Evaluating the Potentiality of Sentinel-2 for Change Detection Analysis Associated to LULUCF in Wallonia, Belgium

: Land Use/Cover changes are crucial for the use of sustainable resources and the delivery of ecosystem services. They play an important contribution in the climate change mitigation due to their ability to emit and remove greenhouse gas from the atmosphere. These emissions/removals are subject to an inventory which must be reported annually under the United Nations Framework Convention on Climate Change. This study investigates the use of Sentinel-2 data for analysing lands conversion associated to Land Use, Land Use Change and Forestry sector in the Wallonia region (southern Belgium). This region is characterized by one of the lowest conversion rates across European countries, which constitutes a particular challenge in identifying land changes. The proposed research tests the most commonly used change detection techniques on a bi-temporal and multi-temporal set of mosaics of Sentinel-2 data from the years 2016 and 2018. Our results reveal that land conversion is a very rare phenomenon in Wallonia. All the change detection techniques tested have been found to substantially overestimate the changes. In spite of this moderate results our study has demonstrated the potential of Sentinel-2 regarding land conversion. However, in this speciﬁc context of very low magnitude of land conversion in Wallonia, change detection techniques appear to be not sufﬁcient to exceed the signal to noise ratio.


Introduction
Land Use/Cover changes (LULCC) lie on a scale of severity that ranges from no alteration through modifications of varying intensity to a full transformation. The rate of change and the nature of the transitions differ in time and space. Some regions are relatively stable (e.g., permanent forest); whereas others areas are subject to rapid and persistent transformation (e.g., urban expansion of previously agricultural land). The increase of human population and technological development has been found to accelerate LULCC [1][2][3]. There is extensive literature on sudden land cover conversion resulting from manmade or natural phenomenon such as forest deterioration, agricultural magnification, natural disaster or urban sprawl. However, few studies focus on subtle land changes. The study of LULCC relies on both subtle and abrupt transitions and an improved understanding of the complex dynamic processes underlying the former would allow for more reliable projections and more realistic scenarios of future changes [4].
According to European statistics [5] only 1.6% of land cover type has changed during the 2006-2012 period. This number covers 39 countries which span over 5.86 million of km 2 . Among European countries, Belgium has one of the lowest mean annual land cover rates. Each year, only 0.1% of the total area (~30 km 2 ) is converted to different land cover classes [6]. As such it is not surprising that many studies focus on African and Asian countries which have undergone major LULCC transformations. Africa has the largest annual rate of forest loss and reports from African countries documented that about 0.82 million km 2

of forest have been converted into other land uses between 1990 and
Land 2021, 10, 55 3 of 23 logic will be tested in the case of the much debated use of per-pixel or per-object techniques to obtain a detailed from-to change information. The validation of this research use harmonized and comparable statistics on land use and land cover across the whole of the EU's territory (Land Parcel Identification System (LPIS), Land Use/Cover Area frame Survey (LUCAS), CORINE Land Cover). This paper is an attempt to fill the gap related to subtle LULCC detection analysis and provides clues for using Copernicus Land Monitoring Services to support the LULUCF regulation. It also highlights the strengths and weaknesses of the most common change detection techniques. Finally, it discusses the use of Sentinel-2 data for measuring changes in carbon stocks resulting from direct human-induced land use.
The paper is organized into four sections. Section 2 gives a brief account of the change detection techniques and the reference data used in the research. Section 3 presents the results of the different techniques. Section 4 discusses the accuracy of the change maps and some challenges related to the use of Sentinel-2 data for LULUCF change detection. Finally, our conclusions are presented in Section 4.

