Detection and Quantification of Vegetation Losses with Sentinel-2 Images Using Bi-Temporal Analysis of Spectral Indices and Transferable Random Forest Model

Abstract

The precise spatially explicit data on land cover and land use changes is one of the essential variables for enhancing the quantification of greenhouse gas emissions and removals, which is relevant for meeting the goal of the European economy and society to become climate-neutral by 2050. The accuracy of the machine learning models trained on remote-sensed data suffers from a lack of reliable training datasets and they are often site-specific. Therefore, in this study, we proposed a method that integrates the bi-temporal analysis of the combination of spectral indices that detects the potential changes, which then serve as reference data for the Random Forest classifier. In addition, we examined the transferability of the pre-trained model over time, which is an important aspect from the operational point of view and may significantly reduce the time required for the preparation of reliable and accurate training data. Two types of vegetation losses were identified: woody coverage converted to non-woody vegetation, and vegetated areas converted to sealed surfaces or bare soil. The vegetation losses were detected annually over the period 2018–2021 with an overall accuracy (OA) above 0.97 and a Kappa coefficient of 0.95 for all time intervals in the study regions in Poland and Norway. Additionally, the pre-trained model’s temporal transferability revealed an improvement of the OA by 5 percentage points and the macroF1-Score value by 12 percentage points compared to the original model.

1. Introduction

Land cover and land use (LCLU) changes play an important role in land management, urban planning, the monitoring of sustainable agriculture, forestry, environmental protection, and achieving the Sustainable Development Goals [1,2]. Moreover, reliable and accurate information on LCLU changes is essential for assessing the ecological risk of the loss of ecosystem services [3] and accounting for carbon emissions and removal from the land use, land use change, and forestry (LULUCF) sector (Decision 529/2013/EU, EC 2013) [4]. The national LULUCF inventories serve as the main source of information for climate change modelling and estimation of greenhouse gas emissions under the IPCC. The accurate assessment, monitoring, and reporting of LCLU changes require precise geospatial data to meet both national and international climate commitments and to understand how these changes impact carbon stocks and emissions.

The growing demand for up-to-date and reliable information on LCLU changes has caused the rapid development of change detection methods and algorithms derived from remotely sensed data. The rapid growth in the volume of Earth Observation data, advanced remote sensing techniques, development of machine learning algorithms, and high-performance computing power has increased the number of studies on LCLU changes and contributed to the development of various change detection approaches [5,6,7,8]. Nevertheless, the post-classification change detection approach, which relies on the comparison of land cover classification maps derived at different time intervals, is still commonly used [6,9,10,11,12]. For example, the Pan-European High-Resolution Change Layers, e.g., imperviousness density or tree cover density available at the Copernicus Land Monitoring Service, have been derived based on the comparison of individual products prepared at three-year time intervals [13]. Another example is the work of Wohlfart, Mack [12], who used Landsat time series data to analyse LCLU changes in China’s four regions and developed 32 models using the Random Forest (RF) algorithm, which is time-consuming and labour-intensive. Furthermore, it should be noted that the post-classification approach is sensitive to error accumulation, which is an effect of the accuracy of input datasets [9,14,15]. Therefore, direct classification methods have been developed, which classify multi-temporal images directly so that there is no accumulation of errors [14]. There has been a rapid development of methods that also rely on a time series analysis of spectral features, for example, BFAST (Breaks For Additive and Seasonal Trend) [16,17] or more advanced multispectral ensembles using a second classification model, e.g., BFAST-Random Forest [18], LandTrendr with RF (Landsat-based detection of Trends in Disturbance and Recovery) [19,20]. These approaches rely on a dense time series of high-quality data; thus, large computing capacities are essential for data processing.

Moreover, there are several studies where a decision tree approach is applied to bi-temporal differences in spectral bands or indices. Cardille, Perez [21] applied a lightweight decision tree combined with Bayesian Updating of Land Cover (BULC) to a Normalised Burn Ratio index (NBR) and SWIR1 band calculated from a time series of Landsat and Sentinel-2 data to detect and categorise stand-replacing forest changes in Canada. The authors distinguished four nested changes and obtained a user’s accuracy (UA) of above 90% for all disturbed classes and much lower producer’s accuracy (PA) of 53.4% to 98.9%. The decision tree approach applied to Sentinel-2-based NDVI and NBR indices was used by Alonso, Picos [22] to detect forest changes smaller in size than those in Cardille, Perez’s study [21]. Reference change pixels were chosen from the subtraction of the land cover maps produced for consecutive years. The limitation of this method is that it is site-specific, as it requires a set of spectral threshold values to be defined, which depends on forest composition and abiotic factors.

