Artificial neural network coupled with robust principal components in near infrared spectroscopic analysis

The objectives of this paper are to demonstrate the use of a combination of artificial neural network and robust principal components (RPCs-ANN) to predict the soluble solid content of intact pineapple, non-invasively, based on near infrared spectral data, and to compare the performance of RPCs-ANN with artificial neural network based on classical principal components (PCs-ANN). First, we implemented second order derivative with first order Savitzky-Golay (SG) smoothing filter to pre-process the spectral data. Second, robust and classical principal component analysis approaches were utilized to reduce the dimension of spectral data and to produce robust principal components (RPCs) and principal components (PCs), respectively. Third, artificial neural network with RPCs as its inputs (RPCs-ANN) was trained based on Bayesian regularization to improve the generalization of the network by optimizing its regularization parameters. The effects of different number of inputs and hidden neurons are discussed. The findings suggest that RPCs-ANN is superior to PCs-ANN.