Access to this full-text is provided by MDPI.
Content available from Remote Sensing
This content is subject to copyright.
Citation: Svoboda, J.; Štych, P.;
Laštoviˇcka, J.; Paluba, D.; Kobliuk, N.
Random Forest Classification of Land
Use, Land-Use Change and Forestry
(LULUCF) Using Sentinel-2 Data—A
Case Study of Czechia. Remote Sens.
2022,14, 1189. https://doi.org/
10.3390/rs14051189
Academic Editor: Michael Sprintsin
Received: 30 December 2021
Accepted: 17 February 2022
Published: 28 February 2022
Publisher’s Note: MDPI stays neutral
with regard to jurisdictional claims in
published maps and institutional affil-
iations.
Copyright: © 2022 by the authors.
Licensee MDPI, Basel, Switzerland.
This article is an open access article
distributed under the terms and
conditions of the Creative Commons
Attribution (CC BY) license (https://
creativecommons.org/licenses/by/
4.0/).
remote sensing
Article
Random Forest Classification of Land Use, Land-Use Change
and Forestry (LULUCF) Using Sentinel-2 Data—A Case Study
of Czechia
Jan Svoboda, Pˇremysl Štych * , Josef Laštoviˇcka , Daniel Paluba and Natalia Kobliuk
EO4Landscape Research Team, Department of Applied Geoinformatics and Cartography, Faculty of Science,
Charles University, 12843 Prague, Czechia; svoboj25@natur.cuni.cz (J.S.); josef.lastovicka@natur.cuni.cz (J.L.);
daniel.paluba@natur.cuni.cz (D.P.); natalia.kobliuk@natur.cuni.cz (N.K.)
*Correspondence: stych@natur.cuni.cz
Abstract:
Land use, land-use change and forestry (LULUCF) is a greenhouse gas inventory sector that
evaluates greenhouse gas changes in the atmosphere from land use and land-use change. This study
focuses on the development of a Sentinel-2 data classification according to the LULUCF requirements
on the cloud-based platform Google Earth Engine (GEE). The methods are tested in selected larger
territorial regions (two Czech NUTS 2 units) using data collected in 2018. The Random Forest method
was used for classification. In terms of classification accuracy, a combination of these parameters was
tested: The Number of Trees (NT), the Variables per Split (VPS) and the Bag Fraction (BF). A total
of 450 combinations of different parameters were tested. The highest accuracy classification with
an overall accuracy = 89.1% and Cohen’s Kappa = 0.84 had the following combination: NT = 150,
VPS = 3
and BF = 0.1. For classification purposes, a mosaic was created using the median method.
The resulting mosaic consisted of all Sentinel-2 bands in 10 and 20 m spatial resolution. Altitude
values derived from SRTM and NDVI variance values were also included in the classification. These
added bands were the most significant in terms of Gini importance.
Keywords: Google Earth Engine; Random Forest; LULUCF; Sentinel-2; Czechia
1. Introduction
The land cover/land-use change (LCLUC) program is one of the most important
sources of information on the development of global environmental change. LCLUC forms
the primary source of data for numerous mathematical models that seek to define future
development scenarios in many areas of the environment, including climate change [
1
,
2
].
The UN Secretariat on Climate Change and the adopted Paris Agreements under the United
States Framework Convention on Climate Change (UNFCCC) have declared the LCLUCs
monitoring to be highly relevant, as LCLUCs have a significant impact on climate change
and the global carbon cycle. For these purposes, the binding regulation is provided for
the inventory and reporting of relevant land use classes, so-called LULUCF—land use,
land-use change and forestry (see Decision 529/2013/EU, European Commission 2013).
LULUCF information is collected and reported on an international scale and is one of the
main input data sources for climate change modeling and GHG (greenhouse gas) emission
estimates within the IPCC (Intergovernmental Panel on Climate Change).
The development of international agreements on climate and climate policy has
been shaping the role of LULUCF. Researchers are increasingly developing sophisticated
research strategies to represent the global dimension of land use and assess its impact
on climate mitigation [
3
]. Full LULUCF integration fits well with ongoing international
efforts to integrate forests and other aspects into the climate policy framework, e.g., the
context of REDD+ (Reduced Emissions from Deforestation and Forest Degradation) [
4
–
6
].
Standardized methods and accurate and harmonized LULUCF data are a key factor in
Remote Sens. 2022,14, 1189. https://doi.org/10.3390/rs14051189 https://www.mdpi.com/journal/remotesensing
Remote Sens. 2022,14, 1189 2 of 26
accounting and evaluating the changes over a long period and modeling climate change
with predictive scenarios [
7
,
8
]. Earth observation (EO) is an effective and promising tool
for monitoring LCLUC [
9
,
10
]. The wide use of satellite data is currently possible mainly
due to the creation of freely available archives of satellite images from different missions
(e.g., Landsat and the Copernicus program). The Copernicus program has brought new
possibilities to EO. ESA is launching new satellite missions called Sentinels specifically
for the operational needs of the Copernicus program. The Sentinel-2 multispectral optical
dataset is now available with the aim to provide data with better resolutions (spatial,
temporal and spectral) than traditional data, such as Landsat images. Sentinel-2 data have
been available since 2015. The images are received via two parallel missions 2A and 2B and,
in the case of the overlapping scenes, the temporal resolution is less than five days [11].
EO has a prospective potential in monitoring LULUCF. In particular, medium-resolution
images, such as the 30 m Landsat resolution and Sentinel-2 (i.e., 10 m) resolution, seem to
be a suitable source of data for LULUCF [
12
]. Based on these opportunities, the EU and
other international institutions are looking for new LULUCF strategies. The use of large
volumes and a wide range of data causes significant difficulties related to the compatibility
and harmonization of input data [
13
,
14
]. Within LULUCF, the status and development
of the area of the following classes are inventoried and reported: Forest Land, Cropland,
Grassland, Wetlands, Settlements and Other Land. The definition and harmonization of
LCLUC inputs according to defined LULUCF classes are one of the most important tasks
within international LULUCF reporting [15].
In the LCLUC classification process, machine learning methods, such as Random
Forest (RF), are currently mainly used and developed. Random Forest was firstly described
by [
16
]. This method is widely used in multitemporal LCLUC classification. For example,
it was applied in [
17
,
18
]. Its essence is the creation of decision trees, where each tree
individually evaluates the class to which each individual pixel belongs. The classification
of a pixel into a class is assigned within the tree based on input parameters [19–21].
