring system, as the success of these CBM systems largely
depends on the knowledge, commitment, feeling of ownership and competencies of these
individuals (Danielsen et al., 2005). The selection process featured in this study is based
on a scheme of collaborative design of monitoring with external interpretation of the data,
one of five schemes of local involvement in monitoring proposed by Danielsen et al. (2009).
A total of 30 local experts were recruited within the frame of the project called “Climate
Protection and Primary Forest Preservation - A Management Model using the Wild Cof-
fee Forests in Ethiopia as an Example” under the Nature and Biodiversity Conservation
Union (NABU). The recruitment was done through the Kafa Zone Bureau of Agriculture
4.2 Materials and Methods 79
and Rural Development (BoARD). The selection was done in such a way that it represents
on average three experts from each of the 10 woredas (administrative units in Ethiopia).
All chosen local experts had at least a secondary level of education and some fundamental
understanding of forest management. This selection procedure was seen as a step towards
greater community involvement in monitoring activities with the representatives involved
from all woredas, assuring the potential for significant enhancement of the monitoring
capacity of the project. Apart from monitoring, these experts also bear responsibilities
for other project activities, such as the development of ecotourism, reforestation, com-
munity plantations, the distribution of energy saving stoves and awareness raising for the
sustainable use of forest resources (e.g., honey and wild coffee).
Data acquisition. Two methods of data acquisition were implemented and tested in
this study. In the first method, paper-based forest disturbance forms with GPS devices
were used by local experts to acquire forest monitoring data. The data collection forms
were designed primarily with project monitoring objectives in mind, but also were com-
pliant with REDD+ MRV requirements. This form focused on capturing forest changes,
including small-scale forest degradation, deforestation and reforestation. In the second
method, mobile devices with integrated GPS and camera functionality were used to in-
crease the ease and simplicity in collection, entering and managing locally acquired data.
For this purpose, a decision-based data collection form (Figure 4.2) was designed in XML
and was deployed on mobile devices using the Open Data Kit (ODK) Collect applica-
tion (Anokwa et al., 2009). This form contains optional input constraints, flows that
depend on previous input, icon-based user-friendly graphics and local language support.
Mobile devices stored the data asynchronously and transferred data to data servers over
GPRS, Wi-Fi or USB, as connectivity was available. An online database management sys-
tem based on ODK Aggregate, postgreSQL and PhP was designed for the proper storage,
analysis and visualization of the acquired data. Further details of the adopted proposed
data acquisition method can be found in Pratihast et al. (2012). A paper-based data
acquisition system was used in 2012, whereas mobile devices were used to collect the data
in 2013. Even though the tools used to acquire data were different, the overall form of the
design was consistent, with a few key differences in terms of multimedia features.
Training and capacity building program. User friendly training materials were
produced for the developed technology and data collection methodology. A series of
training events was conducted before and during the implementation of the monitoring
activities. The main purpose of training was to enhance the capacity of local experts and
to develop approaches and strategies for programme implementation.
80 Combining satellite CBM data
Figure 4.2: Decision-based data acquisition form for local experts; the questions that are
posed in the forms depend on answers given to preceding questions; such a design ensures that
the questions are relevant to the land cover change being described.
4.2.3 Reference Datasets
Local experts are capable of reporting forest change process at a high temporal frequency.
Finding suitable reference data that can thoroughly assess the spatial, temporal and
thematic accuracy of these data is difficult, however. In this study, two types of accurate
reference datasets were acquired to evaluate the accuracy of these local expert data: field-
based reference dataset (FRD) and remote sensing (RS).
Field-based reference dataset (FRD). We conducted a field visit in order to val-
idate the ground data collected by local experts. Due to cost constraints, it was not
possible to visit all locations reported by local experts. We selected six accessible woredas
owing to practical considerations. These woredas contain more than 65% of the local ex-
pert data. Within these woredas, 140 locations (Figure 4.1) were randomly selected and
were revisited during November and December, 2013, by a team of professionals. The
decision-based data acquisition form on the mobile devices (Figure 4.2) was used by the
team of professionals to measure location, size, time, drivers and photographs of change
events.
