Soil organic carbon prediction using vis-NIR spectroscopy with a large dataset

Abstract

Visible and near-infrared reflectance spectroscopy based soil properties estimation is an alternative to traditional laboratory analysis. The calibration model is a main factor influencing predictive performance. In this study, a large scale soil database, which contains 19,036 soil samples, was compared to its subsets to validate the effect of sample size on predictive performance. Four regression techniques based on linear model, namely, multiple linear regression (MLR), principal components regression (PCR), partial least squares regression (PLSR), and stepwise regression (SR) were compared to identify suitable models to predict the content of organic carbon in soil samples. The impact of derivatives or the raw spectra as predictor variables, and the interval of spectra were also studied. The best predictions were obtained using SR and MLR on raw spectra, yielding root mean square of error of cross validation (RMSECV) and coefficient of determination (R2) values of 25.3912, 25.4254 and 0.9227, 0.9225, indicating excellent models.