Illegal Deforestation And Land-Use Change Detection From Satellite Imagery Using U-Net With CNN Backbone And Siamese CNN Architecture

Authors: Lalithavani K, Arulmozhi P, Vaidegi, Kiruthiga P, Nivetha S, Swetha E

Abstract: Background: Illegal deforestation and unlawful land-use change (LUC) represent two of the most severe anthropogenic threats to global biodiversity, carbon sequestration capacity, and ecosystem stability. Traditional field-based monitoring remains logistically impractical at continental scales. Objective: This study proposes a hybrid deep-learning framework that couples a U-Net segmentation architecture enhanced by a ResNet-50 convolutional neural network (CNN) backbone with a Siamese CNN designed explicitly for multi-temporal change detection, enabling automated, high-accuracy identification of illegal forest clearance and LUC events from multi-spectral satellite imagery. Methods: The pipeline is evaluated on three benchmark datasets—Amazon Deforestation Dataset (ADD), Sentinel-2 Global Land Cover (S2GLC), and the DESIS Hyperspectral Forest Dataset—covering more than 200,000 km² of tropical and sub-tropical biomes. Preprocessing integrates radiometric calibration, cloud masking, normalised difference vegetation index (NDVI) computation, and data augmentation to address class imbalance. Results: The combined framework achieves an overall accuracy of 96.8%, an Intersection-over-Union (IoU) of 0.923, and an F1-score of 0.947 on the held-out test partition, outperforming contemporary methods including DeepLab v3+, SegNet, and standalone Siamese networks. Conclusions: The proposed architecture demonstrates operational viability for near-real-time deforestation surveillance and may directly support regulatory agencies in enforcing environmental law under the Convention on Biological Diversity and national REDD+ frameworks.

DOI: https://doi.org/10.5281/zenodo.20352111