Abstract: 【Objective】Newhall navel orange (Citrus sinensis L.) fruits from different producing areas in China, exhibit contrasting quality and market values, due to wide-spreading location of various ecologies. Developing a recognition method based on Near-Infrared (NIR) spectroscopy is very important to identify and distinguish fruits from different producing areas. 【Method】One representative orchard was selected from 17 main producing areas distributed throughout southern China, from which one 100 Newhall navel orange samples were collected. NIR spectra were collected with a SupNIR-1500 spectrograph from the surface of fruit equator and shoulder and the filtered juice for each sample, which were further preprocessed through principal component analysis (PCA) for reduced dimensions and noise. By studying artificial neural network (ANN), a classic three-layer ANN model was established with an input layer, a hidden layer of non-linear activation function and an output layer. By studying support vector machine (SVM) with the radial basis function (RBF) being the kernel function and the principal components of NIR spectra being the input, a one-to-one extended SVM model was established with 126 classifiers. Gene algorithm (GA) with excellent natural selection was used to identify the best Genetic character subset from the principal components as inputs of a SVM classifier, thus a GA-optimized SVM model was composed. These three models were used to classify the NIR spectra of filtered juice, thus the production areas of the oranges, the classification accuracies of which decided the best classifier. Furthermore, the accuracies of the best classifier were tested with the NIR spectra from fruit equator and shoulder surface being the inputs. As a comparison, the best NIR spectra could be identified. 【Result】 Producing area classification was implemented with the three-layer ANN classifier with NIR spectra of Newhall orange juice as the input, where the classifying accuracy reached up to 81.45% when there were 11 input neurons and 13 hidden neurons. The studied one-to-one extended SVM classifier with radial basis function being the core function, exhibited higher accuracy of 86.98% when the number of PC was 20, better than ANN classifier. For GA-SVM classifier took into account the interaction of individual inputs, where the PCA-processed results were optimized by GA. During the experiments, classification accuracy hit 89.72% when the population, generation, mating probability, and mutation probability were 200, 100, 0.7 and 0.01 respectively, surpassed ANN and SVM classifier. Subsequent research found the highest accuracy of GA-SVM classifier was 80% when taking the spectra from the fruit equator, and 69% from the shoulder, not good enough as that of orange juice.【Conclusion】Considering the accuracy, GA-SVM classifier was regarded with the most excellence among three investigated classifiers. Spectra of orange juice were selected as the best data to analyze origins traceability. Accuracy of spectra of fruit equator was inferior to juice but superior to the shoulder, thus had the potential for non-destructive origins classification.

Key words: Newhall Navel orange, producing area recognition, near-infrared spectroscopy, principal components analysis, artificial neural network, support vector machine, genetic algorithm