Abstract
Freshwater resources on islands are critically important, motivating the study of surface freshwater extraction. However, optical remote sensing methods face challenges from cloud cover, coastal background noise, and difficulty in distinguishing saltwater in seawater intrusion zones. To address these issues, we propose a sensing method that fuses optical and radar remote sensing data. Using a bedrock island group outside the Pearl River Estuary as the study area, we employ Chinese GF-1 optical and GF-3 synthetic aperture radar (SAR) images as primary data. A labeled training dataset is constructed through field hydrogeological surveys and chloride ion measurements. We compare the normalized difference water index (NDWI), random forest, support vector machine, and deep learning models for freshwater extraction. Results show that the deep learning model based on GF-1 optical data achieves an AUC of 0.881, and the GF-3 SAR-based model reaches 0.840, both significantly outperforming traditional methods. These two deep learning models exhibit complementary capabilities: they intelligently segment seawater, saltwater from intrusion, and freshwater, identify surface freshwater regions, and fill gaps where optical data are obscured by clouds or adverse weather. This addresses the trade-off between accuracy and scenario robustness, offering significant value for scientific management and sustainable utilization of island water resources.