Evaluating soil moisture content under maize coverage using UAV multimodal data by machine learning algorithms

Abstract

Timely and accurate estimation of soil moisture content (SMC) is essential for precise irrigation management at the farm scale. Unmanned aerial vehicle (UAV) remote sensing with a high spatiotemporal resolution has become a promising method for SMC monitoring. Many existing SMC models have only been tested at a specific crop growth stage using a single type of sensor, and the effects of growth stage and irrigation variation on SMC estimation accuracy remain unclear. To address these limitations, this study used UAV-based multimodal data to quantify SMC in a maize field under various levels of irrigation over two years using three machine learning algorithms (MLA): partial least squares regression (PLSR), K nearest neighbor (KNN), and random forest regression (RFR). The results demonstrated that multimodal data fusion improves the SMC estimation accuracy regardless of the MLA, especially the joint use of thermal and multispectral data. Among three SMC regression models, the RFR model produced the most accurate SMC estimation for the two growing seasons regardless of sensor combinations. The RFR model using all three types of data generated the most accurate and robust SMC estimation at the vegetative stage with R2 of 0.68 and 0.78, and rRMSE of 20.82% and 19.36% for 10- and 20-cm soil depths, respectively; it also produced the best SMC estimation accuracy under well-watered and mild to modest deficit irrigation treatments for both soil depths. The study shows that the high spatial–temporal maps of SMC using UAV-based multimodal data has promising potential for supporting decision-making in irrigation scheduling at the farmland scale.