In contrast to traditional approaches, machine and deep learning algorithms have recently gained more attention in change detection. Afaq and Manocha [23] reviewed change detection techniques, highlighting the challenges in detecting changes despite advancements in artificial intelligence (AI) techniques like machine learning and deep learning. They pointed out that the use of AI increases efficiency and improves accuracy, but it suffers from the limitation of reliable training datasets. Future trends in developing enhanced change detection methods should be focused on algorithms that effectively distinguish changes in diverse geographical locations and varied environmental conditions using transfer learning or domain adaptation as well as the examination of machine learning models with a small sample of reference data [8,14].

To address these challenges, we focused our study on the examination of the machine learning Random Forest model for tracking and quantifying vegetation losses with a small sample of reference data. We aimed to develop the most accurate and operational land cover change detection (LCCD) method, which will not be site-specific and will allow tracking and quantifying vegetation loss on an annual basis over areas with different ecological and topographic conditions and diverse landscapes. The study was conducted at a regional scale in two different climatic zones: warm temperate (continental zone, Poland) and cold temperate (boreal zone, Norway). The proposed method integrates the bi-temporal analysis of the spectral indices and machine learning classification performed on the Sentinel-2 data. A direct classification method in terms of multi-class change detection was used. In addition, we examined the transferability of the pre-trained LCCD model over time, which is an important aspect from the operational point of view, as it may reduce the time required for the preparation of reliable and accurate training data. The knowledge of the temporal transferability of the model prediction is limited [24]. The LCCD was carried out on an annual basis for the period 2018–2021 using the Google Earth Engine (GEE) platform.

2. Materials and Methods

2.1. Study Area and Datasets

2.1.1. Study Area

The study was conducted at the regional scale in Poland over Łódź Province—continental zone, and in Norway over Viken County and Oslo—boreal zone (Figure 1). The Łódź Province with the main city of Łódź is located in central Poland and covers an area of 18,219 km² [25]. It is covered predominantly by agricultural land with a mosaic of arable land, grasslands, orchards, and woody patches. The Łódź agglomeration includes 1 million people and covers an area of 1800 km². The region is developing very dynamically, with many investments and LCLU changes taking place in the city and on its outskirts. This study area represents the typical landscape in Central and Eastern Europe.

Viken County is located in the south-east of Norway, surrounding Oslo, the capital of Norway. The county extends from the Swedish border and the Oslo Fjord with a flat coastal landscape up to the mountainous areas of Hardangervidda (appr. 1900 m above sea level) in the north-western part. The southern region is characterised by farmland and forest. Further towards the north/north-west, this goes over to mountain and valley landscapes. According to the Statistics of Norway [26], Viken County covers an area of 24,592 km², and the city of Oslo, around 454 km².

2.1.2. Sentinel-2 Data and Pre-Processing

In this study, a series of Sentinel-2A and -2B images acquired in the growing season from May to September over the period 2018–2021 was used. The Sentinel-2 Level-2A products available from GEE [27], which consist of atmospherically corrected bands with bottom-of-atmosphere calibrated reflectance, were applied. The Level-2A product additionally provides a Scene Classification map (SCL) at a 20 m spatial resolution, which was applied to mask out pixels saturated or assigned as shadows. A shadow mask was created based on an SCL value of 3. Due to the high omission errors of cloud detection in Sentinel-2 SCL products [28,29,30], the Sentinel-2 Cloud Probability (CLP) available from GEE [31] was applied to mask out cloudy pixels. The CLP is created by the s2cloudless algorithm for automatic cloud detection [28,32]. A cloud mask was created using a threshold of 65% [33]. However, the study by Skakun, Wevers [28] showed that the CLP does not cope well enough with thin cloud masking, which was confirmed by our preliminary results. Therefore, to mask the omitted cirrus clouds, a threshold value of 0.01 was applied to band B10 [34] derived from the top-of-atmosphere Sentinel-2 Level-1C product [35]. According to Skakun, Wevers [28], it is recommended that morphological operations are performed during the conversion of the CLP to the cloud mask. Therefore, morphological operations in GEE on the final mask composed of clouds and shadows were performed: erosion (kernel with a radius of 60 m) followed by dilation (kernel with a radius of 260 m). The kernel radius values were adjusted based on the visual assessment starting from the recommended values. In this way, the potentially contaminated pixels that were not masked using the SCL and CLP products were masked. Furthermore, in Norway, the snow in the mountainous areas was masked using SCL.

