Deep learning mapping of wetland dynamics: a user-friendly web-based decision support tool for Chesapeake Bay

Wetlands are rapidly degrading under land-use change and climate pressures, yet most monitoring tools remain inaccessible to non-specialist decision-makers. This study addresses this gap by developing a web-based Spatial Decision Support System (SDSS) for wetland change detection at Blackwater National Wildlife Refuge, Chesapeake Bay. A Swin-Unet model trained on multi-temporal Sentinel-2 imagery (2017–2022) achieved a validation accuracy of 0.919, an F1-score of 0.868, and an IoU of 0.768 at a 0.7 decision threshold, substantially outperforming a U-Net baseline (F1 = 0.801). Independent validation against 100 manually interpreted high-resolution NAIP points supported real-world reliability, yielding 89% accuracy and an F1-score of 0.84. Between 2017 and 2022, the model detected 6.32 km² of wetland loss and 9.45 km² of gain, representing a net increase of 3.13 km² that masks significant spatial reorganization along vulnerable marsh edges and upland migration corridors. Model outputs are embedded within a Flask-Leaflet dashboard that translates change maps, area statistics, and prioritized management recommendations into an interface requiring no GIS expertise. This study demonstrates that transformer-based deep learning coupled with accessible web technologies can bridge the gap between advanced geospatial analysis and practical decision-making for wetland conservation, supporting adaptive management within the Chesapeake Bay basin.