Weakly supervised seawater intrusion mapping via multi-source remote sensing data fusion

Seawater intrusion is a typical disaster in the coastal critical zone. Assessing its damage to the soil-vegetation continuum faces several bottlenecks. To address these issues, a weakly supervised mapping method for seawater intrusion is proposed. This method utilizes the Google Earth Engine (GEE) cloud platform. It fuses Sentinel-1 Synthetic Aperture Radar (SAR) data, Sentinel-2 optical data, topographic data, and land cover semantic data. First, an automatic seed sample generation algorithm driven by physical rules is developed. Second, the spatial extent of seawater intrusion is stably extracted in complex coastal ecotones. Finally, a three-dimensional feature vector space representing the ‘biological-physical-chemical’ state is constructed by introducing underlying surface heterogeneity features. The seawater intrusion event in the coastal area of Dalian in October 2024 is taken as a case study. Validation results show that the overall accuracy (OA) of the extracted intrusion extent reaches 91.07%, with a Kappa coefficient of 0.87. The F1-score of the core affected category reaches 87.84%. Furthermore, the adaptive grading model accurately identifies 0.37 km2 of severe disaster areas. This work provides reliable technical and data support for coastal ecological early warning and disaster mitigation.