For each Sentinel-2 image, the uncontaminated pixels were selected and used to derive a seasonal mosaic based on the mean value of the spectral reflectance calculated from images acquired between May and September. Carrasco, O’Neil [36] showed that temporal aggregation, which uses some metrics derived from a time series of satellite data, has the potential to efficiently integrate large amounts of data and compensate for the lower quality of automatic image selection and cloud masking. The seasonal mosaics were prepared for the study areas in Poland and Norway for the years 2018, 2019, 2020, and 2021. The mosaics comprised ten spectral Sentinel-2 bands: four bands at 10 m spatial resolution and six bands at 20 m spatial resolution, which were resampled to 10 m using the nearest neighbour resampling method.

2.2. Methodology

The proposed land cover change detection (LCCD) method consisted of two phases (Figure 2). Phase 1 was focused on the derivation of training samples for phase 2, where the machine learning RF algorithm was applied to detect and classify the changes. The following land cover change (LCC) classes were identified: class 0: no change; class 1: woody coverage converted to non-woody vegetation, e.g., clear-cuts; and class 2: vegetated surfaces (woody and non-woody) being converted to sealed surfaces, e.g., newly built-up areas, infrastructure, and construction sites with unvegetated surfaces or agriculture areas with bare soil in Norway. The LCCD method was developed and implemented using the GEE cloud computing platform. GEE provides access to a time series of Sentinel-2 images and products as well as powerful cloud computing facilities, which allows spatial analysis to be caried out anywhere on Earth [37].

2.2.1. Phase 1: Detection of Potential Changes Using Spectral Indices

This phase aimed to derive the training samples for the detection and classification of changes in phase 2. In the first step, differences in the spectral indices were calculated following one of the basic pixel-based change detection techniques [6]. Based on a literature review, three well-known and frequently used spectral indices—the Normalised Difference Vegetation Index (NDVI), Normalised Burn Ratio (NBR), and Normalised Difference Water Index (NDWI)—were calculated [21,38,39,40,41,42].

The NDVI is the most well-known and frequently used vegetation index, applied to analyse the chlorophyll content in leaves, to determine the amount of biomass [40] or to assess changes in woody vegetation cover [43]. The NDVI values range from −1 to 1, where higher values indicate healthy vegetation and a value of around −1 indicates a lack of vegetation. The NBR is commonly used for burned area mapping [38] and recently for the detection of forest disturbance [21,44]. The NBR values range from −1 to 1, where higher values indicate healthy vegetation and lower values indicate recently changed, burned, or bare ground [42]. The NDWI is used to monitor water bodies [39] and ranges from −1 to 1, where higher values indicate the presence of water.

The seasonal maximum values of the NDVI and NBR and a minimum of NDWI were calculated at the pixel level based on images acquired between May and September for each year. These metrics were chosen to extract “pure” values that indicate the state of the vegetation without atmospheric influences and to minimize the effect of cloud coverage [45,46]. Then, the differences between the seasonal values in the year n and year n + 1 were calculated for each index.

Then, to extract changes related to vegetation loss, it was assumed that changes detected by the NDVI and at least one of the other indices have a higher probability of being correctly detected. In both cases, the thresholds of two standard deviations of differential indices were applied.

In addition, to reduce the number of false changes over the agricultural areas caused by the dynamic plant phenology and agricultural practices, the mask of potential agricultural areas had to be created. The standard deviation of the NDVI for the vegetation season was calculated and a value greater than 1.2 was applied to derive the seasonal agricultural mask for each year. The threshold was defined based on the empirical approach, supported by visual assessment.