4.2 Materials and Methods 81
Table 4.1: Summary of the SPOT and RapidEye scenes used in this study.
Sensor Ground resolution Year of acquisition Number of scenes
SPOT4 10m 2005-2006 6
SPOT5 2.5m 2007-2011 8
RapidEye 6.5m 2012-2013 27
Table 4.2: Specific approaches used to assess the spatial, temporal and thematic accuracy of
local expert data.
Category Measured variable local expert data Reference data Measures of accuracy
Spatial Accuracy
Location variables (Qualitative)
Field based Confidence interval (95%)GPS accuracy
Size of forest change
Temporal Accuracy Time of change Remote sensing Time lag
Thematic Accuracy
Presence of forest
Field based Error matrixForest change type
Driver of forest change
Remote sensing (RS). A time series of high resolution remote sensing images ac-
quired between 2005 and 2013 (including pan-sharpened SPOT and RapidEye images)
were available for the analysis of reference data (Table 4.1) in the study area (Figure 4.1).
The SPOT 4 and SPOT 5 imagery have a ground resolution of 10 m and 2.5 m, re-
spectively, whereas RapidEye has a ground resolution of 6.5 m. Locally-reported forest
monitoring locations were visually interpreted based on an approach described by Pohl
and Van Genderen (Pohl & Van Genderen, 1998). Following this approach, images were
systematically examined and pixels representing forest change areas were manually dig-
itized as polygons. The forest change areas were estimated by calculating the polygon
area.
4.2.4 Accuracy Assessment
Several metrics have been proposed by researchers to describe the quality of geographic
data (Haklay, 2010; Devillers et al., 2007; Castro et al., 2013). However, no specific list
of elements with a consistent definition has yet been agreed upon. The latest attempt
to standardize data quality elements was in 2001 with ISO 19113 in 2002 (ISO/TC211,
2002), which proposes the following five elements: completeness, logical consistency, po-
sitional accuracy, temporal accuracy and thematic accuracy. In this study, we limited
the quality assessment to three of these major categories, namely spatial, temporal and
thematic accuracy, since these are essential aspects of forest monitoring datasets (GOFC-
GOLD, 2010). The details of the accuracy measures employed in this study are listed in
Table 4.2.
82 Combining satellite CBM data
Spatial accuracy
In this study, three aspects of the spatial accuracy of the local expert data were assessed,
including categorical location information, GPS location information and the estimated
size of forest change. The categorical location information included categories for repre-
senting the administrative units, like woreda, kebele (administrative sub-unit of a woreda)
and a spatial category representing distance to core forest, nearest village and roads (i.e.,
less than 1 km, 1-2 km, 2-3 km and more than 3 km). To estimate the accuracy of these
responses, comparisons were made between the local expert data and the FRD. From
this sample, the fraction of correct observations in the total population of local expert
reports was estimated using the hypergeometric distribution (Johnson et al., 1992), a
discrete probability distribution that describes the probability of obtaining a correct re-
sponse from a finite population size without replacement. The 95% confidence interval
was calculated by using the 0.025 and 0.975 quantiles of this distribution.
In addition to the categorical location descriptors, local experts provided GPS readings
for each report. Each reading was associated with a measurement error reported by the
GPS receiver. The GPS measurement errors in the local expert dataset were compared
with measurement errors in the FRD using a t-distribution (Johnson et al., 1992). Using
this distribution, the mean bias (with 95% confidence interval) and the standard deviation
between the local expert and FRD GPS errors were calculated.
Finally, the size of forest change polygons mapped by local experts were compared with
change polygons digitized from visually interpreted high resolution SPOT and orthorecti-
fied RapidEye time series imagery. Forty deforestation polygons falling within the spatial
extent of the SPOT and RapidEye time series were selected. The relationship between the
size of field-delineated change areas and polygons digitized from high-resolution imagery
was evaluated using a t-distribution.
