Content uploaded by Rodrigo A. Carrasco
Author content
All content in this area was uploaded by Rodrigo A. Carrasco on Apr 30, 2024
Content may be subject to copyright.
Is the change deforestation? Using time-series analysis of satellite data to1
disentangle deforestation from other forest degradation causes2
Ignacio Fuentesa,b,∗, Javier Lopatinc,d, Mauricio Galleguillosc, Andr´es Ceballos-Comissoc, Susana3
Eyheramendyc, Rodrigo Carrascoe
4
aFacultad de Medicina Veterinaria y Agronom´ıa, Universidad de Las Am´ericas, Sede Providencia, Santiago, Chile5
bN´ucleo en Ciencias Ambientales y Alimentarias (NCAA), Universidad de las Am´ericas, Santiago, Chile6
cFacultad de Ingenier´ıa y Ciencias, Universidad Adolfo Ib´a˜nez, Santiago, Chile7
dFundaci´on Data Observatory, Santiago, Chile8
eInstitute of Mathematical and Computational Engineering & School of Engineering, Pontificia Universidad Cat´olica de9
Chile, Santiago, Chile10
Abstract11
Protecting natural ecosystems requires monitoring approaches that work as early warning systems to avoid12
degradation and protect biodiversity. However, separating forest disturbance causes in change-detection13
pipelines is challenging due to the complex interplay of multiple drivers affecting vegetation. This study14
aims to detect deforestation in highly heterogeneous ecosystems. We used Landsat NDVI time-series data for15
testing three unsupervised change detection methods: 1) the non-parametric phenological anomaly detection16
(npphen), 2) the continuous change detection and classification (CCDC), and 3) the pruned exact linear time17
(PELT) algorithms. We used visual interpretation of Google Earth Pro high-resolution data (<10 m) to18
depict deforestation, and natural-induced changes, like forest browning and fires, evaluating the performance19
of the unsupervised methods. Additionally, a Random Forest model trained with the outputs from detection20
algorithms together with elevation and radar vegetation index data were utilised to depict deforestation in21
a second step. While PELT slightly outperformed other methods for tracking general vegetation changes,22
with overall accuracies (OA) ranging from 0.78 to 0.99, depending on the vegetation type, it also showed23
the slowest deforestation tracking response. CCDC presented the fastest response and an OA between 0.7824
and 0.95. Additionally, we observed a mean OA of only 0.47 when separating deforestation from other25
changes using only the unsupervised models. On the other hand, deforestation was accurately detected (OA26
= 0.93; kappa = 0.83) when using CCDC outputs within a secondary supervised classification, agreeing with27
selected citizen-based complaints from the Environmental Superintendence. The relatively fast response in28
deforestation tracking using CCDC makes it a viable alternative for near real-time monitoring. Commonly29
used unsupervised detection methods may be coupled with supervised techniques to depict vegetation change30
sources robustly. This application constitutes a step forward for managing and monitoring vegetation areas31
in highly complex and dynamic landscapes, like Mediterranean ecosystems.32
Keywords: Deforestation, Remote Sensing, Temporal Segmentation, Structural Breaks33
Preprint submitted to Remote Sensing Applications: Society and Environment March 20, 2024
1. Introduction34
Monitoring forest ecosystems is crucial for addressing global change and improving environmental mon-35
itoring to mitigate ecosystem impairment (Parr et al.,2003;Kofinas,2009;Bakker and Ritts,2018;Dong36
et al.,2019). Forests are essential components of the biosphere functioning and deliver relevant ecosystem37
services, including the provision and regulation of water and carbon (Bengtsson et al.,2000;Nadrowski38
et al.,2010;Masek et al.,2015;Song et al.,2016;L¨of et al.,2019), nutrient cycling (Attiwill and Adams,39
1993;Longo et al.,2020), air purification (Song et al.,2016;Bottalico et al.,2017), biodiversity maintenance40
(de Oliveira Roque et al.,2018), climate regulation (Figueroa and Pasten,2015), and recreation (S´anchez41
et al.,2021).42
Deforestation, a global environmental issue, is causing the reduction and shrinking of forest areas (Ribeiro43
et al.,2011;Hansen et al.,2013;Keenan et al.,2015;Payn et al.,2015;Leblois et al.,2017;Ritchie and Roser,44
2021). Deforestation implies changes from forests to other land use classes, which humans or other agents45
can cause (FAO,2022). Deforestation is driven by various factors, with urban expansion and agricultural46
activities being prominent drivers in many regions (DeFries et al.,2010;Leblois et al.,2017). The conversion47
of forested areas into urban landscapes or agricultural land leads to habitat loss, fragmentation, and the48
disruption of ecological processes (Ribeiro et al.,2011;Zemp et al.,2017). These anthropogenic activities49
significantly affect biodiversity conservation, ecosystem functioning, and sustainable land use (Barlow et al.,50
2016;Rocha-Santos et al.,2020). While extensive research has focused on deforestation in the Amazon51
rainforest due to its profound impact on global climate and biodiversity (Zemp et al.,2017;Ferrante and52
Fearnside,2020;Silva Junior et al.,2021), it is essential to recognise the ecological, economic, and social53
importance of forests in other regions (Myers et al.,2000;Hamilton and Friess,2018). However, there are54
also natural-based drivers of forest and vegetation loss, like fire regimes and browning due to severe and55
prolonged droughts (Garreaud et al.,2020;Miranda et al.,2020;Smith-Ram´ırez et al.,2022).56
At the landscape level, severe water deficits over the last decades (Garreaud et al.,2020;Fuentes et al.,57
2022a), and global warming (Boisier et al.,2016) have led to forest browning (Miranda et al.,2020) and58
tree mortality in some regions with prolonged droughts (Matskovsky et al.,2021). Likewise, extensive forest59
fires are increasing in frequency due to climate change, affecting forest ecosystems and plantations in large60
regions of the globe (Castillo et al.,2020;Canadell et al.,2021;Mansoor et al.,2022), reducing vegetation61
health and vigour. Since all these processes share a decline in vegetation, disentangling deforestation from62
other causes of forest degradation becomes challenging (Sebald et al.,2021). However, it is essential for63
policy-making and management at the landscape level.64
Given the urgency in delivering new monitoring strategies to track forest changes, there has been an65
∗Corresponding author
Email address: ifuentes@udla.cl (Ignacio Fuentes)
2
overtaking development of new advances and technologies that allow the continuous monitoring of forested66
areas to assess forest changes nationwide rapidly (Hansen et al.,2016). However, synoptic and regular data67
are needed to achieve a large-scale monitoring approach (Danielsen et al.,2009). Hence, remote sensing data68
have gained increasing popularity for automatic change detection of Earth’s changes as it allows the periodic69
monitoring of the surface characteristics (Fuentes et al.,2019;Japitana and Burce,2019;Weiss et al.,2020;70
Fuentes et al.,2022b,2024). For example, optical satellites with multispectral instruments evaluate changes71
within the visible and near-infrared wavelength range of the electromagnetic spectrum (Thakur et al.,2020;72
McAllister et al.,2022), while radar satellites use microwave wavelengths (Karthikeyan et al.,2020), offering73
the advantage of cloud penetration and monitoring during all seasons (Kerr et al.,2001;Filipponi,2019).74
Machine learning, segmentation, or statistical approaches have been commonly used to assess changes and75
tendencies in remotely sensed time series data (Lary et al.,2016;Yin et al.,2018).76
Different methods employ distinct architectures to trace breaks or changes in forest time series data77
(Zhu,2017;Housman et al.,2018;Asokan and Anitha,2019). In many cases, changes must demonstrate78
continuity/consistency within the time series and not simply represent a single anomaly, which could be79
caused by clouds, cloud shadows, floods, or artefacts in satellite scenes (Puhm et al.,2020). Detecting80
consecutive outliers can ensure the presence of a structural break in the time series, but this often leads to a81
delay in precisely defining the break, hindering real-time change detection (Verbesselt et al.,2011). Structural82
breaks can be defined as statistically significant changes in time series, implying changes in trend, mean,83
or variance (Muthuramu and Maheswari,2019;Loginova and Mann,2022), and may result in the temporal84
segmentation of a time series (Pasquarella et al.,2022). Alternatively, other change detection methodologies85
may utilise phenological curves that combine multi-annual data into day-of-the-year (DOY) series to detect86
outliers based on observations within the same season, considering historical observations of a particular87
month or week instead of consecutive ordinal observations. These approaches require prior knowledge of88
vegetation behaviour during a period of non-disturbance and seasonal climatic conditions (Estay and Ch´avez,89
2018;Zeng et al.,2020), especially as interannual phenological variation changes between vegetation types90
(Lopatin,2023) and thus can affect the stability of the approach.91
Several algorithms for land change detection have been tested in deforestation studies. Cai et al. (2023)92
applied the continuous change detection and classification (CCDC) algorithm to track forest changes with93
high accuracies. Schultz et al. (2016) applied the break detection for satellite time series data (BFAST) in94
combination with different vegetation indices to track deforestation with robust results. Similarly, Schultz95
et al. (2015) evaluated error sources in deforestation detection using BFAST, highlighting the necessity of96
fine-tuning the algorithm in some areas. LandTrendR is another temporal segmentation algorithm that has97
proven proficient in tracking annual forest disturbances (Cohen et al.,2018;Pasquarella et al.,2022). The98
vegetation change tracker (VCT) algorithm has also been developed to track forest disturbances with annual99
or biannual satellite data (Huang et al.,2010) and has been tested for deforestation in the Yunnan province100
3
of China, with accuracies of up to 82.7% (Pang et al.,2013). However, little consensus has been found in101
the results of all these change detection methods. Cohen et al. (2017) compared different change detection102
algorithms considering the full range of forest disturbance magnitudes and found disagreement on the spatial103
disturbance occurrences between algorithms. While most unsupervised change detection alternatives have104
proven accurate in depicting overall vegetation changes (Zhu and Woodcock,2014;Estay and Ch´avez,2018;105
Wu et al.,2020), less attention has been paid to assessing if their sole use is proficient for separating106
