Ensemble Deep Learning Towards High-Resolution Soil-Moisture Mapping for Enhanced Water Management in California's Central Valley

Abstract

Soil moisture (SM) plays a vital role in both hydrological and agricultural processes and is critical for achieving groundwater sustainability in agriculture through demand management. NASA's Soil Moisture Active Passive (SMAP) satellite measures the SM across the Earth and provides data on both the surface and root zone SM but at a coarse spatial resolution of 9 km, thereby limiting detailed analyses. This study aimed to develop an optimized deep ensemble learning framework to downscale the resolution of SMAP observations of California's Central Valley from 9 km to 30 m for both the surface and root-zone SM. Sensitivity analysis was employed to identify key explanatory variables. The models were then combined into an ensemble DNN trained on multiscale SMAP data and validated against in-situ SM measurements. The results demonstrated that the ensemble model achieved the highest correlations of 0.789 and 0.683 for surface and root-zone SM, respectively, with the lowest root mean square errors of 0.0281 and 0.0814 cm3/cm3, respectively, thereby reliably reproducing the temporal dynamics. Seasonal analysis revealed distinct patterns linked to climate and management practices at a spatial resolution of 30 m, thereby capturing seasonal variations in soil moisture among major crops. Additionally, SM maps can be used to refine the estimated evapotranspiration resulting from applied irrigation water sourced from groundwater pumping, allowing for better monitoring of water use. SM can also be used to inform agronomic practices, such as delayed irrigation in early spring, which can reduce groundwater demand.

Keywords: deep learning, soil moisture mapping, Water management, California's Central Valley, ensemble modeling, SMAP satellite, root-zone soil moisture, downscaling, evapotranspiration, groundwater sustainability, soil water