Modeling soil cation concentration and sodium adsorption ratio using observed diffuse reflectance spectra

Abstract

Spectral analysis is a useful tool for the rapid and accurate prediction of soil properties. Our objective is to select the best model for predicting the three soil cation concentrations ([Na+], [Mg2+], and [Ca2+]) and sodium adsorption ratio (SAR). Three methods were applied, i.e., stepwise multiple linear regression (SMLR), partial least-squares regression (PLSR), and support vector machine (SVM). Estimation models for four soil properties were developed using three different spectral processing and transformation techniques, i.e., reflectance (Re), logarithmof reciprocal Re (LR), and standard normal variable of Re (SNV) were used. A total of 36 models were established. Of these, 27 models for [Na+], [Mg2+], and [Ca2+] were not applicable for subsequent prediction, because the coefficients of determination (R2) were not high (0.224-0.689), and their relative percent deviations (RPD) were all smaller than the 1.4 threshold. However, the models for SAR~R using PLSR (R2 = 0.728 for calibration and 0.661 for validation, RPD = 1.43), SAR~LR using SVM (R2 = 0.791 for calibration and 0.712 for validation, RPD = 1.81), and SAR~SNV using SVM (R2 = 0.878 for calibration and 0.814 for validation, RPD = 2.13) were valid for further prediction. Finally, SAR~SNV using SVM was selected as the best model. There are intrinsic factors resulting in an unsatisfied model performance.