3.2.1. Summer Classification

The "summer classification", as we discussed in the previous section, takes into account, for each of the 302 sites, the summer comparison between 2016 and 2018 for the orthophotos dataset and between 2019 and 2020 for the Pléiades dataset. Again, the performance was assessed by combining the two datasets and computing the TPR, FPR and the F1-score for each class, along with the overall accuracy (see Table 4). The overall performance of the yearly classification based on summer values is satisfactory. The best results were obtained for the "vegetation" class, for which the OA was 90% and the TPR and FPR were 87% and 9%, respectively. The resulting F1-score was 0.80. The performance for the "building" and "soil" classes were slightly lower, with an OA of 76% and 79%, respectively, ye<sup>t</sup> still good, with an F1-score above 0.7.

**Table 4.** "Summer classification" (full dataset): confusion matrix and performance metrics.


To look deeper into the "vegetation" class, Table 5 also shows the results disaggregated by "increase", "decrease" and "no change" types, with the corresponding overall accuracy and omission/commission errors. As can be seen, for both the increase and decrease in vegetation, around 1 in 4 detections was a false alarm, whereas the percentage of missed changes were 20% and 12%, respectively. It is worth noting that there was no confusion between the two classes, as all the errors fell into the "no change" class. For this class, instead, the commission and omission errors were much lower, namely, 4% and 9%, respectively.

**Table 5.** "Summer classification" (full dataset): detailed confusion matrix for the "vegetation" class.

