Statistical and machine learning for improving the utility of soil moisture networks for satellite‐based soil moisture product assessment in the Upper Missouri River Basin

Surface soil moisture is essential for understanding energy exchange between land and atmosphere. Mesoscale modeling and remote sensing are two methods that can provide useful surface soil moisture information for helping characterize these exchanges. While ground‐based sensors are also used, many sensor technologies and network designs lack measurements at the 5‐cm depth often targeted by land surface models and satellite‐based sensors. If deeper soil moisture measurements could be leveraged to accurately estimate near‐surface conditions, the utility of in situ networks operating at various depths would be significantly improved. This study develops a lifted 5‐cm soil moisture in the Upper Missouri River Basin to estimate 5‐cm soil moisture from 10‐cm and 20‐cm sensor data and antecedent precipitation indices, using statistical and machine learning approaches. Model selection was based on three rigorous testing scenarios designed to evaluate different data availability conditions. Across these testing scenarios, we found that statistical multiple linear regression was the most suitable approach for this task due to its simplicity, interpretability, and excellent performance with unbiased root mean square error of 0.018–0.027 m 3 /m 3 . At least 87% of the soil moisture 5‐cm estimates across the three testing scenarios achieved errors under 0.04 m 3 /m 3 . The estimated soil moisture at 5 cm can be effectively used for evaluating Soil Moisture Active and Passive soil moisture accuracy. This study provides a method to significantly increase the availability of 5 cm soil moisture estimates in in situ networks, especially in the Montana Mesonet.