Temporal accuracy
Recording the timing of forest change is essential for the implementation of a robust forest
monitoring system. Assessing the temporal accuracy of local monitoring data remains a
challenge due to a lack of reference time series imagery of sufficient temporal density and
spatial resolution that can describe disturbances in near real-time (Kennedy et al., 2007;
Cohen et al., 2010; Schroeder et al., 2011). To overcome this limitation, only the area for
which time series images of SPOT and RapidEye were available (Table 4.1) was used for
this analysis. Here, a visual interpretation of the time series of satellite images for each
local data set was carried out, and the time of forest disturbance was estimated for each
data set. Furthermore, a temporal lag between the reference satellite datasets and local
expert datasets was calculated to determine the average time delay or temporal lag of
4.3 Results 83
deforestation:
L=Tdist,RS −Tdist,LE (4.1)
where Lis the temporal lag (in years) and Tdist,RS and Tdist,LE are the disturbance times
detected by remote sensing and local experts, respectively.
Thematic accuracy
Attributes, such as the presence or absence of forest, forest change type and drivers of
forest change were included in the assessment of thematic accuracy. The accuracy of these
variables was assessed by comparing local expert dataset with the field-based reference
dataset. An error matrix was produced for each category and used to derive producer’s
accuracy, user’s accuracy and the overall accuracy (Foody & Boyd, 2002).
4.3 Results
4.3.1 Characteristics of Local Monitoring Data
Attributes of the local expert monitoring data
In this study, we focused on deforestation and forest degradation processes to illustrate the
major attributes of the data collected by local people (Figure 4.3, Table 4.3. The results
show that local experts have documented forest change processes, which include spatial
(location and size), temporal (time of change events) and thematic (type of change, driver
of change and photograph from the North, East, West and South directions) information.
Furthermore, deforestation, the conversion from forest to non-forest land (GOFC-GOLD,
2010), and forest degradation, negative changes in forest biomass without conversion to
another land cover type, could be mapped separately using data provided by local experts
(Figure 4.3). In this case, local experts tried to delineate exact deforestation areas from
the ground by recording multiple GPS location around the boundary (Figure 4.3a). On
the other hand, forest degradation is a gradual process without a fixed boundary (GOFC-
GOLD, 2010) and could therefore not be mapped with such precision. In such cases, local
experts provided the central location and approximate area affected rather than an exact
change polygon (Figure 4.3b).
Monitoring frequency
During the period of January, 2012, to December, 2013, a total of 755 locations were
observed (Figure 4.4). Of these, 46% were labelled as forest degradation, 25% as defor-
84 Combining satellite CBM data
estation and 30% as reforestation. All data in 2012 were acquired using paper forms with
hand-held GPS devices, whereas in 2013, data were acquired using mobile devices. In gen-
eral, local observations were spread equally over the whole Biosphere Reserve (Figure 4.1).
However, monitoring efforts were not consistent throughout the year (Figure 4.4). Irreg-
ularities in monitoring activities were influenced by a wide range of factors, including
the timing of training and capacity building programmes and adverse weather conditions.
The number of received monitoring forms (in 2012) and digital observations (in 2013)
increased during training and capacity building programme (January to March), while it
decreased during the rainy season (July to September).
Drivers of forest change
Drivers of forest change were mostly associated with agriculture expansion and settle-
ment expansion, followed by charcoal and firewood extraction, intensive coffee cultiva-
tion, timber harvesting and natural disasters, which mainly included landslides erosion
and windfall. Many of the drivers were found to co-occur at a single location (Table 4.4).
In the case of agricultural expansion, 34 of the events were attributed to agriculture ex-
pansion alone, whereas 185 events were attributed to agriculture expansion together with
charcoal and fire wood collection, and 61 of those changes were found to be due to the
co-occurrence of agriculture expansion and timber harvesting. This observation is logical
considering that agriculture expansion in Kafa Biosphere Reserves is in fact a gradual
process coupled with forest degradation. After demarcation of a portion of forest area for
agricultural development, a farmer commonly keeps much of the forest for the first couple
of years to harvest coffee, spices, fuel wood, charcoal and timber, before the forest is fully
cleared to make way for agricultural activities.
