Minimum variance based-Bayes Combination for prediction of soil properties on Vis-NIR reflectance spectroscopy

Abstract

Traditional laboratory methods for determining soil properties require a great deal of time and expense, while reflectance spectroscopy technology is a fast, inexpensive, and convenient way to predict physical and chemical soil properties. This technology in the spectral range of 400–2500 ​nm (Vis-NIR) as a suitable alternative method to get soil properties are accompanied by problems and challenges to extract considered properties. In this paper, we propose Minimum Variance based-Bayes Combination (MVBC) method to predict the soil properties. In the proposed MVBC method, we design two steps, prediction and combination for the training phase. Firstly, in the prediction step, five regression methods, i.e., partial least squares regression (PLSR), kernel Ridge regression (KRR), linear regression (LR), gradient boosting regression (GBR) and random forest (RF) method used to calculate and estimate nine soil properties, i.e., CaCo3, CEC, Clay, N, OC, PH in CaCl2, PH in H2O, Sand and Silt, separately. Secondly, in the combination step, the estimation errors of all regressions in the prediction step are determined to assign appropriate weight to each of them in the Bayesian framework based on minimum variance. These two steps are repeated until the final estimation error reaches an acceptable minimum value. Finally, these results and the trained system are used for the test phase. Experiments are reported to evaluate the effectiveness of the proposed MVBC method on real soil data, which shows the good performance of the proposed method with better results than other methods.