Comparisons of spatial and non-spatial models for predicting soil carbon content based on visible and near-infrared spectral technology

Abstract

Visible and near-infrared (VNIR) reflectance spectroscopy is a rapid, non-destructive, and cost-effective method for predicting soil properties. Partial least squares regression (PLSR) is a common method used to predict soil properties based on VNIR reflectance spectra. However, PLSR ignores the spatial autocorrelation of soil properties and the assumption of linear regression models, in which explanatory variables and model residuals should be independently and identically distributed. In this study, PLSR, partial least squares–geographically weighted regression (PLS–GWR), partial least squares regression Kriging (PLSRK), and partial least squares–geographically weighted regression Kriging (PLS–GWRK) were constructed to predict soil organic matter (SOM) based on soil spectral reflectance. In addition, this study explores the influence of the spatial non-stationarity of explanatory variables on prediction accuracy. Among the aforementioned models, PLSR was used as a reference model; PLS–GWR considered the spatial autocorrelation of SOM and its auxiliary variables; PLSRK and PLS–GWRK considered the spatial dependence of the model residuals to ensure the usability of PLSR and PLS–GWR. A total of 256 topsoil samples (0–30cm) were collected from Chahe Town, located in Jianghan Plain, China, and the reflectance spectra (400–2350nm) of soil were used. The prediction capabilities of the models were evaluated using the coefficient of determination (R2), the root-mean-square error (RMSE), and the ratio of performance to inter-quartile range (RPIQ). The evaluation indices showed that PLS–GWRK was the optimal model for predicting SOM using VNIR spectra. PLS–GWRK has the lowest values of RMSEC [0.109 ln (g·kg−1)] and RMSEP [0.223 ln (g·kg−1)] and the highest values of R2C (0.933), R2P (0.653), and RPIQ (3.015). PLS–GWR result showed that the spatial dependence of SOM and principal components could improve prediction accuracy compared with the PLSR result. The result of PLSRK showed that the spatial dependence of the model residuals could influence the prediction accuracy of PLSR. The PLS–GWRK approach explicitly addressed the spatial dependency and spatial non-stationarity issues for interpolating SOM at regional scale.