Satellite Image-Based Forest Cover Change Detection Using Siamese U-Net

Deforestation accounts for nearly 10% of global greenhouse gas emissions, yet monitoring forest loss across vast areas remains expensive and slow. Satellite imagery provides repeated, large scale coverage, and deep learning can automate change detection. This paper proposes a Siamese U Net architecture that detects forest cover changes from bi temporal satellite images. The model employs two shared weight U Net encoders to extract features from pre and post event images, computes a difference feature map, and decodes it into a binary change mask. The system was trained on 6,200 pairs of Sentinel 2 images (10 m resolution) from the Amazon and Congo basins, labelled with deforestation polygons from the Global Forest Watch database. The Siamese U Net achieved a mean Intersection over Union (IoU) of 0.82 and an F1 score of 0.89, outperforming a standard U Net (0.74 IoU) and classical change vector analysis (0.61 IoU). The model maintained high accuracy in both dense tropical forests and fragmented agricultural frontiers. Inference on a single image pair completed in 0.8 seconds on a consumer GPU, enabling near real time monitoring. The system was deployed on a Google Earth Engine compatible pipeline, processing 10,000 km² per day. These results demonstrate that Siamese U Net architectures offer a robust, scalable solution for continuous forest change tracking, supporting conservation agencies and climate policy enforcement.