Sentinel-2 Data Processing and Analysis
This study was undertaken in Wallonia, the southern part of Belgium ( Figure 1). The region covers an area of 16,901 km 2 with a population over 3.6 million. Two sets of Sentinel-2 images from 2016 and 2018 have been pre-processed according to the procedure adopted by [26]. Six cloudless and snowless mosaics composed of eight tiles of Sentinel-2 data have been produced. They cover three seasons: winter, spring and summer from both years. The weather conditions during the autumn period of both years did not permit to generate autumn mosaics due to an extended period of cloud cover.
Land 2021, 10, x FOR PEER REVIEW 3 of 23 snowless mosaics of Sentinel-2 from the years 2016 and 2018. The post-classification comparison logic will be tested in the case of the much debated use of per-pixel or per-object techniques to obtain a detailed from-to change information. The validation of this research use harmonized and comparable statistics on land use and land cover across the whole of the EU's territory (Land Parcel Identification System (LPIS), Land Use/Cover Area frame Survey (LUCAS), CORINE Land Cover). This paper is an attempt to fill the gap related to subtle LULCC detection analysis and provides clues for using Copernicus Land Monitoring Services to support the LULUCF regulation. It also highlights the strengths and weaknesses of the most common change detection techniques. Finally, it discusses the use of Sentinel-2 data for measuring changes in carbon stocks resulting from direct human-induced land use. The paper is organized into four sections. Section 2 gives a brief account of the change detection techniques and the reference data used in the research. Section 3 presents the results of the different techniques. Section 4 discusses the accuracy of the change maps and some challenges related to the use of Sentinel-2 data for LULUCF change detection. Finally, our conclusions are presented in Section 4.

Sentinel-2 Data Processing and Analysis
This study was undertaken in Wallonia, the southern part of Belgium ( Figure 1). The region covers an area of 16,901 km 2 with a population over 3.6 million. Two sets of Sentinel-2 images from 2016 and 2018 have been pre-processed according to the procedure adopted by [26]. Six cloudless and snowless mosaics composed of eight tiles of Sentinel-2 data have been produced. They cover three seasons: winter, spring and summer from both years. The weather conditions during the autumn period of both years did not permit to generate autumn mosaics due to an extended period of cloud cover.  (T31UDS, T31UES, T31UFS, T31UGS, T31UER, T31UFS, T31UGR, and T31UFQ)  The processing flow of the change detection analysis is shown in Figure 2. It involves the pre-processing of Sentinel-2 and the production of six mosaics of Sentinel-2  (T31UDS, T31UES, T31UFS, T31UGS, T31UER, T31UFS, T31UGR, and T31UFQ)  The processing flow of the change detection analysis is shown in Figure 2. It involves the pre-processing of Sentinel-2 and the production of six mosaics of Sentinel-2 images. Then, the application of the most commonly used methods in change detection: (a) algebraic and, (b) post-classification [27]. The algebraic methods refer to bi-temporal approach which Land 2021, 10, 55 4 of 23 exploits only the summer mosaic of both years. In opposite, the post-classification methods involve the use of the six mosaics for a multi-temporal approach. The bi-temporal analysis was carried out using the ArcGIS Pro software and its raster calculator tool. A detailed process description of the pixel-based classification can be found in [26]. The object-based classification was also implemented in ArcGIS Pro software (Esri Inc. ArcGIS Pro (version 2.3.3). Software. Redlands, CA, USA: Esri Inc., 2018.).
Land 2021, 10, x FOR PEER REVIEW 4 of 23 images. Then, the application of the most commonly used methods in change detection: (a) algebraic and, (b) post-classification [27]. The algebraic methods refer to bi-temporal approach which exploits only the summer mosaic of both years. In opposite, the post-classification methods involve the use of the six mosaics for a multi-temporal approach. The bi-temporal analysis was carried out using the ArcGIS Pro software and its raster calculator tool. A detailed process description of the pixel-based classification can be found in [26]. The object-based classification was also implemented in ArcGIS Pro software (Esri Inc. ArcGIS Pro (version 2.3.3). Software. Redlands, CA: Esri Inc, 2018.).

Figure 2.
Workflow of the change detection analysis. Two sets of Sentinel-2 data have been pre-processed to produce cloudless and snowless mosaics from winter, spring and summer season of 2016 and 2018 following the procedure described in reference [26]. Then, two approaches of change detection analysis have been tested: (1) algebraic (image differencing, image ratioing, index differencing, principal component analysis) and, (2) post-classification (pixel-based classification and object-based classification). Finally, change maps have been generated for each method.

