Minimising the effect of moisture on soil property prediction accuracy using external parameter orthogonalization

Moisture is one of the most important factors affecting soil reflectance spectra. However, provisional and dynamic behavior of soil moisture (SM) in salty soils reduce the capability of field spectrometry in the estimation of soil properties. This study aims minimising the effect of SM on the accuracy of visible and near-infrared (VNIR) spectra estimation of clay, calcium carbonate (CaCO3), and organic carbon (OC) content in semi-arid soils by adopting External parameter orthogonalization (EPO). Spectral reflectance of disturbed soil samples was measured for seven moisture contents (i.e. air dried, 6%, 12%, 18%, 24%, 30%, and 36%) and five levels of electerical conductivity (EC) (i.e. < 1, 4, 8, 12 and 16 dS/m). The EPO algorithm was customized for four textural classes. The performances of EPO algorithm was evaluated through partial least squares-backpropagation neural network (PLS-BPNN). The optimum number of components for the EPO matrix was determined to be four. Results showed that geometric parameters (area, depth, and width) of diagnostic absorption features (AF) in 550 nm, 2200 nm, and 2340 nm were affected by SM, and the interaction between SM and reflectance was not completely equivalent through different texture classes. The EPO correction showed an improvement of prediction accuracy of clay (RPIQ improved from 1.82 to 2.93), CaCO3 (RPIQ improved from 1.96 to 2.73), and OC (RPIQ improved from 1.94 to 3.44). Also better results have been gained by the customization of EPO for different texture classes. The performances of the EPO algorithm for clay and OC modeling dropped by increasing EC. This suggests that further studies are required to develop a method for eliminating the effects of the external parameters caused by increased salt levels in the soil.