ArticlePDF Available

Representation of tree cover in global land cover products: Finland as a case study area

Springer Nature
Environmental Monitoring and Assessment
Authors:

Abstract and Figures

Forest extent mapping is required for climate modeling and monitoring changes in ecosystem state. Different global land cover (LC) products employ simple tree cover (referred also as “forest cover” or even “vegetation cover”) definitions to differentiate forests from non-forests. Since 1990, a large number of forest extent maps have become available. Although many studies have compared forest extent data, they often use old data (i.e., around the year 2000). In this study, we assessed tree cover representations of three different annual, global LC products (MODIS VCF (MOD44B, Collection 6 (C6)), MCD12Q1 (C6), and CCI LC (v.2.1.1)) using the Finnish Multi-Source National Forest Inventory (MS-NFI) data for the year 2017. In addition, we present an intercomparison approach for analyzing spatial representations of coniferous and deciduous species. Intercomparison of different LC products is often overlooked due to challenges involved in non-standard and overlapping LC class definitions. Global LC products are used for monitoring changes in land use and land cover and modeling of surface fluxes. Given that LC is a major driver of global change through modifiers such as land surface albedo, more attention should be paid to spatial mapping of coniferous and deciduous species. Our results show that tree cover was either overestimated or underestimated depending on the LC product, and classification accuracy varied between 42 and 75%. Intercomparison of the LC products showed large differences in conifer and deciduous species spatial distributions. Spatial mapping of coniferous and deciduous tree covers was the best represented by the CCI LC product as compared with the reference MS-NFI data.
This content is subject to copyright. Terms and conditions apply.
Representation of tree cover in global land cover products:
Finland as a case study area
Titta Majasalmi &Miina Rautiainen
Received: 30 June 2020 /Accepted: 17 January 2021 /Published online: 12 February 2021
#The Author(s) 2021
Environ Monit Assess (2021) 193: 121
https://doi.org/10.1007/s10661-021-08898-2
Abstract Forest extent mapping is required for climate
modeling and monitoring changes in ecosystem state.
Different global land cover (LC) products employ sim-
ple tree cover (referred also as forest coveror even
vegetation cover) definitions to differentiate forests
from non-forests. Since 1990, a large number of forest
extent maps have become available. Although many
studies have compared forest extent data, they often
use old data (i.e., around the year 2000). In this study,
we assessed tree cover representations of three different
annual, global LC products (MODIS VCF (MOD44B,
Collection 6 (C6)), MCD12Q1 (C6), and CCI LC
(v.2.1.1)) using the Finnish Multi-Source National For-
est Inventory (MS-NFI) data for the year 2017. In addi-
tion, we present an intercomparison approach for ana-
lyzing spatial representations of coniferous and decidu-
ous species. Intercomparison of different LC products is
often overlooked due to challenges involved in non-
standard and overlapping LC class definitions. Global
LC products are used for monitoring changes in land use
and land cover and modeling of surface fluxes. Given
that LC is a major driver of global change through
modifiers such as land surface albedo, more attention
should be paid to spatial mapping of coniferous and
deciduous species. Our results show that tree cover
was either overestimated or underestimated depending
on the LC product, and classification accuracy varied
between 42 and 75%. Intercomparison of the LC prod-
ucts showed large differences in conifer and deciduous
species spatial distributions. Spatial mapping of conif-
erous and deciduous tree covers was the best represent-
ed by the CCI LC product as compared with the refer-
ence MS-NFI data.
Keywords Boreal .Canopy cover .Forest cover .
MODIS .CCI LC .MS-NFI
Introduction
Forest extent mapping is required for land cover (LC) and
land use classification, and for monitoring changes (e.g.,
damage, afforestation, deforestation) in ecosystem state.
In addition, many land surface model (LSM) components
of climate models, which simulate surface fluxes of mo-
mentum, heat, and moisture, need a land cover descrip-
tion. The success of classifying forest areas using a
satellite image depends on the applied definition and
estimation accuracy at a pixel level. Since 1990, a large
number of forest extent maps have become available for
different purposes at a spatial ground resolution ranging
from 300 meters to 5 kilometers (e.g., Bartholome and
Belward 2005;Friedletal.2002;Hansenetal.2003,
2000;Lovelandetal.2000; Poulter et al. 2015).
T. Majasalmi (*):M. Rautiainen
School of Engineering, Department of Built Environment, Aalto
University, P.O. Box 14100, 00076 Aalto, Finland
e-mail: titta.majasalmi@aalto.fi
M. Rautiainen
School of Electrical Engineering, Department of Electronics and
Nanoengineering, Aalto University, P.O. Box 14100,
00076 Aalto, Finland
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
The quantification of global forest resources spatially
and temporally relies on the international definition of
forest land area by the Food and Agriculture Organization
of the United Nations (FAO): a forest is land spanning
more than 0.5 hectares with trees higher than 5 meters
and a canopy cover of more than 10 percent, or trees able
to reach these thresholds in situ.(FRA 2015). The
canopy cover (CC) is defined as the proportion of ground
covered by vertical tree crown projection (Jennings et al.
1999). In practice, directly applying the FAO forest def-
inition in remote sensing remains challenging due to an
overlap in LC class definitions. For example, forest area
based on the FAO definition does not include land that is
predominantly under agricultural or urban land use (FRA
2015) even if the forest definition requirements would be
satisfied. Due to challenges in implementing the forest
definition by FAO, different global LC products employ
simple CC (referred also as tree cover,forest cover,
or even vegetation cover) definitions to differentiate
forests from non-forests. For example, the International
Geosphere-Biosphere Programme (IGBP) employs a tree
cover threshold of >60%, and the United Nations Frame-
work Convention on Climate Change (UNFCCC) uses
tree cover thresholds of >10% (e.g., FAO international
forest definition) and >30% (e.g., FAO global forest
ecological zone mapping) to classify an area as a forest.
The European Space Agency Climate Change Initiatives
Land Cover (ESA CCI LC) 19922015 map series has
been developed to provide a complete surface represen-
tation for global-scale modeling studies and employs a
tree cover threshold of >15% (which has two further sub-
classes: 1540% and >40%). The latest release of the
global CCI-LC products covers years 2016, 2017, and
2018. In addition, the binary forestor non-forest
classification, such as that employed by the FAO, has
been found to be insufficient on many occasions, and
thus, there is an ongoing process of moving towards
mapping tree cover as a continuous field (e.g., Sexton
et al. 2016). For instance, an operational vegetation con-
tinuous field (VCF) product (MODIS VCF 2017)from
MODIS (i.e., MOD44B) is known to suffer from
underestimating high cover and overestimating low cover
(e.g., Heiskanen 2008; Sexton et al. 2013).
There are many definitions of tree coverto suit
different needs. The definition of tree crown cover ac-
counts for within-crown gaps as a part of the crown, and
is measured in the vertical direction without double-
counting of overlapping crown projection areas (e.g.,
Gschwantner et al. 2009). Alternative definitions, which
are sometimes used interchangeably, such as canopy
closure (i.e., the fraction of hemispherical sky visibility
at one single spot on the ground (Jennings et al. 1999))
or effective canopy cover (i.e., takes into account both
gaps between crowns and gaps within crowns
(Rautiainen et al. 2005)) also exist. The difference be-
tween definitions is related to the context in which they
are used: the effective CC is preferred when estimating
ecological variables such as fraction of absorbed photo-
synthetically active radiation (fPAR) or leaf area index
(e.g., Chen et al. 1997; Gower et al. 1999), while the CC
is mainly used to map forest extent and dynamics (e.g.,
Poulter et al. 2015). Ignoring whether a tree crown
contains small gaps or not (i.e., application of different
definitions) results in systematic biases. However, it is
often not easy to determine which exact tree cover
definition the satellite-based LC products are based on.
Thus, for simplicity, tree coveris used in this paper to
refer to all tree cover definitions employed by different
LC products.
The increasing importance of spatially explicit data
on forest extent mapping is highlighted by recent papers
such as Bastin et al. (2019). They mapped the global
potential of increasing tree cover to mitigate climate
change and concluded that at least 900 million hectares
of forest could be restored under current climate condi-
tions. Yet, their conclusions rely on existing tree cover
maps, that is, the potential tree cover was obtained after
subtracting the current tree cover from a tree cover map
produced by Hansen et al. (2013). This demonstrates the
necessity of accurate initial mapping of forest extents
and properties. Errors ininitial mapping and ambiguities
in the employed forest areal definition can be expected
to transfer to the results and conclusions of follow-up
studies utilizing the LC product. In addition, although
many recent studies comparing different products exist,
they are using old data, such as from years 19922010
(e.g., Tang et al. 2019;Songetal.2014), and rarely
consider the impacts of employed classification on spa-
tial patterns of conifer and deciduous tree covers.
In any boreal region, the separation between conifer-
ous and deciduous species groups is essential (e.g., to
characterize seasonal courses of surface fluxes in a
climate model). This has been recognized and currently,
all global LC products differentiate between forests
belonging to different phenological groups, such as
evergreen needleleaf forest (ENF) and deciduous broad-
leaf forest (DBF). However, as all LC products employ
their own class labeling systems and different tree cover
121 Page 2 of 19 Environ Monit Assess (2021) 193: 121
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
thresholds, validation and intercomparison of conifer
and deciduous tree cover representations of different
LC products have remained complicated.
Validation of low and medium spatial resolution tree
cover products is challenging with traditional field in-
ventories, and thus, higher spatial resolution optical
satellite data based products provide the best basis for
assessing and evaluating different medium spatial reso-
lution LC classifications. In general, the preference for
finer spatial resolution is supported by the smaller frac-
tion of mixed pixels, that is, containing other LC types
such as built environment or grassland. Furthermore, in
some countries, there already exist nationwide multi-
source (MS) thematic maps produced by National For-
est Inventories (NFIs). These maps quantify local and
regional forest characteristics with a high spatial resolu-
tion, at least when working in a global LSM context.
The concepts of low, medium, and high spatial resolu-
tion are highly dependent on the discipline they are used
in. In this study, low spatial resolution refers to spatial
resolution commonly applied in LSM at 0.05° (i.e.,
approximately 5.6 km), whereas medium and high spa-
tial resolutions denote, for example, 250500 m, and 16-
m map products, respectively.
In boreal Finland, systematic monitoring and classi-
fication of forested areas have been conducted since the
1920s, in regular 510-year cycles (Tomppo et al.
2010). The NFI was developed to produce information
on regional forest resources (i.e., stem volume, growth
and quality of growing stock, and forest carbon stocks
and their changes, forest health, biodiversity of forests,
land use structure, and forest ownership), and satellite
images have been used in NFIs since the 1980s. Finnish
forests are classified into groups of forest land, poorly
productive forest land, and unproductive land depend-
ing on forest wood production capability. All Finnish
boreal forest is secondary forest (i.e., no pristine forest
exists) and under active forest management. Forest floor
understory species have been inventoried using a na-
tionwide sampling grid three times (in the 1950s, 1980s,
and 1995), and the fourth round will take place between
2021 and 2022 (LUKE 2020). The high-resolution MS-
NFI maps are produced based on data from NFI field
plots, remotely sensed data, and other information
sources by national forest authorities and thus provide
the best local estimate on the forest characteristics. They
also carry information regarding the share of deciduous
species and can be used to map geographical variations
in land surface phenology (Moon et al. 2019).
In LSMs, LC products define the spatial locations of
different vegetation types and often employ what is
called tile-based approaches (or subgrid of vegetation
coveror fractional cover). In other words, for each
LC unit, information on cover fractions of different LC
classes is needed (e.g., Masson et al. 2003;Zengetal.