LULUCF reporting in Czechia has been exclusively based on the cadastral land
use information of the Czech Office for Surveying, Mapping and Cadaster (COSMC;
www.cuzk.cz, accessed on 3 September 2021). The Czech land-use representation and the
land-use change identification system use COSMC data. COSMC provides the annually
updated areas for all land-use categories. In addition, data obtained from the Forest Man-
agement Institute (FMI) on forests (harvest, increment, felling, etc.) are used in the LULUCF
categories involving forest land. However, according to many studies, e.g., [
22
,
23
], cadas-
tral data are not able to reflect all changes in time that occur in the landscape and do not
report them fully by the LULUCF classification nomenclature. Thus, the current LULUCF
reporting has several weaknesses that affect the quality of the collected data. Moreover,
there is no database derived from EO data to meet the LULUCF criteria (annual update,
classification nomenclature, minimum mapping unit, etc.). Therefore, this study focuses
on the development of an RF-based classification method that allows the classification
of Sentinel-2 data according to LULUCF requirements. Multispectral satellite data from
the Sentinel-2 mission are used due to their high spatial and temporal resolution. The
methodological procedures are developed and implemented on the freely accessible Google
Earth Engine (GEE) cloud platform. The methodology and results of the study are in
accordance with the LULUCF reporting process and are tested for selected larger territorial
units in Czechia using data collected by Sentinel-2 in 2018. From the research point of
view, the most important task is to find the most suitable combination of Random Forest
classifier input parameters to achieve the highest classification accuracy (Number of Trees,
Variables per Split and Bag Fraction). The LULUCF classification is based on a multitempo-
ral approach, which uses several images in the observed vegetation season. The Stratified
Random Sampling method [24] is used to evaluate the accuracy of the classification.
This study has the following subresearch objectives:
•
Development and testing of methods of mosaicking, accurate clouds detection and
unmasking for Sentinel-2 data in the GEE.
Remote Sens. 2022,14, 1189 3 of 26
•
Creation of the LULUCF classification nomenclature for Czechia with a detailed
semantic description and maximum compatibility with LULUCF.
•
Based on Sentinel-2 data testing RF classification algorithms for LULUCF classification
with the classification accuracy of at least 85% (Kappa index value above 0.75) in
larger territorial units of Czechia, specifically, two NUTS 2 (NUTS are Nomenclature
of Territorial Units for Statistics used in European Union).
•
Accuracy evaluation for individual LULUCF categories: Forest Land, Cropland, Grass-
land, Wetlands, Settlements and Other Land.
•
Discussion on the methodology used, data and achieved results with regard to the
needs of LULUCF reporting.
•
Ultimately, presenting the created methods and outputs in a freely accessible research
platform—GEE.
Research questions:
•
Is it possible to classify large area units with an overall accuracy of more than 85% on
an annual basis using machine learning classification algorithms and high spatial and
temporal resolution satellite data (Sentinel-2)?
•
For which of the LULUCF categories a higher accuracy of Sentinel-2 data classification
could be achieved and which categories appear to be problematic?
•
What methods of cloud mosaic and cloud detection/unmasking are most suitable for
data processing in the GEE cloud environment?
2. Materials and Methods
2.1. Area of Interest
Two NUTS 2 regions were analyzed for the purposes of this study, namely, Jihovýchod
(CZ06) and StˇredníMorava (CZ07), as shown in Figure 1. The region of interest selection
was guided by the criteria of a project, “Developing supports for monitoring and reporting of
GHG emissions and removals from land use, land use change and forestry”, from which this
study originates (https://www.copernicus-user-uptake.eu/user-uptake/details/developing-
support-for-monitoring-and-reporting-of-ghg-emissions-and-removals-from-land-use-land-
use-change-and-forestry-73, accessed on 22 September 2021). The total area of the region is up
to 23,217 km
2
. The region is very heterogeneous: the lowest point is at the confluence of rivers
Morava and Dyje at an elevation of 150 m a. s. l., the highest point is the mountain of Pradˇed,
reaching 1492 m a. s. l. The longest river is Morava, which forms the axis of the region and
flows from north to south. The majority of the land is used for agriculture and vineyards in the
lowlands in the southern parts. From south to west, east and north, the elevation of the area
gradually increases. Forests begin to dominate with increasing altitude. Deciduous forests
are found at lower altitudes (at the confluence of Morava and Dyje, Chˇriby and Moravský
kras) and coniferous forests predominate at higher altitudes (Beskydy, Jeseníky and Vysoˇcina),
which are often formed by monocultures of Norway spruce (Picea abies). The biggest cities
of the area of interest are Brno (381,346 inhabitants), Olomouc (100,663 inhabitants), Zlín
(74,935 inhabitants) and Jihlava (51,216 inhabitants).
2.2. Data
Freely available Sentinel-2 multispectral images from the joint ESA/European Com-
mission Copernicus Mission were used for compositing and classification. The images were
acquired in the late spring and early summer periods of 2018 and preprocessed through the
Sen2Cor algorithm [
11
,
25
]. Therefore, the atmospherically corrected data (L2A) from both
Sentinel-2A and Sentinel-2B satellites were used for this research. These data are provided
in 10 m spatial resolution (B2 Blue, B3 Green, B4 Red, B8 NIR bands) and 20 m spatial
resolution (B5–B7 and B8A Vegetation red edge and B11-B12 SWIR bands) [
11
,
25
]. Bands
with a resolution of 20 m were resampled to a higher resolution of 10 m using the nearest
neighbor method. Sentinel-2 images have a 12-bit radiometric resolution but are provided
in a 16-bit radiometric resolution [
11
], specifically through unsigned integers [
19
] with
values ranging from 0 to 65,535. Classifications were performed in GEE using JavaScript
Remote Sens. 2022,14, 1189 4 of 26
language, where the preprocessed Sentinel-2 Multispectral Instrument Level-2A dataset is
available [25,26].
Remote Sens. 2022, 14, x FOR PEER REVIEW 4 of 27
Figure 1. Area of interest—land cover is derived from CORINE Land Cover 2018 according to the
method documented in Appendix B.
2.2. Data
Freely available Sentinel-2 multispectral images from the joint ESA/European Com-
mission Copernicus Mission were used for compositing and classification. The images
were acquired in the late spring and early summer periods of 2018 and preprocessed
through the Sen2Cor algorithm [11,25]. Therefore, the atmospherically corrected data
(L2A) from both Sentinel-2A and Sentinel-2B satellites were used for this research. These
data are provided in 10 m spatial resolution (B2 Blue, B3 Green, B4 Red, B8 NIR bands)
and 20 m spatial resolution (B5–B7 and B8A Vegetation red edge and B11-B12 SWIR
bands) [11,25]. Bands with a resolution of 20 m were resampled to a higher resolution of
10 m using the nearest neighbor method. Sentinel-2 images have a 12-bit radiometric res-
olution but are provided in a 16-bit radiometric resolution [11], specifically through un-
signed integers [19] with values ranging from 0 to 65,535. Classifications were performed
in GEE using JavaScript language, where the preprocessed Sentinel-2 Multispectral In-
strument Level-2A dataset is available [25,26].
The digital elevation SRTM (The Shuttle Radar Topography Mission) radar data were
used for classification. The dataset is provided within the GEE platform with approxi-
mately 30 m spatial resolution as an SRTM V3 (void-filled) product. The SRTM band was
used as one of the input bands for the classification process.
