**4. Discussion**

The results described in the prior section provide answers to the several challenges that can be encountered when detecting changes on specific sites. Indeed, besides detecting the changes with their dates, there is a need to classify the type of changes and to detect gradual changes. Four main observations may be drawn from this research.

First, the proposed method provided satisfactory results for the change detection and the change classification for both ground truth datasets. As far as the change detection is concerned, thanks to the complementary information provided by the sigma0VH and NDWI features (the former mainly for buildings, and the latter mainly for vegetation/soil), we were able to achieve an overall accuracy for the full dataset of 79%. As far as the change classification is concerned, the OA ranged from 79% to 90%, depending on the type of change that was considered (vegetation, building and soil). The OA of 90% and the F1-score

of 0.80, obtained for the vegetation "summer classification", illustrate the well-known robustness of the selection of the NDVI as a vegetation indicator [25,49,53], especially in summer conditions. As previously shown in [47], the BAI was proven to be useful for soil detection. Regarding the classification of buildings, the results revealed the suitability of combining the BI, BI2 and SBI indices, as an OA of 76% and an F1-score of 0.71 were obtained for the "summer classification". As mentioned in the Methods section, these indices were not used for the building classification rules of the "changepoint classification" and were replaced by the sigma0VH feature. This is due to the fact that the probability of finding cloud-free images in other periods than the summer is lower and the radar backscatter helps improving building discrimination thanks to its sensitivity to variations in height and shape. For this reason, it will be useful to carry out additional tests to investigate whether the use of the sigma0VH feature could be used also for the "summer classification". Moreover, further research could be conducted in regard to the number of Sentinel-2 images used for the "changepoint classification". Although data gaps were filled in through linear interpolation and the time series were smoothed using a Gaussian kernel, the cloud cover limits the number of usable images, especially during winter months. By only selecting the dates for which a certain number of S2 images are available, it is likely that the performance of the change classification would be improved.

Second, the "summer classification" is better suited for the detection of gradual changes. Figure 6 illustrates an ongoing vegetation growth leading to a soil decrease. This was not captured by the changepoint detection method but was classified as a vegetation increase and soil change thanks to the summer 2016–2018 comparison. The "summer classification" also provided better vegetation classification for change dates that occurred during winter, as seasonality strongly impacts the performance, as most vegetation is dormant during the winter. However, when comparing the "summer classification" and the "changepoint classification" results, it should be taken into account that the size of the two datasets is very different (302 vs. 26), and this had an impact on the results both in terms of representativeness and numerical accuracy.

Third, the use of vector polygons originating from the RDSs vector file to group the image pixels in the change analysis constitutes, at the same time, an advantage and a limitation. The fact that we averaged the information over the whole sites, on the one hand, helped reduce the noise (especially as far as Sentinel-1 is concerned) and filter out unnecessary details, but on the other hand, it may have led to the non-detection and/or non-classification of either small changes or bigger changes occurring on large sites, as the scales of the changes do not always match the scales of the vector polygons [14]. To partially overcome these issues, the polygon size could be reduced, for example, by segmenting each site either based on a fixed grid or external sources, such as WALlonie Occupation et Utilisation du Sol (WALOUS) [54,55]. However, this can lead to other problems, such as a significant increase in the computing power and and/or the creation of a large number of objects that would be too small compared to the Sentinel spatial resolution. Moreover, although external sources could, in principle, provide additional information on the type of change, this leads to the challenge of keeping these data up to date.

Fourth, the use of Sentinel data also has its limitations. First, as mentioned above, the spatial resolution reduces the number of RDSs for which the results can be reliable. For example, in total, 90.4% of the RDSs were larger than 400 square meters (roughly one Sentinel-1 pixel and four Sentinel-2 pixels). Moreover, although most of the sites are former industrial facilities with extensive infrastructure, changes may occur on only minor parts of the site, as illustrated in Figure 7. However, Sentinel images offer major advantages compared to orthophotos, which are open access but provided once a year, or Pléiades images, which can be obtained on demand and are costly. In fact, not only can they guarantee a much higher temporal coverage (especially if we consider the Sentinel-1 allweather capabilities), but they are also completely free, which means that the operational costs of the tool are significantly reduced. Moreover, thanks to the Terrascope platform and its cloud computing environment, the method is automated and provides, every

two months, results that are directly usable by regional authorities. Although the use of Sentinel data limits the number of RDSs that can be analyzed and the size of the changes detected, thanks to the results that we have shown, the regional authorities will be able to update the RDS inventory in a more efficient and less expensive way. Indeed, the SARSAR service enables the prioritization of the orthophotos analysis work and drastically limits field efforts. Table 8 shows a sample of bimestrial final change lists, and Figure 8 presents four RDSs, three for which a change date was detected and one with no change.

**Figure 6.** Close-ups of an RDS showing gradual vegetation increase ("Ets Biernaux"), between 2016 and 2018, illustrated at the top with Sentinel-1 images, in the middle with Sentinel-2 images and at the bottom with orthophotos.

**Figure 7.** Close-ups of an RDS showing a building increase, between 2019 and 2020, too small for the Sentinel spatial resolution ("S.A.N.I. Carrelages"), illustrated at the top with Sentinel-1 images, in the middle with Sentinel-2 images and at the bottom with orthophotos.

**Table 8.** Example of bimestrial final change list for a sample of RDSs.


**Figure 8.** Examples of detected and classified changes, on Sentinel-1 images (left) and Sentinel-2 images (right). Details of the changes are explained in Table 8.
