Combining different pre-processing and multivariate methods for prediction of soil organic matter by near infrared spectroscopy (NIRS) in Southern Brazil

In Brazil, most guidelines for soil nitrogen (N) fertilization are based on the soil organic matter (SOM) content, which is one of the most laborious and expensive parameters in routine soil analysis. Recently, several studies have been made to generate and evaluate alternative methods to estimate SOM content, such as those based on near infrared spectroscopy (NIRS). However, the accuracy of NIRS is highly affected by spectral pre-processing techniques and multivariate methods. Moreover, the models are highly site-specific, indicating the need to develop regional calibrations. The objective of this study was to evaluate the effect of pre-processing and multivariate methods in SOM prediction using NIRS. A total of 2388 soil samples from Southern Brazil were analyzed for SOM content by carbon oxidation with a sulfochromic solution, followed by spectrophotometric colorimeter determination (modified Walkley-Black method) and by a NIR-spectrometer (1200–2400 nm). Seven pre-processing techniques were tested, including: Savitzki-Golay derivative (SGD), continuum removal (CR), detrend (DET), binning (BIN), smoothing (SMO), and the standard normal variate (SNV). The pre-processing techniques were combined with four multivariate models including multiple linear regression (MLR), partial least squares regression (PLSR), principal component regression (PCR), support vector machine (SVM), random forest (RF) and gaussian process regression (GPR). The performance of the predictions was evaluated by the coefficient of determination (R²) and the root mean squared error (RMSE). Moreover, SOM predictions obtained from the models were compared with the wet digestion method in order to evaluate the capability to correctly classify the SOM content for N fertilization. Overall, comparing the multivariate models, the best predictions according to R² were found for SVM > GPR > PLSR > MLR > PCR > RF, in decreasing order. The best prediction was obtained with the combination of SVM model and SNV spectral pre-processing (R² = 0.70, RMSE = 0.44, RPIQ = 2.26). For SVM and GPR, only DET, SNV, BIN and CR improved predictions when compared to raw spectra. When using the linear model MLR, the prediction of SOM was improved only with SNV, SMO and DET, compared to raw spectra. For the PLSR model, the prediction of SOM was improved with DET and SNV. Results estimated by NIRS reached 85% accuracy when compared to values of SOM obtained by the modified Walkley-Black method.