The Copernicus CLC (Corine Land Cover) 2018 database provided within the GEE
platform and the ZM 10 map data (“Základní mapa ČR v měřítku 1:10,000”, Basic map of
the Czech Republic at a scale of 1:10,000; WMS from ČÚZK) and LPIS for years 2018
(“Veřejný registr půd”, Public land register available from eAGRI; in shapefile format)
were used for the creation of training and validation datasets. Historical orthophotos from
2017, 2018 and 2019 (WMS from ČÚZK) and historical imageries in Google Earth Pro soft-
ware were used to verify training polygons and validation points. Google Earth Pro soft-
ware provides imagery with very high spatial resolution—Maxar satellite imagery with up
to 0.3 m spatial resolution (from 2015 to 2021) and CNES/Airbus with up to 0.5 m spatial
resolution (from 2015 to 2021), users can examine these data using the internal Time Ma-
chine plugin.
Figure 1.
Area of interest—land cover is derived from CORINE Land Cover 2018 according to the
method documented in Appendix B.
The digital elevation SRTM (The Shuttle Radar Topography Mission) radar data were
used for classification. The dataset is provided within the GEE platform with approximately
30 m spatial resolution as an SRTM V3 (void-filled) product. The SRTM band was used as
one of the input bands for the classification process.
The Copernicus CLC (Corine Land Cover) 2018 database provided within the GEE
platform and the ZM 10 map data (“Základnímapa ˇ
CR v mˇeˇrítku 1:10,000”, Basic map
of the Czech Republic at a scale of 1:10,000; WMS from ˇ
CÚZK) and LPIS for years 2018
(“Veˇrejnýregistr p˚ud”, Public land register available from eAGRI; in shapefile format)
were used for the creation of training and validation datasets. Historical orthophotos
from 2017, 2018 and 2019 (WMS from ˇ
CÚZK) and historical imageries in Google Earth
Pro software were used to verify training polygons and validation points. Google Earth
Pro software provides imagery with very high spatial resolution—Maxar satellite imagery
with up to 0.3 m spatial resolution (from 2015 to 2021) and CNES/Airbus with up to 0.5 m
spatial resolution (from 2015 to 2021), users can examine these data using the internal Time
Machine plugin.
2.3. Legend
The first basic methodological step was the creation of the classification nomenclature.
The classification nomenclature follows the LULUCF regulations [
15
], which distinguish
and report the status and development of areas of the following classes: Forest Land,
Cropland, Grassland, Wetlands, Settlements and Other Land. Within the area of interest,
the following classes were defined:
Forest Land–vegetation can be considered a forest if it covers an area of at least
0.5 ha [
27
] and includes woodlands and clearcut localities where there is no forest present,
but is expected to grow within the next few decades.
Cropland includes agricultural land and permanent crops, including vineyards, hop
fields, gardens and orchards.
Remote Sens. 2022,14, 1189 5 of 26
Grassland includes both natural and managed grasslands (pastures and meadows).
Settlements in addition to built-up areas also include roads, urban greenery, gardens
near houses, landfills and active quarries.
Wetlands include marshlands, bodies of water and watercourses.
Other land mainly includes rocks, subalpine stands of dwarf Norway spruces (Picea abies)
and nonnative shrub mountain pines (Pinus mugo) in the higher parts of the Czech mountains,
as well as woodland/trees outside forest (ToF), such as groves and alleys, which cannot be
considered as a forest according to the LULUCF regulations.
In the initial stage of classification, the Woodland class was created instead of the
Forest Land class to highlight all the forested areas. Due to differences between LULUCF
classes Forest Land and Other Land, the Woodland class was later divided according to the
LULUCF regulations into Forest Land (polygons equal to or greater than 0.5 ha), and the
remaining polygons (areas of less than 0.5 ha) were added to the results of the Other Land
class during the post-classification process.
Detailed information on LULUCF classes, including their content description, is given
in Appendix A.
2.4. Methods
The complete methodological procedure of data processing is shown in Figure 2,
which defines the procedures of preprocessing, mosaicking and classification, as well as
methods for assessing accuracy and post-classification steps. The following parts describe
the individual steps in more detail.
Remote Sens. 2022, 14, x FOR PEER REVIEW 5 of 27
2.3. Legend
The first basic methodological step was the creation of the classification nomencla-
ture. The classification nomenclature follows the LULUCF regulations [15], which distin-
guish and report the status and development of areas of the following classes: Forest Land,
Cropland, Grassland, Wetlands, Settlements and Other Land. Within the area of interest,
the following classes were defined:
Forest Land–vegetation can be considered a forest if it covers an area of at least 0.5
ha [27] and includes woodlands and clearcut localities where there is no forest present,
but is expected to grow within the next few decades.
Cropland includes agricultural land and permanent crops, including vineyards, hop
fields, gardens and orchards.
Grassland includes both natural and managed grasslands (pastures and meadows).
Settlements in addition to built-up areas also include roads, urban greenery, gardens
near houses, landfills and active quarries.
Wetlands include marshlands, bodies of water and watercourses.
Other land mainly includes rocks, subalpine stands of dwarf Norway spruces (Picea
abies) and nonnative shrub mountain pines (Pinus mugo) in the higher parts of the Czech
mountains, as well as woodland/trees outside forest (ToF), such as groves and alleys,
which cannot be considered as a forest according to the LULUCF regulations.
In the initial stage of classification, the Woodland class was created instead of the
Forest Land class to highlight all the forested areas. Due to differences between LULUCF
classes Forest Land and Other Land, the Woodland class was later divided according to
the LULUCF regulations into Forest Land (polygons equal to or greater than 0.5 ha), and
the remaining polygons (areas of less than 0.5 ha) were added to the results of the Other
Land class during the post-classification process.
Detailed information on LULUCF classes, including their content description, is
given in Appendix A.
2.4. Methods
The complete methodological procedure of data processing is shown in Figure 2,
which defines the procedures of preprocessing, mosaicking and classification, as well as
methods for assessing accuracy and post-classification steps. The following parts describe
the individual steps in more detail.
Figure 2. Workflow.
2.4.1. Cloud Masking and Mosaicking
Due to the size of the area of interest, a decision was made to create a mosaic for
classification purposes. The mosaic was created by using the full potential of Sentinel-2
data, i.e., using images taken from both Sentinel-2A and Sentinel-2B. All images for
Figure 2. Workflow.
2.4.1. Cloud Masking and Mosaicking
Due to the size of the area of interest, a decision was made to create a mosaic for
classification purposes. The mosaic was created by using the full potential of Sentinel-2
data, i.e., using images taken from both Sentinel-2A and Sentinel-2B. All images for mosaic
creation were taken in the period from May to the end of July with a total cloud cover
below 75% in the whole scene. This period was used mainly because there are only the last
remnants of snow cover in the peak parts of the area of interest, and the main vegetation
season takes place in the selected months. At the same time, it was the period in which
there seemed to be the lowest cloud cover throughout the year 2018. This set of selected
images was used in a further step—cloud masking.
