Title : Semantic segmentation of satellite imagery to detect deforestation using CNNs
Abstract:
Deforestation remains one of the most critical drivers of climate change, biodiversity loss, and ecosystem degradation. Effective mitigation requires scalable systems capable of accurately monitoring forest cover across large and remote geographic regions. This study presents a comprehensive deep learning based framework for semantic segmentation of satellite imagery to detect deforestation with high accuracy and computational efficiency. Using a multimodal dataset integrating 16 Sentinel-2 spectral bands, vegetation indices, Sobel edge features, and Sentinel-1 SAR data, we designed a robust pre-processing pipeline that enhances vegetation discrimination while preserving critical boundary information. Four convolutional neural network architectures- U-Net, Attention U-Net, a lightweight Attention U-Net, and Attention U-Net with ASPP - were implemented and rigorously evaluated using IoU, F1, Dice, precision, and recall metrics.
Our results show that attention-enhanced architectures consistently outperform the baseline U-Net, with the lightweight Attention U-Net delivering the best balance between accuracy (79.17% IoU) and computational footprint (15.52M parameters). Notably, this model was successfully deployed on a Raspberry Pi 4, achieving real-time inference without requiring cloud connectivity, demonstrating its strong suitability for on-ground monitoring in resource-constrained environments. Statistical validation using paired t-tests and Cohen’s d confirms that the performance improvements over the baseline are both statistically significant and practically meaningful. Beyond accuracy, the study highlights how nonlinear vegetation indices, SAR backscatter features, and physics-informed feature design improve segmentation robustness and enhance generalization. Together, these contributions establish a practical and scalable foundation for near-real-time deforestation monitoring that is technically accessible even to small conservation groups, NGOs, and local authorities.
