Two preprocessing techniques to reduce model covariables in soil property predictions by Vis-NIR spectroscopy

Proximal sensing provides an alternative method to physical and chemical laboratory soil analyses. The aim of this study is to predict soil organic carbon (SOC), clay, sand, and silt content using reduced spectral features as covariables selected by two spectral preprocessing. A total of 299 soil samples were collected in Santa Catarina state, Brazil. Two preprocessing techniques, detrend transformation and continuum removal (CR), were applied to isolate particular absorption features in the reflectance spectrum. Two techniques were used to select the spectral features in the spectrum: hand and mathematical selection. Partial least squares regression (PLSR) and Support vector machines (SVM) were applied to predict the soil properties. The reduction of predictor covariables by hand selection technique contributed in developing a high-level prediction model for SOC. PLSR and SVM presented no statistical difference between the RMSE results, except for clay content, where SVM presented superior performance. The preprocessing techniques were statistically identical based on RMSE results. Overall, the prediction of SOC, clay, sand and silt presented suitable results using reduced spectral features as covariables in modeling process.