For the cloud masking of Sentinel-2 data in GEE, the Sentinel-2: Cloud Probability
dataset (so-called s2cloudless) was used [
28
]. It is constituted of a single band with 20 m
spatial resolution that represents the probability of cloudiness (0–100%) for each pixel
of all Sentinel-2 tiles in the entire archive. The selection of this approach was inspired
by [
29
], who compared different Landsat 8 and Sentinel-2 cloud masking approaches, and
Remote Sens. 2022,14, 1189 6 of 26
the s2cloudless dataset significantly outperformed other methods. Cloud shadow was
detected using an algorithm developed in GEE [
30
], which is based on cloud projection
intersection (defined by the solar azimuth angle obtained in each Sentinel-2 tile metadata)
with low-reflectance near-infrared pixels. The next parameter to detect low-reflectance near-
infrared pixels as the cloud shadow is a distance from the cloud. After testing and visual
inspection, the following parameters were chosen for the used algorithm: cld_prb_thresh
(cloud probability, where higher values were considered as clouds) = 40%, nir_drk_thresh
(reflectance in the NIR band, where lower values were considered as cloud shadows)
= 0.15
and cld_prj_dist (maximum allowed distance in km to search for cloud shadows from
cloud edges) = 1 km; erosion 2 pixels (resolution 20 m/pixel) and dilation 5.5 pixels
(buffer 3.5 pixels) were applied for the elimination of small features and gaps in clouds
and shadows.
Figure 3illustrates the process of cloud masking. Figure 3a shows the initial step of
masked shadows, clouds and the created buffer. Figure 3b shows the final masked area
applied to all the parameters. It is evident that not all pixels that are initially identified as
clouds or cloud shadow (dark pixels) in Figure 3a were included in the final cloud mask.
The final mask does not include objects that were eliminated by erosion, as well as dark
pixels that are not within a defined distance and angle from the detected cloud.
Remote Sens. 2022, 14, x FOR PEER REVIEW 6 of 27
mosaic creation were taken in the period from May to the end of July with a total cloud
cover below 75% in the whole scene. This period was used mainly because there are only
the last remnants of snow cover in the peak parts of the area of interest, and the main
vegetation season takes place in the selected months. At the same time, it was the period
in which there seemed to be the lowest cloud cover throughout the year 2018. This set of
selected images was used in a further step—cloud masking.
For the cloud masking of Sentinel-2 data in GEE, the Sentinel-2: Cloud Probability
dataset (so-called s2cloudless) was used [28]. It is constituted of a single band with 20 m
spatial resolution that represents the probability of cloudiness (0–100%) for each pixel of
all Sentinel-2 tiles in the entire archive. The selection of this approach was inspired by [29],
who compared different Landsat 8 and Sentinel-2 cloud masking approaches, and the
s2cloudless dataset significantly outperformed other methods. Cloud shadow was de-
tected using an algorithm developed in GEE [30], which is based on cloud projection in-
tersection (defined by the solar azimuth angle obtained in each Sentinel-2 tile metadata)
with low-reflectance near-infrared pixels. The next parameter to detect low-reflectance
near-infrared pixels as the cloud shadow is a distance from the cloud. After testing and
visual inspection, the following parameters were chosen for the used algorithm:
cld_prb_thresh (cloud probability, where higher values were considered as clouds) = 40%,
nir_drk_thresh (reflectance in the NIR band, where lower values were considered as cloud
shadows) = 0.15 and cld_prj_dist (maximum allowed distance in km to search for cloud
shadows from cloud edges) = 1 km; erosion 2 pixels (resolution 20 m/pixel) and dilation
5.5 pixels (buffer 3.5 pixels) were applied for the elimination of small features and gaps in
clouds and shadows.
Figure 3 illustrates the process of cloud masking. Figure 3a shows the initial step of
masked shadows, clouds and the created buffer. Figure 3b shows the final masked area
applied to all the parameters. It is evident that not all pixels that are initially identified as
clouds or cloud shadow (dark pixels) in Figure 3a were included in the final cloud mask.
The final mask does not include objects that were eliminated by erosion, as well as dark
pixels that are not within a defined distance and angle from the detected cloud.
Figure 3. Demonstration of cloud masking method and its effectivity based on the selected part of
the image 20180617T095029_20180617T095028_T33UXR with 64.95% cloudiness in the whole image.
(a) shows the initial step of masked shadows, clouds and the created buffer, (b) shows the final
masked area applied to all the parameters.
Figure 3.
Demonstration of cloud masking method and its effectivity based on the selected part of
the image 20180617T095029_20180617T095028_T33UXR with 64.95% cloudiness in the whole image.
(
a
) shows the initial step of masked shadows, clouds and the created buffer, (
b
) shows the final
masked area applied to all the parameters.
At the next step, a median mosaic was created—inspired by [
31
,
32
]. All available
images with lower than 75% cloud cover were selected. All S-2 bands with a resolution of
10/20 m and the NDVI index (calculated from bands B4 and B8) were used. Only pixels
that were identified as cloud free were included in the median calculation. The 75% cloud
cover threshold was chosen to avoid data gaps mainly in mountainous areas, where it was
difficult to detect pixels not infected by clouds or cloud shadows. If high cloud cover is
documented in the metadata of a scene, some areas may not be covered by clouds. This
higher threshold made it possible to work with a larger number of images, which resulted
in a cloud-free mosaic. The median approach was chosen because it is not as affected by
Remote Sens. 2022,14, 1189 7 of 26
outliers as the average value. Figure 4shows a graph comparing the quantile value ranges
and the average values calculated from the available unmasked values (May to July) for
the selected training polygons (ID 331—Cropland; ID 488—Grassland). The mean values
of surface reflectance of training polygon 331 in 6 bands of 10 were higher than 75% of the
values from which these means were calculated. This was caused by outliers.
Remote Sens. 2022, 14, x FOR PEER REVIEW 7 of 27
At the next step, a median mosaic was created—inspired by [31,32]. All available im-
ages with lower than 75% cloud cover were selected. All S-2 bands with a resolution of
10/20 m and the NDVI index (calculated from bands B4 and B8) were used. Only pixels
that were identified as cloud free were included in the median calculation. The 75% cloud
cover threshold was chosen to avoid data gaps mainly in mountainous areas, where it was
difficult to detect pixels not infected by clouds or cloud shadows. If high cloud cover is
documented in the metadata of a scene, some areas may not be covered by clouds. This
higher threshold made it possible to work with a larger number of images, which resulted
in a cloud-free mosaic. The median approach was chosen because it is not as affected by
outliers as the average value. Figure 4 shows a graph comparing the quantile value ranges
and the average values calculated from the available unmasked values (May to July) for
the selected training polygons (ID 331—Cropland; ID 488—Grassland). The mean values
of surface reflectance of training polygon 331 in 6 bands of 10 were higher than 75% of the
values from which these means were calculated. This was caused by outliers.
Figure 4. Spectral signatures of selected two training data polygons; (a) spectral signature of train-
ing polygon 331—Cropland and (b) spectral signature of training polygon 488—Grassland.
The mosaic also includes a band representing the variance of the NDVI values in the
period from May to October. This band helps to distinguish relatively invariant surfaces
such as buildings (small variance) from surfaces dynamically changing during the season,
e.g., arable land, which refers to high variance of the NDVI value.
