ABSTRACT
Soil water content monitoring is an integral part of irrigation water management for efficient crop production. Recent developments in machine learning (ML) algorithms and widely available remote sensing products have made regression-based prediction and forecasting of soil moisture appealing due to lower costs than in-situ measurements. In this chapter, we investigated the efficiency of three machine learning algorithms, i.e., Random Forests (RF), Support Vector Regression (SVR), and Neural Networks (NN), for predicting soil moisture at 5 cm, 10 cm, 20 cm, 51cm, and 102 cm soil depths trained using the in-situ soil moisture retrievals from the Soil and Climate Analysis Network (SCAN) site in the Texas-Gulf regions. Explanatory variables included soil texture, surface, and root zone moisture obtained from Soil Moisture Active Passive (SMAP), vegetation and water indices obtained from MODIS Terra, and climate variables. The Permutation Feature Importance (PFI) method was applied to all three algorithms to identify the importance of the explanatory variables and the top features were reported based on an increase in Root Mean Square Error in the models. The feature importance measures indicated a strong influence of soil texture, elevation, and SMAP-derived root zone soil moisture on model predictions at all depths. Results from this study showed that the predictive performances at most depths for all three algorithms were similar and were satisfactorily able to capture the spatial and temporal variability in soil water content.