2002). In addition, vegetation phenology of deciduous
areas is often estimated based on leaf area index (LAI,
m
2
/m
2
) which is scaled using remotely sensed time-
series of normalized difference vegetation index
(NDVI) (e.g., Masson et al. 2003;Zengetal.2002).
Thus, when assessing spatial patterns of tree cover, in
the context of an LSM, a systematic way to intercom-
pare tree cover definitions and to separate coniferous
and deciduous tree covers need to be established first. In
recent years, numerous processes that are known to
impact the Earth system dynamics have been incremen-
tally added to LSMs (see, e.g., Fisher and Koven 2020):
representations of soil moisture dynamics, land surface
heterogeneity, plant and soil carbon cycling, nitrogen
and phosphorus cycling, among others. However, while
very detailed processes can be represented in site or
regional scale simulations, on a global scale, identifying
the locations of different vegetation types is of utmost
importance. While there have been attempts to represent
forest-age structural properties and to account for forest
management effects in a climate model simulation, these
LC products and models suffice only for regional anal-
yses (see, e.g., Majasalmi et al. 2018,2020).
Due to a large number of different classifications
(and class definitions), direct comparison of different
LC products remains complicated. While the com-
mon approach to validate LC datasets is to aggregate
the finer spatial resolution data to the lower spatial
resolution of the LC dataset to study class bin fre-
quencies and confusion matrices, that does not sup-
port developing the classifications and improving
class definition thresholds. In addition, as the accu-
racy of classification may be expected to be the worst
with the largest number of classes, and the best for
the most simplified classifications, the confusion ma-
trix approach does not allow meaningful intercom-
parison of the products. Alternatively, the categorical
classifications can be transformed into continuous
tree cover estimates based on their class definition
thresholds (bin means). After all products have
been converted to continuous estimates of tree cover,
different metrics can be used to intercompare the tree
covers both statistically and spatially.
Environ Monit Assess (2021) 193: 121 Page 3 of 19 121
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
The aims of this study are: (1) to assess and inter-
compare tree cover classifications currently employed
by three different global LC product series and (2) to
demonstrate an approach which allows a flexible inter-
comparison of different tree cover representations of the
LC products. Our case study is based on the boreal
forests in Finland. For the intercomparison of conifer
and deciduous tree covers, we analyzed the data after it
was aggregated to the spatial resolution employed in
climate modeling studies (i.e., using the Climate Model-
ing Grid).
Materials and methods
Study area
Materials
MS-NFI maps
Since 2012, Finnish MS-NFI data has been publicly
available online in raster map form. The raster maps
are provided for over 40 different forest themes such as
dominant tree species, growing stock volume, and bio-
mass at a spatial resolution of 16 m × 16 m. In this study,
we used the MS-NFI thematic maps of canopy cover
2017 (%),”“deciduous canopy cover 2017 (%),and
Loreys height (the height of the median tree) of H
(dm). In addition, a mask to extract Finnish land surface
area from the global products was prepared using the
MS-NFI thematic map of land class based on FAO
FRA.The uniqueness of the MS-NFI maps comes
from the fact that the algorithm has been trained using
forest inventory data from 53,989 field plots covering
the entirety of Finland.
The MS-NFI 2017 products are available as ready
products (©Natural Resources Institute Finland 2019;
for further details, see, e.g., Mäkisara et al. (2019)). MS-
NFI maps are processed using an improved k-nearest-
neighbor method in which the weights of the features
were sought using an optimization method based on a
genetic algorithm. The other land categories, as well as
water bodies, were delineated out using the elements of
the topographic database of the Land Survey of Finland.
The optical satellite images used to create the products
included eight Sentinel-2A MSI images, six Sentinel-2B
MSI images, and 19 (seven orbits) Landsat 8 OLI im-
ages. The original map coordinate system is EPSG 3047
(ETRS-TM35FIN).
The tree cover definition of MS-NFI tree cover
maps is “…canopy cover of trees is the vertical
projection area on the horizontal plane of the cano-
pies of the individual trees on a field plot.As
canopy cover definition does not separate the con-
tributions of coniferous (evergreen) and deciduous
(broadleaved) species, deciduous canopy cover was
derived based on the basal area (i.e., the cross-
section area of the tree stems of a stand per hectare
and measured at a height of 1.3 m (MS-NFI 2019)).
According to the MS-NFI (2019)manual:the can-
opy cover proportion of broad-leaved trees is de-
rived from the total cover using the basal area.
However, in the seedling stands, the canopy cover
of broad-leaved trees is assessed using the shares of
the stem numbers.The magnitude of the average
error of the estimates of tree cover at pixel level
varies between 14 and 20% and average tree height
(H, m) between 3.5 and 5.9 m (MS-NFI 2019). It is
noteworthy that, although the estimation errors at
pixel level are rather high, they tend to decrease as
the area of the interest increases and contains more
pixels (MS-NFI 2019).
121 Page 4 of 19 Environ Monit Assess (2021) 193: 121
We used Finland (area 338,440 km
2
) as the case study
area because more than 70% of Finnish land surface area
is classified as forest by FAO (FRA FIN 2015)(i.e.,
~222,180 km
2
). Finland is located in the Northern Eu-
ropean boreal zone area (bounding box: 20.6° E, 31.5°
E, 59.8° N, 70.1° N) and Finnish forests have been
under intensive forest management for several decades
and belong to the most intensively studied forest areas in
Europe. Finnish boreal forest is dominated by evergreen
conifers Norway spruce (Picea abies) and Scots pine
(Pinus sylvestris). Broad-leaved deciduous tree species
Silver birch (Betula pendula) and Downy birch (Betula
pubescence) rarely form pure(i.e., single species)
forests. Other broad-leaved or coniferous deciduous tree
species such as European aspen (Pupulus tremula),
alders (Alnus glutinosa,Alnus incana), English oak
(Quercus robur), Norway maple (Acer platanoides), or
larches (Larix spp.) occasionally occur as a mixed spe-
cies but never form pure forests. At the landscape level,
about 10% of tree cover is deciduous. Thus, in this
study, coniferousrefers to evergreen conifers (i.e.,
pine and spruce), while deciduousis used to include
all tree species that shed their leaves for winter (i.e.,
broad-leaved trees and larches).
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
Global land cover products used by different classifications are available from
MODIS MCD12 (2020).
The MOD44B VCF is a yearly product representing
global surface vegetation cover as gradations of three
ground cover components: percent tree cover, percent
non-tree cover, and percent non-vegetated (bare). In this
study, only the component percent tree coverwas
used. The VCF product map crs is EPSG 4326.
The CCI annual land cover product (v.2.1.1) follows
standardized hierarchical classification by the United Na-
tions Land Cover Classification System (UN-LCCS),
which allows conversion from land cover classes into PFTs
using a cross-walking table (e.g., Poulter et al. 2015). The
LC product legend is available from CCI LCCS (2020).
The product was provided in the netcdf file format, and
after rasterization was projected to crs of EPSG 4326.
Climate Modeling Grid
The low-resolution LC product MCD12C1, also known
as the MODIS Climate Modeling Grid (MODIS CMG,
i.e., at 0.05° ~ 5.6 km) (MODIS CMG 2020), was used
here as a basis for spatially comparing the different
medium and high spatial resolution LC product classi-
fications. The MCD12C1 projection is EPSG 4008.
Processing
Preprocessing
Spatial subsets were clipped from the global LC prod-
ucts (i.e., VCF, CCI, and the six MCD12Q1 products) to
cover an area defined by the bounding box of Finland.
Tabl e 1 Land cover/use products used in this study. Abbrevia-
tions: Productis the official product name, Nameis the acro-
nym used to refer to particular dataset, Resolutionis the original
spatial resolution of the product, Type determines if labeling is
continuous or categorical (i.e. classified)
Product Product type Name Resolution Type Reference
MS-NFI Canopy cover MS-NFI 16 m Continuous Mäkisara et al. 2019
MOD44B Tree cover VCF 250 m Continuous MODIS VCF 2020
CCI LC Land cover CCI 300 m Classified Poulter et al. 2015
MCD12Q1 Land cover IGBP 500 m Classified Loveland and Belward 1997
MCD12Q1 Land cover LAI 500 m Classified Myneni et al. 2002
MCD12Q1 Land cover PFT 500 m Classified Bonan et al. 2002
MCD12Q1 Land cover LCCS1 500 m Classified MODIS LCCS 2020
MCD12Q1 Land cover/use LCCS2 500 m Classified MODIS LCCS 2020
MCD12Q1 Land cover/use LCCS3 500 m Classified MODIS LCCS 2020
Environ Monit Assess (2021) 193: 121 Page 5 of 19 121
We used annual Collection 6 (C6) MODIS products: (1)
land cover products suite of MCD12Q1 and (2) MOD44B
vegetation continuous fields (VCF), as well as a land cover
product by the European Space Agency Climate Change
Initiative (ESA CCI), which is the only non-MODIS-based
data product. All products (Table 1) are annual products
for the year 2017. VCF is based on regression trees
(MODIS VCF User guide 2020),andMCD12Q1land
cover products suite is produced using supervised classifi-
cation techniques such as decision trees (MODIS MCD12
2020) and ensemble classification methods (Friedl et al.
2002,2010). The CCI LC is based on machine learning
methods that combine supervised and unsupervised algo-
rithms (Poulter et al. 2015).
The MODIS Land Cover Type product (MCD12Q1)
contains a suite of science data sets which map land
cover globally (and annually) using six different land
cover legends. They include five legacy classifications
(IGBP, UMD, LAI, BGC, and PFT) and three Land
Cover Classification System (LCCS) layers from the
FAO (the first is meant for land cover, the second for
land use, and third for surface hydrology applications).
We used all of these classifications except two: we
excluded the BGC classification from this study because
the tree height data used in MS-NFI field inventories
was measured from a breast-height-diameter of 1.3 m,
whereas the BGC product uses a definition of 1 m. In
addition, as IGBP and UMD schemes are equal for
forest classes, UMD was also excluded from the analy-
ses. The product coordinate reference system (crs) is
SR-ORG 6842 (i.e., MODIS Sinusoidal). The legends
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
These subsets were projected to Finnish UTM (i.e.,
EPSG 32635, zone 35) and masked using land class
based on FAO FRAto include only areas belonging to
Finnish land surface area (i.e., excluding land areas of
neighboring countries). The MS-NFI tree height and
tree cover tiles were mosaicked and projected to Finnish
UTM. The coniferous tree cover fraction was obtained
by subtracting deciduous tree cover from the total tree
cover. The MS-NFI data were resampled to correspond
to the LC product extents and resolutions (i.e., VCF,
CCI, and MCD12Q1).
Assessment of the LC product tree cover classes
To assess the tree cover classifications employed by the
global LC products, the MS-NFI data was classified into
the LC product classes and only tree cover classes were
included (i.e., classes such as LCCS2 class forest/crop-
land mosaicwere excluded). Details regarding the
classes (for original legend class definitions, see
MODIS MCD12 (2020)andCCILCCS(2020)) and
their implementation in our case study are provided in
Table 2. It is noteworthy that the VCF and the CCI
classifications do not employ any tree height thresholds
(i.e., canopy > 2 m) which are used by majority of the
other classifications. The default continuous VCF tree
cover product was binned into five equally spaced bins
to assess the overall classification performance at low
and high tree covers (see details regarding the bins in
Tab le 2).