Figure 5 represents the variance of NDVI in the sample selected area. The map dis-
played on the left side of the image divides these values into three intervals. Forests and
buildings have the lowest variance (see aerial image and ZM 10 in the middle and right
map fields), and grasslands have a higher variance (visualized in yellow on ZM 10). Ara-
ble land shows the highest NDVI variance.
Another band added to the resulting mosaic was SRTM DEM containing altitude
data with a spatial resolution of 30 m. The SRTM band was used together with Landsat 8
multispectral satellite data in [31]. These data were important for distinguishing similar
surfaces in terms of land cover, but different land use approaches for LULUCF purposes.
Examples are stone and paved surfaces, where it is necessary to distinguish blockfields
(Other Land) from paved areas within the Settlements class. The resulting mosaic has a
spatial resolution of 10 m. All input data with a resolution lower than 10 m (S-2 bands
with a resolution of 20 m and SRTM with a resolution of 30 m) were resampled using the
Nearest Neighbor method.
The significance of the bands for classification was recorded using Gini importance
for the 4 parameter combinations in Appendix C. As can be seen in Appendix C, the im-
portance of SRTM elevation and NDVI variance is the most significant, whereas the B8
band is of the least importance.
Figure 4.
Spectral signatures of selected two training data polygons; (
a
) spectral signature of training
polygon 331—Cropland and (b) spectral signature of training polygon 488—Grassland.
The mosaic also includes a band representing the variance of the NDVI values in the
period from May to October. This band helps to distinguish relatively invariant surfaces
such as buildings (small variance) from surfaces dynamically changing during the season,
e.g., arable land, which refers to high variance of the NDVI value.
Figure 5represents the variance of NDVI in the sample selected area. The map
displayed on the left side of the image divides these values into three intervals. Forests and
buildings have the lowest variance (see aerial image and ZM 10 in the middle and right
map fields), and grasslands have a higher variance (visualized in yellow on ZM 10). Arable
land shows the highest NDVI variance.
Remote Sens. 2022, 14, x FOR PEER REVIEW 8 of 27
Figure 5. Comparison of NDVI variance, aerial image 2018 (ČÚZK) and ZM 10 (ČÚZK) in the sam-
ple selected area.
2.4.2. LULUCF Classification
The Random Forest (RF) method was selected, tested and used for classification. This
method has been successfully used in the classification of multitemporal satellite data,
e.g., [18,31]. Compared to other classification algorithms (CART, SVM, kNN and MLC),
this method achieved the best results in many studies [17,32–34]. It is a method of con-
trolled nonparametric classification using machine learning. Its essence is the creation of
decision trees, where each tree individually evaluates to which class each individual pixel
belongs; see [16,34]. The basic parameter is the Number of Trees (NT). Other adjustable
classification parameters are the Variables per Split (VPS), Bag Fraction (BF), Max Nodes
and Min Leaf Population.
2.4.3. Training Polygons
An important aspect of the resulting classification accuracy is the training data. The train-
ing polygons for this study were created by two methods. The first method is the semi-auto-
matic creation of training polygons within the CORINE Land Cover 2018 (CLC 2018) vector
layer. The second method was the manual creation of the additional training polygons.
From the CLC 2018 polygon database, the core areas of the training polygons were
created using the Buffer function with the following parameter: −100 m. Inside these areas,
training polygons of a circle shape with a diameter of 80 m were randomly generated.
This can be seen in Figure 6, where one of these training polygons is visualized. These
polygons/circles were generated with 2500 m minimal distance.
For some evaluated classes/surfaces, no training polygons were used in the proce-
dure above. It is given by both geometric and thematic characters. For example, in the case
of watercourses, this is due to the fact that no polygon in this class has a core area of 100
m inwards. Only one polygon was generated for the other land class, which, however, did
not include some important elements of this class, e.g., no training polygon was created
on the territory of a photovoltaic power plant. Therefore, 7 training polygons were man-
ually created for the Other Land class, 5 of them were located in rubble fields in the Hrubý
Jeseník mountains; in one case, they were rocks in the Suché skály nature reserve, and in
another case, they were scrub mountain pines in the alpine vegetation zone of Hrubý
Figure 5.
Comparison of NDVI variance, aerial image 2018 ( ˇ
CÚZK) and ZM 10 ( ˇ
CÚZK) in the sample
selected area.
Remote Sens. 2022,14, 1189 8 of 26
Another band added to the resulting mosaic was SRTM DEM containing altitude
data with a spatial resolution of 30 m. The SRTM band was used together with Landsat
8 multispectral satellite data in [31]. These data were important for distinguishing similar
surfaces in terms of land cover, but different land use approaches for LULUCF purposes.
Examples are stone and paved surfaces, where it is necessary to distinguish blockfields
(Other Land) from paved areas within the Settlements class. The resulting mosaic has a
spatial resolution of 10 m. All input data with a resolution lower than 10 m (S-2 bands
with a resolution of 20 m and SRTM with a resolution of 30 m) were resampled using the
Nearest Neighbor method.
The significance of the bands for classification was recorded using Gini importance
for the 4 parameter combinations in Appendix C. As can be seen in Appendix C, the
importance of SRTM elevation and NDVI variance is the most significant, whereas the B8
band is of the least importance.
2.4.2. LULUCF Classification
The Random Forest (RF) method was selected, tested and used for classification.
This method has been successfully used in the classification of multitemporal satellite
data, e.g., [
18
,
31
]. Compared to other classification algorithms (CART, SVM, kNN and
MLC), this method achieved the best results in many studies [
17
,
32
–
34
]. It is a method of
controlled nonparametric classification using machine learning. Its essence is the creation of
decision trees, where each tree individually evaluates to which class each individual pixel
belongs; see [
16
,
34
]. The basic parameter is the Number of Trees (NT). Other adjustable
classification parameters are the Variables per Split (VPS), Bag Fraction (BF), Max Nodes
and Min Leaf Population.
2.4.3. Training Polygons
An important aspect of the resulting classification accuracy is the training data. The
training polygons for this study were created by two methods. The first method is the
semi-automatic creation of training polygons within the CORINE Land Cover 2018 (CLC
2018) vector layer. The second method was the manual creation of the additional train-
ing polygons.
From the CLC 2018 polygon database, the core areas of the training polygons were
created using the Buffer function with the following parameter:
−
100 m. Inside these areas,
training polygons of a circle shape with a diameter of 80 m were randomly generated.
This can be seen in Figure 6, where one of these training polygons is visualized. These
polygons/circles were generated with 2500 m minimal distance.
