Semantic water body extraction by the high-quality segment anything model using multiple optical and SAR imagery

Water bodies are an important geographical feature in freshwater security, irrigation, climate regulation, and flood risk management. Thus, monitoring and extracting water bodies are widespread uses in remote sensing. This is the artificial intelligence (AI) age, demonstrated by a variety of AI models developed for a wide range of applications. Additionally, there is an increasing remote sensing data that can be used for AI models’ inputs. The high-quality Segment Anything model (HQ-SAM), a newly improved version of the SAM, is proposed to accurately enable the segmentation of a broad range of objects while maintaining the promptable architecture, efficiency, and zero-shot generalizability of the original SAM. We applied the HQ-SAM, and water indices (NDWI, MNDWI, SWI, AWEI) in the Otsu method for lake/reservoir extractions using optical and Synthetic Aperture Radar (SAR) remote sensing imagery, including Sentinel-1, 2, ALOS-2/PALSAR-2, RadarSAT, Landsat 5 and 8, and Google-based satellite images (Leafmap) for selected lakes in South Korea. The HQ-SAM model is evaluated as working well, exhibiting excellent accuracy (above 95%) of water body masks compared to the measured boundary of the lake. The HQ-SAM results surpassed Otsu’s results applied to four common water indices. Both approaches revealed advantages and disadvantages, where the HQ-SAM worked well with larger, complex lakes but had some mis-segmented small, thin parts of lakes. Nevertheless, the Otsu method did not separate surface water bodies from the snow and ice on the mountains. The HQ-SAM revealed an accurate and promising potential model for water body extraction using remote sensing imagery.