Detecting Marsh-to-Open-Water Transitions Using Dual-Date Sentinel-2 and DeepLabV3

Abstract. Wetland ecosystems, critical for biodiversity and coastal resilience, face increasing threats from marsh-to-open-water transitions driven by environmental change. This study presents a deep learning approach to detect wetland loss in the Blackwater National Wildlife Refuge (BNWR), Maryland, using Sentinel-2 multispectral imagery. We employed the DeepLabV3 model with a ResNet-50 backbone for semantic segmentation to map areas of marsh degradation and conversion to open water. The model was trained on a curated dataset of Sentinel-2 images, leveraging their high spatial and temporal resolution. Preprocessing included atmospheric correction and cloud masking to ensure data quality. Under spatial 5-fold cross-validation, the model achieves F1 (change) = 82.7% (95% CI: 79.1–86.0) and IoU (change) = 70.4% (95% CI: 66.0–74.6) on held-out test tiles, with Overall Accuracy = 94.3% (95% CI: 93.0–95.6) and Balanced Accuracy = 89.0% (95% CI: 87.0–91.0). As an operational check on the full scene, we obtain OA = 95.31%, F1(change) = 83.76%, IoU(change) = 72.07%. The results reveal notable wetland loss within BNWR, identifying priority areas for conservation. This study demonstrates the efficacy of integrating deep learning with Sentinel-2 imagery for high-resolution environmental monitoring and offers a scalable framework for wetland management.