Weakly Supervised Semantic Segmentation of Mangrove Ecosystem Using Sentinel-1 SAR and Deep Convolutional Neural Networks

Mangrove ecosystems are vital blue carbon habitats, and their accurate mapping and ongoing monitoring are essential for conserving their unique ecological and environmental benefits. This study employed a weakly supervised deep learning framework for semantic segmentation of the ecosystem by integrating time-series Sentinel-1 Synthetic Aperture Radar imagery with an existing ecosystem map as weak labels. This approach incorporates a U-Net architecture with an EfficientNet-B7 backbone for robust feature extraction, employs Test-Time Augmentation (TTA) with an ensemble strategy to enhance prediction stability, and validates the method’s temporal transferability across multiple years. The model achieved high accuracy in the baseline year (2020), with 95.56% overall accuracy according to independent reference samples, and successfully generated noise-free maps despite imperfect training data. The TTA approach improved the final map both visually and statistically compared to a simple prediction approach. Visual comparisons with high-resolution satellite imagery, an existing ecosystem map, and a recent global mangrove extent dataset confirmed the superior performance of our proposed workflow in mangrove mapping. The temporal transferability of the model, examined in 2017 and 2023 using the baseline model with no modification, confirmed its applicability by generating the ecosystem maps meticulously, though the predictive uncertainty value increased. The generated multitemporal maps indicated a mangrove extent expansion of about 3.43% between 2017 and 2023. This straightforward workflow facilitates automatic mangrove ecosystem mapping and updating while effectively reducing the need for frequent reference sample collection, especially in cloud-prone areas. The temporal capability of the approach offers practical solutions for large-scale mangrove ecosystem conservation initiatives.

Abstract licence: CC BY-NC-ND

A person counts once a day per paper. We keep no cookie and no address: a reader is a hash that changes every day. Any month under five reads shows no number.