Algebraic Change Detection
The algebraic change detection method involves the transformation of two original images into a new single-band image in which the areas of land cover change are highlighted [28]. The method is based on image algebra [27]. The most popularly techniques include: image differencing, image ratioing, index differencing and principal component analysis (PCA). Threshold selection for finding the change areas is a common procedure in algebra based change detection [29]. These techniques generates only binary change (i.e., change vs. no-change) [27]. They have the advantage of being based on the detection of physical changes between image dates. This avoids the errors introduced in post-classification where inaccuracies in the land cover classification between dates are propagated into the land cover change analysis [28].
Below are the different equations that have been applied on Sentinel-2 mosaics bands to generate the change maps using Raster Calculator tool (ArcGIS Pro): • Image differencing [25] Change map =  Workflow of the change detection analysis. Two sets of Sentinel-2 data have been pre-processed to produce cloudless and snowless mosaics from winter, spring and summer season of 2016 and 2018 following the procedure described in reference [26]. Then, two approaches of change detection analysis have been tested: (1) algebraic (image differencing, image ratioing, index differencing, principal component analysis) and, (2) post-classification (pixel-based classification and object-based classification). Finally, change maps have been generated for each method.

Algebraic Change Detection
The algebraic change detection method involves the transformation of two original images into a new single-band image in which the areas of land cover change are highlighted [28]. The method is based on image algebra [27]. The most popularly techniques include: image differencing, image ratioing, index differencing and principal component analysis (PCA). Threshold selection for finding the change areas is a common procedure in algebra based change detection [29]. These techniques generates only binary change (i.e., change vs. no-change) [27]. They have the advantage of being based on the detection of physical changes between image dates. This avoids the errors introduced in post-classification where inaccuracies in the land cover classification between dates are propagated into the land cover change analysis [28].
Below are the different equations that have been applied on Sentinel-2 mosaics bands to generate the change maps using Raster Calculator tool (ArcGIS Pro): • Image differencing [25] Change map = normalized squared difference facilitates the thresholding since it regroups the change pixels distributed initially in the tails of the distribution curve around the mean to a unique direction.
• Spectral index differencing [25]: In this research, we use four widely used spectral indices to extract land feature: (1) Normalized Vegetation Index (NDVI [30,31]), (2) Normalized Difference Built-up Index (NDBI [32]), (3) Brightness Index (BI [33]) and, (4) the second Brightness Index (BI2 [33])). We use only soil and vegetation indices due to their ability to characterize the most relevant land categories (forest land, cropland, grassland and settlement). We did not use water index because we assumed that this land category did not change much over the time. Image differencing was then applied to all spectral indices. PCA is used to capture the maximum variance in a finite number of orthogonal components based on an eigenvector analysis of the data correlation matrix. It has been used in change detection for many years because of its capacity of enhancing the information on change. The basic premise of PCA is to reduce the dimensionality of a dataset consisting of a large number of interrelated variables, while retaining as much as possible the variation present in the dataset. This is achieved by transforming to a new set of variables, the principal components (PCs), which are uncorrelated, and which are ordered so that the first few retain most of the variation present in all of the original variables [28].
In this study, PCA has been used in two steps. First, it has determined which bands of the 10 bands Sentinel-2 data retain most of the variation ( Table 1). The first four bands (B2, B3, B4, B5) account for 98.78% of the covariance with a percentage of eigen values higher than 1%. For these reasons, we only used B2, B3, B4 and, B5 to carry out the pre-classification analysis.
Second, the resulting PCs from 2018 have been compared with PCA images from 2016 through image differencing in order to generate change maps. Principal Components Analysis was performed using ArcGIS Pro.

