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

摘要

Soil moisture is a key component of water balance models. Physically, it is a nonlinear function of parametersthat are not easily measured spatially, such as soil texture and soil type. Thus, several studies have beenconducted on the estimation of soil moisture using remotely sensed data and data mining techniques such asarti cial neural networks (ANNs) and support vector machines (SVMs). However, all models developed basedon these techniques are limited to site-specific applications where they are trained and their parameters aretuned. Moreover, since the system of non-linear equations produced by and conducted in the machine learningprocess are not accessible to researchers, each application of these machine learning approaches must repeat thesetraining steps for any new study area. The fact that the results of this machine learning, black box approachcannot be easily transferred to different locations for extraction of soil moisture estimates is frustrating, andit can lead to inaccurate comparisons between methods or model performance. To overcome the Black-boxissue, this study employed a powerful technique called genetic programming (GP), which is a combination of anevolutionary 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 eithera linear or nonlinear empirical equation that can be used by others. The performance of GP was comparedwith ANN and SVM modeling results. Several sets of high-resolution aerial imagery captured by the Utah StateUniversity AggieAir UAV system over two experimental pasture sites located in northern and southern Utah wereused for this soil moisture estimation approach. The inputs used to train these models include the reectance forthe visible, near-infrared (NIR), and thermal bands. The results show (1) the performance of GP versus ANNand SVM and (2) the master equation provided by GP, which can be used in other locations and applications.