Cross-domain mangrove change detection and ecological response analysis under typhoon disturbance.

In recent years, global climate change and human activities have significantly degraded mangrove ecosystems. Developing efficient spatio-temporal change detection methods is crucial for assessing mangrove health. In this paper, we propose an automated framework with unsupervised domain adaptation capabilities, combining a Separable Transformer Neural Network with a Residual Stacking strategy (STNN-StackResNet) and an unsupervised spatio-temporal change detection model (USTC-StackNet). This approach achieves efficient modeling of mangrove evolution without manual labeling, attaining an average F1 Score of 83.75% and IoU of 72.32% across cross-domain datasets. The framework was applied to monitor the newly planted mangroves in Dongzhai Port, Hainan Province. Results indicate that the mangrove ecosystem remained ecologically stable, with the total area showing a slight increase from 662.76 hectares to 667.72 hectares. Furthermore, we quantitatively analyzed the ecological dynamics associated with typhoon disturbances for the first time. The study reveals a "pressure-recovery" cycle where ecological indices showed significant declines temporally coincident with typhoon events - specifically, CVI decreased by 9.12% and NDVI by 4.62% in 2022 - followed by a rapid recovery. This study not only provides a high-precision tool for mangrove monitoring but also offers quantitative insights into ecosystem resilience against climatic disasters.