Flood inundation mapping with CYGNSS over CONUS: a two-step machine-learning-based framework

Abstract

Accurate and timely flood inundation mapping is vital for early warning, disaster response, and mitigation planning. Global Navigation Satellite System-Reflectometry (GNSS-R) at L-band shows promise in detecting inundation extent, especially over land and lightly vegetated areas. However, the complex interaction between land surface features and the bistatic configuration of GNSS-R presents challenges for reliable flood mapping. In this work, a machine learning (ML) framework was developed to retrieve fractional inundation as the area proportionally covered by water using bistatic reflectance observations acquired from Cyclone GNSS (CYGNSS) and ancillary variables to characterize land surface conditions. Active C-band synthetic aperture radar (SAR) based high-resolution flood maps derived from Sentinel-1 were used as the reference for ML model training. The Random Forest (RF) model was used to retrieve surface water fraction in two steps through inundated pixel classification and water fraction regression. The sequential two-step (STS) structure was compared with the parallel two-step (PTS) model. Results show that the STS model outperforms both the PTS and the single regressor in deriving daily CYGNSS inundation retrievals at a 3-km resolution across the contiguous United States (CONUS). Cross-validation using a leave-one-year-out approach yields a correlation coefficient of 0.762 [-1 and root-mean-square-error of 0.039 [-1 between the CYGNSS inundation retrievals and the reference SAR-based water fractions. Consistent spatial variations are found between CYGNSS and Sentinel-1 inundated regions, suggesting satisfactory performance of the proposed ML model. In addition, the CYGNSS inundation are compared against several other inundation products, including the official CYGNSS water mask product, one semi-empirical method-based CYGNSS product, and a microwave remote sensing inundation product. Our CYGNSS product shows comparable performance in characterizing flood inundation at a 3-km resolution and daily temporal frequency.