High spatio-temporal resolution soil moisture nowcasting at multiple depths with data-driven approaches

Soil moisture nowcasting provides valuable information for site-specific management in dryland cropping systems. The increasing publicly available data streams have made it possible to capture soil moisture across the profile at fine spatiotemporal resolution. While many studies have applied data-driven approaches, they are generally limited to moderate to coarse spatial resolution and focus on the soil surface. This study investigated the importance of water-related features and showcased a data-driven practice that integrate multi-source water-related data streams for high-resolution soil moisture nowcasting (< 100 m, daily) throughout the soil profile. The models were evaluated with a series of cross-validation experiments, including spatial interpolation, temporal prediction, spatio-temporal prediction, gap-filling and spatial extrapolation. The best performance was observed in the Adelong Creek catchment using RF, with ubRMSE= 0.051 m 3 /m 3 , R= 0.85, and LCCC= 0.82 for spatial interpolation; ubRMSE= 0.041 m 3 /m 3 , R= 0.89, and LCCC= 0.89 for temporal prediction; ubRMSE= 0.051 m 3 /m 3 , R= 0.85, and LCCC= 0.72 for spatio-temporal prediction; and ubRMSE= 0.062 m 3 /m 3 , R= 0.76, and LCCC= 0.73 for spatial extrapolation. Additionally, XGBoost achieved the best performance for gap-filling, with ubRMSE= 0.025 m 3 /m 3 , R= 0.96, and LCCC= 0.96. Our work has the potential to provide an information platform for growers to monitor and understand soil moisture at fine resolution in the future. • Nowcasting soil moisture at fine spatio-temporal resolution (<100 m, daily) for multilayered soil. • Comprehensive cross-validation experiments that evaluate model from multiple aspects. • A unifying analytical framework for data-driven modelling with machine learning. • A benchmarking dataset for building up data-driven models.