Finally, the potential changes were converted into polygons and were subject to the labelling process. For each class in each time interval, around 20–30 change polygons were selected and labelled. The national orthophotos supported by the seasonal Sentinel-2 mosaics were used as the reference data in the labelling process. For class 0—no change, which represented much larger areas, around 40–50 polygons representing various unchanged land cover classes were additionally manually drawn based on the visual interpretation of Sentinel-2 mosaics. In addition, a 20 m buffer equivalent to one Sentinel-2 pixel (SWIR band) was applied to the selected polygons to reduce the edge effect and to assure the homogeneity of changes.

2.2.2. Phase 2: RF Classification and Accuracy Assessment

The labelled polygons from phase 1 were used to generate the reference sampling points for the RF classification using a stratified random sampling method. In each study area, 2000 points for each class were randomly distributed and randomly divided, with 70% used for training and 30% for accuracy assessment within each of the classes [47].

The RF classifier [48,49] implemented in GEE was used to detect and classify the changes. The RF classifier is a machine learning technique where many decision trees are constructed based on a random sub-sampling of the given dataset [50]. The RF classifier has been established by many researchers to be a robust machine learning algorithm for the LULC classification of medium-resolution satellite imagery because it is less sensitive than other streamline machine learning classifiers to the quality of training samples and to overfitting [48,51,52].

The RF classification was performed on a data stack of multispectral seasonal Sentinel-2 mosaics calculated for years n and n + 1 (10 spectral bands for each year), and a set of differential indices: dNDVI, dNBR, and dNDWI from phase 1 (in total, 23 variables). The parameterisation of the RF model was performed on 100 single trees in the forest. The RF probability threshold was set to 50% [53] for the pixels being classified as a change. If the probability did not reach the intended threshold value, a pixel was marked as unclassified.

The classification was performed separately for each time interval: 2018–2019, 2019–2020, and 2020–2021, over both study areas in GEE. The accuracy of the classification models was also assessed in GEE using a confusion matrix [54], OA, Kappa coefficient, UA, and PA. The confusion matrix is an array, which represents the performance of the model. The rows are actual values and the columns are predicted values. The correctly classified elements are located on the main diagonal. There can be four types of elements in the confusion matrix [55]: true positive (TP)—an actual positive class predicted as a positive class; false positive (FP)—an actual negative class predicted as a positive class; false negative (FN)—an actual positive class predicted as a negative class; and true negative (TN)—an actual negative class predicted as a negative class. The UA and PA for each class are expressed as follows:

$$UA={\displaystyle \frac{TP}{TP+FP}},$$

(1)

$$PA={\displaystyle \frac{TP}{TP+FN}}.$$

The Kappa coefficient measures the agreement between the predicted and actual classification [55]. According to Grandini, Bagli [55], the OA and the Kappa statistic for multi-class case are calculated with the following formulas:

$$OA={\displaystyle \frac{c}{s}},$$

(2)

$$Kappa={\displaystyle \frac{c\times s-\sum _k^K{p}_k\times t_k}{s^2-\sum _k^K{p}_k\times t_k}},$$

where $c$ is the total number of elements correctly predicted (on the main diagonal), $s$—the total number of elements, $K$—a number of classes, $p_k$—the total number of elements predicted for class k (column total), and $t_k$—the total number of elements that truly occurs (row total).

The results of the RF classification were converted to a vector format. The changes smaller than 0.2 ha were assumed to be less accurate and potentially false changes. They were added to the unclassified pixels and assigned as uncertain. The final LCC product consists of four categories: three predefined change classes and uncertainty areas.

Additionally, the transferability of the LCCD model between time intervals was examined for the study area in Poland. The pre-trained RF model developed for the interval 2020–2021 was applied to the time interval 2019–2020.