deforestation from other sources of forest decay. As a consequence, Cohen et al. (2018) underscored the107
necessity of a secondary classification to improve disturbance detection.108
Different studies have highlighted the challenge of distinguishing deforestation from other causes of forest109
deterioration. For instance, when studying forest cover changes from satellite imagery, Hansen et al. (2013)110
refer to the “proximate” causes of disturbances. Other studies directly assess the contribution of climate-111
and human-based variables in forest productivity or net primary productivity (Chen et al.,2021). Likewise,112
approaches involving field observations are also used to assess the contribution of drivers of change (Redlich113
et al.,2022), although they may be difficult to scale up to the landscape level. However, incipient citizen114
science alternatives have been implemented to address this scaling issue (Arcanjo et al.,2016). For example,115
Sebald et al. (2021) incorporated the landscape context by estimating the cumulative disturbed forest area116
surrounding forest patches in a secondary classification to identify causal agents of disturbance, resulting117
in an improved depiction of real deforestation. However, few examples of these human- and natural-based118
drives of vegetation change remain, and robust methods are still needed to operationally depict deforestation119
from other drivers of forest change at multiple spatial scales.120
The primary objective of this study is to disentangle deforestation from other forest degradation causes121
with the aim of detecting near real-time human-induced deforestation using satellite data, including time se-122
ries. To achieve this, we seek a) to evaluate and compare three architecturally different algorithms of change123
detection in forest monitoring, discussing their advantages and challenges; b) to apply a secondary classi-124
fication to distinguish deforestation from other forest disturbances, such as large-scale fires and droughts,125
utilising the outputs of the change detection algorithm and ancillary data; c) to compare deforestation re-126
sults against governmental inspection requests raised by the community, evaluating the appropriateness of127
these requests and the temporal gap between them. We will use a large portion of central Chile as a study128
site to achieve these objectives. Central Chile has undergone severe droughts and fires over the last decade,129
presenting alarming tree mortality increases in a complex landscape of forestry and agriculture mosaics130
(Garreaud et al.,2020;Miranda et al.,2020;Fuentes et al.,2021).131
4
2. Materials and methods132
2.1. Study region133
To address our research goals, we selected central Chile between the Valparaiso and El Maule regions (-134
32.3°– -36.5°S; Figure 1), an area with an extent of about 78,000 km2. Forests in this region are important,135
as they have been acknowledged as one of the world’s biodiversity hotspots (Myers et al.,2000). This136
region is mainly dominated by Mediterranean climate types (Csa and Csb) according to the K¨oppen–Geiger137
climate classification (Sarricolea et al.,2017;Beck et al.,2018) and has historically experienced important138
land cover changes (Montoya-Tangarife et al.,2017;Miranda et al.,2017;Fuentes et al.,2021). Moreover,139
these forests face multiple threats, including urban and peri-urban expansion, agriculture expansion, and140
the establishment of exotic plantations (Manuschevich,2018). Consequently, according to the International141
Union of Conservation of Nature (IUCN) (Alaniz et al.,2016), most of these ecosystems have been classified142
as threatened. Furthermore, these ecosystems have experienced severe and prolonged drought conditions,143
resulting in ”browning” events (Garreaud et al.,2020;Miranda et al.,2020). Additionally, the region has144
witnessed large-scale wildfire events over the past decade (Smith-Ram´ırez et al.,2022).145
Vegetation formation classes in the study region and their extent are in Table 1. Forests in the region en-146
compass various vegetation types (Cowling et al.,1996) predominantly dominated by sclerophyllous species.147
Thorny scrublands, mainly characterised by Vachellia caven, are also present, along with deciduous forests148
of Nothofagus species in southern areas with higher precipitation (Donoso and Donoso,2007;Salas et al.,149
2016). Sclerophyllous forest formations are characterised by evergreen trees with hard leaves that reduce150
water loss. These traits make the chlorophyll vegetation resilient to water stress and drought (Yin and151
Bauerle,2017). Plantations and other land cover classes overlap these categories (Zhao et al.,2016). This152
study includes forest categories from Table 1but also incorporates plantation in the analysis.153
2.2. Datasets and pre-processing154
We used the Landsat constellation, including Level 2 Collection 2 tier 1 scenes from TM, ETM+, and155
OLI/TIRS sensors for Landsat 5, 7, 8, and 9 (USGS,2022). These data included 8,804 scenes from 2000-156
01-01 to 2022-06-01. We merged the different collections using cloud and shadow masks based on the157
QA PIXEL band information of the CFMASK algorithm (Foga et al.,2017).158
We depicted the normalised difference vegetation index (NDVI) (Rouse Jr et al.,1973) of the image159
collections using the red and near-infrared bands of images. We further used the 2014 land cover map160
developed by Zhao et al. (2016) in the post-processing step to limit changes detected in forests, plantations, or161
shrubland areas, avoiding agricultural lands that can present artificial structural changes due to management.162
Additionally, we included elevation data from the Multi-Error-Removed Improved-Terrain digital elevation163
model, which was included as a covariate in the secondary classification step alongside the slope of the164
terrain calculated from it.165
5
Figure 1: Study region and its main characteristics in terms of elevation, mean annual rainfall and temperatures, and
vegetation formation classes.
6
Table 1: Proportion of natural vegetation types and plantations in the study region.
Class Area (km2) Coverage (%)
Deciduous forest 9,276.5 11.9
Sclerophyllous forest 29,198.9 37.4
Thorny forest 15,525.1 19.9
Altitude grassland 2,684.2 3.4
Altitude low shrubs 11,423.4 14.6
Sclerophyllous shrubs 2,474.4 3.2
Thorny shrubs 1,555.3 2.0
Plantations* 6,168.0 7.9
Bareland 5,621.4 7.2
*Natural vegetation corresponds to potential habitat distributions. Therefore, coverage including plantations should exceed 100%.
We also used calibrated and ortho-corrected Ground Range Detected scenes from synthetic aperture166
radar Sentinel 1 satellites from February 2016 to July 2022. These were filtered based on the Interferometric167
Wide Swath (IWS) mode using a descending orbit direction, since it allowed to maximise the number of168
scenes. The back-scatter intensity from dual-polarimetric Sentinel-1 images was used to calculate the Radar169
Vegetation Index (Mandal et al.,2020) using:170
RV I =4σ0
V H
σ0
V H +σ0
V V
(1)
being σ0
V H and σ0
V V the dual cross polarisation (vertical transmit/horizontal receive) and single co-polarisation171
(vertical transmit/vertical receive) backscatter intensity bands, respectively. This collection was used to ob-172
tain some covariates in the secondary classification step.173
We used Google Earth Engine (Gorelick et al.,2017) and the Data Cube Chile (https://datacubechile.cl/)174
to acquire and process the data. Subsequent methodologies and analyses were carried out using Python ≥175
3.7.176
2.3. Unsupervised change detection177
We selected three algorithms belonging to structurally different approaches to evaluate methodologies to178
trace structural changes in heterogeneous and complex forest ecosystems in Central Chile. We tested: 1) a179
probabilistic approach based on the phenological characterisation of vegetation considering a bivariate time180
series based on day-of-the-year data; 2) a temporal segmentation method assuming the decomposition of181
time series in intra-annual, inter-annual, and structural changes by evaluating the deviation of observations182
and predictions based on ordinary least squares regressions. This method uses bivariate/multivariate data183
7
containing as reference the date; and 3) a temporal segmentation based on the optimisation of a cost184
function, where the cost is additive in the segmented blocks and requiring a univariate/multivariate time185
series, relaxing the need for temporal references for the change-point detection.186
The outputs of these algorithms correspond to change dates and magnitudes. The change magnitudes187
were calculated as the difference between average NDVI values within stable periods adjacent to the detected188
changes.189
Phenological characterisation: The non-parametric phenological cycle and anomaly detection (npphen) algo-190
rithm191
The first algorithm implies using multi-year phenological stages in forests to characterise the behaviour192
of phenological curves. We used the methodology defined by the ’npphen’ R-package (Estay and Ch´avez,193
2018). This methodology uses multi-annual information and orders them to day-of-the-year (DOY) to have194
a single phenological curve (or a pseudo-DOY in the case of the southern hemisphere). This curve is fitted195
using Kernel Density Estimation (KDE). This allows us to depict a generic phenological response per pixel196
as:197
ˆ
f(x;H) = 1
n
n
X
i=1
KH(x−Xi) (2)
where Xcorresponds to a time series containing paired values of vegetation indices and the DOY, being198
Xithe pair of values for the ith observation, ncorresponds to the number of observations, which can be199
calculated as the multiplication of the number of annual phenological cycles by the observations per cycle,200
xis a generic point in the paired values, His the bandwidth 2 x 2 array, Kis the kernel, in this case201
corresponding to a Gaussian kernel of size defined by H(Wand and Jones,1994), being f(x) the bivariate202
density function of X.203
Npphen can lead to anomaly detection of present values when compared to the historical trend of that204
time of the year (e.g., weekly) using a probabilistic approach based on the following:205
Ai=V Iobs −V Iexp (3)
being Aithe anomaly for the ith DOY, V Iobs the observed vegetation index and V Iexp the expected vege-206
tation index for that DOY.207
We fitted npphen for the period 2000-2015 using the Landsat collection to set each pixel’s historical208
phenological response, consequently estimating anomaly probabilities from 2016 (i.e., see section 2.4.). We209
defined a structural break in the time series when at least five consecutive negative anomalies exceeded two210
standard deviations of the ”training” data. We did this assuming that negative changes or perturbations in211
forest and shrublands structure cause a decline of NDVI values (Hudak and Wessman,2000) and to avoid212
8
single anomalies that may be caused by artefacts that could obscure the structural break detection. By213