All CCI LC product classes (i.e., NET, BDT, NDT,
and mixed; see explanation of acronyms in Table 2)
employ a tree cover definition of >15%.CCI does
not provide a definition for its mixed class and thus, it
was classified using the respective IGBP mixed class
definition. Additionally, a separate three-step classifica-
tion was needed to assign the MS-NFI data into the CCI
LC product classes, because otherwise nearly all pixels
would have been assigned to the most abundant (NET)
class. First, all pixels belonging to the mixed-tree cover
class were classified into the mixed class. Then, after
excluding mixed-classified pixels, all pixels where
BDT+NDT tree cover was larger 15% were classified
into the BDT+NDT class. Finally, after excluding pixels
classified into mixed and BDT+NDT classes, all pixels
where NET tree cover was larger than 15% were classi-
fied into the NETclass. It can be noted that Finland does
not have large NDT dominated forest areas, but the
global LC products do classify some Finnish land area
to belong into NDT class. Thus, we had to assign those
pixels either into either the coniferous or deciduous
species group. Since NDT tree canopy winter albedo is
more similar to that of deciduous trees than that of
evergreen conifers, and because winter albedo has a
strong impact on the land surface energy balance at high
latitudes, NDT was classified as belonging to the decid-
uous species group.
121 Page 6 of 19 Environ Monit Assess (2021) 193: 121
Intercomparison of the tree cover classes in LC
products
In order to allow better intercomparison of different LC
products and classifications, an approach called a trans-
lation legend(Table 3) was developed. The idea of a
translation legend is to translate each categorical forest
class into a numeric tree cover fractionusing a class
mean tree cover value and to separate the tree cover into
coniferous and deciduous tree covers following the
original LC product legend class definitions. Notably,
the number of LC classes that can be included is higher
when using an approach based on a translation legend
than in classifying MS-NFI data into LC product clas-
ses. This is because partially covered tree cover classes
(such as LCCS2 class forest/cropland mosaic)canbe
included.
For most classifications, tree cover and vegetation
height thresholds were used. The only exception was
the CCI product which does not apply a height thresh-
old. The LCCS3 class Woody wetlandswas the only
classification using a height threshold over 1 m (i.e., for
all other classifications employing height threshold, it is
>2 m) and thus, the >2-m height threshold was used also
for that class. Finally, note that as VCF is continuous it
does not count as classification, despite being a tree
cover product.
First, for each LC product pixel, the original legend
class was replaced with a mean tree cover estimate as
defined by its original legend definition. For example, in
the case of the IGBP forest tree cover limit of >60%
would have 80% mean tree cover (i.e., maximum being
100%), and respectively IGBP woody savanna 30
60%would have 45% mean tree cover (Table 3). All
other classes not listed in Table 3were assigned a forest
tree cover value of zero. The mosaicclasses (i.e.,
containing a mixture of LC types) pose challenges,
and here, a 25% rule was applied in the absence of better
information (e.g., in the case of IGBP and CCI). In other
words, 25% of the pixel was assumed to have tree cover.
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
Tabl e 2 LC-product tree cover class definitions and details re-
garding classification implementation of MS-NFI data. Abbrevia-
tions: CC = total tree cover, CC_c = conifer tree cover, CC_d =
deciduous tree cover. Note, in absence of clear definition for the
CCI mixed-class, it was classified using the respective IGBP/
LCCS1 mixed-class definition
Classification Original legend class
(and class number) or bin
Details regarding classification
implementation
VCF CC 10 bin(0.0120)>CC 10
CC 30 bin(20.0140)>CC 30
CC 50 bin(40.0160)>CC 50
CC 70 bin(60.0180)>CC 70
CC 90 bin(80.01100) >CC 90
IGBP (UMD) ENF (1) CC_c> CC_d; CC> 60; h>2
DNF+ DBF (3, 4) CC_d> CC_c; CC> 60; h>2
Mixed (5) CC> 60; CC_c & CC_d= 4060; h>2
Woody savannas (8) 30> = CC <= 60; h>2
Savannas (9) 10> = CC <30; h>2
LAI EBF+ENF (5,7) CC>60; CC_c>CC_d; h>2
DBF+ DNF (6,8) CC> 60; CC_d> CC_c; h>2
Savannas (4) 10> = CC <= 60; h>2
PFT ENT+ EBT (1,2) CC >10; CC_c>CC_d; h>2
DNT+DBT (3,4) CC >10; CC_d > CC_c; h>2
LCCS1 ENF (11) CC> 60; CC_c> CC_d; h>2
DNF+ DBF (13,14) CC> 60; CC_d> CC_c; h>2
Mixed (15,16) CC> 60; CC_c & CC_d= 4060; h>2
Open forest (21) 30 >= CC< = 60; h>2
Sparse forest (22) 10 > = CC<30; h>2
LCCS2 Dense forest (10) CC >60; h>2
Open forest (20) 10 >= CC< = 60; h>2
LCCS3 Dense forest (10) CC >60; h>2
Open forest (20) 10 >= CC< = 60; h>2
CCI NET (70) CC_c>15
BDT+ NDT (60,80) CC_d> 15
Mixed (90) CC> 15; CC_c and CC_d= 4060
Environ Monit Assess (2021) 193: 121 Page 7 of 19 121
As VCF contains a continuous tree cover fraction, no
translations were done. It is noteworthy that we may
expect the forest cover values to saturate at around 80%
cover due to the applied transition legend values.This is,
however, a reasonable assumption in boreal forests, as
boreal forest trees tend to have long and narrow crowns
and have large gaps between individual crowns, where-
as in temperate forests the tree cover may often be close
to 100% (e.g., Horn 1971). For all classifications sepa-
rating coniferous and deciduous tree covers, the classes
without species share information were assumed to ac-
count for 50% of the total tree cover (e.g., IGBP Sa-
vannaclass note markings (50)in Decid. (%) and
Conifer (%) columns). For classifications that do not
separate between species, no separation was attempted.
Spatial aggregation and mapping
The MS-NFI data was first aggregated to correspond to
the LC product resolutions (i.e., VCF, CCI, and
Footnotes: VCF vegetation continuous fields, IGBP International Geosphere-Biosphere Programme, UMD University of Maryland, ENF
evergreen needleleaf forests, DNF deciduous needleleaf forests, DBF deciduous broadleaf forests, LAI leaf area index, EBF evergreen
broadleaf forest, PFT plant functional types, ENT evergreen needleleaf trees, EBT evergreen broadleaf trees, DNT deciduous needleleaf
trees, DBT deciduous broadleaf trees, LCCS1-3 FAO-Land Cover Classification System land cover 13, CCI climate change initiative, NET
needleleaved evergreen trees, BDT broad-leaved deciduous trees, NDT needleleaved deciduous trees
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
MCD12Q1) using aggregation factors (i.e., taking the
mean of X×Xpixel windows to create larger cells) of 9,
15, and 23 for VCF, CCI, and MCD12Q1, respectively.
These factors were obtained after all the data was
projected to Finnish UTM by calculating how many of
the smaller pixels are needed to fill in one larger pixel
and taking the mean of these pixels.
After aggregation, the MS-NFI data was resampled
to LC product resolutions (i.e., VCF, CCI, and
MCD12Q1) and reclassified into three tree cover bins
of 030%, 3160%, and 61100% in order to assess the
classification flexibility to represent areas with high,
moderate, and low tree cover.
After the translation legend had been used to convert
categorical cover classes to continuous tree covers, all
tree cover data was aggregated to MODIS CMG reso-
lution to inspect the spatial pattern in species distribu-
tions. Separate coniferous and deciduous tree cover
maps were created using aggregation factors of 224 for
MS-NFI maps, 15 for CCI, and 9 for MCD12Q1. The
Tabl e 3 Translation legend* for intercomparison of land cover/
use classifications by converting categorical classes into continu-
ous representations of tree cover. For simplicity, we used a 2 m
height threshold (i.e., H-limit-column) for all except the CCI
classification. Pixel total tree (%)shows the amount of within-
pixel tree cover, andDecid.(%)and Conifer (%)the respective
species composition. For classifications that do not separate be-
tween species, no separation was attempted
Classification Original legend class (and class number) CC limit (%) Hlimit (m) Total tree (%) Decid. (%) Conifer (%)
IGBP (UMD) ENF (1) >60 >2 80 100
DNF+ DBF (3, 4) >60 >2 80 100
Mixed savannas (5) >60 >2 80 50 50
Woody savannas (8) 3060 >2 45 (50) (50)
Savannas (9) 1030 >2 20 (50) (50)
Cropland/Natural veg. mosaic (14) 25 (>2) 25 (50) (50)
LAI EBF+ ENF (5,7) >60 >2 80 100
DBF+ DNF (6,8) >60 >2 80 100
Savannas (4) 1060 >2 35 (50) (50)
PFT ENT+EBT (1,2) >10 >2 55 100
DNF+ DBT (3,4) >10 >2 55 100
LCCS1 ENF (11) >60 >2 80 100
DNF+ DBF (13,14) >60 >2 80 100
Mixed (15,16) >60 >2 80 50 50
Open forest (21) 3060 >2 45 (50) (50)
Sparse forest (22) 1030 >2 20 (50) (50)
LCCS2 Dense forest (10) >60 >2 80
Open forest (20) 1060 >2 35
LCCS3 Dense forest (10) >60 >2 80
Open forest (20) 1060 >2 35
Woody wetlands (27) >10 (>2) 45
Tundra (51) <10 (>2) 5
CCI BDT + NDT (60, 80) >15 57.5 100
NET (70) >15 57.5 100
Mixed (90) 50 50 50
Mosaics (30, 40, 100, 110) 25 (50) (50)
Sparse vegetation (150, 160) <15 7.5 (50) (50)
*Footnotes: IGBP International Geosphere-Biosphere Programme, UMD University of Maryland, ENF evergreen needleleaf forests, DNF
deciduous needleleaf forests, DBF deciduous broadleaf forests, LAI leaf area index, EBF evergreen broadleaf forest, PFT plant functional
types, ENT evergreen needleleaf trees, EBT evergreen broadleaf trees, DNT deciduous needleleaf trees, DBT deciduous broadleaf trees,
LCCS1-3 FAO-Land Cover Classification System land cover 13, CCI climate change initiative, NET needleleaved evergreen trees, BDT
broad-leaved deciduous trees, and NDT needleleaved deciduous trees
121 Page 8 of 19 Environ Monit Assess (2021) 193: 121
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
resulting maps allowed spatial intercomparison of co-
niferous and deciduous tree covers at a resolution mean-
ingful for a range of LSMs.
Accuracy assessment
The assessment of the LC classifications included anal-
ysis of pixel counts belonging to different LC classes
and confusion matrices between the LC product classes
and MS-NFI data based classes. For the LC product
intercomparison, we used confusion matrices, the root
mean squared error (RMSE), the mean bias error
(MBE), and coefficient of determination (r
2
).
The MBE and RMSE are defined as:
MBE ¼n
i¼1PiRi
ðÞ
nð1Þ
RMSE ¼ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
n
i¼1PiRi
ðÞ
2
n
sð2Þ
where iis the pixel index, Pis the tree cover from the
LC products (after applying the translation legend), Ris
the tree cover from the MS-NFI data, and nis the sample
size. The RMSE and MBE were calculated also for the
low, moderate, and high tree cover bins (i.e., at 030%,
3160%, and 61100%, respectively).
