Comparison of PLSR, MLR, SVM regression methods for determination of crude protein and carbohydrate content in stored wheat using near Infrared spectroscopy

Abstract

Aim of the study was to predict quality parameters namely protein and carbohydrate content of wheat grain stored for one year at 4 °C and 21 °C using Near Infrared spectroscopy (NIRS) and Chemometrics. The spectra of grains were measured in reflectance mode using lab made filter based pre dispersive NIRS ranging from wavelength 750 nm to 2580 nm. NIR Wavelengths were divided into two sets for all the stored samples depending on its correlation with factors. The reference data and NIR data were analysed by principal component analysis (PCA) scores, partial last squared regression (PLSR) model, Multiple Regression Analysis (MLR) and Support Vector Machine Regression (SVMR). First four factors were sufficient to develop regression models for the above mentioned methods. Prediction performance was compared on the basis of the coefficient of correlation (R2) and Root Mean Square Error (RMSE) for the calibration and validation sets. R2 values for prediction of protein content from PLSR model were 0.955, 0.997 for 4 °C Wavelength Set I and II and 0.903, 0.989 for 21 °C Wavelength Set I and II. For the prediction of carbohydrate content 0.978, 0.951 and 0.946, 0.974 were the R2 values for the 4 °C and 21 °C wheat samples Set I and Set II respectively. MLR model also gave R2 values greater than 0.9 for all the samples. Wavelengths with high β correlation coefficients were defined. SVMR predicted the protein content and carbohydrate content with good accuracy. SVMR and PLS results. All the models developed showed good accuracy for successful prediction of protein and carbohydrate content of stored wheat grains with PLSR and SVMR giving the best results. This study showed that near infrared spectroscopy has potential to distinguish wheat grains stored under different conditions.