Content uploaded by Katarína Onačillová
Author content
All content in this area was uploaded by Katarína Onačillová on Dec 29, 2023
Content may be subject to copyright.
ACTA GEOGRAPHICA UNIVERSITATIS COMENIANAE, Vol. 67, 2023, No. 2, pp. 163-185
AUTOMATIC FOREST COVER CLASSIFICATION
USING SENTINEL-2 MULTISPECTRAL SATELLITE
DATA AND MACHINE LEARNING ALGORITHMS IN
GOOGLE EARTH ENGINE
Katarína Onačillová1,Veronika Krištofová2, Daniel Paluba3
1 Pavol Jozef Šafárik University in Košice, Faculty of Science, Institute of Geography,
e-mail: katarina.onacillova@upjs.sk
2 HRDLIČKA – SLOVAKIA s.r.o., Moyzesova 46, 040 01 Košice, Slovakia, e-mail:
veronika.kristofova@hrdlicka.sk
3 Charles University, Prague, Czech Republic, Faculty of Science, Department of
Applied Geoinformatics and Cartography, e-mail: daniel.paluba@natur.cuni.cz
Abstract: Forest cover plays an essential role in maintaining ecological equilibrium, mitigat-
ing climate change, and securing a sustainable future for both humanity and the planet. Most
countries conduct forest inventory or remote sensing surveys every few years to monitor
changes in forest cover. However, only a few initiatives offer more frequent updates, typically
weekly or monthly, focusing exclusively on areas experiencing high rates of deforestation or
those of significant ecological value. The present study focuses on the classification of forest
cover throughout Slovakia, covering the period 2017-2022, using Sentinel-2 multispectral
satellite imagery along with the machine learning (ML) algorithms Random Forest (RF) and
Support Vector Machine (SVM). The computation was performed in the cloud-based Google
Earth Engine (GEE) platform, which offers a versatile interface for a broad range of computa-
tional capabilities for geospatial analysis and landscape monitoring. Forest cover change pro-
cessing and evaluation is based on the RF classification algorithm, which demonstrated higher
accuracy than the SVM classifier. The results indicate that RF outperformed SVM by 4% and
21% in 2017 and 2020, respectively. The RF algorithm achieved an overall accuracy (OA) of
95% in both classification cases (2017 and 2020) and F1 score of up to 0.95. The selected RF
algorithm revealed an increase in forest cover in Slovakia, particularly notable during the
period 2017-2019, with a slight decrease detected between 2019 and 2020. Furthermore, it
was determined that the current forest cover is lower than that reported in official state statis-
tics and land cover databases. Additionally, a user-friendly automatic tool for forest cover
classification was developed and made freely available in GEE. This tool can benefit foresters,
urban planners, and everyday users by detecting subtle changes in forest cover, crucial for
forest sustainability and human well-being.
Keywords: Google Earth Engine, Sentinel-2, forest cover, image classification, machine
learning, Slovakia
163
1 INTRODUCTION
Forests represent an irreplaceable natural resource providing a wide range of
ecosystem services essential to the cycle of life on Earth. They play an important
role in protecting biodiversity, reducing the effects of global warming in the world
through the carbon cycle process, the hydrological cycle, mitigating soil erosion, and
human recreation (Alkama and Cescatti, 2016). Despite many essential functions,
these ecosystems have been exposed to a great danger of threats to their condition,
longevity, and quality in recent decades (Hansen et al., 2013). Since 1990, the world
has lost more than 178 million forests, which represents approximately 0.20% of the
world's forest cover, although the decline rate in net forest loss has slowed signifi-
cantly between 2010-2020 due to the decline in forest expansion (FAO, 2022). Ac-
cording to official FAO statistics, Europe is the only continent where forest cover in-
creases over a long period of time. However, in the years 2010-2020, significantly
lower values of the net increase rate in forest area were recorded in Europe (and
Asia) than in the years 2000-2010 (FAO, 2020).
Similarly, within Central Europe, according to official inventories, the territory
of Slovakia shows a permanent and long-term increase in forest cover (MŽP, 2021).
There are nine national parks dedicated to the protection of extensive ecological pro-
cesses along with species diversity and ecosystems that are specific to this area
(IUCN, 2020; Šebeň, 2017). However, as stated by the Institute for Environmental
Policy (IEP, 2017), although the area of forests is increasing according to official
data, international satellite images that show the real state of forest land and can ob-
jectively assess the condition of forests indicate that the opposite is true, and that
since the beginning of the 21st century the area of Slovak forests has decreased by
an average of 0.46% annually. The main difference lies in the differences in the
methods used and in the understanding of the forest itself.
Today, both in the world and Slovakia, there is still a noticeable inconsistency
of methodologies for monitoring forests conditions and the lack of sustainable de-
velopment strategies to resolve forest loss and degradation, while balancing the eco-
nomic, social and environmental benefits of forestry. Most countries conduct forest
inventory or remote sensing (RS) surveys only every few years to monitor changes
and forest cover quality. Several models and datasets have been developed to ana-
lyze forest cover changes, but these are usually based on low temporal resolution
discrete data and may therefore ignore key factors that influence forest changes
(Matthews et al., 2007). Therefore, there is an urgent need for models and applica-
tions that help stakeholders and the public to understand the real state, the continu-
ous spatio-temporal dynamics of forest cover, to identify key factors driving forest
change and disrupting forest ecosystems over time, to quickly intervene and reduce
the harmful, perhaps illegal, effects on forest loss and better design possible future
changes (Tang et al., 2019).
Significant potential for investigating forest stands over the past decade has
been provided by the Sentinel-1 and Sentinel-2 satellite imagery of the European
Space Agency (ESA) Copernicus program, which can effectively monitor changes,
164
even in large areas of forests with high temporal and spatial resolution, with the help
of parametric and non-parametric ML classifiers (Viana et al., 2019). The impor-
tance of these satellite data in the classification of forest types using Sentinel-2 data
was also emphasized by Immitzer et al. (2019), or using a combination of Sentinel-1
and Sentinel-2 data by Lechner et al. (2022). Thanks to the GEE tool (Gorelick et
al., 2017), which represents an innovative cloud platform with access to a catalog
with a rich collection of satellite images and other RS data, it is now possible to effi-
ciently perform geospatial analysis at regional and global level in high temporal and
spatial resolution. The use of this platform by the RS scientific community has in-
creased almost exponentially in recent years (Yang et al., 2022).
The aim of this paper was to develop an automated approach for forest cover
classification using Sentinel-2 imagery and supervised ML classification algorithms
RF and SVM in a GEE environment. A tool was developed and applied for the ter-
ritory of Slovakia and its regions for the period 2017-2022. As input data, selected
Sentinel-2 satellite bands, normalized difference vegetation index (NDVI) and the
digital elevation model SRTM (Shuttle Radar Topography Mission) were used. The
combination of the Corine Land Cover (CLC) and Global Forest Change (GFC)
databases was used to create validation training samples. The accuracy of classifica-
tion results using the RF and SVM algorithms was then assesed using the script cre -
ated, and the changes in the state of the forest cover in the selected area were evalu-
ated. The results of image classification in the GEE environment were also com-
pared to the official data obtained from the Statistical Office of the Slovak Republic
(SOSR) and to CLC data on forest area.