Results
Assessment of tree cover estimates in the LC products
Based on pixel counts belonging to different tree cover
classes (i.e., ignoring the spatial distribution of those clas-
ses), tree cover in Finland was either slightly overestimated
or underestimated by different LC products and classifica-
tions (Table 4). Compared to the MS-NFI data, underesti-
mation of the tree cover was noted for VCF, PFT, and CCI
products (at 6%, 2%, and 11%, respectively). For example,
the 11% underestimation observed in the CCI product
indicates that 11% of tree-covered areas in Finland were
not classified as forest by the CCI product. A slight over-
estimation of tree-covered areas was noted for most
MODIS-based classifications. The overestimation of tree
cover was 3% for LAI, LCCS1, and LCCS2, and 2% for
IGBP and LCCS2. Thus, the best performing classifica-
tions in terms of mapping tree-covered areas were PFT and
IGBP, and LCCS3.
Intercomparison of tree cover estimates in the LC
products
Tree cover estimates derived from the LC products
using the transition legend were almost always system-
atically higher than the tree cover values obtained from
the MS-NFI data (Table 6). Using all data, the classifi-
cation with the smallest deviation from the reference
MS-NFI values was VCF (RMSE of 16.7% and MBE
of 3.4), closely followed by CCI (RMSE of 20.4 and
MBE of 9.6) (Table 6). As could be expected, for all
Environ Monit Assess (2021) 193: 121 Page 9 of 19 121
In terms of classification accuracy (i.e., accounting
for the spatial locations of pixels belonging to different
classes), the poorest performance was observed, as ex-
pected, for classification with the largest number of
classes (i.e., IGBP and LCCS1) (Table 5). The accuracy
of these two classifications was approximately 42%.
The continuous VCF product, which was reclassified
into five bins to analyze the overall classification per-
formance, performed equally poorly based on its classi-
fication accuracy. The CCI (49%), and LAI, LCC2, and
LCCS3 classifications (each ~60%) were more accurate
then IGBP, LCCS1, and VCF classifications (each
~42%) (percentage of accurately classified pixels in
parenthesis). Note that LCC2 and LCCS3 employed
the same classification (Table 2), but had differences
in spatial mapping of class open forest.The highest
accuracy was observed for the classification PFT (accu-
racy was ~75%) as it had only two classes.
VCF underestimated most areas with low and high
tree covers (i.e., bins CC 10and CC 70), while it
overestimated the area belonging to the two intermediate
bins (i.e., CC 30and CC 50)(Table5). For IGBP and
LCCS1 (which employ the same classification but have
differences in LC products), the poorest performing class
was mixed, which contained approximately 12% of
pixels but based on MS-NFI data, only 0.8% of pixels
belonged to that class. Deciduous tree cover classification
results of IGBP, LCCS1, and LAI were similar due to
identical class definitions (Table 2). According to Table 5,
the most accurate single-class classifications were woody
savanna for IBGP (29.1%), savanna for LAI (51%), and
open forest for LCCS1 (29.4%). The CCI map product
only correctly detected 0.6% of the deciduous tree cover
whereas the respective MS-NFI-based estimate was 10%
(Table 5). The finer spatial resolution of the CCI product
compared to MCD12Q1 classifications (Table 1) clearly
improved mapping deciduous tree cover.
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
MCD12Q1 based classifications, almost equal RMSE
values (RMSE all”—column in Table 6) were obtained
using different classifications. However, in general, the
tree cover representation was more successful in areas
with high tree cover than in areas with low tree cover.
For VCF, almost equal RMSE values were noted for
the highest and lowest tree cover bins, whereas for all
other classifications, the general pattern was that the
largest RMSE values occurred in the lowest tree cover
bin and got smaller towards higher tree cover bins
Tabl e 4 Validation statistics. LC% and MS-NFI% columns con-
tain fractional covers of pixels belonging to different classes, using
MS-NFI classified total pixel count as a denominator (other
abbreviations are explained in Table 2; note, also LC% column
values were divided with MS-NFI total pixel count)
Classification Original legend class
(and class number) or bin
Pixel count, LC Pixel count, MS-NFI LC% MS-NFI%
VCF CC 10 2,219,830 3,118,988 17.06 23.96
CC 30 4,389,648 3,181,799 33.73 24.45
CC 50 4,956,781 4,131,576 38.08 31.74
CC 70 692,488 2,578,757 5.32 19.81
CC 90 32 4168 0 0.03
sums: 12,258,779 13,015,288 94 100
IGBP ENF (1) 383,769 266,961 18.78 13.07
DNF+ DBF (3, 4) 267 2095 0.01 0.1
Mixed (5) 291,834 17,942 14.28 0.88
Woody savannas (8) 1,105,743 1,186,598 54.12 58.07
Savannas (9) 292,947 569,697 14.34 27.88
sum: 2,074,560 2,043,293 102 100
LAI EBF+ ENF (5, 7) 673,910 279,957 32.98 13.7
DBF+ DNF (6, 8) 5687 7041 0.28 0.34
Savannas (4) 1,429,597 1,756,295 69.97 85.95
sum: 2,109,194 2,043,293 103 100
PFT ENT +EBT (1, 2) 1,988,314 1,923,139 97.31 94.12
DNT+ DBT (3, 4) 19,637 120,109 0.96 5.88
sum: 2,007,951 2,043,248 98 100
LCCS1 ENF (11) 387,342 266,961 18.96 13.07
DNF+ DBF (13,14) 285 2095 0.01 0.1
Mixed (15,16) 292,003 17,942 14.29 0.88
Open forest (21) 1,130,093 1,186,598 55.31 58.07
Sparse forest (22) 303,059 569,697 14.83 27.88
sum: 2,112,782 2,043,293 103 100
LCCS2 Dense forest (10) 679,597 286,998 33.26 14.05
Open forest (20) 1,429,597 1,756,295 69.97 85.95
sum: 2,109,194 2,043,293 103 100
LCCS3 Dense forest (10) 675,903 286,998 33.08 14.05
Open forest (20) 1,402,234 1,756,295 68.63 85.95
sum: 2,078,137 2,043,293 102 100
CCI NET (70) 3,229,056 3,413,809 74.37 78.63
BDT+ NDT (60,80) 90,648 543,029 2.09 12.51
Mixed (90) 525,994 384,789 12.12 8.86
sum: 3,845,698 4,341,627 89 100
121 Page 10 of 19 Environ Monit Assess (2021) 193: 121
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
(Table 6). For VCF, a positive bias was noted for the
lowest tree cover bin which became increasingly nega-
tive towards higher tree covers. In general, positive
biases were noted for both low and moderate tree cover
bins, as well as some negative biases for the highest tree
cover bin (Table 6). The most linear relationship be-
tween the tree cover derived from the LC products and
the MS-NFI data was observed for VCF. As linearity
results from having more tree cover classes with
(possibly) different mean-tree cover values, it can be
used to assess classification skills to represent variation
in tree cover values. In other words, the low r
2
(Table 6)
values indicate that there is little variation in tree cover
values between coarse spatial resolution LC product
pixels compared to that represented by MS-NFI data.
Confusion matrices were used to illustrate some de-
ficiencies in classification performance. For example,
CCI, PFT, and LCCS2-based tree cover estimates are
not able to represent areas belonging to the highest tree
cover bin (Table 7). PFT and LAI classifications, on the
Tabl e 5 Confusion matrix between different tree cover classes (or
bins) at LC product original resolution. NA = no tree cover, W.
Sav. = Woody savannas, Sav. = Savannas, Open fo. = Open forest,
Sparse fo. = Sparse forest, Dense fo. = Dense forest (see Table 2
for other abbreviations)
VCF CCI
NA CC 10 CC 30 CC 50 CC 70 CC 90 Sum: NA ENT DBT/NDT Mixed Sum:
NA 0.0 0.4 0.2 0.0 0.0 0. 0 0.7 NA 17.0 4.0 0.5 0.6 22.1
CC 10 0.9 9.8 7.3 2.6 0.2 0.0 20.6 ENT 9.9 46.6 0.3 4.5 61. 3
CC 30 0.2 6.1 12.1 6.1 0.3 0.0 24.9 DBT/DNT 1.8 4. 7 0. 6 2.6 9.7
CC 50 0.0 1.4 12.7 17.0 1.8 0.0 33.0 Mixed 2. 2 2.7 0. 3 1.8 6. 9
CC 70 0.0 0.1 3.1 14.2 3.3 0. 0 20.7 Sum: 31.0 58. 0 1.6 9.4 49.0
CC 90 0.0 0.0 0.0 0.0 0.0 0.0 0.0
Sum: 1.1 17.9 35. 4 40.0 5. 6 0.0 42. 2
IGBP LCCS2
NA ENF D NF/D BF Mixed W. Sav. Sa v. Sum: NA Dense fo. Ope n fo. Sum:
NA 8.6 0. 2 0.0 0.1 2. 9 2.7 14. 5 NA 8.3 0.3 6.0 14.5
ENF 0. 1 5.1 0. 0 3. 5 2.4 0. 0 11.2 Dens e fo. 0.0 9. 3 2. 7 12.0
DNF/DBF 0.0 0.0 0.0 0.1 0.0 0. 0 0.1 Ope n fo. 3. 7 18. 8 51.0 73.5
Mixed 0.0 0.1 0.0 0.5 0.2 0. 0 0.8 Sum: 12.0 28.4 59.7 60.3
W. Sa v. 1. 2 9.6 0. 0 7. 4 29.1 2.5 49.6
Sav. 3.5 1.0 0. 0 0. 7 11.6 7.0 23.8
Sum: 13.4 16.0 0.0 12.2 46.2 12. 2 41.7
LCCS1 LCCS3
NA ENF DNF/DBF Mixed Open fo. Sparse fo. Sum: NA De nse fo. Ope n fo. Sum:
NA 8.2 0. 3 0.0 0.1 3. 1 2.9 14. 5 NA 8.6 0.3 5.7 14.5
ENF 0. 0 5.1 0. 0 3. 5 2.4 0. 0 11.2 Dens e fo. 0.1 9. 3 2. 7 12.0
DNF/DBF 0.0 0.0 0.0 0.1 0.0 0. 0 0.1 Ope n fo. 4. 6 18. 6 50.2 73.5
Mixed 0.0 0.1 0.0 0.5 0.2 0. 0 0.8 Sum: 13.2 28.2 58.6 59.5
Open fo. 0.7 9.6 0.0 7.4 29.4 2.5 49.6
Spars e fo. 2.9 1. 1 0. 0 0.7 12.0 7.2 23.8
Sum: 11.8 16.2 0.0 12.2 47.2 12. 7 42.2
LAI PFT
NA EBF/ENF DBF/DNF Sav. Sum: NA ENT/EBT DNT/D BT Sum:
NA 8.3 0. 3 0.0 6.0 14.5 NA 9.0 5.4 0. 1 14.5
EBF/ENF 0.0 9.0 0.0 2.6 11.7 ENT/EBT 5.2 74.6 0.6 80.4
DBF/DNF 0.0 0.2 0.0 0.1 0.3 DNT /DBT 1.9 3. 0 0. 1 5.0
Sav. 3.7 18. 6 0. 2 51. 0 73.5 Sum: 16.2 83.0 0. 8 74.7
Sum: 12.0 28.1 0.2 59.7 60.1
LC-product
IFN-SM
Environ Monit Assess (2021) 193: 121 Page 11 of 19 121
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
Table 6 Intercomparison of the total tree cover and binned tree
covers (All = all tree cover data after application of the translation
legend, Low = tree cover bin of 030%, Moderate = tree cover bin of
3160%, and High = tree cover bin of 61100%). Abbreviations:
MBE = mean bias error, RMSE = root mean squared error, and r
2
=
coefficient of determination. Statistics were calculated using the orig-
inal LCproduct resolutions
Tree cover
product
MBE MBE MBE MBE RMSE RMSE RMSE RMSE R
2
Low Moderate High All Low Moderate High All All
VCF 9.0 6.5 15.5 3.4 18.4 14.4 18.6 16.7 0.4
CCI 28.5 8.3 10.4 9.6 32.7 14.4 13.0 20.4 0.1
IGBP 20.8 10.2 7.4 13.0 27.3 19.9 16.5 22.0 0.3
LAI 20.9 4.9 5.2 9.9 25.3 20.8 19.3 22.1 0.3
PFT 37.4 9.3 9.7 15.4 38.4 12.6 10.3 23.7 NA
LCCS1 21.0 10.1 7.4 13.1 27.5 19.9 16.5 22.2 0.3
LCCS2 20.9 4.9 5.2 9.9 25.3 20.8 19.3 22.1 0.3
LCCS3 21.0 4.9 5.2 10.0 25.4 20.7 19.3 22.1 0.3
Tabl e 7 Confusion matrix between different within pixel tree covers at LCproduct original resolution. NA = no tree cover, Low = pixel
tree cover bin of 030%, Moderate = pixel tree cover bin of 3160%, and High = pixel tree cover bin of 61100%. Accuracy is the sum of
correctly classified pixels (diagonal sum, excluding the first NArow)
121 Page 12 of 19 Environ Monit Assess (2021) 193: 121
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
other hand, suffer from misrepresentation of areas be-
longing to the lowest tree cover bin. In MS-NFI data,
areas with moderate tree cover were the most abundant
(i.e., 4656% cover, note that range is provided as
original LC product resolutions were used to calculate
the statistics), followed by areas belonging to the lowest
tree cover bin (i.e., 2532%) and to the highest tree
cover bin (i.e., 1422%) (Table 7). In terms of a classi-
fications ability to represent low, moderate, and high
tree covers, the most flexible classifications were VCF,
IGBP, LCCS1, and LCCS3 (Table 7).
