Performance Evaluation of Water-Body Monitoring Using High-Resolution Optical Satellite Imagery and Deep Learning Models: CAS500-1 Satellite Imagery and Transformer Model

Monitoring water systems using satellite remote sensing data is essential for preventing water disasters across wide areas. CAS500-1 satellite imagery provides high spatial resolution, enabling continuous monitoring of water systems. With recent advancements in artificial intelligence (AI), deep-learning-based object-detection techniques have demonstrated high accuracy in detecting water bodies, even in complex terrains and diverse land-cover environments. Transformer-based deep-learning models can capture global context information, allowing for more detailed and accurate water detection compared to conventional AI models. In this study, the applicability of transformer-based models, including ViT, SegFormer, Swin Transformer, and MaskFormer, was evaluated using CAS500-1 imagery for water-body detection. Each model was trained using RGB bands from CAS500-1 imagery to assess water-body detection performance; the MaskFormer model yielded the highest accuracy, with an Intersection over Union (IoU) of 0.932. Furthermore, additional experiments were conducted with the MaskFormer model using different band combinations (R-G-B, R-G-NIR, NDWI, and G-NIR-NDWI), and the R-G-NIR combination yielded the best performance, with an IoU of 0.951. The findings indicate that continuous water-body detection and hydrological monitoring can contribute to the effective management of water resources and disasters.