Mask R-CNN-based detection and segmentation of Mangrove ecosystems in Lantau Island, Hong Kong

Mangroves play a crucial role in coastal protection and biodiversity but face escalating threats from anthropogenic pressures and climate-driven disturbances. Long-term monitoring remains challenging due to mangrove fragmentation and limited high-resolution historical data. This study presents a deep learning–based approach for mangrove identification, leveraging cloud-free Sentinel-2 MSI imagery (10 m resolution) and Mask R-CNN to map and analyze mangrove dynamics on Lantau Island, Hong Kong, from 2016 to 2024. The model integrates surface reflectance bands, spectral indices (EVI, LSWI, MVI), and elevation data, achieving high accuracy (mean absolute percentage error: 6.91%; root mean square error: 0.04 × 10⁴ ha). Multi-source validation demonstrated its strong generalization capacity across global mangrove ecosystems. Spatiotemporal analysis revealed divergent trends in two key mangrove stands. In Shui Hau, mangrove area declined continuously from 0.77 ha in 2016 to 0.39 ha in 2024, accompanied by shoreline erosion at a rate of 3.07 m/yr. This loss was associated with reduced suspended sediment concentration and persistent high wave energy. In contrast, Tung Chung's mangrove area expanded from 3.28 ha to 3.59 ha, with shoreline accretion at 0.85 m/yr, supported by moderate wave dynamics and higher sediment availability. These findings underscore the value of 10 m resolution Sentinel-2 MSI imagery for historical mangrove mapping, providing critical insights for targeted conservation and management strategies.