Sentinel-2 영상과 딥러닝 기반 객체 탐지 모델을 활용한 하천 모니터링 = River monitoring using Sentinel-2 imagery and deep learning-based object detection models

다국어 초록 (Multilingual Abstract)

Efficient water system monitoring is an essential element for securing stable water resources and preventing water-related disasters. The significance of monitoring has increased in the context of rising frequencies of droughts and floods, a phenomenon attributable to recent climate change. Satellite remote sensing, in particular, has become a core tool in water resource management due to its ability to enable periodic detection across a wide spatial range. Accordingly, the present study was conducted with the objective of detecting water bodies, with a focus on domestic rivers. The study employed Sentinel-2 satellite imagery with a resolution of 10 meters and the Swin-Transformer model. The objective of the study was to identify alterations resulting from watershed monitoring and to assess the model's applicability. Initially, a comparative analysis was conducted among the classical water body detection techniques Otsu, Kittler-Illingworth (KI), and the Convolutional Neural Network (CNN)-based models U-net and High-Resolution Network (HRNet). These techniques were evaluated against a set of manually digitized and labeled images. The Swin-Transformer model demonstrated optimal performance, achieving an accuracy of 0.99, an Intersection over Union (IoU) of 0.88, and an F1-score of 0.87. Furthermore, the Swin- Transformer model, when employed for the purpose of evaluating the performance of watercourse change detection, demonstrated a high degree of efficacy, as evidenced by the high values obtained when evaluating the model's performance against labeled images. The model demonstrated an accuracy of 0.99, an IoU of 0.83, and an F1-score of 0.90. While changes in broad watercourse areas can be detected with high accuracy, limitations existed in detecting water bodies at the boundary edges of sandbars within thin tributaries and watercourse areas with complex structures. The Swin-Transformer model is expected to enable stable monitoring of water systems across diverse terrains, even in medium-to-low resolution satellite imagery. It is anticipated that the utilization of this technology will be effective in detecting changes in water systems and in disaster response.