4.3.2 Results of accuracy assessment
Spatial accuracy
A breakdown of the estimated fraction correct of assigned spatial categories with a 95%
confidence interval is shown in Table 4.5. The spatial accuracy varied considerably across
the various spatial categories included in the monitoring forms. The woreda was recorded
with the highest mean fraction correct of 0.92, whereas the estimated distance to core
forest was found to have the lowest mean fraction correct of 0.71.
A comparison of GPS errors reported by local experts with those reported in the FRD
showed a slight systematic error of 0.65 m between the two datasets (Table 4.6). A
similarly slight bias was found between forest change areas as reported by the local experts
and forest change areas derived from high resolution remote sensing imagery, in cases
where these areas did not exceed 2 ha (Table 4.6). In larger change areas (exceeding
4.3 Results 85
Figure 4.3: Examples of (a) deforestation monitoring and (b) forest degradation monitoring
by local experts. Observations were mapped either as polygons (a) or point (b) features,
depending on the process being described; each form was accompanied by four photos rep-
resenting the north, east, south and west perspectives. The attribute tables associated with
these observations are shown in Table 4.3.
Table 4.3: Attribute tables derived from local expert observations of deforestation and forest
degradation (shown in Figure 4.3a and b, respectively).
Category Measured Variables Value of Deforestation
(Figure 4.3a)
Value of Forest Degrada-
tion (Figure 4.3b)
Spatial
Woreda Gewata Gewata
Kebele Ganity Ona
Distance to Road More than 3km 1-2km
Distance to Nearest Vil-
lage
1-2km 1-2km
Distance to Core Forest More than 3km More than 3km
GPS coordinates (Lat,
Lon)
(7.53, 35.84) (7.54, 35.81)
Temporal Disturbance Date 03-18-2013 03-18-2005
Thematic
Disturbance Type Deforestation Forest degradation
Driver of Disturbance Agricultural expansion,
timber harvesting and
fuelwood
Coffee cultivation, timber
harvesting and fuelwood
Size of Disturbance 2 ha 4 ha
2 ha), however, the absolute bias increased to 1.06, implying that local experts had
systematically underestimated the area of large change polygons.
86 Combining satellite CBM data
Figure 4.4: Number of observations collected by local experts in 2012 and 2013; all obser-
vations in 2012 were acquired using an analogue (paper-based) system, whereas observations
acquired in 2013 were collected using either analogue or digital (smart phone-based) meth-
ods. Periods following training and capacity building activities (*) and rainy periods (+) are
indicated on the plot.
Table 4.4: Number of instances of the co-occurrence of forest change drivers. Numbers along
the diagonal indicate the number of instances that a particular driver was reported alone. AE:
Agricultural expansion. SE: settlement expansion. CF: Charcoal and fuelwood. IC: Intensive
coffee cultivation. TH: Timber harvesting. ND: Natural disaster.
AE SE CF IC TH ND
AE 34
SE 48 42
CF 112 75 57
IC 0 55 76 19
TH 61 70 44 10 15
ND 13 17 2 1 2 2
Table 4.5: Fraction correct of local data assignment to spatial categories.
Spatial category Fraction correct
Woreda 0.92 ±0.04
Kebele 0.78 ±0.06
Distance to nearest village 0.77 ±0.06
Distance to nearest road 0.75 ±0.06
Distance to core forest 0.71 ±0.06
4.3 Results 87
Table 4.6: Positional accuracy of local expert data.
Measure Mean bias Standard deviation
GPS error (m) 0.65 ±0.03 1.79
Size of forest change (<2ha) 0.16 ±0.03 0.29
Size of forest change (>2ha) -1.06 ±0.12 1.26
Temporal accuracy
Each forest change event was recorded by local experts with a time stamp that represents
the time at which the process of change took place. In total, 40 deforestation and 60
degradation locations were visually assessed from high resolution remote sensing (SPOT
and RapidEye) imagery. An example of the visual interpretation of high resolution time
series of SPOT5 (2008-2010) and RapidEye imagery (2012-2013) is shown in Figure 4.5.
