Abstract

Soil water is a critical component of the water balance to make water management decisions at multiple scales. While soil water can be sensed remotely, this is generally at coarse scales (> 12.5 km). In addition, soil moisture products developed at field scale resolutions (< 250 m) have been mostly limited to shallow observations (up to 10 cm depth) and are impacted by land use. The objective of this study was to create an accurate downscaled soil water content product at multiple depths and at a fine 90 m resolution by fusing modelled and remote sensing datasets via deep learning. Reference data was based on the OzFlux and Oznet networks. Covariates included the North America Space Agency (NASA) United States Department of Agriculture (USDA) Soil Moisture Active Passive (SMAP) remote sensing data assimilation model, Sentinel 1 from Copernicus, surface reflectance, land surface temperature and land cover from the Moderate Resolution Imaging Spectroradiometer (MODIS), and gridded soil properties. Two model approaches were used including a multilayer perceptron for the surface (0-10 cm), and recurrent neural networks for the surface/subsurface (0-30 cm and 30-60 cm) soil water content. The surface prediction performance resulted in a root mean square error (RMSE), mean absolute error (MAE) and Pearson's correlation of 0.073, 0.057, and 0.74, degrading in depth to 0.07, 0.062, and 0.5. Overall, these 90 m resolution predictions improve on the performance of NASA-USDA SMAP (10 km resolution) and the Australian Landscape Water Balance (5 km resolution) simulated soil water contents. Land use/land cover (LULC) are important explaining factors for performance and SHapley Additive exPlanations (SHAP) indicate high impor-tance of NASA-USDA SMAP for surface predictions, while soil properties and LULC increase in importance with depth.