using npphen, only the initial change detected is considered, as anomalies in the time series persist until the214
signal aligns with a pattern similar to the undisturbed period.215
Temporal segmentation: The Continuous Change Detection and Classification (CCDC) algorithm216
The second method for defining structural breaks is the Continuous Change Detection and Classification217
(CCDC) algorithm, originally defined to trace land cover changes (Zhu and Woodcock,2014). The algorithm218
is used in time series composed by seasonality, trends, and breaks, estimating coefficients through ordinary219
least squares (OLS) and fitting harmonic functions as:220
ˆρ(i, x)OLS =a0,i +a1,icos(2π
Tx) + b1,isin(2π
Tx) + c1,ix(4)
being x, constrained to the following boundaries:221
τ∗
k−1< x ≤τ∗
k(5)
where ˆρ(i, x)OLS is the estimated value of the time series for the xth day of the year; icorrespond to the222
dimension of the time series evaluated (band of images); Tis the number of days per year; a0,i;a1,i , b1,i ;223
and c1,i are the overall (mean), intra-annual change (seasonal), and inter-annual (trend) change coefficients224
for the iband, respectively; and τ∗
kare the kth break points.225
CCDC can use single or multiband images for change detection and seeks to detect breaks by evaluating226
the difference between observations and predictions, normalised by the root mean square error (RMSE)227
during a defined time window, as:228
1
k
k
X
i=1
|ρ(i, x)−ˆρ(i, x)OLS |
n×RM SEi
>1(zconsecutive times) (6)
where kis the number of bands used, nis a factor that multiplies the RMSE for the selection of the229
break based on the assumption of model prediction ranges, defined as 3 in this case, and zrepresents230
consecutive observations where errors exceed ntimes the RMSE to trace the break. We used the Google231
Earth Engine implementation, and tested several chi-square probability thresholds, setting zto 5 and a232
chi-square probability threshold to 0.9 to improve the change detection (Supplementary materials; Figure233
S1).234
Temporal segmentation: Linearly penalised segmentation (PELT)235
We selected the Pruned Exact Linear Time (PELT) algorithm as the third methodology, which is a236
linearly penalised segmentation implemented in the ’Ruptures ’ Python library (Truong et al.,2020). PELT237
efficiently deals with an unknown number of breaks in each time series. In a time series defined by y1:nof238
9
size n, and mchange points at τ1:mdates, structural breaks can be identified by minimising the following239
(Wambui et al.,2015):240
m+1
X
i=1
[l(y(τi−1+1):τi) + β] (7)
being lthe cost function for the ith segment, and βcorresponds to a penalty to minimise over-fitting.241
PELT mixes partitioning and pruning to achieve computational efficiency by applying the optimal solu-242
tion F(n) of eq. 7:243
F(n) = min
τm
{F(τm) + l(y(τm+1):n)}(8)
being the inner minimisation F(τm):244
F(τm) = min
τ|τm
{
m
X
i=1
[l(y(τi−1+1):τi) + β]}(9)
We used a radial basis function as model input, setting the minimum distance between change points to245
3, the subsample to 5, and the penalty to 30 since it allowed us to empirically improve the change detection246
performance through accuracy metrics (Supplementary materials; Figure S1).247
2.4. Algorithm performance and timeliness analysis248
We used the Global Forest Watch (GFW) map of deforestation records (https://www.globalforestwatch.249
org/) to initially select potential areas of forest change between 2016 and 2022. We then corroborated250
the validation sites by visually interpreting Google Earth Pro images. We selected a total of 382 validation251
sites, from which 142 corresponded to human-based deforestation and 147 to other changes, i.e., 91 polygons252
associated with severe drought conditions during the summer of 2019-2020 that led to vegetation ”browning”253
in the studied region as reported by Miranda et al. (2020) and 56 polygons were selected as affected by fires.254
We also selected 93 sites with stable time series where no changes occurred. We used pixel-based data for the255
analyses, thus having several time series per site (i.e., 23,053 pixel-based time series in total). The selected256
time frame of analysis was chosen to minimise error in methods needing historical data before validation,257
as npphen, and because of the higher frequency of high resolution images available in Google Earth, which258
facilitates the selection of reference polygons for validation. Five examples of non-deforested (above) and259
deforested (below) reference polygons and background reflectance images for the years 2016 and 2021 are260
shown in Figure 2. For the reader’s convenience, Sentinel 2 images are displayed instead of Landsat images.261
We evaluated the performance of the three algorithms in terms of consumer’s accuracy (CA), producer’s262
accuracy (PA), F1-score, overall accuracy (OA), and Cohen’s Kappa coefficient as suggested by Olofsson263
et al. (2014), being PA estimated as the ratio between correctly classified pixels in each category and the264
10
Figure 2: Reference polygons used in this study (above) and examples of non-deforested (below left) and deforested (below
right) forest ecosystems between 2016 and 2022. Black polygons depict validation polygons. A, B, and C show sites without
human- and natural-based changes, while D was classified as affected by drought due to the browning observed, and E was
classified as affected by fires in 2017 through high-resolution satellite images from Google Earth Pro (spatial resolution of less
than 10 m).
11
number of reference pixels within that category, while CA are correctly classified pixels in each category265
divided by the number of pixels classified as such within that category. Additionally, OA is calculated as:266
OA =tp +tn
tp +fn +tn +f p (10)
being tp, tn, fp, fn true positive, true negative, false positive, and false negative observations, respectively.267
Kappa coefficient is estimated using:268
kappa =Po−Pe
1−Pe
(11)
being Pothe probability of actual agreement and Pethe probability of a random agreement. Lastly, the269
F1-score is calculated using:270
F1=2×precision ×recall
precision +recall (12)
being precision and recall estimated as:271
precision =tp
tp +fp (13)
recall =tp
tp +fn (14)
and implying the ability of algorithms to avoid false positives and to identify positives, respectively.272
We finally assessed the model performances of:273
1. all deforested (human-made) areas against stable (undisturbed) forest areas;274
2. all deforested (human-made) areas against stable and natural-made changes (stable+droughts+fire);275
3. same as comparisons 1) and 2), but separating natural forest and plantation deforested areas.276
We divided the analysis into these sections to assess the ability of the unsupervised change detection277
algorithms to depict only deforestation (human-based changes) across vegetation types and natural forest278
alterations.279
Additionally, the rate of algorithm deforestation detection was evaluated in 15 randomly selected refer-280
ence polygons including both, native vegetation and plantations. The reference date of deforestation was281
selected by visually interpreting the first Landsat image where deforestation was evident. The temporal lag282
of deforestation detection was calculated as the difference between the algorithm detection date and the283
reference date. This aims to depict not only the models’ accuracy but also their stability and lag time of284
response to achieve their accuracy.285
12
2.5. Disentangling deforestation from other natural changes286
A supervised classification approach was used to depict deforestation from natural changes in the land-287
scape. To achieve that, we utilised the best of the three unsupervised change detection algorithms in terms288
of accuracy and fast lag response. We masked out pixels with positive change magnitudes as we assumed289
that all deforestation and natural alteration would decrease NDVI values. We then used the spatial patterns290
of the binary detection and its magnitudes to assess if the changes originated from deforestation or natural291
drivers.292
We assumed that vegetation changes associated with droughts and fires lead to broad and heterogeneous293
spatial patterns across the landscape. Contrarily, anthropic deforestation, like logging and urban expansion,294
is expected to have large change magnitude with smaller and homogeneous spatial extents.Therefore, we295
used the gray level co-occurrence matrix (GLCM) (Haralick et al.,1973) approach to generate textural296
features from the magnitude of changes using square kernel neighbourhoods of 15 x 15 pixels. Likewise, we297
calculated the neighbourhood variations using the standard deviation of the change dates and magnitudes298
within a square kernel of 50 x 50 pixels. The selected kernel dimensions were chosen because they yielded the299
best results compared with other configurations explored. We used the entropy, contrast, inverse difference300
moment (IDM), angular second moment (ASM), and sum average (SA) layer metrics from the GLCM301
analysis. Entropy signifies the degree of complexity in the spatial distribution of pixels, contrast reflects the302
local intensity variations between neighbouring pixels, IDM illustrates the local homogeneity, ASM measures303
the uniformity or smoothness of an image, while SA represents the average intensity of the sum of pixel304
pairs. We combined the GLCM data with change magnitudes and the standard deviation layers derived305
from the outputs of the selected unsupervised algorithm as predictors for the supervised classification (Figure306
3). Additionally, slope and elevation were also included as predictors, along with the interquartile range307
and the 25th and 75th quartiles of RVI images as RVI has been recognised as an alternative for monitoring308
vegetation growth (Nasirzadehdizaji et al.,2019), and the interquartile range represents a measure of growth309
variability.310
The reference polygons were split into a 10% for validation (2,283 pixels), while a 10-fold cross-validation311
was applied on the remaining 90% polygons. We used the Random Forest algorithm with 150 trees and an312
out-of-bag fraction of 0.5 to train a pixel-based classifier to depict deforestation from other changes and to313
differentiate forest disturbance sources since these parameters resulted in a good performance using a grid314
search based on the validation subset.315
2.6. Independent verification via Citizen Science316
Finally, we used the citizen-based complaints from the Superintendencia del Medio Ambiente, the or-317
ganism from the Ministry of Environment in charge of monitoring the natural resources, to filter 16 in-situ318
descriptions containing the words ”*forest”, ”fell*”, ”cut*down”, and ”sclerophyllous”. Complaints were319
13
Figure 3: Schema depicting the workflow of the methodology for distinguishing deforestation from other natural-induced
changes.