2 DATA AND METHODS
The analyses in this paper were performed in the GEE environment. To demon-
strate the effective use of this platform for image classification and analysis of large-
volume satellite data, even for large areas, the entire territory of Slovakia was
chosen as the area of interest. In the first step, input data were prepared, namely Sen-
tinel-2 multispectral satellite data, SRTM elevation model, CLC and GFC database.
From the Sentinel-2 dataset, images with a cloud cover below 5% were chosen for
a selected time range. The applied input bands were then enhanced with NDVI and
SRTM. From the selected time series of images, a median mosaic was created for
each pixel and each input band. A forest mask was created with the intersection of
the CLC and GFC databases, which was also used to generate 2000 randomly gener-
ated training data. In the next step, two ML algorithms, RF and SVM, were tested by
evaluating accuracy using error matrices. The algorithm with better results, RF, was
used to produce land cover maps and quantify the year-to-year changes in forest area
in the period 2017-2022 for the entire Slovakia and individual regions, and then
compared with official SOSR data. Figure 1 depicts the work's methodology. A de-
tailed description of the individual steps can be found in the following chapters.
165
Figure 1 Flowchart of the methodology used in this work. Source: processed by the
authors
2.1 Area of interest
Forest cover classification using ML algorithms and Sentinel-2 multispectral
data was carried out for the territory of Slovakia with an area of 49,036 km2 located
in Central Europe (Figure 2). Although Slovakia is one of the smaller countries in
terms of size, its geographical location and regional differences in altitude allow to
observe a variety of natural and climatic conditions, diverse ecosystems and land
cover types.
Slovakia’s forest cover today reflects long-term development and the effects of
both natural and man-made factors. With 22% and 18% of Slovakia’s total forest
area, respectively, beech and spruce forests make up the largest proportion of the
country’s forest cover. Compared to the primary forest structure in Slovakia, there is
currently a decline in species such as oaks, beeches and firs, but an increase in
spruce and pine can be observed (Minďáš et al., 2006). However, this period is also
characterized by the intentional planting of new forests, mainly spruce monocul-
tures. The State of Europe’s Forests report by Forest Europe (2020) states that Slo-
vakia ranks 13th out of 43 European countries in terms of forest cover. One of the
main current challenges of forest management is the need to face the increasing risk
of harmful factors that affect overall forest cover. The loss of forest cover due to hu-
man activity persists at the expense of the economy, and forests are frequently
turned into agricultural land, which not only negatively affects the soil and water
cycle but also reduces the ability of forests to regulate the environment. In addition,
biotic factors (fungi, insects and bacteria) and, to a large extent, abiotic factors such
as weather extremes also affect the forest. Wind calamities, which have occurred
166
frequently in Slovakia’s history and are still observed annually, have a particularly
detrimental effect on the forests.
Figure 2 Location of the area of interest within Europe. Source: ©ESRI basemap
(2022)
2.2 Input data
The main source of data for the supervised classification was Sentinel-2 multis-
pectral satellite imagery at the Level-2A preprocessing level. The Copernicus Sen-
tinel-2 mission consists of a constellation of two satellites, Sentinel-2A and Sentinel-
2B, providing optimal coverage and data for the entire planet. Due to its high spatial,
temporal (data acquired on the same location by two satellites approximately every
five days), and spectral resolution (10-60 m) this mission provide valuable informa-
tion particularly for agricultural research and forest vegetation monitoring (Drusch
et al., 2012). Sentinel-2 Level-2A data have been available in GEE since 28 March
2017, therefore images from 2017 were used in this study.
Consequently, the time period for satellite image retrieval was defined only for
the growing season, i.e. April-October, as this setting allows for a better differenti-
ation of forest cover from other land cover classes with a lower probability of cloud
and snow cover. Another important factor was setting the maximum cloud cover cri-
terion at 5%, which eliminated the input of cloud-covered scenes in image classifica-
tion that could bias the classification result. Testing and then selecting those spectral
bands that had the greatest impact on image classification also contributed to achiev-
ing high accuracy in forest cover classification. To classify forest cover, only the
bands listed in Table 1 were used.
167
Table 1 Properties of selected bands of Sentinel-2 satellite.
Band Name / description Central wavelength (µm) Spatial resolution (m)
Band 2 Visible blue 0.490 10
Band 3 Visible green 0.560 10
Band 4 Visible red 0.665 10
Band 5 Vegetation Red Edge 0.705 20
Band 6 Vegetation Red Edge 0.740 20
Band 8 Near Infrared (NIR) 0.842 10
Band 11 Short Wavelength Infrared 1 (SWIR) 1.610 20
Band 12 Short Wavelength Infrared 2 (SWIR) 2.190 20
Source: data based on Drusch et al. (2012), European Space Agency (2015)
Based on the testing of the image classification results, we discovered that
adding the SRTM and NDVI to the input feature set improves the accuracy of image
classification. Also in the study of Svoboda et al. (2022) for Czechia, the results of
the land cover classification were improved when NDVI and SRTM were included.
The NDVI is calculated as a normalized difference of visible infrared (B8) and visi-
ble red (B4) bands, using the following equation (Rouse et al. 1974):
NDVI = (B8 – B4)/(B8 + B4)
Elevation information from SRTM was added as an input feature for image
classification primarily to represent heights that affect the occurrence of forest ve-
getation. The SRTM data are provided at a resolution of 1 arcsecond (approximately
30×30 m) (Farr et al., 2007). We have used the latest version available in GEE,
SRTM90_V4, which, unlike previous versions, has been processed to fill data gaps
to ensure ease of use.
The so-called median composite was then used as input into the classification,
which was created by calculating the median value for each pixel and each input
band across Slovakia from the time series of available Sentinel-2 images for a selec-
ted period in a given year. Visual analysis showed that the period between 20 April
and 10 October yields the least cloudy composites for each year, so we used this
range in our analyses. The median composite was chosen because it is less suscep-
tible to possible outliers due to potential cloud and snow cover effects compared to
the mean composite.