2.2.3. Independent Verification of the Land Cover Changes (LCCs)

The independent verification of the final LCC product was carried out using the national aerial orthophotos with the support of Sentinel-2 seasonal mosaics and images available in Google Earth Pro. In Poland, the LCC products for the periods of 2019–2020 and 2020–2021 were verified against the national aerial orthophotos from 2019, 2021, and partially from 2020, Sentinel-2 mosaics, and Google Earth Pro. In Norway, the national aerial orthophotos captured mainly during April and May 2020 and 2021 covering Viken County and Sentinel-2 mosaics were applied in the verification of the LCC for the interval 2020–2021. There was a lack of reference data over the mountainous region in the north-west. Almost 500,000 random points were distributed within the study area in Poland and partially over the common area in Norway. The points were assigned to three classes of the LCC products and approximately 200 random points for each class were selected and examined. To evaluate the LCC products, the confusion matrices together with OA, UA, PA, Kappa coefficient, and F1-Score were calculated. In addition, the macroF1-Score, which is the harmonic mean of the arithmetic mean of UA for single classes and the arithmetic mean of PA for single classes [55], was calculated as follows:

$$macro\mathrm{F}1-\mathrm{Score}=2\times {\displaystyle \frac{(meanUA\times meanPA)}{({meanUA}^{-1}+{meanPA}^{-1})}}.$$

(3)

3. Results

3.1. LCCs Detected in Poland and Norway

The two-step approach allowed for the accurate detection of two classes of vegetation loss on a yearly basis using the Sentinel-2 data. Figure 3 presents the variation in the size of the polygons detected as LCCs in classes 1 and 2 for Poland and Norway for three time intervals. In general, the large proportion of detected changes in both countries was related to class 1. The median size of the change polygon in class 1 ranged from 0.5 to 0.7 ha for Poland and was around 0.8 ha for Norway. There were many more large-scale forest changes in Norway compared to Poland.

Of interest, the number of changes for class 2 decreased in Poland and increased twice in Norway over the three-year period. In both countries, the average size of the changed polygons reached around 0.4 ha; however, there were many more large-scale changes in Poland compared to Norway (Figure 3). In Poland, the biggest changes were related to the development of road infrastructure, the expansion of the mine area, and newly built warehouses. In Norway, the large-scale changes in class 2, reaching a total area of around 28 ha, were related to a massive clay landslide that occurred in 2020–2021. The examples of LCCs detected over the period 2018–2021 within both study areas were presented in Figure 4 and Figure 5.

3.2. Accuracy of the LCC Classification Models

The OA in all intervals for both study areas reached above 0.96 and the Kappa coefficient was above 0.95 (Figure 6). The UA values were equal to or greater than 0.98, and the PA was above 0.94 for both study areas. The accuracy of class 2 was slightly lower than class 1. In general, the accuracy of the classification models was very high; however, some pixels from class 0 (no change) were misclassified as class 2 or left unclassified (this class contained pixels with the RF probability less than 50%, which were assumed to be less reliable).

3.3. Independent Verification of LCC Products

The results of the independent verification of the LCC products for the periods 2020–2021 and 2019–2020 are presented in Figure 7a–c. The OA of the independent verification of the LCC products was equal to 0.93 or 0.94 (macroF1-Score of 0.82 and 0.84) for Poland and Norway for the period 2020–2021, and slightly lower, equal to 0.91 (macroF1-Score = 0.76), for Poland for the period 2019–2020. Class 0 was predicted correctly in Poland and Norway, with just 1–2 points being misclassified as class 1 or class 2. Comparable results were obtained for class 1—woody coverage converted to non-woody vegetation in both study areas for 2020–2021. Slightly lower agreement was observed in Poland for the period 2019–2020, where, out of 200 verified points, 10 were misclassified as no change and 15 as class 2—vegetated surfaces converted to sealed surfaces or bare soil. For class 2, around 29–33 points out of 200 were misclassified as no change in both study areas and both analysed periods. This overestimation was mostly related to the seasonal changes in agricultural land and coal mines in Poland. Class 0 showed the highest UA (0.99) and the lowest PA (0.84–0.88). In Poland, the highest UA and PA were obtained for class 1 for 2020–2021 and slightly lower for 2019–2020. For class 1, the F1-Score value was over 0.93. In general, the accuracy of class 2 was lower compared to classes 0 and 1 and reached an F1-Score value of around 0.90 for the verified periods.