. Post-Classification Change Detection
The post-classification method is the comparative analysis of two independently produced classifications from different dates [25]. The post-classification comparison can be done in a pixel-or object-based manner. In the pixel based approach, the classification is performed at the raster cell level whereas the object-based approach groups pixels into homogenous units based on local variance criteria objects are created using local homogeneity criteria, merging spatially contiguous pixels [34][35][36]. It is generally argued that the object-based classification is more suitable for Very High Resolution (VHR) images where the pixel-based approach faces the challenge posed by higher spectral variation and mixed pixels [27,35]. Contrary to the image algebra method, these techniques provide from-to change information. They have the advantage to bypass the difficulties in change detection associated with the analysis of images acquired at different times of the year [37][38][39]. However, as mentioned before, they are highly sensitive to the individual classification accuracies [28] and the comparison of the classifications inevitability leads to overstating the extent of changes [40].
In this research, the pixel-based classification was performed according to the work of [26]. The object-based classification was conducted using ArcGIS Pro software. The segmentation was carried out using Segment Mean Shift of the Spatial Analyst toolbox and the classifier was Maximum Likelihood Classification. Training and validation datasets were the same as the work of [26]. The comparison of each independent classification was executed using Raster Calculator tool.

Land Parcel Identification System (LPIS)
In Wallonia, the land parcel identification system (LPIS) is called anonymous agricultural plot (AAP). The LPIS indicates the use of land in agricultural areas managed within the framework of the Common Agricultural Policy. The AAP is publicly available through the geoservices of Wallonia (https://geoservices.wallonie.be/arcgis/rest/ services/AGRICULTURE/SIGEC_PARC_AGRI_ANON__2018/MapServer). This dataset gives the delineations of boundaries of agricultural fields, as well as the other relevant information assigned by farmers for each claim year. For this study, the AAP has been converted through a conversion table (Appendix D) to have only the distribution of grassland and cropland and thereby corresponds to the definitions of the categories of land as defined by reference [12].

CORINE Land Cover (CLC)
CORINE Land Cover is a land cover database that has been produced for 1990, 2000, 2006, 2012 and 2018 [41]. This inventory consists of 44 land cover classes and uses a minimum mapping unit of 0.25 km 2 . This classes are grouped into 5 land cover classes in the land cover change and statistics 2000-2018 (available at https://www.eea.europa.eu/ Land 2021, 10, 55 7 of 23 data-and-maps/dashboards/land-cover-and-change-statistics. It is an interactive viewer that displays land cover statistics per country (Table 2).

Reference Points
The Land Use/Cover Area frame Survey (LUCAS) and the Water and Wetness (WAW) layer of the Copernicus land monitoring HLR have been used to produce the reference points of this study. The LUCAS database is a survey conducted by Eurostat which provides harmonized statistics on LUC across European Union. LUCAS is based on statistical calculations that interpret observations in the field. It is based on a standardized survey methodology in terms of a sampling plan, classifications, data collection processes and statistical estimators that are used to obtain harmonized and unbiased estimates of land use and land cover [42]. The database was converted into the 5 categories of land (forest land, cropland, grassland, wetland and settlement) as defined by Reference [12] through a conversion table (Appendix C). Both definitions of land categories were difficult to align perfectly because the classes nomenclature were not the same and the LULUCF land categories are really rigorous (e.g., a forest has to be of at least of 0.5 ha, 20% of trees and a height of 5 m). In addition, LUCAS nomenclature has been made to be harmonized and comparable at the EU scale and is not specific to the particularities of Belgium landscape. To take into account these limitations, all the points were further validated by means of the interpretation of the Sentinel-2 mosaics and aerial orthophotography available at ( https://geoportail.wallonie.be/catalogue/647e383d-c74b-4ee6-bf48-a5ebc746e8 bf.html) from years 2016 and 2018. The Water and Wetness (WAW) layer of 2015 was used to address the lack of points in the "wetland" category of the LUCAS database. Points have been randomly allocated in the classes' permanent water and permanent wet area. This stratified random sampling design enables to satisfy the accuracy assessment [43,44]. The resulting reference points are depicted in Figure 3 and the statistics are available in Table 3. In total, eight points have changed between 2016 and 2018. This gives a percentage of change of 0.37%. The area of land converted ranges from 0.02 to 0.33 km 2 .

