Removing the moisture effect on predicting soil organic matter using vis-NIR spectroscopy with external parameter orthogonalization

Abstract

External parameter orthogonalization (EPO) can remove the moisture effect on the in situ soil reflectance spectra. However, when applied to a sample space with high changeability in structure of mineral, EPO cannot fully remove the in situ effect on spectra due to the complex interactions between both soil moisture and mineral composition on the spectra. This study aims to minimize the effect of soil moisture on the accuracy of in situ visible and near-infrared (Vis-NIR) spectra on estimating of soil organic matter (SOM) content by using strategies of firstly classifying the samples according to absorbance length ∼ 2200 nm (AL), texture (TC) and land use (LC), and then implementing the EPO through each class in each strategy. The samples were split into the total EPO development, where a serial subset was selected for the EPO dataset, and validation dataset. For every subset selected from the total EPO development, an EPO corrected matrix, developed from the spectra from each class of each classification strategy, was implemented with the spectra from the validation dataset and local spectra library. The performances of the EPO algorithm were evaluated through partial least squares regression (PLSR). The results showed that the classifications of AL, LC and TC combined with EPO led to a greater reduction in the principal component space of in situ subsets and improved the similarity between in situ spectra and dry subset distribution. The combination of classification with EPO leads to a reduction of 11% - 25% prediction RMSE of SOM. Classification with EPO can minimize the moisture effect on in situ spectra from a sample space with high variability in mineral composition.