Fusing SAR image and CYGNSS data for monitoring river water level changes by machine learning

Abstract

Accurate river water level estimation is essential for effective flood monitoring and water resources management. However, traditional techniques and single satellite observations have low accuracy and resolution. In this paper, we propose a novel method to enhance river water level estimation by fusing Cyclone Global Navigation Satellite System (CYGNSS) data and Sentinel-1 Synthetic Aperture Radar (SAR) imagery based on advanced machine learning (ML) techniques. SAR provides high-resolution, all-weather surface imagery, while the GNSS-Reflectometry from the eight-satellite CYGNSS mission offers frequent and wide-coverage observations. Dynamic river water levels are obtained at a daily temporal resolution by extracting changes in Sentinel-1 backscattering coefficients and integrating them with the CYGNSS data's high temporal resolution feature. To guarantee the model's robustness, a ten-fold cross-validation (CV) procedure is used with incorporating 15 uniformly distributed gauge sites. Experimental results show that the data fusion method significantly improved the temporal resolution, and more importantly the precision of water level estimation. As opposed to the model without data fusion, the optimized fusion algorithm achieved a 50.74 % reduction in RMSE from 0.341 to 0.168 m, while the R was improved from 0.876 to 0.936. An improvement of over 35 % in RMSE was observed at 8 out of 15 stations. To further validate the model's generalizability, we tested it using data from 8 spatially and temporally independent hydrological stations. The fusion method reduced the RMSE from 0.479 to 0.202 m and increased the R from 0.848 to 0.927, further confirming its effectiveness in enhancing water level estimation. The findings indicate that integrating SAR imagery and CYGNSS time series data has complementary effects and enables better water level estimation.