14
also filtered based on forests, shrublands, and plantation land cover classes from the Zhao et al. (2016) map.320
These complaints (Supplementary materials, Table S1) were compared with deforestation maps, and the321
dates of the change detection were used as an independent verification of the method.322
3. Results323
3.1. Performance of unsupervised change detection methods324
Table 2presents the performance of different change detection algorithms used between stable (i.e., no325
changes between 2016-2022) and deforested classes and between all other classes (stable, fire, and drought326
affected) and deforested classes. The algorithms show robust performances when comparing stable and327
deforested classes, with OA ranging from 0.78 to 0.99 and kappa coefficients from 0.56 to 0.95. Under native328
forest types, npphen outperforms other algorithms, but under plantations, PELT has the best performance.329
Likewise, when combining native vegetation and plantations, PELT slightly outperforms other algorithms.330
Overall, we observed that all algorithms perform less accurately on native forests than on plantations, given331
the vegetation heterogeneity. In general, commission errors are slightly higher than omission errors, which332
is reflected in slightly lower CA (CA = 100 - commission errors) compared to PA (PA = 100 - omission333
errors).334
We depicted very low accuracies when comparing deforestation with the rest of the categories (includ-335
ing natural-based changes), with OA and kappa values ranging from 0.27 to 0.61 and from 0.03 to 0.27,336
respectively. However, again plantations show a better performance than native forest types. In this case,337
large errors are caused, among others, by natural-induced changes detected by algorithms other than defor-338
estation, implying that further processing is required to distinguish between human- and natural-induced339
vegetation changes.340
Figure 4shows the averaged NDVI time series for non-deforested (left) and deforested (right) examples341
with their corresponding change detection by the tree algorithms (vertical lines). PELT and CCDC algo-342
rithms depicted a structural change in the summer of 2020 associated with drought in Figure 4D. All three343
algorithms detect a fire-based change in example E from Figure 4during 2017.344
Deforested examples (right-column panel of Figure 4) are correctly identified by PELT and CCDC345
algorithms in all cases. On the other hand, npphen detects structural changes in four of the five examples.346
Example I is associated with a change in vegetation in the period used for training the algorithm (2000-347
2016). In this example, a growing forest can be observed between 2008 and 2014, being the previous years348
characterised by early-stage forest vegetation. This results in the spread of NDVI low-probabilities observed349
in the phenological curve of example I (Supplementary materials, Figure S3), affecting the change detection.350
Figure 5shows spatiotemporal maps of structural breaks since 2016, applying a 100-200 m buffer around351
the example polygons. The npphen algorithm tracked deforestation changes, except for example I, incurring352
15
Table 2: Performances of the three unsupervised change detection algorithms, with metrics estimated between de-
forested and stable reference classes and between deforested and all other classes (stable, drought-affected, and fire-
affected). The comparison was performed on every time series (combined) and for native vegetation and plantations
separately. Bold numbers depict local maxima in the models
Forest Algorithm OA* PA CA kappa f1 OA PA CA kappa f1
——– stable v/s deforested ——– ——— all v/s deforested ——–
Native
npphen 0.83 0.83 0.83 0.67 0.84 0.39 0.57 0.53 0.05 0.24
CC DC 0.78 0.78 0.78 0.56 0.81 0.27 0.50 0.50 0.03 0.23
P ELT 0.78 0.78 0.79 0.57 0.76 0.55 0.60 0.54 0.09 0.27
Plantation
npphen 0.96 0.97 0.88 0.85 0.97 0.61 0.64 0.71 0.27 0.68
CC DC 0.95 0.93 0.97 0.87 0.98 0.53 0.58 0.69 0.14 0.64
P ELT 0.99 0.98 0.96 0.95 0.99 0.55 0.59 0.73 0.18 0.66
Combined
npphen 0.89 0.90 0.87 0.76 0.92 0.48 0.61 0.61 0.15 0.46
CC DC 0.88 0.86 0.86 0.71 0.92 0.37 0.55 0.57 0.06 0.41
P ELT 0.90 0.91 0.86 0.77 0.93 0.55 0.66 0.63 0.21 0.49
*OA: Overall accuracy; PA: Producer’s accuracy (mean of classes); CA: Consumer’s accuracy (mean of
classes).
16
Figure 4: Non-deforested (left) and deforested (right) averaged NDVI pixels sampled from reference examples. Structural
changes detected using npphen, CCDC, and PELT algorithms are also presented (vertical lines). A, B, and C polygons were
classified as stable, while D and E were classified as affected by drought and fires, respectively. F-J polygons are classified as
deforested.
17
Figure 5: Spatio-temporal maps of last break detections using three approaches for non-deforested (left) and deforested
(right) polygon examples. Letters A-J correspond to non-deforested/deforested example polygons as shown in Figure 2. Red
polygons depict the selected validation areas. Non-deforested A, B, and C polygons were classified as stable, while D and E
were classified as affected by drought and fires, respectively. Land cover maps correspond to Zhao et al. (2016)
18
in more false negatives. Poorly defined boundaries for phenological curves (Figure S3 associated with353
vegetation changes between 2000 and 2016, i.e. training period in npphen) limit phenology-based algorithms’354
structural breaks detection capacities, preventing structural breaks tracking in Figure 5example I. On the355
other hand, CCDC and PELT changes are correctly delimited.356
Changes in non-deforested examples only occur within example D (Figure 5) using PELT, corresponding357
to a drought event (Figure 4). Meanwhile, example E, corresponding to a fire event, was identified by all358
unsupervised algorithms. CCDC detected changes in sparse pixels from examples C and D, but most pixels359
in those cases do not present changes.360
3.2. Latency between reference deforestation and change detection361
The temporal lag between reference deforestation dates and average change detection dates is shown in362
Figure 6. Although PELT leads to the best cumulative change detection, it also presents the most extended363
temporal lag (in percentage) between average change detection and reference deforestation dates. While364
PELT deforestation dates in Figure 4are accurately tracked, its cost function optimisation results in several365
months of delay in tracking changes, making it more suitable for offline applications. In contrast, CCDC366
demonstrates the fastest detection response, with over 50% of pixel changes detected within the first two367
months after the reference deforestation date. This lagged response may also be adjusted by modifying z368
in equation (5). CCDC is followed by npphen in the latency between reference deforestation dates and the369
algorithm detection dates. Thus, the latency of CCDC and npphen makes them suitable for near real-time370
applications. Given the high performance and fast response of CCDC to depict forest changes, we selected371
it for the secondary classification step.372
3.3. Disentangling deforestation and performance373
Figure 7shows the date and magnitude of general changes across central Chile. The landscape changes374
depicted distinct patterns of natural- and human-based changes. For example, the above-zoomed area375
represents natural changes associated with browning and fire. Here, the changes are widespread with het-376
erogeneous shapes and low change intensities. Contrary, the under-zoomed area depicts deforestation or377
human-made alteration in the landscape. These changes have unnatural homogeneous shapes, like squares378
and rectangles, and often high change magnitudes.379
Change and magnitude distribution differences between disturbed classes are in Figure 8. A bimodal380
distribution of change magnitudes is observed (upper panel) and a large dispersion (middle panel) of change381
dates that increases during the spring and summer seasons. Distinctively larger frequencies were depicted382
in the summer of 2017, associated with large fire events, and between 2019 and 2020, associated with383
severe drought conditions (Supplementary materials, Figure S4). We also observed that fires presented the384
strongest change magnitudes (i.e., lower negative values) followed by deforestation events, while drought385
19
Figure 6: Temporal lag between reference deforestation and average detected changes using unsupervised algorithms.
Figure 7: Change detection magnitudes (left) and dates (right) in central Chile using PELT. Changes correspond to the last
break detected on the time series. Two zoomed-in areas are also depicted.
20
events led to higher values. Also, change dates for drought are very limited to specific periods of time, while386
deforestation and fires showed more spread alteration throughout the study period.387
Table 3show the classification performance using the Random Forest model for disentangling deforesta-388
tion from other classes and for determining the source of forest disturbance. Results depict a substantial389
agreement between reference and predicted classes. Deforestation is tracked correctly, with similar values for390
consumer and producer accuracies and a kappa of 0.83 in validation. Additionally, the disturbance sources391
are also tracked quite correctly, with a kappa value of 0.88 in validation.392
Table 3: Deforestation classification performance using the supervisor Random Forest classifier.
