High accuracy water level estimation using super-resolution Sentinel-1 data

Water level can be effectively monitored through satellite remote sensing. However, a persistent challenge lies in balancing spatial coverage and resolution. High-resolution data provide detailed information but suffer from limited spatial coverage, while low-resolution data are inadequate for accurate large-scale monitoring. In this paper, a super-resolution approach with machine learning is proposed to enhance the accuracy of water level monitoring as a solution to this problem. By leveraging Real-world Enhanced Super-Resolution Generative Adversarial Network (Real-ESRGAN) and a high-quality pretrained model, this study enhances the spatial resolution of Sentinel-1 SAR imagery, whose features are then used to develop a water level estimation model, enabling dynamic monitoring of river water levels. Results show that after super-resolution, the Root Mean Squared Error (RMSE) is reduced to 0.322 m (∼30.6% decrease), the Nash-Sutcliffe Efficiency (NSE) increases to 0.957, and the Pearson Correlation Coefficient (R) improves to 0.987. The proposed super-resolution approach improves both spatial resolution (to 2.5 m/pixel) and water level estimation accuracy. Moreover, the results from representative stations, including densely urbanized regions, agricultural irrigation zones, and areas near sluices, show that our super-resolution approach provides accurate water level estimates and reliable trend predictions. Our experiments confirm its usefulness across a diverse spectrum of high accuracy hydrological monitoring scenarios.