High-Resolution Mangrove Land Cover Mapping in the Sundarbans Using U-Net and Multi-Source Satellite Data

Accurate land use and land cover (LULC) classification is essential for the conservation and sustainable management of ecologically sensitive regions such as the Sundarbans, a UN-ESCO World Heritage Site and the largest contiguous mangrove forest in the world. This study presents a deep learning framework for LULC classification in the Sundarban region of West Bengal, India, using Sentinel-2 multispectral imagery and the high-resolution ESRI 10m Land Cover dataset as ground truth. A U-Net convolutional neural network (CNN) was trained on spatially aligned and normalized 32×32 pixel patches, incorporating four Sentinel-2 bands (visible and near-infrared) to classify seven land cover classes, including mangroves, water, wetlands, and built-up areas. The workflow addressed key challenges in data preprocessing, including CRS alignment, resampling, and patch extraction, leveraging datasets derived from Google Earth Engine for enhanced reproducibility. The validation results demonstrated strong overall model performance, achieving an accuracy of 0.9532, precision of 0.9529, recall of 0.9532, and an F1-score of 0.9465. The Intersection over Union (IoU) score was 0.9126, underscoring the model’s robustness in distinguishing land cover classes. While dominant classes such as water and mangroves achieved high F1-scores (0.9465), challenges remain in classifying minority classes, such as wetlands, due to class imbalance. This study highlights the potential of integrating open-access satellite data with deep learning for ecological applications while also identifying limitations in rare-class recognition. The proposed framework offers a scalable approach for LULC mapping in complex coastal ecosystems, with significant implications for biodiversity conservation and climate resilience planning in the Sundarbans and similar mangrove-dominated regions.