Intercomparison of coniferous and deciduous tree cover
Coniferous and deciduous tree cover estimates derived
from the LC products using the transition legend
showed clear differences in error values between the
two species groups (Table 8). The deviations from the
reference MS-NFI values were, in general, higher for
coniferous species compared to deciduous species.
However, the highest RMSE among all classifications
was observed for the deciduous PFT classification
(Table 8). For coniferous tree covers, the smallest
RMSEs and biases were noted in the IGBP and LCCS1
classifications, which had the largest number of tree
cover classes. For deciduous tree covers, the smallest
RMSE was in the LAI classification which only used
three tree cover classes. For deciduous species, positive
biases were noted for the lowest tree cover bins, whereas
negative biases were observed for moderate or high tree
cover bins (except for PFT) (Table 8). For coniferous
species, positive biases were also observed for the low-
est tree cover bins but for the moderate or high tree
cover bins, the biases were either positive or negative
depending on the classification.
Intercomparison of coniferous and deciduous tree
covers using CMG grid
Intercomparison of the maps (Fig. 1) revealed clear
differences in spatial patterns of the coniferous and
deciduous tree covers. At CMG resolution, the MS-
NFI-based coniferous tree cover (Fig. 1a) varied be-
tween 20 and 40%, being less than 10% (which is the
threshold used by the FAO forest definition) in only
some parts of the country. However, the coniferous tree
cover maps based on the IGBP, LAI, and PFT classifi-
cations based (Fig. 1e, g, i) showed too high or low
coniferous tree cover values for large areas. The CCI
classification-based coniferous tree cover map (Fig. 1c)
appeared the most similar to the MS-NFI-based map
the tree cover values were fairly similar to those of the
MS-NFI map, and the CCI map does not show any
anomalies in spatial distributions of tree covers, which
are present in other maps. For deciduous tree cover, the
MS-NFI-based tree cover (Fig. 1b) was often less than
Tabl e 8 Intercomparison of the coniferous and deciduous tree
covers and binned tree covers with the reference tree cover esti-
mates from MS-NFI (All all tree cover data after application of the
translation legend, Low tree cover bin of 030%, Moderate tree
cover bin of 3160%, High tree cover bin of 61100%). Abbrevi-
ations: MBE mean bias error, RMSE root mean squared error, r
2
coefficient of determination. Statistics were calculated using the
original LC product resolutions
Tree cover product MBE MBE MBE MBE RMSE RMSE RMSE RMSE
Low Moderate High All Low Moderate High All
Coniferous
CCI 24.5 7.5 9.0 13.4 31.6 17.3 14.8 23.6
IGBP 6.5 2.2 3.4 1.7 18.4 23.5 22.6 21.4
LAI 9.4 3.5 10.0 6.2 23.5 29.6 22.9 27.0
PFT 37.6 12.0 7.2 22.5 38.6 14.3 7.4 27.2
LCCS1 6.6 2.2 3.3 1.8 18.5 23.5 22.7 21.4
Deciduous
CCI 10.1 8.9 29.4 9.1 16.1 17.5 34.4 16.1
IGBP 15.5 2.3 26.5 15.3 17.5 10.9 27.4 17.5
LAI 11.9 9.8 8.1 11.8 13.5 21.8 30.8 13.5
PFT 45.7 19.3 8.7 45.0 46.2 20.0 8.8 45.7
LCCS1 15.5 2.5 26.5 15.3 17.5 11.0 27.4 17.4
Environ Monit Assess (2021) 193: 121 Page 13 of 19 121
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
Fig. 1 Spatial distribution of coniferous and deciduous tree covers based on four different global LC products and reference data from the Finnish MS-NFI. All data was aggregated and
resampled to MODIS Climate Modeling Grid (CMG) resolution. The top row (a,c,e,g,i) shows tree cover values for coniferous and the lower row (b,d,f,h,j) for deciduous species. Note:
black color is used to denote areas below the 10% CC threshold employed by the international forest definition by FAO
121 Page 14 of 19 Environ Monit Assess (2021) 193: 121
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
10%, and, for some regions, varied between 10 and
20%. For the IGBP, LAI, and PFT classifications (Fig.
1f, h, j), some clear spatial patterns were observed: all
three LC classification-based maps showed either far
too high deciduous tree cover values for large areas
and/or no deciduous tree cover values. The CCI
classification-based map of deciduous tree cover (Fig.
1b) showed the smallest spatial variations in deciduous
tree cover values and was also the most similar to the
MS-NFI map of deciduous tree cover. These observed
differences in the spatial mapping of the tree cover
values demonstrate the need to assess also spatial vari-
ations present in the map data.
Discussion
Tod ays LC products are often created based on polar-
orbiting satellite sensor data. As the areas to be mapped
are large, this poses challenges in terms of defining land
cover classes to be separated from the satellite sensor
data (e.g., Ustin and Gamon 2010). However, as the
spatial resolution of the observation unit becomes coars-
er, the probability of mixed pixels (i.e., different LC
classes) increases, which further reduces the number of
(forest) classes that have separable spectral and structur-
al properties. An obvious solution would be to achieve a
better separation of different LC classes (i.e., forest
types) by using higher spatial resolution satellite sensor
data from Landsat 8 (L8) and Sentinel 2 (S2), for exam-
ple. Alternatively, satellite-borne (e.g., GEDI or ICESat-
2) or airborne laser scanning data (e.g., from national
land surveys) could be used to delineate areas with tree
cover and quantify their structural properties (such as
canopy cover and leaf area index, e.g., Korhonen et al.
2011; Majasalmi et al. 2017). However, as satellite or
airborne laser scanning data is not necessarily free nor
readily available for large areas, utilizing optical satellite
data remains currently the only operationally feasible
solution for tree cover mapping in regional or global
applications. Although the higher spatial resolution L8
and S2 data would solve some of the problems associ-
ated with low spatial resolution mixedpixels, ambi-
guity would still remain surrounding the forestor
tree coverdefinitions.
The international forest definition by FAO does not
allow direct quantification from space. The obvious
problems in applying the FAO forest definition in clas-
sifying tree cover from satellite data result from
Environ Monit Assess (2021) 193: 121 Page 15 of 19 121
expectations regarding forest life cycle and future de-
velopment trajectories. In addition, the strict application
of the FAO forest classification would mean excluding
northern tundra (i.e., dominated by stunted deciduous
trees) from forest area, although the growth of these
forest areas (e.g., Heiskanen 2008) may be expected to
increase due to climate change. Thus, there might be
room for redefining forestin a way that would allow
better mapping of forested areas using remote sensing
data, as well as applying the maps as an input for
different regional LSM modeling frameworks dealing
with forest management or land surface hydrology.
Such classification should ideally be based on thresh-
olds that can be retrieved from optical satellite data, and
other globally available auxiliary data. A systematic
forest definition that allows measuring and monitoring
from space is needed for the global quantification of
forests, and benchmarking of the national/regional
estimates.
In the boreal region, separation between coniferous and
deciduous species groups is essential, for example, for
better quantification of seasonal courses of surface fluxes
in LSMs. In this study, an assessment and intercomparison
of todays most used LC products was conducted, and a
new approach, called a translation legend, was developed
and applied. This was necessary as there is very limited
information available regarding the algorithms (and data)
which are used to produce these global LC products. Thus,
direct methodological comparisons are not possible. In
addition, since each LC product classification uses its
own class definitions, there are very few approaches avail-
able for validating or intercomparing the informational
content of the products.
Application of the translation legend allows system-
atic intercomparison between different categorical tree
cover (i.e., forest type) classifications and analysis of
spatial distributions of conifer and deciduous species.
While traditional LC product assessments based on
pixel counts belonging to different classes and incor-
rectly classified pixels are important, they do not allow
developing better LC classifications and class defini-
tions. Application of the translation legend allows
converting categorical classes into continuous tree cover
values while respectingthe original LC class definitions,
and intercomparison of either total, or conifer and de-
ciduous tree covers, in a variety of different ways, which
provides new insight for product developers and users.
However, as the transition legend is based on
converting categorical LC classes into continuous by
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
121 Page 16 of 19 Environ Monit Assess (2021) 193: 121
using bin means (i.e., the true LC distribution that was
used to prepare the LC product remains unknown),
using a translation legend does not replace the need for
traditional LC product validation studies. Rather, it is an
extension of conventional validation exercises, a sys-
tematic intercomparison tool. The transition legend al-
lows the calculation of statistics (such as R
2
,RMSE,
MBE) and reclassification of the continuous values for
further analysis. In this study, we reclassified the data
into three groups of low, moderate, and high tree cover
to observe which classifications are flexible enough to
represent these variations in the tree cover values. Al-
though the MBEs and RMSEs obtained using the trans-
lation legend are not the truth as such, they reflect the
classification skill to map different tree cover values.
Due to regional differences in conifer and deciduous tree
species distributions, we acknowledge the difficulty in
developing classifications that would suit all geograph-
ical areas.
It is noteworthy that the number of LC classes that
can be included in LC product analyses is higher when
using the translation legend approach than in traditional
assessment because partially covered tree cover classes,
such as the LCCS2 class forest/cropland mosaic,can
be included using the translation legend. This can be
accomplished by assigning a partially forested pixels
respective tree cover fraction (e.g., 25%) adapted from
the original legend class definition, and by assuming the
share of conifer and deciduous tree species groups is
equal. The assumption of equal shares of conifers and
deciduous groups of these mixed classes is justified by
our study region, as the surroundings of the cultivated
areas are often outside the most active forest manage-
ment operations (i.e., deciduous trees are not harvested
as often). In addition, locations close to cultivated areas
have a good supply of light, water, and nutrients, all of
which benefit the growth of deciduous tree species.
