Deforestation Detection from Satellite Images Using an Enhanced Residual Attention U-Net

Abstract

Deforestation is one of the most serious environmental threats to ecological balance, and accurate monitoring of forest-cover change remains challenging. Deep-learning models have increasingly been applied to remote-sensing imagery for automated forest mapping. Although U-Net is widely used for semantic segmentation, its conventional architecture has limited ability to suppress irrelevant skip-connection features and capture multiscale context in heterogeneous satellite scenes. This study proposes an Enhanced Residual Attention U-Net that integrates a Channel Attention Layer (CAL) within the skip connections and a Dilated Convolutional Attention Layer (DCAL) at the bottleneck. CAL emphasizes informative channel responses, whereas DCAL expands the receptive field to capture fine-and coarse-scale contextual features. Single-run experiments were conducted on two public Sentinel-2 benchmark datasets representing the Amazon and Atlantic Forest biomes. The experiment summary reported accuracies of 96.0% and 96.5% and IoU values of 97.0% and 96.9% on the Amazon and Atlantic Forest datasets, respectively. Compared with standard U-Net, the reported accuracy gains were 6.8 and 7.37 percentage points; compared with Attention U-Net, they were 2.0 and 1.5 percentage points. The architecture therefore provides a practical framework for forest/nonforest segmentation and automated deforestation monitoring.