*3.1. Change Detection*

The performance was assessed in terms of true positive rate (TPR) and false positive rate (FPR). The overall problem can be in fact seen as a binary classification where either a "change" (1) or a "no change" (0) has to be detected. In order to compare the results with the ground truth, the latter was coded so that any change in any of the three classes (building, vegetation and soil) was assigned the value 1; in the case of no change for all three classes, the ground truth was given the value 0. A confusion matrix was then generated so that the number of true positives (TPs), true negatives (TNs), false positives (FPs) and false negatives (FNs) could be used to compute the TPR and FPR. To provide a unique measure that takes into account both detection and miss rates, the F1-score was also calculated. For the sake of completeness, the overall accuracy (OA) is also reported.

It is worth mentioning that, due to the specific way in which the ground truth is constructed, in order to generate the confusion matrix, we made the arbitrary assumption that only one change per site occurred in the considered period of time. This is a simplification that helped us to compare the results in a more straightforward way, but might not fully reflect the real situation, especially for the sites belonging to the orthophotos ground truth, as for a certain number of them it is more likely that multiple changes occurred at different times.

The change detection was performed using the sigma0VH and NDWI features, which amongs<sup>t</sup> the other features ultimately provided the highest accuracy. The use of both Sentinel-1 and Sentinel-2 data, which provide complementary information (the VH band mostly about buildings and the NDWI index mostly about vegetation and soil), allowed a more effective identification and classification of changes. The results for the entire dataset are shown in the first row of Table 3. The number of sites for which we had an estimated change is 108, 91 of which were correctly classified. Among the unchanged sites, we missed 46 of them, resulting in an OA of 79%. In terms of correct and miss detection rates, we, therefore, obtained a TPR of 66% and an FPR of 10%, with an F1-score of 0.74.


**Table 3.** Changepoint analysis: confusion matrix and performance metrics.

In order to better understand the results of the following block, the change classification, it was helpful to separate the Pléiades detections from the full dataset. The results are provided in the second row of Table 1. For this dataset, the number of sites that were flagged as changed was 26, with nine FPs, whereas the correct detections of the unchanged sites were 125. As a result, the TPR and FPR decreased to 55% and 7%, respectively, and, consequently, the F1-score dropped to 0.59. The OA, instead, increased to 87%, mainly due to the fact that the dataset was rather unbalanced.