2.3 Training data preparation
A sufficient number and quality of training samples are another prerequisite for
successful classification. Based on the extent of the area of interest, training plots
were created for two classification classes to validate the resulting forest cover clas-
168
sifications. Sites covered by forest vegetation were classified into the forest class,
and sites without forest cover into the non-forest class. For the creation of training
samples, we relied on the research by Paluba et al. (2021), in which a forest mask
was created with an overall accuracy of more than 90%. Specifically, we used two
freely available land cover databases in GEE for this purpose, namely the CLC by
Büttner et al. (2017) and GFC by Hansen et al. (2013). The CLC product provides
land cover layers classified into 44 classes, in the form of five updated databases for
1990, 2000, 2006, 2012 and 2018. These databases were created using satellite im-
age classification and in-situ data, and use a minimum mapping unit (MMU) of
25 hectares for areal phenomena and a minimum width of 100 m for linear phenom-
ena. In this work, the most up-to-date CLC layer (CLC2018) was used, of which
3 classes were selected to represent forest cover: deciduous (311), coniferous (312)
and mixed forest (313). Forests are defined in the CLC as forest vegetation covers
taller than 5 m with a minimum canopy cover of 30%. For young plantations, a min-
imum threshold of 500 subjects per hectare is considered (Kosztra, 2019).
Another database used to generate the training data was the GFC global-scale
database, which is the result of a time series analysis of Landsat imagery and de-
scribes forest cover, loss and gain at a spatial resolution of 30 m. Similar to the CLC
database, trees are defined in this database as vegetation taller than 5 meters. Forest
loss is defined in the GFC database as a disturbance of the forest area, where the
forest-covered area has changed to an area without forests. This loss should repre-
sent a change of at least 50% of the canopy cover at pixel level. Forest gain is de-
fined as the exact opposite of the forest loss. It is the phenomenon where a non-
forest area is converted to forest (Hansen et al., 2013). For the purpose of this work,
GFC version 1.10 (Hansen et al., 2023), containing data from 2000 to 2022, was used.
We selected pixels with a tree cover greater than 50% using the tree cover layer
from 2000 (treecover2000) from the GFC database. Subsequently, the treecover2000
layer was masked using the layer representing forest cover loss for each year
(lossyear) to produce the resulting GFC layer. The final forest mask (with two
classes: “forestˮ and “non-forestˮ) was created using the intersection of the resulting
GFC layer with CLC2018. The “non-forestˮ class contained all types of land cover
except forest. Using the intersection of the two databases, i.e. selecting pixels that
belong to the “forestˮ category in both databases, helps to reduce errors that may ap-
pear in each database.
In the next step, the training dataset was created by generating 2000 random
points. The values of all input bands were extracted to each point. Using the final
forest mask, it was automatically determined for all points whether they represented
forest or non-forest area. This is an automatic process whose advantage is that the
input parameters for the classification can be changed quickly and efficiently; that is,
the GFC forest loss layer can be used for any year period since 2000.
2.4 Classification algorithm selection
To select an appropriate classification algorithm, it was necessary to review all
factors that influence the classification. These include the spatial resolution of the
169
RS data, the quality and number of training samples, availability of classifiers in the
selected environment and the source of the reference data (Lu and Weng, 2007).
The GEE platform has a variety of classification algorithms that can be imple-
mented according to the user’s needs. To obtain relevant results, two supervised ML
classification algorithms, RF and SVM, were used and compared. Since the begin-
ning of the 21st century, the SVM algorithm has represented a possible alternative
with higher classification accuracy compared to the numerous algorithms used until
then (Pal and Mather, 2005). It is a nonparametric classification algorithm that can
achieve good classification results even from complex data that contain noise
(Chuvieco, 2020) and also when using fewer training data (Mantero et al., 2005).
The RF classification algorithm operates on the basis of building multiple decision
trees using a randomly selected subset of training samples and variables (Belgiu and
Drăguţ, 2016). It is becoming increasingly used for its high accuracy, reliability and
efficiency in processing high-dimensional and non-linearly distributed data. For ex-
ample, it is often used to map forest stands and complex ecosystems (Waśniewski et
al., 2020; Li et al., 2022), arable land (Phalke et al., 2020), mountainous areas (Hoś-
ciło and Lewandowska, 2019), and has also shown its potential in simulating water
salinity (Khan et al., 2020).
In GEE, RF number of trees was set due to the computational speed to 10,
which is the default setting of this classifier in the GEE environment. However, the
user should keep in mind, that setting greater number of trees tends to stabilize the
out-of-bag error, thus also the resulting classification accuracy, while being compu-
tationally more expensive. The other parameters were left to the default, while SVM
was used with the default parameters, i.e., with a linear kernel. Since no other kernel
setting could efficiently classify the forest cover of the entire territory of Slovakia in
the GEE, the linear kernel for the SVM was selected. This is due to the computa-
tional limitation of GEE – a large area is associated with a greater computational
complexity of other types of kernels – this in the free GEE version caused that other
types of kernel were out of the computational capacity.
2.5 Accuracy assessment
Evaluation of classification accuracy was particularly important to determine
a more reliable classifier for forest cover classification. The accuracy assessment
was performed separately for 2017 and 2020 classification results and for the two
classification algorithms used. We selected 2017 as the first year with Sentinel-2
data available in GEE and also because the CLC2018 layer was built on 2017 data,
so the accuracy of CLC2018 should be most accurate in this year. The year 2020
was chosen due to the noticeable high forest loss recorded between 2017 and 2020
in the Veľká Fatra and Malá Fatra National Parks (MŽP, 2021).
For each chosen year, two independent reference datasets generated based on
the high-resolution open-access satellite imagery in Google Earth Pro (GEP) were
used to verify the classification’s accuracy. The reference dataset consisted of 200
reference points, representing 10% of the training dataset, with 100 points represent-
ing the ‘forest’ class and another 100 points representing the non-forest class. To
170
validate the reference points for 2017, GEP mosaics consisting mostly of imagery
from 2017 were used. The reference points for 2020 were validated using GEP mo-
saic consisting mostly of imagery from 2020. The reference datasets for both years
were exported in.kml format, further converted to shapefile (.shp) and then imported
into the GEE environment, where the classifications’ accuracy of the algorithms
were evaluated and expressed in terms of error matrix and metrics such as overall
accuracy (OA), User’s Accuracy (UA), Producer’s accuracy (PA) and kappa coeffi-
cient. In addition, the overall F1 score was computed according to the formula by
Long et al. (2015):
F1−score=2∗precision∗recall
precision+recall
.
3 RESULTS AND DISCUSSION
3.1 Evaluation of the accuracy of the SVM and RF algorithms
First, the accuracy of classifications using the SVM and RF algorithms was
compared for 2017 and 2020. Classification outputs are shown in Figure 3. Accura-
cy metrics of both algorithms using error matrices for both years are displayed in the
Table 2.
