Temporal Observation Fusion for Deforestation Detection in Amazonian Legal Reserves: Insights from Reserve- and Pixel-Level Evaluation

Authors:

Abstract
Monitoring deforestation within Legal Reserves (LRs) requires methods capable of integrating spectral, spatial, and temporal information over large areas. This study developed and evaluated a semantic segmentation model based on a U-Net++ architecture with an EfficientNet-B0 backbone to identify deforestation in LRs declared in the Brazilian Rural Environmental Registry (CAR) in the state of Pará. Temporal differences in NDVI, NBR, and NDRE derived from Harmonized Sentinel-2 MSI imagery were stacked as three input channels. The model-development dataset comprised 6,949 LRs, divided into 70% for training and 30% for validation. The best model was selected according to the lowest validation loss, and the final decision threshold was calibrated using the Matthews Correlation Coefficient (MCC) at the LR level. During validation, the model achieved an IoU of 0.6820 and a Dice coefficient of 0.8110 at the pixel level, as well as an F1-score of 0.9743 and an MCC of 0.9492 at the LR level. The independent test included 2,500 LRs balanced between deforested and non-deforested classes, each represented by two distinct temporal combinations (CT1 and CT2), in addition to a Maximum Fusion strategy based on pixel-wise probabilities. At the LR level, Maximum Fusion achieved the best overall performance, with an accuracy of 0.9216, F1-score of 0.9183, MCC of 0.8460, and sensitivity of 0.8808. At the pixel level, however, CT1 achieved the highest IoU (0.2294), indicating lower geometric agreement with PRODES polygons during independent testing. Overall, the model was more consistent in identifying the occurrence of deforestation within LRs than in reproducing its exact spatial geometry. The results also highlight the influence of temporal image selection and the potential of integrating multiple observations to support environmental monitoring.

This publication has 43 references indexed in Scilit: