Fine-Tunned Segment Anything Model (SAM) for Reservoir Extractions Compared With Popular CNNs: An Experiment for Space-Borne Synthetic-Aperture Radar Images

• Published in: IEEE access 2025, Vol.13, p.1727-1750
• Main Authors: Quang, Nguyen Hong, Lee, Hanna, Kim, Eui-Myoung, Kim, Gihong
• Format: Article
• Language: English
• Subjects: Accuracy; Artificial neural networks; Data mining; Data models; Datasets; Feature extraction; Fine-tunning; Ground truth; Image segmentation; Lakes; Machine learning; Masks; Radar data; Radar imaging; Radar polarimetry; Remote sensing; Republic of Korea; reservoir; Reservoirs; Residual neural networks; SAM; Segments; Semantic segmentation; space-borne SAR; Surface water; Synthetic aperture radar
• ISSN: 2169-3536; 2169-3536
• DOI: 10.1109/ACCESS.2024.3516519

The freshwater resource is invaluable and indispensable for any nation like the Republic of Korea. Recently, deep learning (DL), AI models have become more popular and applied frequently for surface water studies. The Segment Anything Model (SAM) has been developing sharply and takes an adaptable approach with the ability to perform zero-shot inference. Although, pre-trained SAM was trained with millions of images (a billion masks), applying it to remote sensing data reveals limitations of inaccurate results and unlabeled classes, particularly for the more complex and noise data from Synthetic-aperture Radar Images. Hence, we fine-tune the SAM model and other popular CNN models of YOLOv8, U-net(ResNet50), and DeepLab(ResNet50, EfficientNet) for lake semantic segmentation using multi-SAR RS datasets of Kompsat-5, ALOS-2, Sentinel-1, and a combination of the three datasets (data link) for model result comparisons. This study's accuracy assessment showed the SAM was the most precise model (accuracy overall \approx ~0.95 ) followed by the DeepLab(ResNet50), YOLOv8, U-net(ResNet50), and DeepLab(EfficientNet) model. Almost all models segmented highly accurate lake areas fitted well with ground-truth masks, nevertheless, the SAM (code link) presented the most well-performance model. The YOLO deals well with larger datasets and requires deeper trains to gain higher accuracy outputs. In addition, this research investigated the responses of each DL model to the SAR dataset proving a need for fine-tuning model results on SAR RS for better lake segmentations.