The locally mapped polygon is displayed at the center of each subset of image. The
interpretation shows that the forest cover was significantly reduced after 2012.
The histogram of the temporal accuracy of locally determined change dates compared to
high resolution imagery for deforestation and forest degradation is shown in Figure 4.6.
Here, a positive temporal lag indicates that local experts indicated a change date earlier
than that determined using remote sensing data, and a negative time lag indicates the
reverse situation. The results reveal that 33% of deforestation events reported by local
experts corresponded accurately to the dates observed in the remote sensing data. In other
cases, 25% and 20% of total deforestation events as observed from remote sensing were
detected one and two years earlier than the local reported time, respectively (Figure 4.6).
On the other hand, the comparison of dates associated with forest degradation as reported
by local experts shows that the majority of these signals were recorded one (32%) to two
(22%) years earlier than dates detected by remote sensing.
Thematic accuracy
Thematic information is one of the added values of the local expert dataset compared
to remote sensing. Summaries of the accuracy assessment of three thematic elements
(the presence of forest, forest change type and drivers of forest change) are shown in
Table 4.7.
The results show an overall accuracy of 82% for thematic elements compared to the field-
based reference dataset. The presence of forest was found to have a producer’s accuracy
of 92%, a user’s accuracy of 93% and an overall accuracy of 94%. The drivers of forest
change had a comparatively lower producer’s accuracy of 71%, a user’s accuracy of 68%
and an overall accuracy of 69%.
88 Combining satellite CBM data
Figure 4.5: Example of visual interpretation to assess the temporal accuracy of the local
expert dataset; the image subset is based on SPOT5 data from 2008 to 2011 (red = Band 3,
green = Band 1, blue = Band 2) and two RapidEye images from 2012 and 2013 (red = Band
3, green = Band 2, blue = Band 1); a ground photograph taken by a local expert in 2013 is
also shown; the red polygon is the forest change mapped by a local expert; the forest change
occurred between 2012 and 2013.
Figure 4.6: Histogram of time lags in capturing deforestation (left) and forest degradation by
remote sensing (SPOT and RapidEye) imagery (right); a time lag is defined as the difference
between change dates observed from remote sensing image interpretation and those dates
recorded by local experts.
4.4 Discussion 89
Table 4.7: Accuracy assessment of local expert data compared to field-based reference dataset
in the thematic domain.
Elements User’s accuracy Producer’s accuracy Overall accuracy
Presence of forest 93% 92% 94%
Forest change type 83% 84% 83%
Driver of forest change 71% 68% 69%
4.4 Discussion
4.4.1 Local Expert-Based Forest Monitoring System
The establishment of robust and reliable NFMS in developing countries is an expensive
and challenging task. Several studies have shown that CBM has the potential to increase
the saliency, credibility and legitimacy of such forest monitoring systems (Danielsen et al.,
2011, 2013; Topp-Jørgensen et al., 2005; Shrestha, 2010; Danielsen et al., 2014). However,
current studies do not clearly describe the following aspects of forest change monitor-
ing (related to activity data): (1) the long-term operational procedures of community
involvement; (2) technology selection; (3) consistency of local datasets; and (4) comple-
mentarity with remote sensing data. In this regard, we demonstrate an operational forest
monitoring system that includes local expert activity monitoring data in the UNESCO
Kafa Biosphere Reserve, Southern Nations, Nationalities and People’s Region (SNNPR),
Ethiopia. In general, our monitoring setup allows local experts to collect forest change
variables, such as geo-location, size of forest change, time of forest change and proximate
drivers behind the change, in more detail. Similar to previous studies (Pratihast et al.,
2012; Bowler et al., 2012), we also found that the use of mobile devices has a clear ad-
vantage over a paper-based system in capturing photographs and multimedia information
from the ground and improves the local capacity in data collection, transmission and
visualization procedures (Figure 4.3 and Table 4.3). Furthermore, our results show that
these datasets are fully structured in terms of spatial, temporal and thematic detail and
capable of describing the forest change process well. While our results are based on a local
case study, these monitoring activities have the potential to be scaled up to the national
level and integrated with an NFMS.