Although the area belonging to these partially forested
pixels is small in forested landscapes such as in Finland
(i.e., forests are fragmented by differently structured
forests rather than different LC classes), in other geo-
graphical areas fragmented by agriculture and forestry,
the fraction of pixels belonging to these mixed classes
may be significant. Thus, mixed pixels should be
accounted for in tree cover assessments.
As the VCF is continuous by default and the finest
spatial resolution tree cover product, it managed to
describe variations in tree cover values well. However,
it does not separate between conifer and deciduous
species, which is an obvious limitation. The CCI prod-
uct was found to perform well in representing areas with
low, moderate, and high tree covers and in separating
coniferous and deciduous areas, especially at the final
stage in which data was aggregated to the CMG grid
resolution. Notably, it is the only classification that did
not employ a tree height threshold. Thus, the necessity
of the height threshold in forest LC classification may be
questioned; we acknowledge that while there is a clear
need for a height threshold in forest field inventories, it
is perhaps not necessary in global LC classifications,
especially as tree height is challenging to retrieve from
optical satellite data.
As the LC product maps are independent of the class
definitions (i.e., the same classification can be used by
several LC products with differences in their spatial
class distributions), both must be assessed to evaluate
classification performance. For example, in LSMs (or
more generally in climate modeling), the climate data
(e.g., maps of temperature, and precipitation) are pro-
vided as maps and thus, the spatial patterns of conifer-
ous and deciduous tree covers must also be sufficiently
mapped to predict vegetation fluxes correctly. Especial-
ly in a boreal region, the separation of coniferous and
deciduous tree cover areas is necessary due to large
effects of tree phenology and snow on surface albedo
(Bright et al. 2018).
Direct evaluation of the impacts of varying tree
cover descriptions in a climate model requires cli-
mate simulations due to the simultaneous usage of
different LC products. For example, the community
land surface model (CLM) uses MODIS-based
monthly mean leaf area index data and IGBP-based
fractional covers in its simulations (Zeng et al. 2002).
More precisely, while CLM employs IGBP classifi-
cation to derive vegetation fractional cover values, it
uses six alternative biomes (grasses and cereal crops,
shrubs, broadleaf crops, savanna, broadleaf forests,
and needle forests) defined based on vegetation struc-
ture to derive the monthly mean leaf area index
values (Myneni et al. 1997).Asaresult,thevegeta-
tion cover fraction and LAI are not constant for each
pixel (Zeng et al. 2002) and thus, the impact of
fractional cover (i.e., LC product) on climate simula-
tion outcomes cannot be directly evaluated. More
attentionshouldbepaidtoLCdata(andtherefore
also the underlying forest class definitions) that are
being used to parametrize the LSMs, as the expected
improvements in predicted surface fluxes will rely on
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
Environ Monit Assess (2021) 193: 121 Page 17 of 19 121
the quality of the employed tree cover mapping (e.g.,
Bright et al. 2018; Majasalmi et al. 2018; Majasalmi
et al. 2020).
Conclusions
We used Finnish MS-NFI data to assess tree cover
representations of eight annual global LC classifications
for the year 2017 and developed and applied a transla-
tion legend approach for better intercomparison of their
tree cover representations. The benefits of the devel-
oped approach are that it is transparent to apply, can be
adapted to any classification and across any spatial
scale, and allows the calculation of different statistical
metrics. We observed large differences in classification
skills of representing variations in tree cover values,
and in their spatial mapping of conifer and deciduous
tree covers. Based on our analyses, the tree cover was
either overestimated or underestimated depending on
the LC product, and classification accuracy varied be-
tween 42 and 75%. Intercomparison of the LC products
revealed clear differences in spatial distributions of
conifer and deciduous species. In general, the CCI LC
product had the most realistic spatial mapping of conif-
erous and deciduous tree covers compared to the refer-
ence MS-NFI data. As the differences in tree cover
mapping may be expected to translate into differences
in predicted surface fluxes, users and developers of the
LSMs relying on prescribed land cover information are
encouraged to pay attention to what type of LC product
and classification their analysis is based on. Ideally, the
next generation of LC products will be based on a forest
definition that facilitates measuring and monitoring
from space, and classification that accurately represents
coniferous and deciduous species tree covers.
Author contribution T. Majasalmi: conceptualization, method-
ology, formal analysis, investigation, writingoriginal draft,
writingreview & editing, visualization. M. Rautiainen:
writingreview & editing
Funding Open Access funding provided by Aalto University.
The MS-NFI data were provided by ©Natural Resources Institute
Finland, 2019 as The Multi-source National Forest Inventory
Raster Maps of 2017. T. Majasalmi has received funding from
Aalto ENG postdoctoral funds. M. Rautiainen has received
funding from the European Research Council (ERC) under the
European Unions Horizon 2020 research and innovation program
(grant agreement No 771049).
Data availability All data is publicly available (see Table 1,and
reference list).
Material availability All data is publicly available (see Table 1,
and reference list).
Code availability Code is available from the author.
Declarations
Conflict of interest The authors declare no competing interests.
Disclaimer The text reflects only the authorsview and the
Agency is not responsible for any use that may be made of the
information it contains.
Open Access This article is licensed under a Creative Commons
Attribution 4.0 International License, which permits use, sharing,
adaptation, distribution and reproduction in any medium or format,
as long as you give appropriate credit to the original author(s) and
the source, provide a link to the Creative Commons licence, and
indicate if changes were made. The images or other third party
material in this article are included in the article's Creative Com-
mons licence, unless indicated otherwise in a credit line to the
material. If material is not included in the article's Creative Com-
mons licence and your intended use is not permitted by statutory
regulation or exceeds the permitted use, you will need to obtain
permission directly from the copyright holder. To view a copy of
this licence, visit http://creativecommons.org/licenses/by/4.0/.
References
Bartholome, E., & Belward, A. S. (2005). GLC2000: a new
approach to global land cover mapping from Earth observa-
tion data. International Journal of Remote Sensing, 26(9),
19591977.
Bastin, J. F., Finegold, Y., Garcia, C., Mollicone, D., Rezende, M.,
Routh, D., et al. (2019). The global tree restoration potential.
Science, 365(6448), 7679.
Bonan, G. B., Levis, S., Kergoat, L., & Oleson, K. W. (2002).
Landscapes as patches of plant functional types: an integrat-
ing concept for climate and ecosystem models. Global
Biogeochemical Cycles, 16(2), 51.
Bright, R. M., Eisner, S., Lund, M. T., Majasalmi, T., Myhre, G.,
& Astrup, R. (2018). Inferring surface albedo prediction error
linked to forest structure at high latitudes. Journal of
Geophysical Research: Atmospheres, 123(10), 49104925.
CCI LCCS, (2020). https://maps.elie.ucl.ac.be/CCI/viewer/downl
oad/CCI-LC_Maps_Legend.pdf. Accessed 10 Jan 2020.
Chen, J. M., Rich, P. M., Gower, S. T., Norman, J. M., &
Plummer, S. (1997). Leaf area index of boreal forests: theory,
techniques, and measurements. Journal of Geophysical
Research: Atmospheres, 102(D24), 2942929443.
Fisher, R. A., & Koven, C. D.(2020). Perspectives on the future of
Land Surface Models and the challenges of representing
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
121 Page 18 of 19 Environ Monit Assess (2021) 193: 121
complex terrestrial systems. Journal of Advances in
Modeling Earth Systems, 12(4), e2018MS001453.
FRA, (2015). Forest Resource Assessment working paper 180,
http://www.fao.org/3/ap862e/ap862e00.pdf.Accessed10
Feb 2020.
FRA FIN, (2015). Global forest Resources Assessment 2015.
Country report Finland. Available: http://www.fao.org/3/a-
az213e.pdf. Accessed 10 Feb 2020.
Friedl, M. A., McIver, D. K., Hodges, J. C., Zhang, X. Y.,
Muchoney, D., Strahler, A. H., et al. (2002). Global land
cover mapping from MODIS: algorithms and early results.
Remote sensing of Environment, 83(1-2), 287302.
Friedl, M. A., Sulla-Menashe, D., Tan, B., Schneider, A.,
Ramankutty, N., Sibley, A., & Huang, X. (2010). MODIS
Collection 5 global land cover: algorithm refinements and
characterization of new datasets. Remote Sensing of
Environment, 114(1), 168182.
Gower, S. T., Kucharik, C. J., & Norman, J. M. (1999). Direct and
indirect estimation of leaf area index, fAPAR, and net pri-
mary production of terrestrial ecosystems. Remote sensing of
environment, 70(1), 2951.
Gschwantner, T., Schadauer, K., Vidal, C., Lanz, A., Tomppo, E.,
Di Cosmo, L., et al. (2009). Common tree definitions for
national forest inventories in Europe. Silva Fennica, 43(2),
303321.
Hansen, M. C., DeFries, R. S., Townshend, J. R., & Sohlberg, R.
(2000). Global land cover classification at 1 km spatial res-
olution using a classification tree approach. International
journalofremotesensing,21(6-7), 13311364.
Hansen, M. C., DeFries, R. S., Townshend, J. R. G., Carroll, M.,
Dimiceli, C., & Sohlberg, R. A. (2003). Global percent tree
cover at a spatial resolution of 500 meters: first results of the
MODIS vegetation continuous fields algorithm. Earth
Interactions, 7(10), 115.
Hansen,M.C.,Potapov,P.V.,Moore,R.,Hancher,M.,
Turubanova, S. A. A., Tyukavina, A., et al. (2013). High-
resolution global maps of 21st-century forest cover change.
Science, 342(6160), 850853.
Heiskanen, J. (2008). Evaluation of global land cover data sets
over the tundrataiga transition zone in northernmost
Finland. International Journal of Remote Sensing, 29(13),
37273751.
Horn,H.S.(1971).Adaptive geometry of trees (MPB-3).
Princeton University Press. isbn:0-691-08089-5.
Jennings, S. B., Brown, N. D., & Sheil, D. (1999). Assessing forest
canopies and understorey illumination: canopy closure, can-
opy cover and other measures. Forestry: An International
Journal of Forest Research, 72(1), 5974.
Korhonen, L., Korpela, I., Heiskanen, J., & Maltamo, M.
(2011). Airborne discrete-return LIDAR data in the
estimation of vertical canopy cover, angular canopy
closure and leaf area index. Remote Sensing of
Environment, 115(4), 10651080.
Loveland, T. R., & Belward, A. S. (1997). The international
geosphere biosphere programme data and information sys-
tem global land cover data set (DISCover). Acta
Astronautica, 41(4-10), 681689.
Loveland, T. R., Reed, B. C., Brown, J. F., Ohlen, D. O., Zhu, Z.,
Yang,L.W.M.J.,&Merchant,J.W.(2000).Development
of a global land cover characteristics database and IGBP
DISCover from 1 km AVHRR data. International Journal
of Remote Sensing, 21(6-7), 13031330.
LUKE, (2020). https://www.luke.fi/tietoa-luonnonvaroista
/metsa/metsien-monimuotoisuus/operaatio-mustikka/.
Accessed 8 Dec 2020.
Majasalmi, T., Korhonen, L., Korpela, I., & Vauhkonen, J. (2017).
