Lake detection and semantic segmentation using a deep learning model and Kompsat-5 images

• Published in: IOP conference series. Earth and environmental science Vol. 1501; no. 1; pp. 12015 - 12025
• Main Authors: Hong Quang, Nguyen, Lee, Hanna, Thi Thu Hue, Ha, Park, Sumin, Kim, Gihong
• Format: Journal Article
• Language: English
• Published: Bristol IOP Publishing 01.05.2025
• Subjects: Accuracy; Deep learning; Ground truth; Image resolution; Kompsat-5; lake; Lakes; Machine learning; Noise reduction; Radar imaging; Remote sensing; Resource management; Satellite imagery; Semantic segmentation; South Korea; Surface water; Synthetic aperture radar; Water resources management; YOLO
• ISSN: 1755-1307; 1755-1315; 1755-1315
• DOI: 10.1088/1755-1315/1501/1/012015

Surface water detection and extraction from remote sensing data is a favorable study topic due to the critical importance of this indispensable resource. For water resource management implementation, lake and reservoir detection and segmentation can supply valuable information. Korea Aerospace Research Institute (KARI) has developed and been operating the Kompsat-5 satellite acquiring high-resolution Synthetic-aperture radar (SAR) images. The Kompsat-5 sensor transmits a short radar wavelength of band-x (3.2 cm) obtaining more details of observed objects. However, that also results in more noise on the water surfaces when they are affected by floating vegetation and waves. Hence, the noise will reduce the accuracy of lake surface extraction using traditional methods such as thresholding and even some machine learning models. Therefore, this study aims to fine-tune a modern deep learning model of You Only Look Once (YOLO) version 8 for detecting lakes and segmenting their boundaries using the Kompsat-5 images for regions of the Republic of Korea. This study’s results show the robustness of the YOLOv8 with its accuracy greater than 80% of lake detection and extraction compared to ground truth masks (link data).