For some evaluated classes/surfaces, no training polygons were used in the procedure
above. It is given by both geometric and thematic characters. For example, in the case of
watercourses, this is due to the fact that no polygon in this class has a core area of 100 m
inwards. Only one polygon was generated for the other land class, which, however, did
not include some important elements of this class, e.g., no training polygon was created on
the territory of a photovoltaic power plant. Therefore, 7 training polygons were manually
created for the Other Land class, 5 of them were located in rubble fields in the Hrubý
Jeseník mountains; in one case, they were rocks in the Suchéskály nature reserve, and
in another case, they were scrub mountain pines in the alpine vegetation zone of Hrubý
Jeseník. Polygons for specific areas of mountain meadows in the Beskydy mountains
were collected manually. Training polygons for peatbogs and reeds were added to the
Wetlands category. Due to the high heterogeneity of the Cropland class, some polygons
were added to cover some specific types of land cover. It was found during preliminary
classification testing that some areas within the Cropland class were misclassified as other
classes. As a result, some additional training polygons were manually added at these
localities. Polygons that were deforested due to droughts and bark-beetle disturbances
(Ips typographus) and are currently in the initial stages of forest growth were also created.
Cropland and Grassland polygons were also manually added to better differentiate these
Remote Sens. 2022,14, 1189 9 of 26
surfaces. A total of 299 training polygons were created; their distribution within the area
of interest can be seen in Figure 7. The number and structure of manually added training
polygons are shown in Table 1.
Remote Sens. 2022, 14, x FOR PEER REVIEW 9 of 27
Jeseník. Polygons for specific areas of mountain meadows in the Beskydy mountains were
collected manually. Training polygons for peatbogs and reeds were added to the Wet-
lands category. Due to the high heterogeneity of the Cropland class, some polygons were
added to cover some specific types of land cover. It was found during preliminary classi-
fication testing that some areas within the Cropland class were misclassified as other clas-
ses. As a result, some additional training polygons were manually added at these locali-
ties. Polygons that were deforested due to droughts and bark-beetle disturbances (Ips ty-
pographus) and are currently in the initial stages of forest growth were also created.
Cropland and Grassland polygons were also manually added to better differentiate these
surfaces. A total of 299 training polygons were created; their distribution within the area
of interest can be seen in Figure 7. The number and structure of manually added training
polygons are shown in Table 1.
Figure 6. Scheme of creating training polygons.
Figure 6. Scheme of creating training polygons.
Remote Sens. 2022, 14, x FOR PEER REVIEW 10 of 27
Figure 7. Distribution of training polygons within the area of interest.
Table 1. Amount of training polygons for individual training classes.
Class LULUCF
Semi-Automatically Created
Polygons
Manually Created Polygons
Total
Settlements
16
18
34
Cropland
122
24
146
Woodland
60
12
72
Grassland
18
13
31
Wetlands
2
6
8
Other land
1
7
8
Total sum
219
80
299
A conversion table of CLC to LULUCF classes was created to systematically deter-
mine the LULUCF class. The individual training polygons determined from CLC belong
to the LULUCF class, which was decided based on this conversion table, as presented in
Appendix B.
All created training polygons from the CLC were verified using an orthophoto to
check if their declared land cover matched the LULUCF class. This verification was per-
formed using orthophotos from ČÚZK available for the area of interest. Most of the im-
ages of the area were captured in 2018, and only the western parts of the area were missing
images from the same year; therefore, images from 2017 and 2019 were used. If the land
cover of the checked training polygon did not change in aerial photographs in this time
interval (2017–2019), there was no reason to consider the class incorrectly assigned. If the
specified orthophoto land cover did not match the declared CLC land cover, the training
polygon was deleted or manually adjusted to be within the declared land cover. Therefore,
emphasis was placed on the polygon lying in its entirety in one class without interfering
with other classes. The affiliation of training polygons to the Grassland class was checked
using LPIS data and ZM 10 maps from ČÚZK. This detailed inspection was carried out due
to the difficult to distinguish Cropland and Grassland classes using orthophotos.
2.4.4. Parameters of Classification
Figure 7. Distribution of training polygons within the area of interest.
Remote Sens. 2022,14, 1189 10 of 26
Table 1. Amount of training polygons for individual training classes.
Class LULUCF Semi-Automatically Created Polygons Manually Created Polygons Total
Settlements 16 18 34
Cropland 122 24 146
Woodland 60 12 72
Grassland 18 13 31
Wetlands 2 6 8
Other land 1 7 8
Total sum 219 80 299
A conversion table of CLC to LULUCF classes was created to systematically determine
the LULUCF class. The individual training polygons determined from CLC belong to
the LULUCF class, which was decided based on this conversion table, as presented in
Appendix B.
All created training polygons from the CLC were verified using an orthophoto to check
if their declared land cover matched the LULUCF class. This verification was performed
using orthophotos from ˇ
CÚZK available for the area of interest. Most of the images of the
area were captured in 2018, and only the western parts of the area were missing images
from the same year; therefore, images from 2017 and 2019 were used. If the land cover of
the checked training polygon did not change in aerial photographs in this time interval
(2017–2019), there was no reason to consider the class incorrectly assigned. If the specified
orthophoto land cover did not match the declared CLC land cover, the training polygon
was deleted or manually adjusted to be within the declared land cover. Therefore, emphasis
was placed on the polygon lying in its entirety in one class without interfering with other
classes. The affiliation of training polygons to the Grassland class was checked using LPIS
data and ZM 10 maps from ˇ
CÚZK. This detailed inspection was carried out due to the
difficult to distinguish Cropland and Grassland classes using orthophotos.
2.4.4. Parameters of Classification
In the classification process, the most important task was to define a combination of
parameter settings that would deliver the highest accuracy. The Number of Trees parameter
was tested from 50 to 400 at 25-tree intervals, the Variable per Split parameter was tested
from 1 to 6 at 1-variable interval and the Bag Fraction parameter was tested from 0.1 to 0.5
at 0.1-fraction intervals. The other Max Nodes parameters were left with the default value
‘NULL’, that is, without limits, as well as the default value of 1 for the min Leaf Population.
A total of 450 combinations of the parameters Number of Trees, Variables per Split and Bag
Fraction were generated and evaluated.
Per-pixel classification often brings a ‘salt-and-pepper’ effect. This effect was elim-
inated by filtering and replacing isolated pixels with neighboring values. At this step,
the areas represented by one pixel were eliminated and replaced by the majority value
of the pixels in the 3
×
3 kernel window filtering. The point of this step is documented
in Figure 8a,b. The main complication is the pixels on the borders between two classes,
or in the case of tree growth, trees can cast shadows into their immediate surroundings,
which are mostly incorrectly classified as Wetlands. These lonely pixels were filtered. The
minimum mapping unit of the classification is 2 pixels, i.e., 200 m2.
Remote Sens. 2022,14, 1189 11 of 26
Remote Sens. 2022, 14, x FOR PEER REVIEW 12 of 27
Figure 8. Post-processing Classification; (a) map of original classification, (b) map with replaced
isolated pixels, (c) map with distinguished Forest land and ToF and (d) same area in ortophoto from
2018 (ČÚZK).
3. Results
3.1. Influence of Parameter Selection on the Resulting Accuracy of Classification
One of the main goals of the study was to develop an RF-based classification method
that allows the classification of Sentinel-2 data in GEE according to the LULUCF require-
ments. In terms of classification, the most important task was to find the most suitable
combination of RF classifier input parameters that will lead to the highest accuracy of the
LULUCF classification. The evaluated parameters were Number of Trees, Variables per
Split and Max Nodes. For this purpose, an innovative script was developed in the GEE
environment, which can evaluate hundreds of combinations of input parameters in a short
time and use the overall accuracy and Kappa index to select the combination with the
highest accuracy achieved.