The local expert-based forest monitoring system in this study faced some critical barriers,
such as systematic coverage and consistency in monitoring frequency. Our results show
that 53% of the local data were collected within 1 km of the local road network, hindering
systematic coverage of the study area. This restriction is a result of poor road infras-
tructure or a lack of transportation means. A recent study in Southwestern Ethiopia has
shown that most forest change occurs in remote locations far from urban areas (Getahun
et al., 2013), suggesting that much of these changes could not be fully captured by local
90 Combining satellite CBM data
experts alone. This mobility barrier could be overcome by engaging local communities
who live near the forest areas of interest.
We also observed that the frequency of local data collection depends largely on weather
conditions and motivations towards monitoring activities. A decrease in data acquisition
was seen during the rainy season, indicating that weather has a significant impact on the
mobility of local people. This reduction in data frequency may also be due to a decrease
in disturbance activities by farmers during this time. The motivation can be triggered by
providing local experts with adequate incentives for conducting monitoring activities even
during adverse weather conditions and also providing them with the necessary accessories
and travel means. Regular training and capacity building programmes should also be
conducted to keep the local experts updated. While such initiatives in motivating the
local experts towards efficient monitoring may not fill the data gap completely, they
could help to substantially increase the commitment and long-term engagement of local
people towards monitoring.
4.4.2 Critical Review of the Accuracy of Local Datasets
In this study, we assessed the spatial, temporal and thematic accuracy of the local expert
dataset. Identifying the factors influencing these accuracies is important to understanding
the role that this dataset can play in a forest monitoring system. The main influencing
factors are explained in detail below.
Spatial accuracy
Spatial accuracy was influenced by three main factors: interpretation of administrative
boundaries, GPS errors and failure to map full polygons. First, the administrative bound-
aries are not always visible on the ground. Local experts may incorrectly interpret these
boundaries when they are away from their own villages. This error might be solved by
providing base maps prepared by an Ethiopian mapping agency and regional governments
during field work, which may contain the updated information regarding these adminis-
trative layers.
Second, GPS location error arises due to the weak signal caused by dense forests and high
slopes. Mobile devices used in this study achieve maximum GPS accuracy by taking the
average measurement from all available satellites reached in a given time. GPS accuracy
could be improved by using averaging positional measurements over a longer period of
time (Sigrist et al., 1999).
Third, the area of change estimated by local experts was found to be biased due to
difficulties in mapping large change polygons in the field. When an insufficient number
of polygon vertices was mapped by the local experts, resulting polygons were smaller
4.4 Discussion 91
than those delineated by visual interpretation from remote sensing imagery, giving rise
to a negative bias in field-based area estimations. These errors could be avoided by
implementing a visualization feature in the mobile device-based forms, whereby local
experts can see the polygon they have mapped while in the field. Based on observed errors
that arise in the mapping process, these can be corrected by the local experts.
Temporal accuracy
To assess temporal accuracy of the local dataset, temporal lag was calculated based on
forest disturbance dates determined using remote sensing time series data. The temporal
lag in detecting deforestation and degradation (Figure 4.5) is not necessarily a direct result
of inaccuracies in the local dataset, but rather highlights differences in the interpretation
of change between ground-based and satellite-based methods in the case of deforestation
and forest degradation.
Evidence from our study indicates that deforestation is detected earlier using higher res-
olution SPOT and RapidEye imagery compared to local expert observations. This time
lag in deforestation detection is likely due to differences in the interpretation of change
events. Since optical remote sensing observes changes in the canopy cover of forests,
changes delineated by visual interpretation of remote sensing time series were directly
related to land cover changes. Local experts, on the other hand, reported changes in
land use (e.g., the conversion of forest land to agricultural land; Verburg et al., 2011).
