Quang et al. (2025) Semantic water body extraction by the high-quality segment anything model using multiple optical and SAR imagery

Identification

• Journal: Acta Geophysica
• Year: 2025
• Date: 2025-12-08
• Authors: Nguyễn Hồng Quang, Namhoon Kim, Hanna Lee, Seunghyo Ahn, Gihong Kim
• DOI: 10.1007/s11600-025-01732-0

Research Groups

• Institute for Smart Infrastructure, Gangneung-Wonju National University, South Korea
• Department of Civil Engineering and Environmental Sciences, Korea Military Academy, South Korea
• Department of Smart Infrastructure Disaster Prevention, Gangneung-Wonju National University, South Korea
• Department of Civil and Environmental Engineering, Gangneung-Wonju National University, South Korea

Short Summary

This study evaluates the high-quality Segment Anything Model (HQ-SAM) for semantic water body extraction using diverse optical and Synthetic Aperture Radar (SAR) imagery. The HQ-SAM demonstrated high accuracy (above 95%) and outperformed traditional water index-based methods for delineating water bodies in South Korea.

Objective

• To investigate the ability and robustness of the high-quality Segment Anything Model (HQ-SAM), specifically GeoSAM, for accurately segmenting water bodies (lakes/reservoirs) using multiple optical and SAR remote sensing data sources.
• To compare HQ-SAM's performance against the traditional Otsu threshold method applied to common water indices (NDWI, MNDWI, SWI, AWEI).

Study Configuration

• Spatial Scale: Gyeongpo Lake in Gangneung City, Gangwon-do, South Korea, and five other selected lakes/reservoirs across South Korea.
• Temporal Scale: Remote sensing data acquired primarily in December. Field survey conducted on 2024-01-08. Time-series analysis of Gyeongpo Lake from 1988 to 2023 at 5-year intervals.

Methodology and Data

• Models used: High-quality Segment Anything Model (HQ-SAM), specifically Segment-Geospatial SAM (GeoSAM). Compared with the Otsu method applied to four water indices: Normalized Difference Water Index (NDWI), Modified Normalized Difference Water Index (MNDWI), Sentinel-2 Water Index (SWI), and Automated Water Extraction Index (AWEI).
• Data sources:
  • SAR Imagery: Sentinel-1 (European Space Agency), ALOS-2/PALSAR-2 (Japan Aerospace Exploration Agency), RadarSAT (MDA, Canada).
  • Optical Imagery: Sentinel-2 (European Space Agency), Landsat 5 TM (United States Geological Survey), Landsat 8 OLI (United States Geological Survey), Google-based satellite images (Leafmap).
  • Ground Truth: Field survey using a Trimble R12i Global Navigation Satellite System (GNSS) with 8 mm horizontal and 15 mm vertical precision for Gyeongpo Lake boundary measurement.

Main Results

• HQ-SAM achieved excellent accuracy (overall accuracy above 98%) for water body masks of Gyeongpo Lake when compared to GNSS-measured boundaries.
• HQ-SAM results consistently surpassed the Otsu method applied to water indices by approximately 4% in accuracy.
• Image resolution significantly influenced HQ-SAM's accuracy, with higher resolution data (e.g., Leafmap, Sentinel-1, Sentinel-2) yielding better results.
• HQ-SAM performed well with larger, complex lakes but showed some mis-segmentation in small, thin parts of intricate water bodies.
• Unlike the Otsu method, HQ-SAM successfully differentiated surface water bodies from snow and ice on mountains.
• HQ-SAM was robust against speckle noise commonly found in SAR images.
• The model successfully delineated Gyeongpo Lake's surface area changes over 3.5 decades (1988-2023), demonstrating its potential for lake monitoring and change assessment.
• Among the water indices, MNDWI consistently provided the most stable and high-precision results (above 95% overall accuracy).

Contributions

• This study is the first to thoroughly investigate and demonstrate the HQ-SAM's capability for semantic water body extraction using a combination of both optical and SAR remote sensing imagery, addressing a gap in existing literature.
• It provides a quantitative comparison, showing HQ-SAM's superior performance (approximately 4% higher accuracy) over traditional water index-based methods for water body delineation.
• The research highlights HQ-SAM's robustness across diverse lake sizes, shapes, and various remote sensing data types, while preserving georeference information, which is often lost in other deep learning models.
• It demonstrates HQ-SAM's utility for multi-temporal analysis and monitoring of lake dynamics, linking observed changes to flood management activities and natural factors.
• The study identifies practical considerations for HQ-SAM application, including its strict 8-bit 3-band RGB input requirement and high computational resource demands.

Funding

• Basic Science Research Program through the National Research Foundation of Korea (NRF) funded by the Ministry of Education (2021R1A6A1A03044326).

Citation

@article{Quang2025Semantic,
  author = {Quang, Nguyễn Hồng and Kim, Namhoon and Lee, Hanna and Ahn, Seunghyo and Kim, Gihong},
  title = {Semantic water body extraction by the high-quality segment anything model using multiple optical and SAR imagery},
  journal = {Acta Geophysica},
  year = {2025},
  doi = {10.1007/s11600-025-01732-0},
  url = {https://doi.org/10.1007/s11600-025-01732-0}
}