From the achieved results of individual combinations (Appendix E), it was seen that
the highest value of the Kappa index (κ) was achieved in the case of a combination of the
settings of the parameters NT: 150, VPS: 3 and BF: 0.1 with the value κ = 0.8383. At the
same time, this setting of the examined parameters achieved the highest overall accuracy
of all combinations, with a value of 89.01%. The results of κ and the overall accuracy for
the individual parameter combinations are given in Appendix E.
Upon closer inspection of the settings and relevance of individual parameters, the
average values of κ (calculated from testing control points) for individual values of input
parameters are documented in Figure 9. In the case of the NT parameter, each average
value was calculated from a total of 30 κ values (6 combinations of the VPS parameter and
5 combinations of the Bag Fraction parameter), the Variables per Split parameter was cal-
culated from 75 values and the Bag Fraction was evaluated from 90 κ values. The param-
eters values used for the final (most accurate) classification are highlighted in red. If we
look at the values obtained from individual parameters, then the average value κ of the
NT parameter had an ascending character, together with the number of trees, and the
highest value of κ was reached at the maximum number of trees (400). However, the in-
crease in the value of κ from 150 was no longer as significant. The value of 150 trees was
evaluated as the most suitable in combination with other parameters. For the VPS param-
eters, the highest average value of κ was reached at the parameter value of 2. A
Figure 8.
Post-processing Classification; (
a
) map of original classification, (
b
) map with replaced
isolated pixels, (
c
) map with distinguished Forest land and ToF and (
d
) same area in ortophoto from
2018 ( ˇ
CÚZK).
2.4.5. Accuracy Assessment
The validation points were created in the ESRI ArcGIS Pro software, where the Create
Accuracy Assessment Points tool (Spatial Analyst) was used. A total of 2235 points were
created (in WGS 84/UTM zone 33N EPSG: 32633 coordinate system), and the Stratified
Sampling method based on preliminary classification testing was used for the accuracy
assessment and random creation of control points [
24
]. The affiliation of control points
to the LULUCF class was performed in the same way as in the case of training polygons;
see Section 2.4.3. Due to the low number of control points generated randomly for the
Other Land, 31 points in this category were manually created. For effective validation, an
innovative algorithm was created in the cloud-based platform GEE. The control points
were uploaded to GEE, where the Classifier package was used to validate the classifications.
Specifically, the classifier.confusionMatrix() function was used for confusion matrices
and the errorMatrix() function for overall accuracy [
35
] and the ConfusionMatrix.kappa()
function for the Kappa index by [
36
]. The Kappa index value (Cohen’s Kappa) was
calculated for each combination of input parameters. The combinations of input parameters
that achieved the highest Kappa index value were selected, and validation matrices and
overall accuracy were subsequently generated for them. The combinations of parameters
with the highest accuracy were selected as the most suitable for classification.
2.4.6. Post-Processing Classification
According to the definitions of LULUCF, tree growth with an area of less than 0.5 ha
cannot be considered as a forest, but as trees outside forest (ToF). For this reason, all
Woodland growths with an area of less than 5000 m
2
(less than 50 pixels) were converted
to the Other Land class. Figure 8b the state before the division of the Woodland class into
Forest Land and ToF and in Figure 8c the state after the division. Based on this step, the
minimum mapping unit for Forest Land differed from the other categories and was 5000 m
2
(0.5 ha).
Remote Sens. 2022,14, 1189 12 of 26
3. Results
3.1. Influence of Parameter Selection on the Resulting Accuracy of Classification
One of the main goals of the study was to develop an RF-based classification method
that allows the classification of Sentinel-2 data in GEE according to the LULUCF require-
ments. In terms of classification, the most important task was to find the most suitable
combination of RF classifier input parameters that will lead to the highest accuracy of the
LULUCF classification. The evaluated parameters were Number of Trees, Variables per
Split and Max Nodes. For this purpose, an innovative script was developed in the GEE
environment, which can evaluate hundreds of combinations of input parameters in a short
time and use the overall accuracy and Kappa index to select the combination with the
highest accuracy achieved.
From the achieved results of individual combinations (Appendix E), it was seen that
the highest value of the Kappa index (
κ
) was achieved in the case of a combination of the
settings of the parameters NT: 150, VPS: 3 and BF: 0.1 with the value
κ
= 0.8383. At the
same time, this setting of the examined parameters achieved the highest overall accuracy
of all combinations, with a value of 89.01%. The results of
κ
and the overall accuracy for
the individual parameter combinations are given in Appendix E.
Upon closer inspection of the settings and relevance of individual parameters, the
average values of
κ
(calculated from testing control points) for individual values of input
parameters are documented in Figure 9. In the case of the NT parameter, each average
value was calculated from a total of 30
κ
values (6 combinations of the VPS parameter
and 5 combinations of the Bag Fraction parameter), the Variables per Split parameter was
calculated from 75 values and the Bag Fraction was evaluated from 90
κ
values. The
parameters values used for the final (most accurate) classification are highlighted in red.
If we look at the values obtained from individual parameters, then the average value
κ
of the NT parameter had an ascending character, together with the number of trees, and
the highest value of
κ
was reached at the maximum number of trees (400). However, the
increase in the value of
κ
from 150 was no longer as significant. The value of 150 trees
was evaluated as the most suitable in combination with other parameters. For the VPS
parameters, the highest average value of
κ
was reached at the parameter value of 2. A
comparatively lower average value was obtained by the value of parameter 3, which
was selected for the combination of the final classification. This parameter also had the
largest range of the minimum and maximum average values of
κ
, which may indicate that
this parameter has a significant effect on the resulting combination for the most accurate
classification. For the Bag Fraction parameter, the highest average value of
κ
matched the
finally selected value of 0.1. Other BF parameter settings evaluated had lower κvalues.
Figure 10 describes in detail the average values of
κ
for the pair of evaluated param-
eters. The average values of the combination of VPS and NT (VPS/NT) were calculated
from 5 values of
κ
, the combinations of BF and NT (BF/NT) were calculated from 6 values
of κand the combinations of VPS and BF (VPS/BF) were calculated from 15 values.
When evaluating a VPS/NT combination, the VPS parameter values are expressed
in rows and NT in columns. The highest values of
κ
were reached by the combination
NT = 275
and VPS = 2 (NT = 150, VPS = 2 are the values used in the final combination of
three parameters for classification). The values of
κ
did not change much in the rows, unlike
the evaluated NT parameters in the columns. Therefore, the average value of
κ
is more
affected by the VPS parameter than the number of trees (especially obvious when setting
the number of trees above 50). The high relevance of the VPS parameter is also evidenced
by the evaluation of the VPS/BF parameter combination; in this case, the VPS influence
appears even stronger. The combination with the highest value of κ(BF = 1, VPS = 3) was
the same, which was selected as the most suitable in the combination of all three parameters.