Figure 3 Forest classification using SVM and RF classification algorithms for 2017
and 2020. Source: prepared by the authors
171
Table 2 Error matrix for image classification using the SVM and RF classifiers for 2017-2022
SVM 2017
Class Reference data Total UA(%)
Non-forest Forest
Non-forest 91 8 99 91.92
Forest 9 92 101 91.09
Total 100 100 200
PA (%) 91 92 0.91 OA
F1-score 0.91 Kappa 0.83
RF 2017
Class Reference data Total UA(%)
Non-forest Forest
Non-forest 98 9 107 91.59
Forest 2 91 93 97.85
Total 100 100 200
PA (%) 98 91 0.95 OA
F1-score 0.95 Kappa 0.89
SVM 2020
Class Reference data Total UA(%)
Non-forest Forest
Non-forest 96 48 144 66.67
Forest 4 52 56 92.86
Total 100 100 200
PA (%) 96 52 0.74 OA
F1-score 0.79 Kappa 0.48
RF 2020
Class Reference data Total UA(%)
Non-forest Forest
Non-forest 100 11 111 90.09
Forest 0 89 89 100
Total 100 100 200
PA (%) 100 89 0.95 OA
F1-score 0.95 Kappa 0.89
Source: processed by authors
The OA of the four evaluated classifications using the SVM and RF algorithms
varied from 0.74% to 0.95%. The values of the Kappa coefficient ranged from
0.48% to 0.89%. The RF algorithm achieved higher overall accuracy and behaved
approximately equally stably, showing an OA of 95% in both classification cases
(2017 and 2020). For the forest class, the RF results in 2017 and 2020 produced
PA’s of 91% and 89%, respectively.
Although the OA using RF algorithm is high, the forest cover layers show some
EOs (e.g. non-contiguous areas of the forest) and commission (e.g. dense crops,
172
grasslands or dwarf mountain pine classified as forest). Similar results were obtained
by Waśniewski et al. (2021), in whose study RF achieved the lowest accuracy in
places where the forest was disturbed (UA = 61% and PA = 71%) and also in the
study by Conette et al. (2016) where RF showed the lowest accuracy in classifying
the degraded forest class, where up to 24% of forest reference points were misclassi-
fied as non-forest classes.
Using the same training samples, the SVM classification algorithm showed sig-
nificant classification inaccuracy for 2020 (OA of 74%, Kappa of 0.48) compared to
the RF algorithm. The SVM algorithm’s classification results inaccurately expressed
significant differences in the forest cover. These errors consisted of incorrectly clas-
sified areas, e.g. water bodies or meadows and pastures, for example, in the vicinity
of Liptovská Mara and Starina, which the SVM mistakenly classified as “forestˮ
(Figure 4), as can be seen in comparisons of true colour (RGB: B4-B3-B2) and false
colour compositions (CIR: B8-B4-B3) from Sentinel-2 satellite imagery and results
for image classification using SVM and RF classifier. The higher classification error
can also be attributed to the fact that SVM was used with the so-called linear kernel.
The choice of this SVM was conditioned by the fact that only this specific type of
kernel worked successfully in GEE for the entire territory of Slovakia.
The classification using SVM for 2020 revealed 1.085 million. ha, that is, by
more than 1 mil. ha of forest lower status compared to the RF classification. In con-
trast, according to the RF algorithm, there was an increase in forest cover in Slova-
kia by almost 24 thousand ha (from 1.810 to 1.834 million ha).
Considering the accuracy assessment, the more reliable was RF algorithm, since
it worked with higher accuracy than SVM in both cases. Due to its reliability, the RF
algorithm has been preferred in many previous studies. In the work of Li et al.
(2022), RF demonstrated a strong potential for forest vegetation classification with
an OA of 97.57% and a Kappa value of 0.95. OA equal to 95%, similar to our case,
was also achieved by the RF algorithm in the study of Nomura and Mitchard (2018),
who monitored tropical forest vegetation using Sentinel-2 satellite images.
This accuracy assessment method has its limitations. The number of reference
points set at 200 is not so representative of the entire territory of the Slovak Repub -
lic. To enhance the model’s generalization, a greater quantity of testing points
should be used in the next study for a more comprehensive evaluation.
3.2 Evaluation of changes in forest area in Slovakia
In the further analysis, we exclusively used the outcomes of a more reliable
classification using the RF algorithm. To determine the state of forests, we created
a layer from the image classification – a mask, which contained only the “forestˮ
class. Subsequently, with the help and flexibility of the created tool, we were able to
determine the exact classified forest area and monitor changes at the level of the en-
tire Slovakia, but also at the regional level. However, the calculation of the forest
area can be applied to any time period and territory within Europe, the GEE environ-
ment also offers a number of datasets containing boundary layers (e.g. FAO district
boundaries, boundaries of protected areas NATURA 2000, etc.) and the import of
173
own data that can also be used for the classification of areas of interest and the cre-
ation of training samples, even for places that are difficult to access or too large for
in-situ data collection.
Figure 4 SVM classification results compared to RF for the areas of Liptovská Mara
and Starina. From the left: true colour composition (RGB: B4-B3-B2), false colour
composition (CIR: B8-B4-B3) (both converted to greyscale), SVM classification
results, RF classification results. Source: background colour compositions of
Sentinel-2 images from 08/22/2020, processed by the authors
Based on the RF classification results, we can observe the increase in the forest
area between 2017 and 2018, by 17,890.48 ha, i.e. 0.99% (Table 3, Figure 5A). The
largest increase occurred in 2018-2019 (24,831.68 ha). In 2019, one of the highest
levels of forest cover in the monitored period was observed, 1,852,785.6 ha. The de-
crease by 19,285.11 ha can be observed from 2019 to 2020, resulting in a decrease
in forest cover by 1.04% from 2019. In the following years 2021-2022, we can ob-
serve an increase in forest area again, with the year 2022 marking the end of our
monitoring period and exhibiting the highest forest cover during the 2017-2022
period. This increase can be explained by the effective elimination of calamity wood
(a remnant of past wind disasters) and bark beetles. Comparing the beginning and
end of the examined period (years 2017 and 2022), the largest gains in the forest
174
cover can be observed southwest of the town Brezno, east of the city Nové Mesto
nad Váhom and Trenčín, and also in the area of the High Tatras National Park. On
the other hand, the highest losses are visible in the area of Slovak Paradise National
Park, west of the city of Košice and near the Prešov city and Prievidza (Figure 5B).
Table 3 Forest area in Slovakia for the period 2017-2022 based on the RF classification
Year Forest area [ha]
2017 1,810,063.44
2018 1,827,953.92
2019 1,852,785.60
2020 1,833,500.49
2021 1,844,562.41
2022 1,855,912.20
* the median composite was calculated for the period April 1 - October 31 due to the higher effects of
cloudiness in the original range.
Source: processed by authors
These classification results are accessible also through the forest cover database
we created for the entire territory of Slovakia covering all years from 2017 to 2022.
The database stored in raster format has a spatial resolution of 30 m and is freely
available as a GEE Image Collection and can be imported using the following code:
ee. ImageCollection('users/danielp/Slovakia_forests_2017-2022').
3.3 Comparison of the classification results with official
SOSR data on forest land
This section compares the results of our RF classification with the official
SOSR data on forest land for each region and the entire SR during the monitored
period (Table 4).