3.4. Applicability of the Pre-Trained Model

For the study area in Poland, the results of the pre-trained model were compared with the results of the independent model from phase 2. The independent verification of the LCC product derived from the pre-trained model applied for the period 2019–2020 revealed an improvement of the OA by 5 percentage points and the macroF1-Score value by 12 percentage points compared to the independent model (Figure 7d). The most pronounced improvement was observed for class 2, where the F1-Score value increased from 0.89 to 0.95, and the commission error decreased from 14.5% to 8.5%. The number of points of the actual class 0 wrongly predicted as class 2 was reduced from 29 to 17. The omission error for this class decreased from 8.1% to 0.5%. Slightly less of an improvement was reported for class 1, where the F1-Score increased from 0.93 to 0.98; however, the commission error decreased from 12.5% to 4.5%, with the number of wrongly predicted points reduced from 25 to 9.

Furthermore, the application of the pre-defined model reduced the omission error in the no-change class from 16% to 11%, with 39 to 25 points being misclassified as other LCC classes.

4. Discussion

The methodology outlined in this study is able to accurately detect two types of LCCs related to vegetation loss at the inter-annual time scale. We proposed a two-step approach, which firstly flags the potential changes based on simple thresholding of the difference in spectral indices; secondly, the RF classifier is used to assign the direction of changes. The operator’s involvement is minor and limited to the manual indication of a small number of reference polygons (we identified around 30 polygons per change class), which served as reference data for further classification. Accurate and reliable reference data are crucial to obtain high-accuracy results in the machine learning classification process [56]. The process of sample collection is labour-intensive and time-consuming, especially when a thousand points must be checked manually or automatically against aerial or satellite imagery [57]. Johnson and Iizuka [58] proposed a promising approach to training data collection based on the use of crowdsourced data, but obtained low accuracies for the classifications performed (OA = 72–84%). Therefore, we proposed the application of simple automatically defined threshold values, providing an efficient way of identifying the provisional changes. We demonstrated that the RF model parametrised with a small number of samples of reference data gave highly accurate results. However, more study is needed to determine the smallest number of reference polygons required to obtain superior LCCD model performance. The approach proposed in our study was successfully tested at a regional scale in Poland and Norway and gave accurate and reliable results beyond the specific geographical context and different landscape patterns.

This approach provides a reliable alternative to the decision tree approach proposed recently by Alonso, Picos [22] and Cardille, Perez [21]; however, their studies focused on bi-annual changes. Decision-tree-based methods rely on threshold values, which are more site-specific and depend on forest characteristics and conditions [21]. Alonso, Picos [22] developed a decision tree for the NDVI and NBR for forest disturbance in Spain and obtained an OA of 0.82 for the NDVI and 0.7 for the NBR. They found that the NBR is less accurate than the NDVI due to a higher number of false positives. Cardille, Perez [21] achieved higher PA and UA values (above 0.75) with the NBR compared to those obtained by Alonso, Picos [22]. We used the combination of the NDVI and NBR or NDWI as the input data for the rapid detection of potential changes, which served as reference data for the classification process. Positive values of dNDVI and dNBR indicated vegetation decrease, while positive values of dNDVI and negative values of dNDWI indicated a higher probability of biomass decrease due to water body occurrence. However, such cases of new water bodies have proven to be negligible in both study areas. In Poland, the combination of all three indices worked best, while for Norway, due to the excessive variability in the NDWI index, a combination of dNDVI and dNBR worked better.

In general, the independent verification revealed that changes related to class 1 (woody coverage converted to non-woody vegetation) were detected with higher accuracy (PA and UA above 0.9 in all analysed periods) in both study areas compared to class 2 (vegetated surfaces being converted to sealed surfaces), except for the period 2019–2020 in Poland (UA = 0.88). These values are generally higher than those achieved by Alonso, Picos [22] and Cardille, Perez [21]. Our results are also more accurate than those achieved by Polykretis, Grillakis [40], who used a set of spectral indices derived from Landsat data and change vector analysis (CVA) to delineate four categories of LCCs over the island of Crete, Greece. They obtained the highest OA, over 0.9 for change/no-change maps, and below 0.9 for change direction categories.

