Estimation of soil moisture at different soil levels using machine learning techniques and unmanned aerial vehicle (UAV) multispectral imagery

Soil moisture is a key component of water balance models. Physically, it is a nonlinear function of parameters that are not easily measured spatially, such as soil texture and soil type. Thus, several studies have been conducted on the estimation of soil moisture using remotely sensed data and data mining techniques such as artificial neural networks (ANNs) and support vector machines (SVMs). However, all models developed based on these techniques are limited to site-specific applications where they are trained and their parameters are tuned. Moreover, since the system of non-linear equations produced by and conducted in the machine learning process are not accessible to researchers, each application of these machine learning approaches must repeat these training steps for any new study area. The fact that the results of this machine learning, black box approach cannot be easily transferred to different locations for extraction of soil moisture estimates is frustrating, and it can lead to inaccurate comparisons between methods or model performance. To overcome the Black-box issue, this study employed a powerful technique called genetic programming (GP), which is a combination of an evolutionary algorithm and artificial intelligence, to simulate soil moisture at different levels using high-resolution, multispectral imagery acquired with an unmanned aerial vehicle (UAV). The output of this approach is either a linear or nonlinear empirical equation that can be used by others. The performance of GP was compared with ANN and SVM modeling results. Several sets of high-resolution aerial imagery captured by the Utah State University AggieAir UAV system over two experimental pasture sites located in northern and southern Utah were used for this soil moisture estimation approach. The inputs used to train these models include the reflectance for the visible, near-infrared (NIR), and thermal bands. The results show (1) the performance of GP versus ANN and SVM and (2) the master equation provided by GP, which can be used in other locations and applications.