Regression-based surface water fraction mapping using a synthetic spectral library for monitoring small water bodies

Small water bodies (SWBs), such as ponds and on-farm reservoirs, play a crucial role in agriculture irrigation, carbon storage, and biogeochemical cycle. Medium-spatial-resolution satellite imagery such as Sentinel-2 imagery has been widely promoted to monitor SWBs, due to its relatively fine spatial and temporal resolution. However, the small size and diverse spectral characteristics of SWBs present significant challenges, particularly the mixed-pixel problem, where both water and land classes contribute to the observed spectral response of the image pixel. To address this issue, we propose a novel regression-based surface water fraction mapping method (RSWFM) that leverages a random forest regression model and a synthetic spectral library to generate 10 m spatial resolution surface water fraction maps from Sentinel-2 imagery. RSWFM incorporates a compact set of endmembers, representing water, vegetation, impervious surfaces, and soil, to simulate a spectral library using both linear and nonlinear mixture models, while accounting for spectral variability across diverse SWBs. Additionally, to enlarge the number of pure spectra and enhance their representativeness for training, RSWFM applies data augmentation based on Gaussian noise. The performance of RSWFM was assessed across ten study sites with hundreds to thousands of SWBs smaller than 1 ha and was compared with fully constrained least squares (FCLS) linear spectral mixture analysis, multiple endmember spectral mixture analysis (MESMA), and random forest (RF) regression without data augmentation. Results indicated that RSWFM generates a low root mean square error (RMSE) of less than 0.09, reducing by approximately 30%, 15%, and 11% compared to FCLS, MESMA, and nonlinear RF regression without data augmentation, respectively. Furthermore, RSWFM achieves an R² of approximately 0.85 in estimating the area of SWBs smaller than 1 ha. This study demonstrates the potential of RSWFM for addressing the mixed pixel problem in SWB monitoring across large areas.