Application of 3D triangulations of airborne laser scanning
data to estimate boreal forest leaf area index. International
journal of applied earth observation and geoinformation, 59,
5362.
Majasalmi, T., Eisner, S., Astrup, R. A., Fridman, J., & Bright, R.
M. (2018). An enhanced forest classification scheme for
modeling vegetationclimate interactions based on national
forest inventory data. Biogeosciences,15 (2), 399-412.
Majasalmi, T., Allen, M., Antón-Fernández, C., Astrup, R., &
Bright, R. M. (2020). A simple grid-based framework for
simulating forest structural trajectories linked to transient
forest management scenarios in Fennoscandia. Climatic
Change, 162,21392155.
Mäkisara, K., Katila, M. & Peräsaari, J. (2019). The Multi-Source
National Forest Inventory of Finland methods and results
2015. Natural resources and bioeconomy studies 8/2019,
Natural Resources Institute Finland. 57 p. http://urn.
fi/URN:ISBN:978-952-326-711-4 http://jukuri.luke.
fi/handle/10024/543826. Accessed 15 Oct 2019.
Masson, V., Champeaux, J. L., Chauvin, F., Meriguet, C., &
Lacaze, R. (2003). A global database of land surface param-
eters at 1-km resolution in meteorological and climate
models. Journal of climate, 16(9), 12611282.
MODIS CMG, (2020). https://lpdaac.usgs.gov/products/mcd12c1
v006/. Accessed 10 Feb 2020.
MODIS LCCS, (2020). https://lpdaac.usgs.gov/products/mcd12
q1v006/. Accessed 23 Jan 2020.
MODIS MCD12, (2020). User guide to collection 6 MODIS Land
Cover (MCD12Q1 and MCD12C1) Product. Sulla-Menashe
D,, & Friedl M.A. https://modis.ornl.
gov/documentation/guides/MCD12_User_Guide_V6.pdf.
Accessed 23 Jan 2020.
MODIS VCF, (2017). The MOD44B Version 6 Vegetation
Continuous Fields (VCF). https://lpdaac.usgs.
gov/products/mod44bv006/. Accessed 10 Feb 2020.
MODIS VCF, (2020). https://modis-land.gsfc.nasa.gov/vcc.html.
Accessed 23 Jan 2020.
MODIS VCF User guide, (2020). User Guide for the MODIS
vegetation continuous fields product Collection 6, version 1.
https://lpdaac.usgs.gov/documents/112/MOD44B_User_
Guide_V6.pdf. Accessed 8 Dec 2020.
Moon, M., Zhang, X., Henebry, G. M., Liu, L., Gray, J. M.,
Melaas, E. K., & Friedl, M. A. (2019). Long-term continuity
in land surface phenology measurements: a comparative
assessment of the MODIS land cover dynamics and VIIRS
land surface phenology products. Remote sensing of environ-
ment, 226,7492.
MS-NFI, (2019). Multi-source national forest inventory (MS-NFI)
raster maps of 2017. Readme-file.
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
Myneni, R. B., Ramakrishna, R., Nemani, R., & Running, S. W.
(1997). Estimation of global leaf area index and absorbed
PAR using radiative transfer models. IEEE Transactions on
Geoscience and remote sensing, 35(6), 13801393.
Myneni, R. B., Hoffman, S., Knyazikhin, Y., Privette, J. L.,
Glassy,J.,Tian,Y.,etal.(2002).Globalproductsof
vegetation leaf area and fraction absorbed PAR from
year one of MODIS data. Remote sensing of environ-
ment, 83(1-2), 214231.
Poulter, B., MacBean, N., Hartley, A., Khlystova, I., Arino, O.,
Betts, R., etal. (2015). Plant functional type classification for
earth system models: results from the European Space
Agencys Land Cover Climate Change Initiative.
Geoscientific Model Development, 8,23152328.
Rautiainen, M., Stenberg, P., & Nilson, T. (2005). Estimating
canopy cover in Scots pine stands. Silva Fennica, 39(1),
137142.
Sexton, J. O., Song, X. P., Feng, M., Noojipady, P., Anand, A.,
Huang, C., et al. (2013). Global, 30-m resolution continuous
fields of tree cover: Landsat-based rescaling of MODIS veg-
etation continuous fields with lidar-based estimates of error.
International Journal of Digital Earth, 6(5), 427448.
Sexton, J. O., Noojipady, P., Song, X. P., Feng, M., Song, D. X.,
Kim, D. H., et al. (2016). Conservation policy and the mea-
surement of forests. Nature Climate Change, 6(2), 192.
Song, X. P., Huang, C., Feng, M., Sexton, J. O., Channan, S., &
Townshend, J. R. (2014). Integrating global land cover prod-
ucts for improved forest cover characterization: an applica-
tion in North America. International Journal of Digital
Earth, 7(9), 709724.
Tang, H., Song, X. P., Zhao, F. A., Strahler, A. H., Schaaf, C. L.,
Goetz, S., et al. (2019). Definition and measurement of tree
cover: a comparative analysis of field-, lidar-and landsat-
based tree cover estimations in the Sierra national forests,
USA. Agricultural and forest meteorology, 268,258268.
Tomppo, E., Gschwantner, T., Lawrence, M., McRoberts, R. E.,
Gabler, K., Schadauer, K., et al. (2010). National forest
inventories. Pathways for Common Reporting. European
Science Foundation, 1,541553.
Ustin, S. L., & Gamon, J. A. (2010). Remote sensing of plant
functional types. New Phytologist, 186(4), 795816.
Zeng, X., Shaikh, M., Dai, Y., Dickinson, R. E., & Myneni, R.
(2002). Coupling of the Common Land Model to the NCAR
Community Climate Model. Journal of Climate, 15, 1832
1854.
PublishersnoteSpringer Nature remains neutral with regard to
jurisdictional claims in published maps and institutional
affiliations.
Environ Monit Assess (2021) 193: 121 Page 19 of 19 121
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
1.
2.
3.
4.
5.
6.
Terms and Conditions
Springer Nature journal content, brought to you courtesy of Springer Nature Customer Service Center GmbH (“Springer Nature”).
Springer Nature supports a reasonable amount of sharing of research papers by authors, subscribers and authorised users (“Users”),
for small-scale personal, non-commercial use provided that all copyright, trade and service marks and other proprietary notices are
maintained. By accessing, sharing, receiving or otherwise using the Springer Nature journal content you agree to these terms of use
(“Terms”). For these purposes, Springer Nature considers academic use (by researchers and students) to be non-commercial.
These Terms are supplementary and will apply in addition to any applicable website terms and conditions, a relevant site licence or
a personal subscription. These Terms will prevail over any conflict or ambiguity with regards to the relevant terms, a site licence or
a personal subscription (to the extent of the conflict or ambiguity only). For Creative Commons-licensed articles, the terms of the
Creative Commons license used will apply.
We collect and use personal data to provide access to the Springer Nature journal content. We may also use these personal data
internally within ResearchGate and Springer Nature and as agreed share it, in an anonymised way, for purposes of tracking,
analysis and reporting. We will not otherwise disclose your personal data outside the ResearchGate or the Springer Nature group of
companies unless we have your permission as detailed in the Privacy Policy.
While Users may use the Springer Nature journal content for small scale, personal non-commercial use, it is important to note that
Users may not:
use such content for the purpose of providing other users with access on a regular or large scale basis or as a means to
circumvent access control;
use such content where to do so would be considered a criminal or statutory offence in any jurisdiction, or gives rise to civil
liability, or is otherwise unlawful;
falsely or misleadingly imply or suggest endorsement, approval , sponsorship, or association unless explicitly agreed to by
Springer Nature in writing;
use bots or other automated methods to access the content or redirect messages
override any security feature or exclusionary protocol; or
share the content in order to create substitute for Springer Nature products or services or a systematic database of Springer
Nature journal content.
In line with the restriction against commercial use, Springer Nature does not permit the creation of a product or service that creates
revenue, royalties, rent or income from our content or its inclusion as part of a paid for service or for other commercial gain.
Springer Nature journal content cannot be used for inter-library loans and librarians may not upload Springer Nature journal
content on a large scale into their, or any other, institutional repository.
These terms of use are reviewed regularly and may be amended at any time. Springer Nature is not obligated to publish any
information or content on this website and may remove it or features or functionality at our sole discretion, at any time with or
without notice. Springer Nature may revoke this licence to you at any time and remove access to any copies of the Springer Nature
journal content which have been saved.
To the fullest extent permitted by law, Springer Nature makes no warranties, representations or guarantees to Users, either express
or implied with respect to the Springer nature journal content and all parties disclaim and waive any implied warranties or
warranties imposed by law, including merchantability or fitness for any particular purpose.
Please note that these rights do not automatically extend to content, data or other material published by Springer Nature that may be
licensed from third parties.
If you would like to use or distribute our Springer Nature journal content to a wider audience or on a regular basis or in any other
manner not expressly permitted by these Terms, please contact Springer Nature at
onlineservice@springernature.com
... An important example is automated detection of trees and green spaces that are significant contributors to ecosystem services such as air purification and carbon sequestration. Recent studies include [1] and [2] for global monitoring of environment and forest cover using Sentinel-2 imagery. A Copernicus Sentinel-2B satellite, launched in 2017 provides 13 bands with spatial resolution from 10 m to 60 m. ...
... The high spatial and temporal resolution of data from this satellite is specifically designed for vegetation monitoring. For tree cover estimation, a broad range of methodologies have been presented in the literature, e.g., [1,3,4,5]. The authors in [3] proposed a data fusion method of different spatial resolution satellite imagery for forest-type mapping. ...
... change using GIS and remote sensing techniques in Ethiopia valley is presented in [5]. In [1], a simple tree cover (referred to as 'forest cover') approach using three different land cover (LC) products is employed in Finland. Clearly, most of these approaches focus on forest mapping -a gap exists in urban tree cover estimation in developing countries with low resolution imagery. ...
Preprint
Full-text available
This paper proposes a multi-spectral random forest classifier with suitable feature selection and masking for tree cover estimation in urban areas. The key feature of the proposed classifier is filtering out the built-up region using spectral indices followed by random forest classification on the remaining mask with carefully selected features. Using Sentinel-2 satellite imagery, we evaluate the performance of the proposed technique on a specified area (approximately 82 acres) of Lahore University of Management Sciences (LUMS) and demonstrate that our method outperforms a conventional random forest classifier as well as state-of-the-art methods such as European Space Agency (ESA) WorldCover 10m 2020 product as well as a DeepLabv3 deep learning architecture.
... Such vertical errors are significant compared to the amplitude of most river flood waves, which typically range from 1-2 m up to ∼ 12 m for the Amazon River at Manaus in Brazil (Trigg et al., 2009;Bates et al., 2013). Whilst global InSAR DEM errors can be reduced by intelligent processing (O'Loughlin et al., 2016;Yamazaki et al., 2017;Archer et al., 2018;Liu et al., 2021;Hawker et al., 2022) and by aggregating to coarser grid resolutions to mitigate random errors, they remain distinctly sub-optimal for a lot of flood inundation modelling (Schumann and Bates, 2018). Instead, inundation modelling is best conducted with DEMs generated using airborne lidars for most applications. ...
... The model solves the local inertial form of the shallow water equations in two dimensions across the model domain. For pluvial flood modelling, the model takes the terrain elevation and rainfall data as inputs and uses a raster-on-grid approach to calculate the velocity, water depth, and inundation (Bates et al., 2021). The input DEMs were aggregated to 10 m by averaging before being used in the flood simulation. ...
