DiffWater: A Conditional Diffusion Model for Estimating Surface Water Fraction Using CyGNSS Data

Recent advances in cyclone global navigation satellite system (CyGNSS) data have significantly improved the extraction of monthly surface water fraction (SWF), with neural networks being widely used for large-scale water body mapping based on global navigation satellite system-reflectometry (GNSS-R) signals. However, inherent noise in CyGNSS signals, such as multipath effects and interference, presents substantial challenges to the accuracy of SWF estimation. Diffusion models, an emerging class of generative deep learning techniques, have shown remarkable capabilities in capturing complex data distributions. By leveraging an iterative process of noise addition and removal, these models demonstrate significant advantages in processing low signal-to-noise ratio data, offering a novel methodology for precise SWF estimation from CyGNSS data. This study introduces DiffWater, a framework designed to address the unique characteristics of CyGNSS data and systematically explore the applicability of conditional diffusion models for remote sensing tasks. Utilizing a composite reference dataset, which includes the global surface water (GSW) dataset and the global surface water dynamics (GLAD) dataset as training targets, DiffWater enhances the objectives of conditional diffusion models by integrating advanced conditional feature extractors and implementing multilevel fusion of conditional and temporal features, thereby achieving significant improvements in SWF estimation performance. Comprehensive experimental evaluations on the reference dataset demonstrate that DiffWater achieved the best performance, with a root-mean-squared error (RMSE) of 4.987% and a correlation coefficient (R) of 0.946. Compared to state-of-the-art SWF estimation methods, the proposed approach demonstrated significant improvements in both quantitative and qualitative results.