Determination of Soil Nitrogen and Organic Carbon from Near-infrared Reflectance with Variable Selection

Abstract

<abstract> <bold>Abstract.</bold> The aims of this work were: to select informative variables for modeling near-infrared spectra to soil nitrogen (N) and organic carbon (OC) and to provide interpretation for the selected variables. The dataset that consisted of 225 soil samples was spilt into calibration set, validation set and prediction set by Knennard-Stone algorithm. Spectra in calibration set were conducted for variable selection by the method of Monte Carlo uninformative variable elimination (MC-UVE) and successive projections algorithm (SPA). Partial least squares regression (PLSR) and multiple linear regression (MLR) were used to construct calibration models for each property based on the selected variables. The proposed method MC-UVE-PLSR achieved the optimal performance for soil N and OC comparing with full-spectrum PLSR and SPA-MLR. The coefficients of determination (R²), residual prediction deviation (RPD) were respectively 0.86, 0.87, and 2.8, 2.7 for N and OC. The results indicate that MC-UVE is an effective tool for spectral variable selection, and is able to promote model prediction accuracy and efficiency. Analysis of the feature variables show that some of the selected variables of soil N and OC are directly related to their functional groups, while some influence the results by measuring the correlation constituents of N and OC. It is also suggested to adopt its own spectral response for the prediction of soil N instead of calculation by the autocorrelation of soil OC.