Development of Soil Properties Nonlinear Prediction Model with Variable Region Selection and Applicability Domain

Abstract

In precision agriculture, near-infrared (NIR) spectroscopy is a useful tool to predict soil properties. When applying statistical learning methods to near-infrared spectroscopy, wavelength selection becomes an inevitable issue due to the wide range of wavelengths measured. By comparing linear regression method and nonlinear regression method combined with variable region selection, it could be confirmed that some variables in NIR spectroscopy are nonlinearly related to soil properties.　Additionally, soil properties are quite different for each area, so this large inter-area variability makes the prediction of new areas difficult. From this premise, multiple Bayesian ensemble regression is proposed for solving this applicability domain problem. For NIR spectroscopy data coming from different sources, prediction results of the proposed method were shown to be significantly superior to traditional modeling methods.