Stage OA PA CA kappa f1 OA PA CA kappa f1
— deforested v/s non-deforested — ——— Change type ——–
calibration 0.99 0.99 0.99 0.99 0.99 0.99 0.99 0.99 0.99 0.99
validation 0.93 0.92 0.92 0.83 0.92 0.92 0.93 0.92 0.88 0.92
The classification results of zoomed-in examples from Figure 7are in Figure 9, showing the changes393
caused only by deforestation (human-made). False colour images from years 2016 and 2022 are included394
to evaluate differences, and the mean and standard deviation of deforestation changes depicted from the395
10-fold repetitions show stable estimation (i.e., low variations). Likewise, different classes of disturbance396
sources for the zoomed-in examples evaluated are in Figure 10.397
3.4. Evaluation using citizen complains398
Figure 11 depicts the comparison between user complaints and temporal segmentation algorithms. From399
the 16 filtered complaints, only 11 presented deforestation events detected (i.e., 69%). However, some of400
these complaints do not show significant changes in the associated NDVI time series, and in other cases401
changes are subtle. Furthermore, the methodology often detected forest changes before the occurrence of402
citizen complaints (Supplementary materials, Figure S5), hence improving the detection in time and space403
from the status quo. Statistically, comparing the complaints where deforestation was detected, the algorithm404
detection occurs on average 94 days before the complaint is made.405
4. Discussion406
4.1. Performance of the unsupervised forest change detection algorithms407
The PELT temporal segmentation algorithm yielded slightly higher overall accuracies than other algo-408
rithms when depicting deforestation against stable vegetation. PELT stands out for its flexibility, as it allows409
for the adoption of different linear and non-linear models to reduce the cost functions, hence eliminating the410
21
Figure 8: Distribution of changes based on magnitudes and dates. Histograms of change magnitudes and dates for the entire
study region are in the above and middle panels, respectively. The distribution of change magnitudes and dates for reference
polygons based on change classes are in the lower panel.
22
Figure 9: Deforestation changes tracked. Two zoomed-in regions are depicted with mean false colour (NIR-red-green) images
for the years 2016 and 2022. Mean deforestation changes and standard deviation deforestation changes from the 10-fold
validation are also shown.
need for prior knowledge of seasonal vegetation patterns (Wambui et al.,2015). While PELT is computation-411
ally efficient, flexible (Killick et al.,2012), and has demonstrated faster and more consistent performance412
than visual assessments of breaks in oceanographic wave height time series (Killick et al.,2011), it also413
demonstrated to work offline, leading to a large lag between deforestation and its detection. On the other414
hand, changes that occurred during the phenological characterisation using the Landsat collection (2000415
- 2015) affected the results obtained using npphen. By relying on the phenological curve characteriation416
through a period of non-disturbance, npphen leads to the detection of continuous anomalies after a change417
occurs, making it difficult to detect subsequent changes. These factors reduce its versatility. Nevertheless,418
these drawbacks can be mitigated by employing a time series with minimal disturbance during training419
and selecting only the first change. We make sure to use undisturbed polygons between 2000 and 2015 for420
evaluating deforestation between 2016 and 2022, except for one reference polygon used for evaluation, which421
experienced changes during the ”training” period (Figure 4I). This approach has also proven helpful in422
detecting ecosystems with high dynamic seasonality, such as the blooming desert in northern Chile (Ch´avez423
et al.,2019). Unlike npphen, which utilises non-parametric methods and does not assume any predefined424
phenological cycle shape (Estay and Ch´avez,2018), CCDC assumes harmonic/seasonal cycles with different425
orders among its components (Zhu and Woodcock,2014). CCDC leads to an overall good performance in426
the unsupervised change detection and led to the faster deforestation detection response, which implies it427
may be used for near-real deforestation applications.428
The use of different optical satellite datasets can yield varied results due to differences in temporal reso-429
23
Figure 10: Deforestation changes tracked. Two zoomed-in regions are depicted with mean false colour (NIR-red-green)
images for the years 2016 and 2022. Mean deforestation changes and standard deviation deforestation changes from the 10-fold
validation are also shown.
24
Figure 11: Different series for locations of citizen-based complaints raised to the Environmental Superintendence and defor-
estation detection results.
25
lution, the time span of available data, spatial resolution, and the robustness of cloud and shadow detection430
methods. Notably, we observed noise in the data collections, mainly caused by cloud and shadow detection431
errors, that increased deviations in the time series (Baret et al.,2007;Griffiths et al.,2020). This, together432
with artefacts in satellite images such as those caused by the failure of the scan line corrector in Landsat433
7, can cause missed structural breaks, impacting npphen and CCDC, which rely on anomaly deviations or434
deviations of residuals to track significant changes. The frequency of satellite image acquisition also may435
play a role in timely change detection (Fuentes et al.,2019), and it should be fully considered when aiming436
for continuous environmental monitoring. For instance, Zhu and Woodcock (2014) highlights the influence437
of observation frequency on the speed of change detection using temporal segmentation algorithms, empha-438
sizing the importance of frequent clear-sky observations. However, there is a dearth of studies investigating439
the effects of observation frequency on change detection. Lunetta et al. (2004) explored frequencies of 3,440
7, and 10 years using Landsat images to track land cover changes, finding that a frequency of at least 3441
years is needed to appropriately detect changes in North Carolina, USA. However, further research is needed442
to transfer these findings to other site conditions and satellites. Harmonising and combining Landsat and443
Sentinel 2 datasets can potentially increase observation frequency (Claverie et al.,2018), and their effects444
on change detection should be further examined.445
4.2. Disentangling human- and natural-induced forest changes446
As discussed previously, optical data solely is often insufficient to determine the nature of the detected447
changes (Bannari et al.,1995). While NDVI values can serve as a proxy for vegetation presence or greenness448
(Chapungu et al.,2020), the NDVI values cannot directly provide information about the cause or driver of449
the disturbances. Therefore, a significant change in the NDVI time series can be caused by varying natural-450
and human-made drivers (Jackson and Huete,1991). For instance, we found widespread systematic distur-451
bances throughout central Chile during the summer of 2020, which resulted from browning and prolonged452
drought conditions (Figure 8) (Miranda et al.,2020,2022). The systematic breaks identified during the453
summer of 2020 were associated with a ”browning” process affecting native vegetation. This indicates a454
systematic decline in vegetation health or vegetation death, which can be observed through changes in NDVI455
or productivity measures (Koulgi et al.,2019). Other studies have observed a similar phenomenon in other456
areas of the globe (Hao et al.,2022).457
We identified deforestation changes by leveraging spatial and temporal occurrence patterns of the overall458
unsupervised change detection. We used, in this case, a secondary supervised classification approach as459
suggested by Cohen et al. (2018) when studying the potential extrapolation of change detection methods,460
incorporating neighbourhood data and textural analysis of change detections and magnitudes as covariates461
to disentangle deforestation from other forest disturbances. We also included topographic and radar data462
to integrate structural changes in the vegetation canopies as additional information aiding the identification463
26
of deforestation (Reiche et al.,2018). However, further research is needed to explore the importance of464
covariates in the deforestation detection.465
4.3. Policy-making and management466
Overall, we found that in various cases the remotely sensed change detection approach significantly re-467
duced the response time compared to citizen complaints managed by the Environmental Superintendence468
(Figure 9). This further highlights the potential for prompt action if automated methodologies like this one469
are integrated into governance schemes. Depicting change detection at the landscape level is crucial for mon-470
itoring unauthorized activities, such as logging and specific urban expansions, reducing the environmental471
impacts and the expenses associated with in-situ monitoring and control by the government. However, it is472
essential to approach public complaints with caution. While some complaints contain specific information473