Compared to the SOSR data, the forest cover value calculated from the RF clas-
sification was lower in all cases. SOSR reports that the forest cover in Slovakia
gradually increased during the monitoring period, by 0.05% in total. Our RF classi-
fication shows that the amount of forest cover increased between 2017 and 2019, de-
creased slightly in 2020, and then increased again at the end of the monitoring
period. In 2022, according to the RF classification, forest areas accounted for
37.86% of Slovakia’s total area – approximately 4% less than according to SOSR
data. According to the RF classification results, forest cover increased by 0.95%
between 2017 and 2022, with the highest increase in 2017-2019 (0.88%).
175
Figure 5 A) Change in forest cover in Slovakia according to RF classification during
the period 2017-2022. B) The largest areas of forest loss and forest gain – difference
raster for years 2017 and 2022. Source: processed by the authors
The reason for the different forest cover values between the RF classification
results and SOSR data is primarily due to the fact that the official SOSR data actu-
ally provide information on forest land, not forest cover. As forest land, the SOSR
define an area that is used for forest production that can also be temporarily defores-
ted but restored and takes into account also areas that indirectly fulfill forestry func-
tions, such as forest roads, warehouses and nurseries, as yet unforested land that was
taken from the agricultural land fund and subsequently assigned to forestry (SOSR,
176
2022). The input of our classification was freely available satellite images that con-
trary to SOSR data, provides objective information on the actual extent of the forest
cover. Naturally, even the results from RS data work with a certain degree of accura-
cy (95% in our case), which can be affected by snow and cloud cover, or other phe-
nomena that prevent effective analysis. However, we also must draw attention to the
spatial resolution of our results (30 m), due to which it is not possible to detect forest
areas below 0.1 ha (900 m2).
Table 4 Forest cover of Slovakia and its regions according to the RF classification and SOSR
data
Forest cover for the monitored year (%)
Region 2017 2018 2019 2020 2021 2022
SOSR RF SOSR RF SOSR RF SOSR RF SOSR RF SOSR RF
Bratislava 36.59 29.97 36.59 30.16 36.59 30.08 36.58 30.36 36.59 28.83 36.58 29.07
Trnava 15.79 11.85 15.77 12.48 15.77 12.46 15.77 12.45 15.77 12.24 15.77 12.67
Trenčín 49.95 46.36 49.43 45.73 49.45 47.10 49.46 46.98 49.46 46.71 49.46 46.72
Nitra 15.28 12.52 15.28 12.85 15.29 13.00 15.29 12.82 15.29 12.83 15.29 13.18
Žilina 56.36 49.09 56.39 46.91 56.34 46.01 56.35 46.16 56.39 45.69 56.41 45.32
Banská
Bystrica 49.40 47.94 49.43 46.98 49.47 45.96 49.49 47.22 49.50 46.24 49.52 46.11
Prešov 49.53 47.02 49.65 46.51 49.74 46.42 49.77 47.60 49.80 44.56 49.82 47.00
Košice 39.85 37.53 39.87 35.50 39.88 35.03 39.92 35.92 39.93 35.79 39.92 37.00
SR 41.33 36.91 41.32 37.28 41.34 37.79 41.36 37.36 41.37 37.62 41.38 37.86
Source: RF data processed by the authors, SOSR (2017-2022)
Thanks to effective calculation of forest cover with the help of the developed
tool, it is also possible to monitor the forest cover e.g. at the level of individual re -
gions. Forest cover is expressed by the share of the area represented by forests in the
total area of the area area under study – the region (NUTS 3). The district layer was
added using the dataset FAO GAUL (2015) (note that the FAO GAUL boundary
area layer may differ slightly from the district boundary layer used by SOSR).
The largest share of the total forest area in 2022 was concentrated in the Banská
Bystrica (432,206.63 ha) and Prešov region (427,956.76 ha). We observed a higher
than 40% share of forests in the regions of Banská Bystrica and Prešov for the entire
2017-2022 period. The highest forest coverage with 49.09% was observed in the
Žilina region in 2017. Six out of nine national parks in Slovakia are located on the
territory of these three regions. Throughout the analyzed period, the lowest forest
cover was observed in the Trnava region in 2017 (11.85%). The highest forest area
gain (by 4,269.48 ha) was observed between 2017 and 2022 in the Nitra region,
while the largest decrease occurred in the Žilina region (by 25,888.38 ha). MŽP SR
(2021) states that large deforestation occurred in this region (Žilina) between 2017-
2020 with the greatest forest loss manifested in the national parks of Veľká Fatra
177
and Malá Fatra. Additionally, the MŽP SR emphasized the high rate of forest loss in
the area of Kysuce.
3.4 Comparison of classification results with the CLC data-
base
To compare and highlight the potential of the developed forest research method,
we compared the classification results with all the CLC classes containing informa-
tion on forests (classes 311, 312, 313). These classes were then combined into a uni-
form mask for the forest, and the forest area and the percentage of forest cover in
Slovakia was calculated. Since the CLC2018 database is mainly composed of 2017
satellite imagery and the availability of Sentinel-2 data in the GEE archive begins
from 2017, we used the RF classification for the first available year, 2017, for com -
parison (Table 5).
Table 5 Comparison of the RF classification results with the CLC database
Data Forest area (ha) Forest cover (%)
RF (2017) 1,810,063.44 36.89
CLC (2018) 2,053,389.72 41.84
Difference 243,326.28 4.96
Source: RF data processed by the authors, CLC (2018)
Figure 6 illustrates the differences between the CLC data and the generated re-
sults of the RF classification for the forest area and forest cover. The difference in
values can be mainly attributed to the resolution of the input data entering the classi -
fications (MMU 25 ha for CLC2018 and 10×10 m, i.e. 0.01 ha for our results) and
the accuracy of the maps (Figure 7).
Figure 6 Forests of Slovakia according to the CLC2018 classification (mainly based
on satellite data from 2017) and the classification using the RF algorithm. Source:
CLC (2018), RF classification processed by the authors
178
Figure 7 Forest area of RF classification compared to CLC2018. Source: CLC
(2018), colour compositions of Sentinel-2 images from 28 May 2017 – processed
by the authors
The created RF classification classified only areas that are covered by forest ve-
getation, while the CLC also listed areas that are not covered by forest vegetation as
forests due to the high MMU. According to CLC data, the forest area was
243,326.28 ha larger than according to the RF classification. This is also related to
a higher forest coverage according to CLC.
The CLC database provides sufficient information for the creation of training
areas, but the spatial and temporal resolution of this inventory is insufficient to
monitor forest status. With this comparison, we wanted to point out the potential of
the image classification method used in the GEE cloud environment and the high
performance of the RF algorithm and Sentinel-2 satellite images, which provided us
with detailed information about the forest cover and could be used for detailed ana-
lyses.