The difference between the land cover and land use-based definition of deforestation is
important in this case, because actual land use change typically follows several years of
gradual canopy cover change. Whereas deforestation was understood by local experts
to mean the conversion of forest land to cropland, changes in the canopy cover in the
years preceding this change were often interpreted as land cover change (deforestation)
by the remote sensing analyst, thus giving rise to the temporal lag observed in this study
(Figure 4.6).
Interestingly, a reverse temporal lag was found in the case of forest degradation reported
by local experts. Optical remote sensing data are known to have limitations with re-
gards to the detection of low-level degradation, especially when driven by fuelwood col-
lection (Skutsch et al., 2011), as was found in this study (Table 4.4). This low level
degradation generally takes place underneath the forest canopy and is thus not detectable
using remote sensing data until degradation rates are such that canopy openings begin to
appear. For this reason, a delay in degradation detection by remote sensing was found in
this study. In many cases, low-level degradation is not at all detectable with optical re-
mote sensing data when degradation fails to result in canopy openings. In this case, local
datasets convey a clear advantage when combined with remote sensing data to achieve a
comprehensive description of the degradation processes.
92 Combining satellite CBM data
Thematic accuracy
While analysis of the thematic accuracy of the local expert dataset showed a high overall
accuracy (82%), the drivers of forest change were reported with a relatively lower accuracy
(69%). One possible explanation for this lower accuracy could be due to differences in
perceiving the proximal drivers of forest change by local experts and the team of profes-
sionals who were involved in collecting FRD. Another explanation for this lower accuracy
could be the complexity of multiple drivers and dynamic nature of land use changes, which
make categorization of forest change drivers difficult. In the case of Ethiopia, multiple
drivers, such as fuelwood extraction, grazing, timber harvesting and agriculture expan-
sion, operate together, and choosing the most prominent driver for such a situation is
difficult (Table 4.4). The reporting of drivers could be improved through improved form
design (e.g., using simplified classes and iconography).
4.4.3 Role of Local Datasets in an Integrated Monitoring System
Complementarity with remote sensing analysis
The local data stream presented in this paper is not an investigation to replace or com-
pete with remote sensing-based monitoring data, which is conventionally used in forest
area change analyses, but is rather envisioned to be complementary to these data. The
complementarity between remote sensing and community-observations is described below
in the context of several key REDD+ MRV questions (Figure 4.7).
The first question for REDD+ MRV is the location of change. Remote sensing approaches
are highly suitable for answering this question. The value of remote sensing data and their
successful implementation to monitor forest change on various scales (global, regional, na-
tional, etc.) and at various resolutions is well established (Hansen et al., 2013; Achard
et al., 2010). The advantages of these methods include consistent data acquisitions, au-
tomated data processing and large area coverage (De Sy et al., 2012; Roy et al., 2014;
Hansen & Loveland, 2012). A main shortcoming is the need for spatially-explicit ground
(in situ) data to enhance the reliability of these remote sensing products (Li et al., 2013).
There is always a lack of spatially-explicit and statistically representative ground data,
because this information is expensive and time consuming to acquire. To address this defi-
ciency, local data streams proposed in this study may provide a useful way to complement
remote sensing data. The spatial accuracy results of the local expert data (Tables 5 and
6) show that local datasets can be used to better understand information related to local
administration (e.g., the name of the district and village) or geographical characteristics
(distance to roads, nearest village and core forest). Similarly, remote sensing may help to
add value to local data streams by providing wall-to-wall coverage, which can be used to
validate local data streams. The synergies of both methods may lead to a more efficient
4.4 Discussion 93
Figure 4.7: Contributions of remote sensing and community-observation for REDD+ MRV
monitoring objectives related to location, size, timing and drivers of forest change; black arrows
indicate a very strong contribution; dark grey arrows indicate a reasonably strong contribution;
and light grey arrows indicate a limited contribution to these monitoring objectives.
monitoring system for data acquisition and to rendering reliable information.
