Evaluation of Deep Learning-Based Water Bodies and Flooded Area Detection with Nanosatellites: The PlanetScope Satellite Imageries and HRNet Model

The continuous monitoring of water body is essential for efficient water resource management and the prevention of water-related disasters.Microsatellite and nanosatellite imageries provide a tool for continuous and accurate monitoring of water bodies at high spatial and temporal resolution.In this study, PlanetScope imagery with a resolution of 3.7 m and High-Resolution Network (HRNet) model were used to detect water bodies at dams and rivers in Korea, with the objective of evaluating the utility of water surface area monitoring.The HRNet model and the optimal band combinations of PlanetScope imagery which were R+G+B, R+G+B+NIR, Normalized Difference Water Index (NDWI), and Green+NIR+ NDWI, were initially evaluated.The Green+NIR+NDWI combination performed the best, with an accuracy of 0.91 and loss function of 0.05 for the validation set.Water body detection was performed using the HRNet model with the optimal band combination and models from previous studies (Otsu, K-means, U-net) The performance was evaluated through quantitative validation using labeled images.The HRNet model showed the best performance with an Intersection over Union (IoU) of 0.96, compared to models K-means (Krishna and Murty, 1999).