Optimizing soil moisture and water content estimation for wheat based on machine learning models

Abstract Soil moisture is essential for various fields, including agriculture, climate research, and the management of water resources. Nevertheless, conventional techniques for assessing soil moisture, such as the gravimetric method and ground-based sensors, frequently face challenges related to their scalability and efficiency. Therefore, the objective of this research is to enhance the estimation of soil moisture (SM) and Vegetation Water Content (VWC) for wheat through the application of single and stacking ensemble machine learning models based on different input scenarios. Four single machine learning algorithms were employed: Random Forest (RF), Extreme Gradient Boosting (XGB), Support Vector Regression (SVR), and Multilayer Perceptron Regression (MLPR), the input variables consisted of vegetation indices derived from Sentinel 2 data. The stacking model achieved R² values of 0.929 and 0.928, with RMSE values of 6.57% and 6.604% for VWC in scenarios Sc3 and Sc7, respectively. For SM, the stacking model yielded R² values of 0.676 and 0.675, with RMSE values of 3.98% for scenario Sc8. Moreover, by knowing the VWC, SM can be predicted by R² of 0.66 which can lead to irrigation scheduling. In general, the stacking model can enhance the accuracy of SM and VWC predictions compared to more widely used machine learning models, as it effectively addresses the limitations of the individual base algorithms while keeping the parameter count low and ensuring quick computation times. Also, the stacking ensemble model showed good applicability during the main growth stages of wheat. This study suggests that the application of Sentinel 2 images along with a stacking model may enhance the accuracy of SM estimation in regions where wheat is cultivated.