3.2.2. Changepoint Classification

The "changepoint classification" takes into consideration only the RDSs for which at least one changepoint date has been estimated within the change detection process. As multiple changes can occur in the same site during the considered time period, a yearly comparison was required for each estimated change date. This was only possible using the Pléiades dataset, as only for this ground truth are the exact change dates available. A performance assessment (Tables 6 and 7) was carried out for all the changepoint dates knowing that the overall accuracy of the changepoint dates themselves was shown in a previous section.




**Table 7.** "Changepoint classification" (Pléiades dataset): detailed confusion matrix for the "vegetation" and "building" classes.

Although some dates were during winter months, the results for the vegetation changes remained good, with an OA of 84% and a F1-score of 0.75. With respect to the "summer classification", the main difference here was in the TPR, which was lower by 20 percentage points (87% for "summer classification" and 67% for Pléiades dataset). As regards the "building" class, there was the opposite trend for the Pléiades dataset, with both a higher OA and F1-score than those obtained for the "summer classification". Although the TPR was slightly lower, the significant drop in the FPR improved the performance. Finally, for the "soil" class, all the metrics showed a drop in the performance, especially as far as the FPR is concerned.

To complete the analysis, the detailed confusion matrices for the classes "vegetation" and "building" are provided in Table 7. Once again, the results are disaggregated by "increase", "decrease" and "no change" types. For the "vegetation" class, no increase was reported within any site of the ground truth; therefore, no metric was calculated. Instead, out of nine "decrease" changes, six were correctly identified, resulting in a commission error of 14% and an omission error of 33%. If we look at the "no change" class, we had a similar false alarm rate, but a much lower miss rate. For the "building" class, half of the "increase" changes in buildings were missed (50% omission error). However, all the changes that were flagged as an increase were correct (0% commission error). Instead, the classification of a decrease was more accurate, with only one false alarm and one missed detection. Finally, the "no change" classification was the one providing the best performance, with a commission error of 17% and an omission error of 6%.
