Abstract / Description
Abstract Even though nutrient and irrigation management could be improved by accurate soil moisture monitoring, a cost‐effective approach is not available for obtaining this information. The objective of this study was to use soil‐ and weather‐derived predictors to estimate volumetric water content at multiple depths. Two multi‐site data sets were used in this study. One dataset was derived from preplant soil samples collected from 19 sites located in soils with an ustic soil moisture regime between 2019 and 2022, and the second dataset was collected at sites that had udic soil moisture regime. At the second dataset, the soil sensors collected volumetric soil information near‐continuously between 2021 and 2023. In the study, the performance of multiple linear regression, support vector regression, random forest, extreme gradient boosting, and a stacked feed‐forward neural network meta‐learner model were compared. The machine learning models for the preplant gravimetric soil moisture collected from sites that had an ustic soil moisture regime had low predictability (R2 < 0.3) and high root mean square errors. These results suggest that the models did not adequately capture the nonlinear and spatio‐temporal drivers governing soil moisture variability. Machine learning models that estimated in‐season soil moisture at sites located in the udic soil moisture regime had higher predictability, and the best models had R2 values as high as 0.99. The high performance of these models was attributed to the use of temporal predictors, and in general stacking provided consistent gains (up to ΔR2=+0.114 at 10 cm; mean R2¯=0.618 across depths) but offered limited benefit when temporal features are included. Overall, the results show that incorporating temporal context and, when needed, stacking complementary learners can produce more reliable soil moisture estimates for field‐scale management.