Potato Crisp moisture determination using NIR data and a Back Propagation Neural Network

Abstract

Near infrared analysis is a tool used for non-destructive determination of material properties and the potato crisp production sector has been using the technique for determination of moisture content however near infrared spectral models suffer from problems associated with light scatter. Light scatter results from geometric irregularities in the samples geometry and this reduces the accuracy of near infrared calibration models without preprocessing for scatter removal. Quantitative calibration models have benefited from the development of artificial intelligence methods and the neural network is now a popular tool for quantitative calibration model formation. In this paper we compare the performance of a back propagation neural network calibration model using 3 forms of preprocessed data, orthogonal signal correction, standard normal variate and data with no scatter preprocessing prior. The correlation coefficient was used to determine the neural networks methods performance and it was found that a neural network using data with no scatter preprocessing yielded the best results.