Integrating residual, recurrent, and transformer models for deforestation monitoring with multitemporal SAR imagery

The Amazon rainforest, twice the size of Western Europe and the largest forest on Earth, plays a crucial role in controlling the global climate. However, detecting deforestation in this area through optical satellite images is tough due to clouds blocking the view for much of the year. We introduce four deep learning models designed to track deforestation more accurately using synthetic aperture radar (SAR) data, which are unaffected by adverse weather conditions, such as clouds. The proposed models, named recurrent residual CNN-1 (RRCNN-1), RRCNN-2, RRCNN-3, and RRSwin, extend existing architectures by combining residual and recurrent units in different configurations to improve spatiotemporal feature extraction. RRSwin further integrates a transformer decoder to enhance long-range context modeling. Residual learning facilitates gradient flow and feature reuse, recurrent layers capture temporal dependencies across image sequences, and the Swin Transformer enables hierarchical attention across spatial patches. The models were evaluated on SAR data from the Brazilian Amazon, using deforestation maps from the PRODES project as reference. Experiments considered both bitemporal and multitemporal inputs to assess the impact of temporal depth on performance. The RRSwin model achieved the best results, with metric values of 85.3% precision and 83.7% F1-score when using a seven-image sequence. Further analysis showed that extending the temporal input improved detection metrics, achieving an 85.5% F1-score with a 25-image sequence, but improvements became marginal after a certain point, despite higher computational costs. The results highlight that combining local features, temporal modeling, and attention mechanisms, alongside multitemporal inputs, improved the performance of change detection in SAR imagery. These findings support the development of more reliable deforestation monitoring systems using radar imagery, especially in persistently cloud-covered regions such as the Amazon.