Pacheco-Pascagaza, Gou [59] developed a Sentinel-2-based near-real-time (NRT) change detection system. Their system accurately detected forest loss and other vegetation loss and gain in two test sites in pan-tropical areas: Mexico and Colombia. For the forest cover loss, they obtained PA values of 0.75 and 0.88 for the Columbia and Mexico sites and UA values of 0.99 and 0.97, respectively. In comparison, we obtained higher PA values and similar UA values for the same class, except for the 2019–2020 period for Poland (however, using the pre-trained model, the UA was 0.96). The NRT system is based on the direct classification of changes through a chronologically stacked image pair (two composites), which is similar to the algorithm we proposed, except for the step with the analysis of the spectral indices that we proposed to generate the potential changes and speed up the process of the preparation of the reference dataset. Their system requires the collection of local training data as well as the generation of a baseline forest map. The generation of the training sample in particular can be time-consuming and labour-intensive.

Although the LCCD approach was demonstrated to be highly accurate in terms of RF model performance, the independent verification of the results revealed quite high commission error for class 2, related to vegetated surfaces converted to sealed surfaces or bare soil, at around 15% for Poland and 17% for Norway. This is predominantly caused by the seasonal variability of agriculture fields and the presence of bare soil after harvesting the crop residues. This could potentially be overcome by applying the mask of agricultural land. Furthermore, it is worth exploring whether adding statistical variables like the seasonal standard deviation of spectral values into the RF model could improve the accuracy over these areas. In addition, the uncertainty areas, where the probability of the RF model was lower than 50% or the cluster of detected changes smaller than 0.2 hectares, were located on the edges of the identified change areas. This was especially noticeable for class 1, in the forest clear-cuts or along the newly constructed road, not wide enough to be delineated with Sentinel-2 data [60]. Yu, Zhou [61] largely eliminated the boundary problem using an object-based method. However, the disadvantages of object-based methods [8], i.e., computational intensity, sensitivity to the selection of the segmentation algorithm, and potential ineffectiveness in detecting small changes, make this approach unsuitable for operational purposes. The further detailed analysis of the probability of the RF models is essential to set the threshold for pixels being assigned as changes with high and lower probability.

As this study was focused on the regional scale, we applied temporal cloud-free seasonal mosaics derived from uncontaminated pixels using temporal aggregation [36]. To capture the changes in the boreal latitude, a wider seasonal composite window (May–September) was required to prepare a seasonal mosaic not contaminated by clouds and shadows. The seasonal mosaic was calculated using the mean values, which caused a problem with the detection of changes that occurred at the end of the growing season. In addition, changes that occurred after the composite window were not able to be detected until the following year [21]. The length of the image composition should be determined by latitude and adjusted to the level of cloudiness. The proposed approach could also be tested for the bi-annual changes using, for example, transfer learning models.

Furthermore, the pre-trained model transferability demonstrates a significant improvement in the accuracy of the LCC classification compared to the results obtained from the original model. For both LCC classes, a large reduction in commission error was obtained. The temporal transferability was performed over the same spatial domain and the same input variables were used, which may explain the good performance of the model. This is in line with the temporal transferability of the model for LCLU mapping by Praveen, Mustak [62], where the accuracy of the transferred model decreased by about 1% compared to the reference data. By contrast, Wijesingha, Dzene [24] tested the spatial–temporal model transferability for agricultural land cover mapping and concluded that the temporal transferability reduced the model performance by 15%. However, both studies dealt with the model for LCLU mapping, not change detection. There is a need for further research on quantifying the performance of the temporal and spatial transferability of the models trained on data from different spatial and temporal domains as well as exploring different machine learning algorithms.

5. Conclusions

This study was focused on developing a semi-automated LCCD algorithm, which is able to accurately quantify vegetation loss on an annual basis at the regional scale and is not site-specific. The identified changes related to woody coverage converted to non-woody vegetation and vegetated surfaces (woody and non-woody) converted to sealed surfaces or bare soil are crucial for forest management and spatial planning needs. The LCCD algorithm enabled the detection of changes with high accuracy for both regions in Poland and Norway representing different climatic zones, which was confirmed by the independent verification of the results. Further work is needed to examine the changes smaller in size than 0.2 ha, which could be important for change detection in urban areas and essential from a spatial planning perspective. Although the temporal and spatial transferability of the models should be examined further, our results regarding the temporal transferability of the model revealed the high performance of the pre-trained model and demonstrated an improvement in the accuracy of the LCC classification compared to the results from the original model.