Estimation of soil moisture content under high maize canopy coverage from UAV multimodal data and machine learning

An accurate in-field estimate of soil moisture content (SMC) is critical for precision irrigation management. Current ground methods to measure SMC were limited by the disadvantages of small-scale monitoring and high cost. The development of unmanned aerial vehicle (UAV) platforms now provides a cost-effective means for measuring SMC on a large scale. However, previous studies have considered only single-sensor estimates of SMC, so the combination of multiple sensors has yet to be thoroughly discussed. Additionally, the way in which soil depth, canopy coverage, and crop cultivars affect the SMC-estimation accuracy remains unclear. Therefore, the objectives of this study were to (1) evaluate the SMC-estimation accuracy provided by multimodal data fusion and four machine learning algorithms: partial least squares regression, K nearest neighbor, random forest regression (RFR), and backpropagation neural network (BPNN); (2) discuss the accuracy of the remote-sensing approach for estimating SMC at different soil depths, and (3) explore how canopy coverage and crop cultivars affect the accuracy of SMC estimation. The following results were obtained: (1) Data from multispectral sensors provided the most accurate SMC estimates regardless of which of the four machine learning algorithms was used. (2) Multimodal data fusion improved the SMC estimation accuracy, especially when combining multispectral and thermal data. (3) The RFR algorithm provided more accurate SMC estimates than the other three algorithms, with the highest accuracy obtained by combining data from RGB, multispectral, and thermal sensors with an R2 = 0.78 (0.78) and a relative root-mean-square error of 11.2% (9.6%) for 10-cm-deep (20-cm-deep) soil. (4) UAV-based SMC-estimation methods provided similar, stable performance for SMC estimates at various depths and even yielded smaller relative error for deeper estimates (20 cm). (5) The RFR and BPNN machine learning algorithms both provided relatively accurate SMC estimates for modest canopy coverage (0.2–0.4) but relatively poor estimates for higher (>0.4) or lower (