related to deforestation (Table S1, Supplementary materials), they often lack information about the nature,474
magnitude, or extent of the disturbance. Moreover, the georeferencing of complaints can be inaccurately475
provided by users, leading to cases of low NDVI values, making categorising the locations as forests or476
closed shrublands challenging. Therefore, while these complaint datasets can serve as a reference for further477
investigation and comparison, they should not be considered as real validation sources.478
5. Conclusions479
Forest ecosystems were monitored using phenological characterisation and temporal segmentation algo-480
rithms. While the temporal segmentation through PELT slightly outperformed other methods for defor-481
estation tracking using NDVI calculated from the Landsat dataset, it also led to the a high latency between482
deforestation and its detection. CCDC leads to a generally good performance and a fast deforestation de-483
tection response. We observed the general tendencies of the three methods to depict forest decline due to484
severe drought conditions and large fires. This implies that the separation of these events from human-based485
interventions is challenging. Disentangling deforestation from other changes through a secondary classifi-486
cation using neighbourhood statistics and textural analysis applied to the detected changes, together with487
topographic and radar data, led to robust results. Additionally, deforestation detection was assertive when488
evaluated against citizen complaints raised to the Environmental Superintendence, leading, on average, to489
a faster detection (94 days before than complaints). Further research is needed to assess other alternatives490
depicting human-based forest disturbances through, for example, deep learning methods.491
Acknowledgements492
This project was supported by the ’SAMSARA’ FONDEF IdeA I+D ID21I10102 project, ANID, Chile.493
Ignacio is also funded by the ANID FONDECYT Postdoctoral Project N°3220317.494
27
Code availability495
The code associated with this project will be made available in the following repository: https://github.496
com/IFuentesSR/samsara deforestation497
References498
Alaniz, A. J., Galleguillos, M. and Perez-Quezada, J. F. (2016), ‘Assessment of quality of input data used to classify ecosystems499
according to the iucn red list methodology: The case of the central chile hotspot’, Biological Conservation 204, 378–385.500
Arcanjo, J. S., Luz, E. F., Fazenda, ´
A. L. and Ramos, F. M. (2016), ‘Methods for evaluating volunteers’ contributions in a501
deforestation detection citizen science project’, Future Generation Computer Systems 56, 550–557.502
Asokan, A. and Anitha, J. (2019), ‘Change detection techniques for remote sensing applications: a survey’, Earth Science503
Informatics 12(2), 143–160.504
Attiwill, P. M. and Adams, M. A. (1993), ‘Nutrient cycling in forests’, New phytologist 124(4), 561–582.505
Bakker, K. and Ritts, M. (2018), ‘Smart earth: A meta-review and implications for environmental governance’, Global envi-506
ronmental change 52, 201–211.507
Bannari, A., Morin, D., Bonn, F. and Huete, A. (1995), ‘A review of vegetation indices’, Remote sensing reviews 13(1-2), 95–508
120.509
Baret, F., Hagolle, O., Geiger, B., Bicheron, P., Miras, B., Huc, M., Berthelot, B., Ni˜no, F., Weiss, M., Samain, O. et al.510
(2007), ‘Lai, fapar and fcover cyclopes global products derived from vegetation: Part 1: Principles of the algorithm’, Remote511
sensing of environment 110(3), 275–286.512
Barlow, J., Lennox, G. D., Ferreira, J., Berenguer, E., Lees, A. C., Nally, R. M., Thomson, J. R., Ferraz, S. F. d. B., Louzada, J.,513
Oliveira, V. H. F. et al. (2016), ‘Anthropogenic disturbance in tropical forests can double biodiversity loss from deforestation’,514
Nature 535(7610), 144–147.515
Beck, H. E., Zimmermann, N. E., McVicar, T. R., Vergopolan, N., Berg, A. and Wood, E. F. (2018), ‘Present and future516
k¨oppen-geiger climate classification maps at 1-km resolution’, Scientific data 5(1), 1–12.517
Bengtsson, J., Nilsson, S. G., Franc, A. and Menozzi, P. (2000), ‘Biodiversity, disturbances, ecosystem function and management518
of european forests’, Forest ecology and management 132(1), 39–50.519
Boisier, J. P., Rondanelli, R., Garreaud, R. D. and Mu˜noz, F. (2016), ‘Anthropogenic and natural contributions to the southeast520
pacific precipitation decline and recent megadrought in central chile’, Geophysical Research Letters 43(1), 413–421.521
Bottalico, F., Travaglini, D., Chirici, G., Garf`ı, V., Giannetti, F., De Marco, A., Fares, S., Marchetti, M., Nocentini, S.,522
Paoletti, E. et al. (2017), ‘A spatially-explicit method to assess the dry deposition of air pollution by urban forests in the523
city of florence, italy’, Urban Forestry & Urban Greening 27, 221–234.524
Cai, Y., Shi, Q., Xu, X. and Liu, X. (2023), ‘A novel approach towards continuous monitoring of forest change dynamics in525
fragmented landscapes using time series landsat imagery’, International Journal of Applied Earth Observation and Geoin-526
formation 118, 103226.527
Canadell, J. G., Meyer, C., Cook, G. D., Dowdy, A., Briggs, P. R., Knauer, J., Pepler, A. and Haverd, V. (2021), ‘Multi-decadal528
increase of forest burned area in australia is linked to climate change’, Nature communications 12(1), 6921.529
Castillo, M., Plaza, ´
A. and Garfias, R. (2020), ‘A recent review of fire behavior and fire effects on native vegetation in central530
chile’, Global Ecology and Conservation 24, e01210.531
Chapungu, L., Nhamo, L. and Gatti, R. C. (2020), ‘Estimating biomass of savanna grasslands as a proxy of carbon stock using532
multispectral remote sensing’, Remote sensing applications: Society and environment 17, 100275.533
Ch´avez, R. O., Moreira-Mu˜noz, A., Galleguillos, M., Olea, M., Aguayo, J., Lat´ın, A., Aguilera-Betti, I., Mu˜noz, A. A. and534
Manr´ıquez, H. (2019), ‘Gimms ndvi time series reveal the extent, duration, and intensity of “blooming desert” events in535
28
the hyper-arid atacama desert, northern chile’, International Journal of Applied Earth Observation and Geoinformation536
76, 193–203.537
Chen, S., Wen, Z., Ma, M. and Wu, S. (2021), ‘Disentangling climatic factors and human activities in governing the old and538
new forest productivity’, Remote Sensing 13(18), 3746.539
Claverie, M., Ju, J., Masek, J. G., Dungan, J. L., Vermote, E. F., Roger, J.-C., Skakun, S. V. and Justice, C. (2018), ‘The540
harmonized landsat and sentinel-2 surface reflectance data set’, Remote sensing of environment 219, 145–161.541
Cohen, W. B., Healey, S. P., Yang, Z., Stehman, S. V., Brewer, C. K., Brooks, E. B., Gorelick, N., Huang, C., Hughes, M. J.,542
Kennedy, R. E. et al. (2017), ‘How similar are forest disturbance maps derived from different landsat time series algorithms?’,543
Forests 8(4), 98.544
Cohen, W. B., Yang, Z., Healey, S. P., Kennedy, R. E. and Gorelick, N. (2018), ‘A landtrendr multispectral ensemble for forest545
disturbance detection’, Remote sensing of environment 205, 131–140.546
Cowling, R. M., Rundel, P. W., Lamont, B. B., Arroyo, M. K. and Arianoutsou, M. (1996), ‘Plant diversity in mediterranean-547
climate regions’, Trends in Ecology & Evolution 11(9), 362–366.548
Danielsen, F., Burgess, N. D., Balmford, A., Donald, P. F., Funder, M., Jones, J. P., Alviola, P., Balete, D. S., Blomley,549
T., Brashares, J. et al. (2009), ‘Local participation in natural resource monitoring: a characterization of approaches’,550
Conservation biology 23(1), 31–42.551
de Oliveira Roque, F., Menezes, J. F., Northfield, T., Ochoa-Quintero, J. M., Campbell, M. J. and Laurance, W. F. (2018),552
‘Warning signals of biodiversity collapse across gradients of tropical forest loss’, Scientific Reports 8(1), 1–7.553
DeFries, R. S., Rudel, T., Uriarte, M. and Hansen, M. (2010), ‘Deforestation driven by urban population growth and agricultural554
trade in the twenty-first century’, Nature Geoscience 3(3), 178–181.555
Dong, J., Metternicht, G., Hostert, P., Fensholt, R. and Chowdhury, R. R. (2019), ‘Remote sensing and geospatial technologies556
in support of a normative land system science: Status and prospects’, Current Opinion in Environmental Sustainability557
38, 44–52.558
Donoso, P. and Donoso, C. (2007), Forest species and stand types of chile, in F. W. Cubbage, ed., ‘Forests and Forestry in559
the Americas: An Encyclopedia’, Society of American Foresters and International Society of Tropical Foresters. Accessed560
2022-06-04.561
Estay, S. and Ch´avez, R. (2018), ‘npphen: an r-package for non-parametric reconstruction of vegetation phenology and anomaly562
detection using remote sensing’, BioRxiv p. 301143.563
FAO (2022), The State of the World’s Forests 2022. Forest pathways for green recovery and building inclusive, resilient and564
sustainable economies, Rome.565
Ferrante, L. and Fearnside, P. M. (2020), ‘The amazon’s road to deforestation’, Science 369(6504), 634–634.566
Figueroa, E. and Pasten, R. (2015), ‘The economic value of forests in supplying local climate regulation’, Australian Journal567
of Agricultural and Resource Economics 59(3), 446–457.568
Filipponi, F. (2019), Sentinel-1 grd preprocessing workflow, in ‘Multidisciplinary digital publishing institute proceedings’,569