Algebraic Change Detection Results
The algebraic change detection analysis applied algebraic operations (differencing and ratioing) on the mosaic of summer 2018 and 2016. These mosaics have been chosen because they were acquired in the best period of the year for executing the image analysis (homogeneous vegetation status, good atmospheric conditions, good illumination and viewing angle). The selection of the bands of interest has been made through a principal component analysis (PCA) (Section 2.1.1). Figure 4 is an illustration representing an example of land use change in Wallonia. The location of the observed changes is delimited by black polygons. In this area of interest, two polygons of forest land (A and B) in 2016 have been converted into settlement and one polygon of grassland in 2016 (C) has been converted into settlement in 2018.

Algebraic Change Detection Results
The algebraic change detection analysis applied algebraic operations (differencing and ratioing) on the mosaic of summer 2018 and 2016. These mosaics have been chosen because they were acquired in the best period of the year for executing the image analysis (homogeneous vegetation status, good atmospheric conditions, good illumination and viewing angle). The selection of the bands of interest has been made through a principal component analysis (PCA) (Section 2.1.1). Figure 4 is an illustration representing an example of land use change in Wallonia. The location of the observed changes is delimited by black polygons. In this area of interest, two polygons of forest land (A and B) in 2016 have been converted into settlement and one polygon of grassland in 2016 (C) has been converted into settlement in 2018. Figure 5 shows the results of the algebraic method and Table 4 presents the Walloon' statistics of change associated to each algebra technique.  Table 4 presents the Walloon' statistics of change associated to each algebra technique.   Land 2021, 10, x FOR PEER REVIEW Figure 5 shows the results of the algebraic method and Table 4 presents the Wa statistics of change associated to each algebra technique.     Table 7 reports the errors matrix in terms of estimated area proportion for the pixel-based classification.    Table 7 reports the errors matrix in terms of estimated area proportion for the pixel-based classification.

Comparison with Validation Datasets
The change maps have been compared with three different datasets: (1) CORINE Land Cover, (2) Anonymous Agricultural Plot and (3) the reference points. Table 8 consists of a summary of the different results of the three validation datasets.

Accuracy Assessment
Appendices A and B present the confusion matrices of each change map using the 2141 reference points based on LUCAS and WAW database for the year 2016 and 2018. In addition, the overall accuracy (OA) was generated from the confusion matrix [36]. Figures 8-10 show a graphical representation of the results for the OA, errors of omission (changed erroneously) and errors of commission (not changed erroneously). We also evaluated the statistical significance of the difference between the pixel-based classification and the object-based classification using the chi-square distribution with one degree of freedom. The test equation may be expressed as [45]:         Table 9. Chi squared test for evaluating the statistical significance between pixel-based classification and object-based classification [45].

Discussion
As mentioned by reference [25], the selection of a suitable method of change detection for a given research is not straightforward. It depends on the remote sensing data, the study area and the type and magnitude of change. Four observations may be drawn from the results of this research.
First, the three validation datasets have highlighted the fact that the rate of LUC change in Belgium is very low. According to reference [6], Belgium is a country with one of the lowest mean annual land cover change rates in Europe. Each year, only 0.1% (~30 km 2 ) of the total area is converted to different land cover classes whereas the European mean rate is 1.6%. The reference points give a land conversion rate of 0.4% in Wallonia (~70 km 2 )) and enable the identification of the most converted land areas in Wallonia. They are grassland (−0.41%) and settlement (+1.21%) ( Table 3). This is not surprising since grassland is the main source for artificial land take in the country. The AAP also identifies grassland as a category of land which undergoes a notable conversion (2.55%). However, this dataset does not provide the direction of changes. Meanwhile, the agricultural area of CLC shows a change of −0.13% (Tables 2 and 7). Unlike the other validation datasets, it points out a major wetland conversion which is in fact the result of the minimum mapping unit of CLC (0.25 km 2 ) which is not sufficient to properly map most of wetland areas in Belgium.
Second, when comparing the algebraic and post-classification methods, the algebraic methods provide a percentage of change closer to the reality of LULUCF changes The resulting matrix is presented in Table 9. The Chi squared test shows a relationship between both classifications. This is not surprising since both classifications were trained with the same training sample and the same classifier (Maximum Likelihood). Table 9. Chi squared test for evaluating the statistical significance between pixel-based classification and object-based classification [45].

