Neural Network and Extreme Gradient Boosting in Near Infrared Spectroscopy

Abstract

Near infrared spectroscopy is a secondary measurement approach that aims to quantitatively or qualitatively estimate the components of interest from the acquired near infrared spectrum using computation methods e.g. machine learning algorithms. After decades of investigation, neural network has been accepted as a nonlinear benchmark model in near infrared spectroscopy. Although a recent work reported that Extreme Gradient Boosting (XGBoost) outperformed neural network in groundwater level prediction, the optimization process and the learning algorithm of the neural network were not reported. This implies that the neural network might not be the optimal. Thus, this study aims to compare the performance of the optimal Bayesian regularized neural network and XGBoost in a regression application using more than one thousand of near infrared spectral data that were acquired throughout different years. The regression models were established to predict the dry matter content (DMC) of mangoes using the respective spectral data. Results show that even though XGBoost could achieve a satisfactory accuracy with RMSEV, RMSEP, R2V, and R2P of 1.16%, 1.22%, 0.73, and 0.80, respectively, the Bayesian regularized neural network achieved substantially better RMSEV, RMSEP, R2V, and R2P of 0.83%, 0.86%, 0.86, and 0.90, respectively. Thus, a Bayesian regularized neural network is recommended to be tested when more than one thousand near infrared spectral data were available.