The second REDD+ MRV question is the area of forest change. Both remote sensing and
local datasets have their own difficulties when used to map the area of forest change. In
general, remote sensing plays a promising role for mapping larger areas, because of its
ability to map wall-to-wall changes (Achard et al., 2010). However, the trade-offs between
the spatial and temporal capabilities of remote sensing limits their use to monitoring
small-scale forest change (De Sy et al., 2012). Since we have shown that local datasets
are sufficiently accurate to track small forest changes, the overall mapping of forest change
area can be enhanced by exploiting the synergy between these datasets.
The third REDD+ MRV question is related to the timing of forest change. Historical
archives of remote sensing imagery and the prospect of a continuous data stream based on
new satellites, such as Landsat 8 and Sentinel-2, offer a possibility to analyze the temporal
patterns of forest change and the impact of human activities (Hansen & Loveland, 2012;
Drusch et al., 2012). However, the temporal accuracy of detected changes based on this
imagery depends on: (1) the availability of cloud-free observations; (2) the seasonality
and climate trends; and (3) the spatial scales of land cover change phenomena. In areas
with high persistent cloud cover, the detection of actual changes can be delayed due
to missing observations, and the seasonality of vegetation can obscure actual changes.
Climate events, such as major droughts, can result in temporal signals that resemble
actual change, thus contributing to errors. Finally, the scale of change can influence the
time at which a change is detected from space. Specifically, we have seen in this study
that higher resolution SPOT and RapidEye imagery detect deforestation earlier than
local experts, whereas the detection of forest degradation using remote sensing data is
94 Combining satellite CBM data
delayed compared to that of local experts. Reports of small-scale deforestation and forest
degradation from local experts can therefore contribute to an improved understanding
of change processes, and the integration of both methods should lead to a more efficient
system to signal new changes in near real-time.
The final REDD+ MRV question is related to the driver of forest change. NFMS for
REDD+ needs to be designed to track and completely document the drivers of forest
change processes (UNFCCC, 2013). Drivers vary across regions (Hosonuma et al., 2012),
leading to different dominant forest change processes and different approaches needed to
tackle these drivers (Skutsch et al., 2011). In general, remote sensing has limited capa-
bilities to track forest change drivers, whereas community-observations are very accurate
in reporting these drivers. These drivers of change can be better understood with an
intimate knowledge of forest change processes, and this information has the potential
to enhance the pertinence of the remote sensing data analysis. Information on drivers
collected by local experts thus presents new opportunities for monitoring forest change
events.
Link to the National Forest Monitoring System: “Up-Scaling”
The UNFCCC encourages developing countries to establish an NFMS in support of
REDD+ MRV (UNFCCC, 2013). The NFMS needs to monitor forest carbon and changes
in compliance with the five IPCC principles: consistency, transparency, comparability,
completeness and accuracy (Herold & Skutsch, 2011; Penman et al., 2003). However,
most developing countries have a low monitoring capacity, and the development of these
capacities will take considerable time and resources (Romijn et al., 2012). In this research,
we found that local communities can monitor forest changes in a cost-effective way. By
scaling up CBM activities to the national level, these capacity gaps can be addressed in
an efficient and cost-effective way. Developing countries should therefore give priority to
CBM in developing their NFMS and MRV systems.
The UNFCCC REDD+ also offers an opportunity for safeguards, biodiversity conserva-
tion and other ecosystem services beyond carbon sequestration (Dickson & Kapos, 2012;
Chhatre et al., 2012; Torres & Skutsch, 2012). Monitoring all of these elements within
REDD+ is a challenge. Our proposed local monitoring system is based on well-established
monitoring principles and experiences. The main advantage of the system is the flexibil-
ity in design. The data acquisition side of the system can be easily modified, and it can
incorporate other types of environmental monitoring variables. Thus, the integration of
other environmental monitoring variables may lead to long-term benefits (Defries et al.,
2007) and shape the future of REDD+ monitoring and implementation efforts (Visseren-
Hamakers et al., 2012).
4.5 Conclusions 95
4.4.4 Future Research Directions
Although our study is founded on the argument that considerable progress can be made
towards community-based forest monitoring in REDD+, there is a clear need for improve-
ments to the monitoring set-up. The first area of improvement is the engagement of local