Vol. 18, p. 11.570
Foga, S., Scaramuzza, P., Guo, S., Zhu, Z., Jr, D., R.D., Beckmann, T., Schmidt, G., Dwyer, J., Hughes, M. and Laue, B.571
(2017), ‘Cloud detection algorithm comparison and validation for operational landsat data products’, Remote Sensing of572
Environment 194, 379–390.573
Fuentes, I., Fuster, R., Avil´es, D. and Vervoort, W. (2021), ‘Water scarcity in central chile: the effect of climate and land cover574
changes on hydrologic resources’, Hydrological Sciences Journal 66(6), 1028–1044.575
Fuentes, I., Padarian, J., Van Ogtrop, F. and Vervoort, R. W. (2019), ‘Spatiotemporal evaluation of inundated areas using576
modis imagery at a catchment scale’, Journal of Hydrology 573, 952–963.577
Fuentes, I., Padarian, J. and Vervoort, R. W. (2022a), ‘Spatial and temporal global patterns of drought propagation’, Frontiers578
29
in Environmental Science 10.579
URL: https://www.frontiersin.org/article/10.3389/fenvs.2022.788248580
Fuentes, I., Padarian, J. and Vervoort, R. W. (2022b), ‘Towards near real-time national-scale soil water content monitoring581
using data fusion as a downscaling alternative’, Journal of Hydrology 609, 127705.582
Fuentes, I., Vervoort, R. W. and McPhee, J. (2024), ‘Global evapotranspiration models and their performance at different583
spatial scales: Contrasting a latitudinal gradient against global catchments’, Journal of Hydrology 628, 130477.584
Garreaud, R. D., Boisier, J. P., Rondanelli, R., Montecinos, A., Sep´ulveda, H. H. and Veloso-Aguila, D. (2020), ‘The central585
chile mega drought (2010–2018): a climate dynamics perspective’, International Journal of Climatology 40(1), 421–439.586
Gorelick, N., Hancher, M., Dixon, M., Ilyushchenko, S., Thau, D. and Moore, R. (2017), ‘Google earth engine: Planetary-scale587
geospatial analysis for everyone’, Remote sensing of Environment 202, 18–27.588
Griffiths, P., Nendel, C., Pickert, J. and Hostert, P. (2020), ‘Towards national-scale characterization of grassland use intensity589
from integrated sentinel-2 and landsat time series’, Remote sensing of Environment 238, 111124.590
Hamilton, S. E. and Friess, D. A. (2018), ‘Global carbon stocks and potential emissions due to mangrove deforestation from591
2000 to 2012’, Nature Climate Change 8(3), 240–244.592
Hansen, M. C., Krylov, A., Tyukavina, A., Potapov, P. V., Turubanova, S., Zutta, B., Ifo, S., Margono, B., Stolle, F. and Moore,593
R. (2016), ‘Humid tropical forest disturbance alerts using landsat data’, Environmental Research Letters 11(3), 034008.594
Hansen, M. C., Potapov, P. V., Moore, R., Hancher, M., Turubanova, S. A., Tyukavina, A., Thau, D., Stehman, S. V., Goetz,595
S. J., Loveland, T. R., Kommareddy, A., Egorov, A., Chini, L., Justice, C. O. and Townshend, J. R. G. (2013), ‘High-596
resolution global maps of 21st-century forest cover change’, Science 342(6160), 850–853.597
URL: https://www.science.org/doi/abs/10.1126/science.1244693598
Hao, H., Chen, Y., Xu, J., Li, Z., Li, Y. and Kayumba, P. M. (2022), ‘Water deficit may cause vegetation browning in central599
asia’, Remote Sensing 14(11), 2574.600
Haralick, R. M., Shanmugam, K. and Dinstein, I. H. (1973), ‘Textural features for image classification’, IEEE Transactions on601
systems, man, and cybernetics (6), 610–621.602
Housman, I. W., Chastain, R. A. and Finco, M. V. (2018), ‘An evaluation of forest health insect and disease survey data603
and satellite-based remote sensing forest change detection methods: Case studies in the united states’, Remote Sensing604
10(8), 1184.605
Huang, C., Goward, S. N., Masek, J. G., Thomas, N., Zhu, Z. and Vogelmann, J. E. (2010), ‘An automated approach for606
reconstructing recent forest disturbance history using dense landsat time series stacks’, Remote Sensing of Environment607
114(1), 183–198.608
Hudak, A. T. and Wessman, C. A. (2000), ‘Deforestation in mwanza district, malawi, from 1981 to 1992, as determined from609
landsat mss imagery’, Applied Geography 20(2), 155–175.610
Jackson, R. D. and Huete, A. R. (1991), ‘Interpreting vegetation indices’, Preventive veterinary medicine 11(3-4), 185–200.611
Japitana, M. V. and Burce, M. E. C. (2019), ‘A satellite-based remote sensing technique for surface water quality estimation’,612
Engineering, Technology & Applied Science Research 9(2), 3965–3970.613
Karthikeyan, L., Chawla, I. and Mishra, A. K. (2020), ‘A review of remote sensing applications in agriculture for food security:614
Crop growth and yield, irrigation, and crop losses’, Journal of Hydrology 586, 124905.615
URL: https://www.sciencedirect.com/science/article/pii/S0022169420303656616
Keenan, R. J., Reams, G. A., Achard, F., de Freitas, J. V., Grainger, A. and Lindquist, E. (2015), ‘Dynamics of global forest617
area: Results from the fao global forest resources assessment 2015’, Forest Ecology and Management 352, 9–20.618
Kerr, Y. H., Waldteufel, P., Wigneron, J.-P., Martinuzzi, J., Font, J. and Berger, M. (2001), ‘Soil moisture retrieval from space:619
The soil moisture and ocean salinity (smos) mission’, IEEE transactions on Geoscience and remote sensing 39(8), 1729–1735.620
Killick, R., Eckley, I. A., Jonathan, P. and Chester, U. (2011), Efficient detection of multiple changepoints within an oceano-621
30
graphic time series, in ‘Proceedings of the 58th world science congress of ISI’, pp. 4137–4142.622
Killick, R., Fearnhead, P. and Eckley, I. A. (2012), ‘Optimal detection of changepoints with a linear computational cost’,623
Journal of the American Statistical Association 107(500), 1590–1598.624
Kofinas, G. P. (2009), Adaptive co-management in social-ecological governance, in ‘Principles of ecosystem stewardship’,625
Springer, pp. 77–101.626
Koulgi, P. S., Clinton, N. and Karanth, K. K. (2019), ‘Extensive vegetation browning and drying in forests of india’s tiger627
reserves’, Scientific reports 9(1), 1–12.628
Lary, D. J., Alavi, A. H., Gandomi, A. H. and Walker, A. L. (2016), ‘Machine learning in geosciences and remote sensing’,629
Geoscience Frontiers 7(1), 3–10.630
Leblois, A., Damette, O. and Wolfersberger, J. (2017), ‘What has driven deforestation in developing countries since the 2000s?631
evidence from new remote-sensing data’, World Development 92, 82–102.632
L¨of, M., Madsen, P., Metslaid, M., Witzell, J. and Jacobs, D. F. (2019), ‘Restoring forests: regeneration and ecosystem function633
for the future’, New Forests 50(2), 139–151.634
Loginova, D. and Mann, S. (2022), ‘Measuring stability and structural breaks: Applications in social sciences’, Journal of635
Economic Surveys .636
Longo, M., Saatchi, S., Keller, M., Bowman, K., Ferraz, A., Moorcroft, P. R., Morton, D. C., Bonal, D., Brando, P., Burban,637
B. et al. (2020), ‘Impacts of degradation on water, energy, and carbon cycling of the amazon tropical forests’, Journal of638
Geophysical Research: Biogeosciences 125(8), e2020JG005677.639
Lopatin, J. (2023), ‘Interannual variability of remotely sensed phenology relates to plant communities’, IEEE Geoscience and640
Remote Sensing Letters 20, 1–5.641
Lunetta, R. S., Johnson, D. M., Lyon, J. G. and Crotwell, J. (2004), ‘Impacts of imagery temporal frequency on land-cover642
change detection monitoring’, Remote Sensing of Environment 89(4), 444–454.643
Mandal, D., Kumar, V., Ratha, D., Dey, S., Bhattacharya, A., Lopez-Sanchez, J. M., McNairn, H. and Rao, Y. S. (2020), ‘Dual644
polarimetric radar vegetation index for crop growth monitoring using sentinel-1 sar data’, Remote Sensing of Environment645
247, 111954.646
Mansoor, S., Farooq, I., Kachroo, M. M., Mahmoud, A. E. D., Fawzy, M., Popescu, S. M., Alyemeni, M., Sonne, C., Rinklebe,647
J. and Ahmad, P. (2022), ‘Elevation in wildfire frequencies with respect to the climate change’, Journal of Environmental648
management 301, 113769.649
Manuschevich, D. (2018), ‘A critical assessment of the adaptive capacity of land use change in chile: A socio-ecological approach’,650
Land Use-Assessing the Past, Envisioning the Future .651
Masek, J. G., Hayes, D. J., Hughes, M. J., Healey, S. P. and Turner, D. P. (2015), ‘The role of remote sensing in process-scaling652
studies of managed forest ecosystems’, Forest Ecology and Management 355, 109–123.653
Matskovsky, V., Venegas-Gonz´alez, A., Garreaud, R., Roig, F. A., Guti´errez, A. G., Mu˜noz, A. A., Le Quesne, C., Klock, K.654
and Canales, C. (2021), ‘Tree growth decline as a response to projected climate change in the 21st century in mediterranean655
mountain forests of chile’, Global and Planetary Change 198, 103406.656
McAllister, E., Payo, A., Novellino, A., Dolphin, T. and Medina-Lopez, E. (2022), ‘Multispectral satellite imagery and machine657
learning for the extraction of shoreline indicators’, Coastal Engineering p. 104102.658
Miranda, A., Altamirano, A., Cayuela, L., Lara, A. and Gonz´alez, M. (2017), ‘Native forest loss in the chilean biodiversity659
hotspot: revealing the evidence’, Regional Environmental Change 17(1), 285–297.660
Miranda, A., Lara, A., Altamirano, A., Di Bella, C., Gonz´alez, M. E. and Camarero, J. J. (2020), ‘Forest browning trends in661
response to drought in a highly threatened mediterranean landscape of south america’, Ecological Indicators 115, 106401.662
Miranda, A., Ovalle, J., Syphard, A., Berdugo, M., Carrasco, J., Mentler, R., Squeo, F., Delpiano, C., Vargas, S., Miranda, M.,663