When comparing these two parameters, the largest difference between the maximum and
minimum average values of
κ
was evident. On the contrary, the smallest differences in the
variability of
κ
values could be seen by comparing combinations of BF/NT parameters.
Remote Sens. 2022,14, 1189 13 of 26
The values of
κ
were found in the relatively narrow range of 0.811–0.820. BF = 0.1 and
NT = 150 are the values contained in the resulting combination of three parameters.
Remote Sens. 2022, 14, x FOR PEER REVIEW 13 of 27
comparatively lower average value was obtained by the value of parameter 3, which was
selected for the combination of the final classification. This parameter also had the largest
range of the minimum and maximum average values of κ, which may indicate that this
parameter has a significant effect on the resulting combination for the most accurate clas-
sification. For the Bag Fraction parameter, the highest average value of κ matched the
finally selected value of 0.1. Other BF parameter settings evaluated had lower κ values.
Figure 9. Average Kappa index values for individual parameters.
Figure 10 describes in detail the average values of κ for the pair of evaluated param-
eters. The average values of the combination of VPS and NT (VPS/NT) were calculated
from 5 values of κ, the combinations of BF and NT (BF/NT) were calculated from 6 values
of κ and the combinations of VPS and BF (VPS/BF) were calculated from 15 values.
Figure 10. Average Kappa index values for individual combinations of parameters.
When evaluating a VPS/NT combination, the VPS parameter values are expressed in
rows and NT in columns. The highest values of κ were reached by the combination NT =
275 and VPS = 2 (NT = 150, VPS = 2 are the values used in the final combination of three
Figure 9. Average Kappa index values for individual parameters.
Remote Sens. 2022, 14, x FOR PEER REVIEW 13 of 27
comparatively lower average value was obtained by the value of parameter 3, which was
selected for the combination of the final classification. This parameter also had the largest
range of the minimum and maximum average values of κ, which may indicate that this
parameter has a significant effect on the resulting combination for the most accurate clas-
sification. For the Bag Fraction parameter, the highest average value of κ matched the
finally selected value of 0.1. Other BF parameter settings evaluated had lower κ values.
Figure 9. Average Kappa index values for individual parameters.
Figure 10 describes in detail the average values of κ for the pair of evaluated param-
eters. The average values of the combination of VPS and NT (VPS/NT) were calculated
from 5 values of κ, the combinations of BF and NT (BF/NT) were calculated from 6 values
of κ and the combinations of VPS and BF (VPS/BF) were calculated from 15 values.
Figure 10. Average Kappa index values for individual combinations of parameters.
When evaluating a VPS/NT combination, the VPS parameter values are expressed in
rows and NT in columns. The highest values of κ were reached by the combination NT =
275 and VPS = 2 (NT = 150, VPS = 2 are the values used in the final combination of three
Figure 10. Average Kappa index values for individual combinations of parameters.
3.2. Accuracy Assessment of the Classification
In addition to the calculation of
κ
, the values of overall accuracy (OA) and validation
matrices were calculated for a detailed evaluation of the accuracy of the classification.
The validation matrix for the resulting combination of classification parameters (NT: 150,
VPS: 3 and BF: 0.1) is documented in Table 2. The overall accuracy reached 89.01%,
and Cohen’s Kappa was 0.8383 for this combination. When a closer inspection of the
producer accuracy and user accuracy values was performed, Settlements were most often
misclassified as the Cropland class. On the contrary, the Other Land class was most often
misclassified as Settlements. More validation points of Settlements were classified as
another class than validation points of other classes classified as Settlements. It follows
that the Settlements class should have a slightly undervalued classified area. Cropland
Remote Sens. 2022,14, 1189 14 of 26
was most often misclassified as Grassland and, conversely, Grassland was most often
misclassified as Cropland. Changing the Grassland class to the Cropland class with an
amount of 54 points was the most common change in the classification and accounted for
about one-fifth of all errors (2.24% of all validation points). The Cropland class appeared
to be slightly overvalued (especially at the expense of Grassland). The Woodland class
was most often misclassified as Grassland and vice versa. This could be caused by grass-
like clearings that formed in the forests after the trees were logged. This type of forest
could be observed within the area of interest due to bark beetle calamities and droughts
consequences. The Wetlands control points were most often misclassified as the Cropland
class. The main cause of this phenomenon could be seen in the location of control points in
places where there are swamps, reeds and peat bogs, i.e., wetlands with a more substantial
representation of the vegetation component. When comparing user accuracy and producer
accuracy, the Wetlands class appeared to be slightly underestimated. The Other Land class
seemed to be clearly underestimated, as less than half of the points belonging to this class
were correctly classified (confusion mainly with Settlements). On the other hand, no control
point from any other class was classified as Other Land.
Table 2. Validation matrix of the final classification.
Settlements Cropland Woodland Grassland Wetlands Other Land User Accuracy
Settlements 107 26 10 18 1 0 66.05%
Cropland 8 882 6 26 1 0 95.56%
Woodland 1 11 728 32 0 0 94.30%
Grassland 1 54 23 243 1 0 75.47%
Wetlands 0 5 1 1 17 0 70.83%
Other Land 16 1 0 1 0 14 43.75%
Producer Accuracy 80.45% 90.09% 94.79% 75.70% 85.00% 100.00% Overall accuracy 89.01%
The LULUCF classification was performed in the area of interest with a heterogeneous
character both from a physical–geographical and socio-economic point of view. The hetero-
geneous character of the area was determined by the diverse representation of individual
LULUCF classes. From the results documented in Table 3and Figure 11, it is clear that the
Cropland class had the highest area in 2018 with more than 42% of the total area. Grassland
class occupied over 15%. The total area of agricultural land, i.e., Cropland and Grassland,
accounted for 58%. Around 36% of the area of interest was classified as Forest Land, which
is close to the average forest area in the Czech Republic (34%). The Settlements area was
close to 5%. The remaining two categories did not exceed 1%, the Wetlands class occupied
0.77% and the Other Land occupied 0.96% of the area.
Table 3.
Areas and representations of individual LULUCF classes according to the final classification.
Class Area Percentage
Settlements 1059.8 km24.56%
Cropland 9936.4 km242.80%
Forest Land 8259.6 km235.58%
Grassland 3560.7 km215.34%
Wetlands 178.6 km20.77%
Other Land 222.1 km20.96%
Remote Sens. 2022,14, 1189 15 of 26
Remote Sens. 2022, 14, x FOR PEER REVIEW 15 of 27
heterogeneous character of the area was determined by the diverse representation of in-
dividual LULUCF classes. From the results documented in Table 3 and Figure 11, it is
clear that the Cropland class had the highest area in 2018 with more than 42% of the total
area. Grassland class occupied over 15%. The total area of agricultural land, i.e., Cropland
and Grassland, accounted for 58%. Around 36% of the area of interest was classified as
Forest Land, which is close to the average forest area in the Czech Republic (34%). The
Settlements area was close to 5%. The remaining two categories did not exceed 1%, the
Wetlands class occupied 0.77% and the Other Land occupied 0.96% of the