Deep Learning-Based Methods for Mapping Mangrove Forests in Bohol, Philippines Using Sentinel-2 Imagery

Abstract. This study presents a deep learning-based approach for mapping mangrove forests in Bohol, Philippines using high-resolution Sentinel-2 imagery. Given the limitations of traditional mapping techniques and the ecological importance of mangroves, four convolutional neural network (CNN) architectures—U-Net, Attention U-Net, MSNet, and SegNet—were trained and evaluated. The preprocessing pipeline included patch generation, normalization, and random sampling to ensure spatial representativeness. Hyperparameter tuning explored combinations of loss functions and learning rates to optimize model performance. Results showed that U-Net consistently achieved the highest accuracy across all evaluation metrics, with an IoU of 0.93, accuracy of 0.98, precision of 0.966 and F1-score of 0.963. Visual inspections confirmed U-Net and Attention U-Net’s superior ability to delineate mangrove boundaries, particularly in complex coastal zones. In contrast, SegNet produced coarser edges but trained significantly faster, offering a practical alternative for rapid assessments or resource-constrained deployments. These findings emphasize the value of skip connections and attention mechanisms not just for performance enhancement but for improving the usability of outputs in real-world conservation. The study recommends U-Net for integration into local government monitoring systems, supporting disaster risk reduction, marine zoning, and restoration planning. Future work may incorporate drone imagery and transfer learning to improve adaptability across other Philippine coastal ecosystems.