Discussion
As mentioned by reference [25], the selection of a suitable method of change detection for a given research is not straightforward. It depends on the remote sensing data, the study area and the type and magnitude of change. Four observations may be drawn from the results of this research.
First, the three validation datasets have highlighted the fact that the rate of LUC change in Belgium is very low. According to reference [6], Belgium is a country with one of the lowest mean annual land cover change rates in Europe. Each year, only 0.1% (~30 km 2 ) of the total area is converted to different land cover classes whereas the European mean rate is 1.6%. The reference points give a land conversion rate of 0.4% in Wallonia (~70 km 2 )) and enable the identification of the most converted land areas in Wallonia. They are grassland (−0.41%) and settlement (+1.21%) ( Table 3). This is not surprising since grassland is the main source for artificial land take in the country. The AAP also identifies grassland as a category of land which undergoes a notable conversion (2.55%). However, this dataset does not provide the direction of changes. Meanwhile, the agricultural area of CLC shows a change of −0.13% (Tables 2 and 7). Unlike the other validation datasets, it points out a major wetland conversion which is in fact the result of the minimum mapping unit of CLC (0.25 km 2 ) which is not sufficient to properly map most of wetland areas in Belgium.
Second, when comparing the algebraic and post-classification methods, the algebraic methods provide a percentage of change closer to the reality of LULUCF changes (Tables 8 and 9). The change maps of the algebraic methods show a change percentage ranging from 1.6% (ratio B4) to 15.76% (PC3) and an overall accuracy (OA) ranging from 82.6% (BI2 differencing) to 98.1% (ratio B4). According to the classification standard of [46], most of these overall accuracies are considered as satisfactory because they are higher than 85%. Although, the algebraic methods overall accuracies are high, these numbers are mainly driven by the large proportion of unchanged points. The results of the postclassification methods differ further from the real change percentage (from 16.6% to 32.8%) and have lower overall accuracies (Tables 5 and 6). As mentioned by reference [47] determining land changes by overlaying maps that have the same categories from two points in time makes sense when the map are perfectly accurate. In this study, the maps are not perfectly accurate (OA pixel-based = 91.9% and 91.7%; OA object-based = 84.7% and 76.6%) and the amount of error is too large to ignore. Moreover, according to the reference points, the amount of change is 0.4%, while the errors in maps is significantly higher (Error pixel-based = 8.1% and 8.3%; Error object-based = 15.3% and 23.4%. Hence, errors in each individual map result in differences between the two maps. Despite having more misclassification and misregistration errors, Figure 10 shows that the post-classification methods are the most sensitive change detection technique. Among them, the object-based technique gives the most satisfying results when looking at identifying the location of observed changes (6 reference point of "change" have been correctly attributed to "change" in the change map). However, we did not observe a reduction of the small spurious change within the extent of each object that should results in a high spectral variability in the pixel-based classification [35]. Furthermore, the objectbased technique has also the most important commission errors (652 reference points of "no change" have been erroneously attributed to "change" in the change map). In conclusion, all of the change detection techniques substantially overestimated the changes.
Third, the use of Sentinel-2 data for LULCC detection can be summarized by the following points. In terms of spatial scale, the 10 m spatial resolution is sufficient to delineate individual geographic objects of interest. The visualization of change maps has shown that the converted land areas in Wallonia range from 20 pixels to 3300 pixels. Regarding the temporal scale, Sentinel-2A is available since June 2015 and should have a lifespan of 7 years. A second generation should follow for 7 additional years. Sentinel-2A and 2B have a high revisit time of 5 days ensuring the production of several cloud-free mosaics per year that minimizes the seasonal phenological differences. Furthermore, the twin satellites are deployed in polar sun-synchronous orbit which ensures that the angle of sunlight upon the Earth's surface is consistently maintained which limits the shadow effects. Consequently, Sentinel-2 provides high resolution images for the operational monitoring of land and the production of land-change detection maps.
Finally, the results of the change detection applied in the Walloon context of land conversion associated to the LULUCF sector shows its limits in precisely identifying the changes. On account of the low rate of land conversion in Wallonia (~0.4%; corresponding to~70km 2 of change), we reach a critical point where all techniques face difficulties to properly identify land conversion. As mentioned in the above point, Sentinel-2 data are not responsible for these moderate results. In addition, changing the temporal window from 2 years to 5 or more years would not improve significantly the results since the CLC data from 2006 to 2018 (Table 2) has not shown any increase of the magnitude of change. Similarly, the possibility of increasing the classification accuracy is very limited when reaching the 92% of overall accuracy. And if so, improving a few percent would still be too few to properly map the changes. As an example, two classifications of 98% of overall accuracy would make 96% of land correctly allocated in the change map and 4% of errors (~700 km 2 ) for only 70 km 2 of real changes.
In future, similar research should concentrate on (1) post-processing, (2) the combination of methods and (3) AI-based change detection. Nevertheless, it is essential to bear in mind that the post-processing could interfere with the automatic nature of the approach as well as its wide-scale implementation e.g., through the use of regional databases. Regarding the combination of methods, it is likely to propagate errors which would impede the final results. In recent years, integrated artificial intelligence technology has become a research focus in developing new change detection methods. Several studies have suggested that they could outperform the traditional change detection methods.

