Linear and non-linear pattern recognition models for classification of fruit from visible–near infrared spectra

Environment and genotype affect the composition, quality, storability and sensory properties of plant-based products. Visible–near infrared (NIR) spectral measurements are used increasingly to monitor fruit properties such as maturity, sensory properties and storability non-destructively both prior to harvest and during storage. To explore this problem, at harvest and after storage, visible–NIR spectra containing 1024 individual data points were measured on kiwifruit berries sourced from six pre-harvest fruit management treatments. These raw spectra were processed by principal component analysis (PCA), or by Fourier, Hartley, Haar, Hurst, range renormalisation or polar coordinate transforms (PCT) in order to extract a smaller set of features selected independently of treatment. In order to reduce their dimensionality further, the extracted features were processed by canonical variate analysis. The ability of various connectionist and linear discrimination pattern recognition models to predict the treatment source of unknown fruit on the basis of these features was evaluated. Thus far, this work has established that the performance of the non-linear model was shown to be significantly better in comparison to the linear model. From these results, it has also been shown that both the feature extraction and selection techniques have a marked effect on the ability to classify fruit by treatment source and storage date. In general, the best classifications were based on features extracted using the Fast Fourier Transform (FFT) method, but the best performance in any single classification was given by the Haar transform (HT) in conjunction with the scaled conjugated gradient learning method.