179
3.5 A tool for automatic forest cover classification in GEE
In this paper, we also created a GEE tool that enables automatic forest cover
classification using Sentinel-2 satellite imagery. A great advantage of this tool and
the GEE platform is the possibility of changing the input parameters and modifying
individual parts of the code by the users for their own needs. The tool was created
using JavaScript using GEE’s Code Editor interface and is available in the GitHub
repository via the link:
https://github.com/palubad/Automatic-Forest-Classification-GEE
In GEE:
https://code.earthengine.google.com/c2f07a9161037480b5fbf8f11a6acaf2
In the first step, the user of the tool chooses the area of interest for automatic
forest cover classification – the user can select any country or lower administrative
unit in the EU from the LSIB 2017 database or create its own area in GEE. As train-
ing data are prepared using the CLC database, which only covers EU countries, the
selection is limited to EU territory. In the next step, the year of analysis is selected.
Currently, the years 2017-2022 can be selected due to the availability of data in the
GFC database. The range of months for the input data for the median composite can
also be specified. Another advantage of the created tool is the possibility to choose
the upper limit of cloud cover for Sentinel-2 scenes and any number of training
samples to be used in image classification. Advanced users can also modify other
parts of the code, e.g. input bands, spectral indices (e.g. NDVI) and additional inputs
to the classification (e.g. SRTM).
In the next parts, the methodology described in this paper is carried out, i.e.,
automatic creation of training dataset using the intersection of the CLC2018 and
GLC databases, supervised classification using RF using the generated training data.
The following layers are displayed in the map window: a median composite of the
Sentinel-2 time series as an RGB composite, the intersection of the GLC and
CLC2018 databases, and the resulting classification for the entire area of interest.
The classified result can be exported as a so-called asset to GEE or downloaded as
a GeoTiff to Google Disk via the “Tasksˮ tab on the right.
All the codes used in this work, including the comparison of the classification
algorithms, the accuracy assessment and the validation points, are available in the
mentioned GitHub repository.
3.6 Direction of further research
The following steps will focus on the thorough verification of the obtained out-
comes and the implementation of the created technique beyond Slovakian borders.
Application of sophisticated methods to mask out clouds in Sentinel-2 imagery, such
as S2Cloudless (Zupanc, 2017) or Cloud Score+ (Pasquarella et al., 2023), would
probably contribute to achieving even more accurate results. Another goal is to im-
prove the method for automatic forest cover classification outside the EU by using
global land cover databases, such as Copernicus Global Land Cover Layers (Buch-
horn et al., 2020), ESA WorldCover (Zanaga et al., 2022), etc. It will be important to
180
focus not only on the binary classification of forest/non-forest, but also on the classi-
fication of forest types. The high potential for forest monitoring is currently also in
the use of freely available Sentinel-1 radar satellite data, which could be used in fu-
sion with Sentinel-2 multispectral data.
4 CONCLUSIONS
Forests are important prerequisites for preserving biodiversity and protecting
the climate, soil and water resources of every country. Preventive measures, such as
monitoring forest conditions and early detection of factors disrupting forest cover,
are necessary to protect forest ecosystems. Nowadays, the image classification using
advanced RS methods, especially multispectral satellite data, offers a fast and effec-
tive way of obtaining information on the condition and changes in land cover.
The aim of this paper was to analyze changes in forest cover in Slovakia using
methodology developed applying Sentinel-2 satellite data and ML algorithms in the
GEE environment. In the article, we pointed out the potential of combining freely
available Sentinel-2 imagery with other freely available data, such as SRTM,
CLC2018 and the GFC database.
The main product of our work is the forest cover classification for Slovakia,
while the developed tool for automatic classification of the forest cover can be easily
applied for individual regions or districts of Slovakia, or modified for any other area
within the EU. The ML algorithms RF and SVM were chosen for image classifica-
tion, while the RF algorithm showed higher reliability with OA equal to 95% for
both validation years. In a further analysis using the RF algorithm, we found that the
most forest cover increased in Slovakia in the period 2018-2019 (by 24,831.68 ha).
On the contrary, there was a 1.04% decrease in the amount of forest cover in 2019-
2020. On a regional level, the highest levels of forest cover were observed in the
Žilina, Banská Bystrica and Prešov region.
In comparison to official SOSR and CLC2018 data, the forest cover classifica-
tion using ML algorithm RF and Sentinel-2 satellite data in GEE demonstrated high
classification accuracy and efficiency in monitoring forest cover changes in high
temporal and spatial resolution. We have made the mentioned procedure of auto-
matic forest cover classification in the form of a GEE code freely available so that it
can be applied to any other territory within Europe, from a local to global scale. In
addition, the forest cover raster layers of Slovakia for 2017-2022 created in this
work are also freely available in GEE.
Acknowledgements
This research was funded by the Slovak Academy of Sciences (VEGA) under the
contract no. VEGA 1/0085/23 “Modeling urban heat islands using geospatial
tools”, and by the Charles University Grant Agency (GAUK): “Evaluation of forest
disturbances and recovery using radar and optical satellite data in Czechia”, nr.
412722.
181
References
ALKAMA, R., CESCATTI, A. 2016. Biophysical climate impacts of recent changes in global
forest cover. Science, 351, 600-604.
BELGIU, M., DRĂGUŢ, L. 2016: Random forest in remote sensing: A review of applications
and future directions. ISPRS Journal of Photogrammetry and Remote Sensing, 114, 24-31.
BUCHHORN, M., SMETS, B., BERTELS, L., ROO, B. D., LESIV, M., TSENDBAZAR, N.
E., HEROLD, M., FRITZ, S. 2020. Copernicus Global Land Service: Land Cover 100m:
collection 3: epoch 2017: Globe (V3.0.1) [Data set]. Zenodo. https://doi.org/10.5281/zen-
odo.3518036
BÜTTNER, G., KOSZTRA, B., SOUKUP, T., SOUSA, A., LANGANKE, T. 2017.
CLC2018 Technical Guidelines. Wien: European Environment Agency.
CONETTE, G., OSWALD, P., SOGNER, M., LEIMGRUBER, P. 2016: Mapping distinct
forest types improves overall forest identification based on multi-spectral landsat imagery
for myanmar’s Tanintharyi region. Remote Sensing, 8, 882.
DRUSCH, M., BELLO, U. D., CARLIER, S., COLIN, O., FERNANDEZ, V., GASCON, F.,
HOERSCH, B., ISOLA, C., LABERINTI, P., MARTIMORT, P., MEYGRET, A., SPOTO,
F., SY, O., MARCHESE, F., BARGELLINI, P. 2012. Sentinel-2: ESA's Optical High-Res-
olution Mission for GMES Operational Services. Remote Sensing of Environment, 120, 25-
36.
EUROPEAN SPACE AGENCY, 2015. Sentinel-2 User Handbook. [online] [cit. 2023-12-06].