Conclusions
Gaining a better understanding of carbon cycle and climate change requires accurate information on land conversion. The recent launch of Sentinel-2 satellites provides new opportunities for studying LUC changes on a regional and global scale. A wide variety of studies have analyzed significant LUC changes such as massive forest deterioration or rapid urbanization in developing countries. Only a few have focused on more developed countries undergoing a low land conversion rate such as EU countries. In research explored the effectiveness of Sentinel-2 data to detect changes related to the LULUCF sector in Wallonia, Belgium. The approach tested the most commonly accepted change detection techniques in order to evaluate the capability of Sentinel-2 data to account for low land conversion. Our results suggest that the rate of conversion is too low to precisely identify changes. All the change detection techniques have been found to overestimate the change. We consider that Sentinel-2 data have a great potential for LUC change detection analysis. However, change detection capabilities are largely determined by whether the applied change magnitude exceeds the signal to noise ratio [48].
Author Contributions: O.C. is the main author of this manuscript. She processed the data and conducted the change detection analysis and the validation. S.P. made a substantial contribution in the processing of the Sentinel-2 data for the multi-temporal analysis and offered valuable comments on the methods and manuscript. B.B. conceived the original idea of the study. E.H. supervised the research. Please turn to the CRediT taxonomy for the term explanation. Authorship must be limited to those who have contributed substantially to the work reported. All authors have read and agreed to the published version of the manuscript.
Funding: This research was conducted in the framework of the "EO4LULUCF" project, which was funded by an internal fund of Institut Scientifique de Service Public Moerman (ISSeP).

Informed Consent Statement:
Informed consent was obtained from all subjects involved in the study.

Acknowledgments:
The authors would like to thank the European Union's Earth Observation Programme Copernicus and Eurostat for the provision of the LUCAS database.

Conflicts of Interest:
The authors declare no conflict of interest.   Appendix B Table A2. Confusion matrix of the post-classification methods. The object-based is the most sensitive technique with 6 location of change correctly labelled whereas the pixel-based has only 1 location of change correctly attributed. The overall accuracy (OA) of each technique is expressed at the right side of the   Cultures fruitières annuelles-Fraises 2 9520