Vis-NIR spectroscopy coupled with machine learning algorithms to predict soil gypsum in calcareous soils, southern Iran

Abstract

The use of soil spectral reflectance, which has been introduced as a new method in soil science, is widely used in estimating the physicochemical properties of soil. The aim of this study was to estimate the amount of gypsum in surface soils of Fars province. Based on random sampling method, 100 soil samples were collected and measured by standard method. Spectral analysis of soil samples was performed using a spectrophotometer in the range of 2500-400 nm. After this stage, various preprocessing methods were evaluated and finally the percentage of soil gypsum was modeled using two models of partial least squares regression (PLSR) and support vector regression (SVR). The results showed that the best results for estimating the percentage of soil gypsum are related to the SVR model with Preprocessing Savitzky- Golay Filter with the first derivative. Also, according to RPIQ statistics, the estimation of PLSR model for the percentage of soil gypsum in the weak class is 1.02% and for the SVR model in the moderate class is 1.54%.