Dobbs, C., G´omez-Gonz´alez, S., Lara, A. and Garreaud, R. D. (2022), Widespread synchronous decline in a mediterranean664
31
forest driven by accelerated aridity.665
Montoya-Tangarife, C., De La Barrera, F., Salazar, A. and Inostroza, L. (2017), ‘Monitoring the effects of land cover change666
on the supply of ecosystem services in an urban region: A study of santiago-valpara´ıso, chile’, PloS one 12(11), e0188117.667
Muthuramu, P. and Maheswari, T. U. (2019), ‘Tests for structural breaks in time series analysis: A review of recent develop-668
ment’, Shanlax International Journal of Economics 7(4), 66–79.669
Myers, N., Mittermeier, R. A., Mittermeier, C. G., Da Fonseca, G. A. and Kent, J. (2000), ‘Biodiversity hotspots for conservation670
priorities’, Nature 403(6772), 853–858.671
Nadrowski, K., Wirth, C. and Scherer-Lorenzen, M. (2010), ‘Is forest diversity driving ecosystem function and service?’, Current672
Opinion in Environmental Sustainability 2(1-2), 75–79.673
Nasirzadehdizaji, R., Balik Sanli, F., Abdikan, S., Cakir, Z., Sekertekin, A. and Ustuner, M. (2019), ‘Sensitivity analysis of674
multi-temporal sentinel-1 sar parameters to crop height and canopy coverage’, Applied Sciences 9(4), 655.675
Olofsson, P., Foody, G. M., Herold, M., Stehman, S. V., Woodcock, C. E. and Wulder, M. A. (2014), ‘Good practices for676
estimating area and assessing accuracy of land change’, Remote sensing of Environment 148, 42–57.677
Pang, Y., Zhang, L., Huang, C., Yu, X. and Li, Z. (2013), Automatic deforestation detection using time series landsat images678
in a tropical forest of china, in ‘2013 IEEE International Geoscience and Remote Sensing Symposium-IGARSS’, IEEE,679
pp. 3872–3875.680
Parr, T. W., Sier, A. R., Battarbee, R., Mackay, A. and Burgess, J. (2003), ‘Detecting environmental change: science and681
society—perspectives on long-term research and monitoring in the 21st century’, Science of the total environment 310(1-682
3), 1–8.683
Pasquarella, V. J., Ar´evalo, P., Bratley, K. H., Bullock, E. L., Gorelick, N., Yang, Z. and Kennedy, R. E. (2022), ‘Demystifying684
landtrendr and ccdc temporal segmentation’, International Journal of Applied Earth Observation and Geoinformation685
110, 102806.686
Payn, T., Carnus, J.-M., Freer-Smith, P., Kimberley, M., Kollert, W., Liu, S., Orazio, C., Rodriguez, L., Silva, L. N. and687
Wingfield, M. J. (2015), ‘Changes in planted forests and future global implications’, Forest Ecology and Management688
352, 57–67.689
Puhm, M., Deutscher, J., Hirschmugl, M., Wimmer, A., Schmitt, U. and Schardt, M. (2020), ‘A near real-time method for690
forest change detection based on a structural time series model and the kalman filter’, Remote Sensing 12(19), 3135.691
Redlich, S., Zhang, J., Benjamin, C., Dhillon, M. S., Englmeier, J., Ewald, J., Fricke, U., Ganuza, C., Haensel, M., Hovestadt, T.692
et al. (2022), ‘Disentangling effects of climate and land use on biodiversity and ecosystem services—a multi-scale experimental693
design’, Methods in Ecology and Evolution 13(2), 514–527.694
Reiche, J., Hamunyela, E., Verbesselt, J., Hoekman, D. and Herold, M. (2018), ‘Improving near-real time deforestation moni-695
toring in tropical dry forests by combining dense sentinel-1 time series with landsat and alos-2 palsar-2’, Remote Sensing of696
Environment 204, 147–161.697
Ribeiro, M. C., Martensen, A. C., Metzger, J. P., Tabarelli, M., Scarano, F. and Fortin, M.-J. (2011), ‘The brazilian atlantic698
forest: a shrinking biodiversity hotspot’, Biodiversity hotspots: distribution and protection of conservation priority areas699
pp. 405–434.700
Ritchie, H. and Roser, M. (2021), ‘Forests and deforestation’, Our World in Data .701
Rocha-Santos, L., Mayfield, M. M., Lopes, A. V., Pessoa, M. S., Talora, D. C., Faria, D. and Cazetta, E. (2020), ‘The loss of702
functional diversity: A detrimental influence of landscape-scale deforestation on tree reproductive traits’, Journal of Ecology703
108(1), 212–223.704
Rouse Jr, J., Haas, R., Schell, J. and Deering, D. (1973), Paper a 20, in ‘Third Earth Resources Technology Satellite-1705
Symposium: The Proceedings of a Symposium Held by Goddard Space Flight Center at Washington, DC on’, Vol. 351,706
p. 309.707
32
Salas, C., Donoso, P. J., Vargas, R., Arriagada, C. A., Pedraza, R. and Soto, D. P. (2016), ‘The forest sector in chile: an708
overview and current challenges’, Journal of Forestry 114(5), 562–571.709
S´anchez, J. J., Marcos-Martinez, R., Srivastava, L. and Soonsawad, N. (2021), ‘Valuing the impacts of forest disturbances on710
ecosystem services: An examination of recreation and climate regulation services in us national forests’, Trees, Forests and711
People 5, 100123.712
Sarricolea, P., Herrera-Ossandon, M. and ´
Oliver Meseguer-Ruiz (2017), ‘Climatic regionalisation of continental chile’, Journal713
of Maps 13(2), 66–73.714
Schultz, M., Clevers, J. G., Carter, S., Verbesselt, J., Avitabile, V., Quang, H. V. and Herold, M. (2016), ‘Performance of715
vegetation indices from landsat time series in deforestation monitoring’, International journal of applied earth observation716
and geoinformation 52, 318–327.717
Schultz, M., Verbesselt, J., Avitabile, V., Souza, C. and Herold, M. (2015), ‘Error sources in deforestation detection using bfast718
monitor on landsat time series across three tropical sites’, IEEE Journal of Selected Topics in Applied Earth Observations719
and Remote Sensing 9(8), 3667–3679.720
Sebald, J., Senf, C. and Seidl, R. (2021), ‘Human or natural? landscape context improves the attribution of forest disturbances721
mapped from landsat in central europe’, Remote Sensing of Environment 262, 112502.722
Silva Junior, C. H., Pessˆoa, A. C., Carvalho, N. S., Reis, J. B., Anderson, L. O. and Arag˜ao, L. E. (2021), ‘The brazilian723
amazon deforestation rate in 2020 is the greatest of the decade’, Nature Ecology & Evolution 5(2), 144–145.724
Smith-Ram´ırez, C., Castillo-Mandujano, J., Becerra, P., Sandoval, N., Fuentes, R., Allende, R. and Acu˜na, M. P. (2022),725
‘Combining remote sensing and field data to assess recovery of the chilean mediterranean vegetation after fire: Effect of time726
elapsed and burn severity’, Forest Ecology and Management 503, 119800.727
Song, C., Lee, W.-K., Choi, H.-A., Kim, J., Jeon, S. W. and Kim, J. S. (2016), ‘Spatial assessment of ecosystem functions and728
services for air purification of forests in south korea’, Environmental Science & Policy 63, 27–34.729
Thakur, S., Mondal, I., Ghosh, P., Das, P. and De, T. (2020), ‘A review of the application of multispectral remote sensing730
in the study of mangrove ecosystems with special emphasis on image processing techniques’, Spatial Information Research731
28(1), 39–51.732
Truong, C., Oudre, L. and Vayatis, N. (2020), ‘Selective review of offline change point detection methods’, Signal Processing733
167, 107299.734
USGS (2022), ‘Landsat 8-9 collection 2 (c2), level 2 science product (l2sp) guide’.735
Verbesselt, J., Zeileis, A. and Herold, M. (2011), ‘Near real-time disturbance detection in terrestrial ecosystems using satellite736
image time series: Drought detection in somalia’, Working Papers in Economics and Statistics p. 18.737
Wambui, G. D., Waititu, G. A. and Wanjoya, A. (2015), ‘The power of the pruned exact linear time (pelt) test in multiple738
changepoint detection’, Am. J. Theor. Appl. Stat 4(6), 581–586.739
Wand, M. P. and Jones, M. C. (1994), Kernel smoothing, CRC press.740
Weiss, M., Jacob, F. and Duveiller, G. (2020), ‘Remote sensing for agricultural applications: A meta-review’, Remote Sensing741
of Environment 236, 111402.742
Wu, L., Li, Z., Liu, X., Zhu, L., Tang, Y., Zhang, B., Xu, B., Liu, M., Meng, Y. and Liu, B. (2020), ‘Multi-type forest change743
detection using bfast and monthly landsat time series for monitoring spatiotemporal dynamics of forests in subtropical744
wetland’, Remote Sensing 12(2), 341.745
Yin, H., Prishchepov, A. V., Kuemmerle, T., Bleyhl, B., Buchner, J. and Radeloff, V. C. (2018), ‘Mapping agricultural land746
abandonment from spatial and temporal segmentation of landsat time series’, Remote sensing of environment 210, 12–24.747
Yin, J. and Bauerle, T. L. (2017), ‘A global analysis of plant recovery performance from water stress’, Oikos 126(10), 1377–1388.748
Zemp, D., Schleussner, C.-F., Barbosa, H. d. M. J. and Rammig, A. (2017), ‘Deforestation effects on amazon forest resilience’,749
Geophysical Research Letters 44(12), 6182–6190.750
33
Zeng, L., Wardlow, B., Xiang, D., Hu, S. and Li, D. (2020), ‘A review of vegetation phenological metrics extraction using751
time-series, multispectral satellite data’, Remote Sensing of Environment 237, 111511.752
Zhao, Y., Feng, D., Yu, L., Wang, X., Chen, Y., Bai, Y., Hern´andez, H. J., Galleguillos, M., Estades, C., Biging, G. S.,753
Radke, J. D. and Gong, P. (2016), ‘Detailed dynamic land cover mapping of chile: Accuracy improvement by integrating754
multi-temporal data’, Remote Sensing of Environment 183, 170–185.755
Zhu, Z. (2017), ‘Change detection using landsat time series: A review of frequencies, preprocessing, algorithms, and applica-756
tions’, ISPRS Journal of Photogrammetry and Remote Sensing 130, 370–384.757
Zhu, Z. and Woodcock, C. E. (2014), ‘Continuous change detection and classification of land cover using all available landsat758
data’, Remote sensing of Environment 144, 152–171.759
34