Available at: <https://sentinel.esa.int/documents/247904/685211/sentinel-2_user_hand-
book>
FAO, 2020. Global Forest Resources Assessment 2020. Rome: FAO.
FAO, 2022. The State of the World’s Forests 2022. Rome: FAO.
FAO GAUL, 2015. The Global Administrative Unit Layers (GAUL). Available at: <https://de-
velopers.google.com/earth-engine/datasets/catalog/DataLicenseGAUL2015.pdf>
FARR, T. G., ROSEN, P. A., CARO, E., CRIPPEN, R. et al. 2007. The Shuttle Radar Topo-
graphy Mission. Reviews of Geophysics, 45, RG2004.
FOREST EUROPE, 2020. State of Europe’s Forests 2020. Bratislava: Ministerial Conference
on the Protection of Forests in Europe. [online] [cit. 2023-06-21]. Available at: <https://for-
esteurope.org/wp-content/uploads/2016/08/SoEF_2020.pdf>
GORELICK, N., HANCHER, M., DIXON, M., ILYUSHCHENKO, S., THAU, D., MOORE,
R., 2017. Google Earth Engine: Planetary-Scale Geospatial Analysis for Everyone. Remote
Sens. Environ., 202, 18-27.
HANSEN, M. C., POTAPOV, P. V., MOORE, R., HANCHER, M., TURUBANOVA, S. A.,
TYUKAVINA, A., THAU, D., STEHMAN, S. V., GOETZ, S. J., LOVELAND, T. R.,
KOMMAREDDY, A., EGOROV, A., CHINI, L., JUSTICE, C. O., TOWNSHEND, J. R.
G., 2023. High-Resolution Global Maps of 21st-Century Forest Cover Change. Science,
342, 850-853.
HANSEN, M., POTAPOV, P., MARGONO, B., STEHMAN, S., TURUBANOVA, S., TY -
UKAVINA, A., 2013. User Notes for Version 1.6 of Global Forest Change 2000–2022.
[online] [cit. 2023-08-20]. Available at: <https://developers.google.com/earth-engine/data-
sets/catalog/UMD_hansen_global_forest_change_2022_v1_10>
HOŚCIŁO, A., LEWANDOWSKA, A., 2019. Mapping Forest Type and Tree Species on
a Regional Scale Using Multi-Temporal Sentinel-2 Data. Remote Sensing, 11, 929.
IEP, 2017. V národných parkoch bola strata lesa oproti iným územiam dvojnásobná. Bratis-
lava: Ministerstvo životného prostredia SR.
IMMITZER, M., NEUWIRTH, M., BÖCK, S., BRENNER, H., VUOLO, F., ATZBERGER,
C., 2019. Optimal Input Features for Tree Species Classification in Central Europe Based
on Multi-Temporal Sentinel-2 Data. Remote Sensing, 11, 2599.
182
IUCN, 2020. Protected Areas. [online] [cit. 2023-05-07]. Available at: <https://
www.iucn.org/theme/protected-areas/about/protected-areas-categories/category-ii-national-
park>
KHAN, M. A., SHAH, M. I., JAVED, M. F., KHAN, M. I., RASHEED, S., EL-
SHORBAGY, M. A., 2022. Application of random forest for modelling of surface water sa-
linity. Ain Shams Engineering Journal, 13, 101635.
KOSZTRA, B., BÜTTNER, G., HAZEU, G., ARNOLD, S. 2019. European topic centre on
urban, land and soil systems: Updated CLC illustrated nomenclature guidelines. Kodaň:
European Environment Agency.
LECHNER, M., DOSTÁLOVÁ, A., HOLLAUS, M., ATZBERGER, C., IMMITZER, M.,
2022. Combination of Sentinel-1 and Sentinel-2 Data for Tree Species Classification in
a Central European Biosphere Reserve. Remote Sening, 14, 2687.
LI, R., FANG, P., XU, W., WANG, L., OU, G., ZHANG, W., HUANG, X., 2022. Classifying
Forest Types over a Mountainous Area in Southwest China with Landsat Data Composites
and Multiple Environmental Factors. Forests, 13, 135.
LONG, J., SHELHAMER, E., DARNELL, T. 2015. Fully convolutional networks for se-
mantic segmentation. In: Proceedings of the IEEE Conference on Computer Vision and
Pattern Recognition, 3431-3440.
LU, D., WENG, Q., 2007. A survey of image classification methods and techniques for im-
proving classification performance. International Journal of Remote Sensing, 28, 823-870.
MANTERO, P., MOSER, G., SERPICO, S. B., 2005. Partially Supervised classification of
remote sensing images through SVM-based probability density estimation. IEEE Transac-
tions on Geoscience and Remote Sensing, 43, 559-570.
MATTHEWS, R. W., ROBERTSON, K., MARLAND, G., MARLAND, E., 2007. Carbon in
wood products and product substitution. Forestry and climate change, 91-14.
MINĎÁŠ, J., KONÔPKA, J., NOVOTNÝ, J., JENDEK, S., 2006. Lesy Slovenska. Zvolen:
Národné lesnícke centrum.
MŽP, 2021. Správa o stave životného prostredia Slovenskej republiky v roku 2021. Bratislava:
Ministerstvo životného prostredia Slovenskej republiky.
NOMURA, K., MITCHARD, E. T. A., 2018. More Than Meets the Eye: Using Sentinel-2 to
Map Small Plantations in Complex Forest Landscapes. Remote Sensing, 10, 1693.
PAL, M., MATHER, P. M., 2005. Support vector machines for classification in remote sens -
ing. International Journal of Remote Sensing, 5, 1007-1011.
PALUBA, D., LAŠTOVIČKA, J., MOURATIDIS, A., ŠTYCH, P. 2021. Land Cover-Spe-
cific Local Incidence Angle Correction: A Method for Time-Series Analysis of Forest Eco-
systems. Remote Sensing, 13, 1743.
PASQUARELLA, V. J., BROWN, C. F., CZERWINSKI, W., RUCKLIDGE, W. J., 2023.
Comprehensive Quality Assessment of Optical Satellite Imagery Using Weakly Supervised
Video Learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and
Pattern Recognition, 2124-2134.
PHALKE, A., OZDOGAN, M., THENKABAIL, P., ERICKSON, T., GORELICK, N., 2020.
Mapping croplands of Europe, Middle East, Russia, and Central Asia using Landsat 30-m
data, machine learning algorithms and Google Earth Engine. ISPRS Journal of Photogram-
metry and Remote Sensing, 167, 104-122.
ROUSE, J. W., HAAS, R. H., SCHELL, J. A., DEERING, D. W., 1974. Monitoring vegeta-
tion systems in the Great Plains with ERTS, In: Freden, S. C., Mercanti, E. P., Becker, M.
(eds.) Third Earth Resources Technology Satellite–1 Symposium. Volume I: Technical
Presentations, NASA SP-351. Washington, D. C.: NASA, 309-317.