Article
Full-text available
In urban areas, topography data without above-ground objects are typically preferred in wide-area flood simulation but are not yet available for many locations globally. High-resolution satellite photogrammetric DEMs, like ArcticDEM, are now emerging and could prove extremely useful for global urban flood modelling; however, approaches to generate bare-earth DEMs from them have not yet been fully investigated. In this paper, we test the use of two morphological filters (simple morphological filter – SMRF – and progressive morphological filter – PMF) to remove surface artefacts from ArcticDEM using the city of Helsinki (192 km2) as a case study. The optimal filter is selected and used to generate a bare-earth version of ArcticDEM. Using a lidar digital terrain model (DTM) as a benchmark, the elevation error and flooding simulation performance for a pluvial scenario were then evaluated at 2 and 10 m spatial resolution, respectively. The SMRF was found to be more effective at removing artefacts than PMF over a broad parameter range. For the optimal ArcticDEM-SMRF the elevation RMSE was reduced by up to 70 % over the uncorrected DEM, achieving a final value of 1.02 m. The simulated water depth error was reduced to 0.3 m, which is comparable to typical model errors using lidar DTM data. This paper indicates that the SMRF can be directly applied to generate a bare-earth version of ArcticDEM in urban environments, although caution should be exercised for areas with densely packed buildings or vegetation. The results imply that where lidar DTMs do not exist, widely available high-resolution satellite photogrammetric DEMs could be used instead.
... Annual VCF data was acquired from MOD44B at 250 m resolution (Majasalmi and Rautiainen 2021). The data represents the sub-pixel-level surface vegetation cover and provides a gradient of three ground cover components: percent tree cover, percent nontree cover, and percent non-vegetated. ...
Article
Full-text available
Context In the past decades, several ecological engineering (eco-engineering) programs have been conducted in China, leading to a significant increase in regional carbon sink. However, the contribution of different eco-engineering programs to carbon uptake is still not clear, as the location of different programs is difficult to identify, and their impacts are concurrent with climate change. Objectives We aim to detect the location of eco-engineering programs and attribute the impacts of eco-engineering and climate change on vegetation dynamics and carbon uptake in Northeastern China during 2000–2020. Methods We developed a new framework to detect the location of eco-engineering programs by combining a temporal pattern analysis method and Markov model, and to attribute the impacts of eco-engineering and climate change on vegetation greenness and carbon uptake by combining a neighbor contrast method within a sliding window and trend analysis on the normalized difference vegetation index (NDVI) and gross primary production (GPP). Results We identified four main forestry eco-engineering programs: croplands to forest (CtoF), grasslands to forest (GtoF), savannas to forest (StoF), and natural forest conservation (NFC) programs, whose areas accounted for 2.11%, 1.89%, 3.41%, and 1.72% of the total study area, respectively. Both eco-engineering and climate change contributed to the increase in greenness and carbon uptake. Compared to climate change effect, eco-engineering increased NDVI and GPP by 121% and 21.43% on average, respectively. Specifically, the eco-engineering-induced increases in GPP were 54.1%, 9.46%, 8.13%, and 24.20% for CtoF, GtoF, StoF, and NFC, respectively. Conclusions These findings highlight the important and direct contribution of eco-engineering on vegetation greening with positive effects on carbon sequestration at a fine scale, providing an important implication for eco-engineering planning and management towards a carbon-neutral future.
... Satellite Earths observations provide data to develop land use land cover, specifically TC products for monitoring TC changes. Previous studies have used the MODIS Vegetation Continuous Field (VCF), Landsat, Sentinel 1-2, and Hansen global forest cover and change data to quantify TC from local to global scales (Majasalmi & Rautiainen, 2021;Pickering et al., 2021;Potapov et al., 2008;W. Zhang et al., 2019;Wimberly et al., 2022). ...
Article
Trees provide vital ecosystem services, yet anthropogenic activities have continued to reduce tree cover (TC) globally, leading to disparities in the distribution of ecosystem services at various scales. Similarly, empirical analysis of indicators that measure variations in TC at various temporal and spatial scales is limited in the semi-arid regions where global environmental challenges such as climate change are already affecting TC distribution. This paper quantified recent (2017 - 2019) TC changes in the semi-arid region of Ghana using Sentinel-2 times series data and the application of landscape metrics. The results showed that TC declined by 30% within two years. Urbanizing areas showed more rapid TC depletion rates, while some rural areas experienced a slight increase in TC. The findings suggest the need for local environmental governance by consolidating actions towards conserving semi-arid TC. Sustainable charcoal and fuelwood production are imperative for a long-term sustainable supply of ecosystem services.
... There is at the moment no reason to believe that the proposed methods should be suited for coniferous canopies, and the development of RTM satisfactory for needle and canopy modeling would help significantly. Coniferous forests constitute a major portion of the global land cover (Friedl et al., 2010), and in Europe, 16% of forests consist in coniferous open canopies , while conifers dominate boreal ecosystems (Caldwell et al., 2001;Majasalmi and Rautiainen, 2021). Identifying a retrieval method appropriate for conifers is therefore critical for the monitoring of vegetation traits at the global scale. ...
Thesis
Due to increasing climatic and anthropic pressures, the Mediterranean ecoregions, that are recognized plant biodiversity hotspots, are some of the most endangered ecosystems. Airborne and satellite remote sensing methods, which can provide information about large landscapes in a regular fashion, are most adapted for a future global monitoring. However, difficulties arise when retrieving information from open forests, largely distributed in the Mediterranean-climate regions, as the contribution of tree crowns to the measured radiative signal is limited. This thesis aims at developing method for the estimation of vegetation traits of open canopies, when field knowledge is insufficient to calibrate regression models. Initially, 18 m GSD airborne hyperspectral images were considered. First, using the DART model, a simplified modeling, with a flat lambertian ground and ellipsoidal tree crowns, was identified as suitable for physically-based estimations of LAI and leaf pigment contents. Then, exploratory works were undertaken to identify a method to estimate EWT and LMA with acceptable accuracy, first by considering refinements in the sampling schemes, then by assessing the effects of the 3D modeling on trees' radiative behavior. Finally, the findings were used to estimate all multiple vegetation traits (gap fraction, leaf chlorophyll and carotenoid contents, EWT, LMA) from synthetic hyperspectral satellite images at 8 an 30 m GSD using a hybrid method. This thesis had demonstrated that physically-based and hybrid methods were adequate for the estimation of multiple canopy and leaf traits from hyperspectral satellite images of open canopies in an operational context, using little if any a priori knowledge. To consolidate the results, efforts are required to test the methods over more various sites that would present a higher diversity in terms of traits, trait variation ranges, or species. Moreover, identifying methods that would work for periods during which the understory is photosynthetically active would be necessary to allow for a global monitoring over the complete annual phenological cycle.
Article
Climate change increases fire-favorable weather in forests, but fire trends are also affected by multiple other controlling factors that are difficult to untangle. We use machine learning to systematically group forest ecoregions into 12 global forest pyromes, with each showing distinct sensitivities to climatic, human, and vegetation controls. This delineation revealed that rapidly increasing forest fire emissions in extratropical pyromes, linked to climate change, offset declining emissions in tropical pyromes during 2001 to 2023. Annual emissions tripled in one extratropical pyrome due to increases in fire-favorable weather, compounded by increased forest cover and productivity. This contributed to a 60% increase in forest fire carbon emissions from forest ecoregions globally. Our results highlight the increasing vulnerability of forests and their carbon stocks to fire disturbance under climate change.
Preprint
Full-text available
Climate change is increasing the frequency and intensity of wildfires globally, with significant impacts on society and the environment. However, our understanding of the global distribution of extreme fires remains skewed, primarily influenced by media coverage and regional research concentration. This inaugural State of Wildfires report systematically analyses fire activity worldwide, identifying extreme events from the March 2023–February 2024 fire season. We assess the causes, predictability, and attribution of these events to climate change and land use, and forecast future risks under different climate scenarios. During the 2023–24 fire season, 3.9 million km2 burned globally, slightly below the average of previous seasons, but fire carbon (C) emissions were 16 % above average, totaling 2.4 Pg C. This was driven by record emissions in Canadian boreal forests (over 9 times the average) and dampened by reduced activity in African savannahs. Notable events included record-breaking wildfire extent and emissions in Canada, the largest recorded wildfire in the European Union (Greece), drought-driven fires in western Amazonia and northern parts of South America, and deadly fires in Hawai’i (100 deaths) and Chile (131 deaths). Over 232,000 people were evacuated in Canada alone, highlighting the severity of human impact. Our analyses revealed that multiple drivers were needed to cause areas of extreme fire activity. In Canada and Greece a combination of high fire weather and an abundance of dry fuels increased the probability of fires by 4.5-fold and 1.9–4.1-fold, respectively, whereas fuel load and direct human suppression often modulated areas with anomalous burned area. The fire season in Canada was predictable three months in advance based on the fire weather index, whereas events in Greece and Amazonia had shorter predictability horizons. Formal attribution analyses indicated that the probability of extreme events has increased significantly due to anthropogenic climate change, with a 2.9–3.6-fold increase in likelihood of high fire weather in Canada and a 20.0–28.5-fold increase in Amazonia. By the end of the century, events of similar magnitude are projected to occur 2.22–9.58 times more frequently in Canada under high emission scenarios. Without mitigation, regions like Western Amazonia could see up to a 2.9-fold increase in extreme fire events. For the 2024–25 fire season, seasonal forecasts highlight moderate positive anomalies in fire weather for parts of western Canada and South America, but no clear signal for extreme anomalies is present in the forecast. This report represents our first annual effort to catalogue extreme wildfire events, explain their occurrence, and predict future risks. By consolidating state-of-the-art wildfire science and delivering key insights relevant to policymakers, disaster management services, firefighting agencies, and land managers, we aim to enhance society’s resilience to wildfires and promote advances in preparedness, mitigation, and adaptation.
Article
Full-text available
Land degradation is a global environmental issue, affecting human sustainable development. The trend of land cover change (LCC) is one of the core sub-indicators of land degradation assessment in Sustainable Development Goal 15.3.1, which is closely related to the quality of land resources and regional ecological environment. Sustainable management of land resources is the basis and premise of sustainable development of human society and economy. In order to more comprehensively assess the land degradation situation from LCC, optimize the land cover (LC) structure and land resource utilization. This paper combines land cover qualitative change (LQaC) (LC type change) and land cover quantitative change (LQtC) (land vegetation cover change) to comprehensively assess land degradation, and applies it to the Pakistan part of the China-Pakistan Economic Corridor. The results showed that land degradation area in Pakistan from 2001 to 2020 was 1.99% (17,550.07km²), of which 1.90% (16,756.35km²) was caused by LQaC, and 0.09% (793.72km²) was caused by LQtC. Land improvement was 10.75% (94,805.65 km²). Land degradation is mainly caused by LQaC, and it mainly in the Indus River Valley and the central and eastern plains, which is significantly affected by human activities. At the same time, we found that in Pakistan: the continuous agricultural expansion (grassland to cropland) and vegetation establishment (unused land to grassland) are accompanied by serious withdrawal of agriculture (cropland to grassland) and vegetation loss (grassland to unused land). The agricultural expansion and the intensification of human activities lead to land degradation. Our assessment suggests that the government should focus on formulating effective land use policies, while protecting the ecological environment and vegetation: improving the productivity of cropland, optimizing the use of cropland resources, and reasonably reclaiming cropland to realize “Land Degradation Neutrality” and sustainable development of land resources in Pakistan.