Protecting Forests Through Intelligent Edge Computing: CBAM-UNet for Real-Time Deforestation Detection

Abstract

The accelerating loss and degradation of forest ecosystems necessitate accurate and efficient monitoring approaches. This study presents a lightweight, attention-enhanced semantic segmentation framework for binary forest canopy segmentation using satellite imagery. The proposed model integrates a U-Net architecture with a MobileNetV2 encoder and a Convolutional Block Attention Module (CBAM) to improve feature representation while maintaining computational efficiency. The model is evaluated on the Bragagnolo Amazon Rainforest dataset for forest-versus-non-forest segmentation. The CBAM-enhanced model achieved a mean Intersection over Union (mIoU) of 0.6339 compared with 0.6282 for the baseline U-Net–MobileNetV2, while maintaining a comparable Dice score (0.6141 vs. 0.6187), indicating more consistent class-wise segmentation. The exported ONNX model sustained approximately 5.9 frames per second on an NVIDIA Jetson Nano while consuming under 80 MB of memory, confirming its feasibility for resource-constrained edge deployment. This work focuses strictly on binary canopy segmentation and provides insights into the trade-offs between accuracy and efficiency in lightweight deep-learning models for remote-sensing applications.