SVOBODA, J., ŠTYCH, P., LAŠTOVIČKA, J., PALUBA, D., KOBLIUK, N., 2022. Ran-
dom Forest Classification of Land Use, Land-Use Change and Forestry (LULUCF) Using
Sentinel-2 Data – A Case Study of Czechia. Remote Sens., 14, 1189.
STATISTIC OFFICE SR 2022. [online] [cit. 2022-03-08]. Available at: <https://slovak.stat-
istics.sk>
183
ŠEBEŇ, V., 2017. Národná inventarizácia a monitoring lesov Slovenskej republiky 2015-
2016. Lesnícke štúdie, 65, 43.
TANG, X., BULLOCK, E. L., OLOFSSON, P., ESTEL, S., WOODCOCK, C. E., 2019. Near
real-time monitoring of tropical forest disturbance: New algorithms and as-sessment frame -
work. Remote Sensing of Environment, 224, 202-218.
VIANA, C. M., GIRÃO, I., ROCHA, J., 2019. Long-Term Satellite Image Time-Series for
Land Use/Land Cover Change Detection Using Refined Open Source Data in a Rural Re-
gion. Remote Sensing, 11, 1104.
WAŚNIEWSKI, A., HOŚCIŁO, A., ZAGAJEWSKI, B., MOUKÉTOU-TARAZEWICZ, D.,
2020. Assessment of Sentinel-2 Satellite Images and Random Forest Classifier for Rain-
forest Mapping in Gabon. Forests, 11, 941.
YANG, L., DRISCOL, J., SARIGAI, S., WU, Q., CHEN, H., LIPPITT, C. D., 2022. Google
Earth Engine and Artificial Intelligence (AI): A Comprehensive Review. Remote Sens., 14,
3253.
ZANAGA, D., VAN DE KERCHOVE, R., DAEMS, D., DE KEERSMAECKER, W.,
BROCKMANN, C., KIRCHES, G., WEVERS, J., CARTUS, O., SANTORO, M., FRITZ,
S., LESIV, M., HEROLD, M., TSENDBAZAR, N.-E., XU, P., RAMOINO, F., ARINO,
O., 2022. ESA WorldCover 10 m 2021 v200 (Version v200) [Data set]. Zenodo. https://
doi.org/10.5281/zenodo.7254221
ZUPANC, A., 2017. Improving Cloud Detection with Machine Learning. [online] [cit. 2023-
10-25]. Available at: <https://medium.com/sentinel-hub/improving-cloud-detection-with-
machine-learning-c09dc5d7cf13>
Automatická klasifikácia lesnej pokrývky pomocou multispektrálnych
satelitných dát Sentinel-2 a algoritmov strojového učenia v Google
Earth Engine
Súhrn
S rastúcimi hrozbami a tlakom na lesné zdroje v posledných desaťročiach začína
dochádzať k strate a degradácii lesov. Úbytok lesov je celosvetovým fenoménom
a jedným z kľúčových faktorov globálnej zmeny klímy. V súčasnosti sa preto v re-
gionálnom, ale aj celosvetovom meradle, zvyšujú požiadavky na získavanie včas-
nejších a presnejších informácií o stave lesa, jeho fungovaní a udržateľnosti, pričom
na monitorovanie časopriestorových zmien lesa sa využívajú v posledných desaťro-
čiach aj metódy diaľkového prieskumu Zeme.
Z hľadiska lesnatosti sa Slovenská republika radí medzi popredné krajiny v Európe.
Avšak aj lesy na tomto území čelia v posledných rokoch rastúcemu riziku škodli-
vých činiteľov. K poklesu lesnatosti dochádza vplyvom človeka, ktorý premieňa les
na poľnohospodárske plochy za účelom hospodárstva, ale aj vplyvom biotických
a abiotických faktorov, akými sú extrémy počasia a veterné kalamity.
Cieľom tohto článku bolo vyvinúť automatizovaný postup klasifikácie lesnej po-
krývky pomocou snímok Sentinel-2 a klasifikačných algoritmov Random Forest
(RF) a Support Vector Machine (SVM) v prostredí Google Earth Engine (GEE).
Pre monitorovanie zmien lesa na území Slovenska a v jeho krajoch v časovom ob-
dobí rokov 2017-2022 bol vytvorený a aplikovaný skript. Ako vstupné údaje boli
použité vybrané pásma družice Sentinel-2, normalizovaný diferenčný vegetačný in-
dex NDVI a digitálny výškový model SRTM. Na vytvorenie validačných trénova-
cích vzoriek bola použitá kombinácia databáz Corine Land Cover (CLC) a Global
Forest Change (GFC). Následne boli porovnané výsledky klasifikácie s použitím
klasifikačných algoritmov riadenej klasifikácie RF a SVM. Pomocou vytvoreného
184
skriptu sa vyhodnotila správnosť klasifikácií a vyhodnotili sa zmeny stavu lesnej
pokrývky na vybranom území, ktoré boli porovnané aj s oficiálnymi údajmi získa-
nými zo Štatistického úradu SR (ŠÚ SR) a údajmi CLC o výmere lesov.
Hlavným produktom sú klasifikácie lesnej pokrývky SR, pričom vytvorený nástroj
– kód – je možné jednoducho aplikovať na klasifikáciu lesnej pokrývky pre jednot-
livé kraje alebo okresy SR, a tiež modifikovať pre akékoľvek iné záujmové územie
v rámci EÚ. Algoritmus RF vykazoval vyššiu spoľahlivosť s celkovou presnosťou
95 % pre oba validačné roky (2017 a 2020), v ktorých sa správal približne rovnako
stabilne. Pri ďalších analýzach pomocou RF algoritmu sme zistili, že najviac lesnej
pokrývky pribudlo na území SR v období rokov 2018 až 2019. Výmera lesa sa za
toto obdobie zvýšila o 24 831,68 ha. Naopak, úbytok lesa bolo možné pozorovať
medzi rokmi 2019 a 2020, kedy sa stav lesa znížil o 1,04 %. Na regionálnej úrovni
bola v sledovanom období najvyššia lesnatosť pozorovaná v Žilinskom, Banskobys-
trickom a Prešovskom kraji.
Vykonaná klasifikácia lesnej pokrývky pomocou algoritmu strojového učenia RF
a satelitných dát Sentinel-2 v GEE preukázala vysokú presnosť klasifikácie a efekti-
vitu pri sledovaní zmien lesnej pokrývky na území Slovenska vo vysokom časovom
a priestorovom rozlíšení v porovnaní s oficiálnymi údajmi ŠÚ SR a CLC. Uvedený
postup automatickej klasifikácie lesnej pokrývky a samotné klasifikované rastrové
vrstvy lesnej pokrývky Slovenska pre roky 2017 až 2022 sú dostupné vo forme voľ-
ne dostupných kódov pre GEE.
185