A Study on Improving Water Body Detection Accuracy in CAS500-1 Satellite Imagery Using Deep Learning SeoJin Kong(a), Wonwoo Seo(a), SooAhm Rhee(a*)
(a) Image Eng. Research Center, 3DLabs Co. Ltd, Republic of Korea,
(*)ahmkun[at]3dlabs.co.kr
Abstract
Efficient management and continuous monitoring of water resources is essential for agriculture, urban, disaster response and other sectors. Accordingly, the demand for automated water body detection techniques is steadily increasing. CAS500-1, a high-resolution satellite developed in Korea, provides Analysis Ready Data (ARD), including surface reflectance images and additional information such as water body and cloud masks. The water body mask is currently created through manual digitization, which is time-consuming, costly, and limited in reflecting temporal changes. This study aims to automate and improve the accuracy of water body detection in CAS500-1 satellite imagery through deep learning models. To this end, deep learning models for semantic segmentation models-U-Net and HRNet-were applied and their performance was compared. And the resulting water body masks were evaluated against existing labeled data. The results showed high performance: U-Net achieved an F1-score of 0.95 and IoU of 0.91, while HRNet achieved an F1-score of 0.92 and IoU of 0.86. Notably, the models were able to distinguish small objects such as ships and bridges from water bodies, even when such details were absent in the label data. This study overcomes the limitations of the existing manual method and enables automated detection of water body using high-resolution satellite imagery. It facilitates continuous and precise monitoring of surface water areas, and is expected to contribute meaningfully to decision-making processes related to water resource utilization.
Keywords: ARD, CAS500-1, Deep Learning, Semantic Segmentation, Water body detection
Topic: Topic C: Emerging Technologies in Remote Sensing
