A support vector machine discriminator for tobacco growing areas based on near-infrared spectrum

The tobacco growing area is of great importance in the quality control of cigarette, because the fragrance of tobacco leaves would be divergent for different climates planting environments. Currently, most of discrimination processes are manually operated, which are time-consuming and inevitably limited by the subjective evaluation. In this paper, an automatic growing area discrimination method is presented based on tobacco near-infrared (NIR) spectrum using support vector machine (SVM). The Savitzky-Golay smoothing method and principle component analysis are used for tobacco NIR spectra preprocessing. A SVM model is established to investigate the characteristics of growing areas. The developed SVM classifier produces the best prediction accuracy of 80.3% in testing subset with 14 principle components as the inputs. It is 6% and 2% higher than that of artificial neuron network and Mahalanobia distance model respectively, which were developed for comparison. It demonstrates the effectiveness and robustness of SVM for growing area discrimination. The prediction ability for each growing region is further analyzed by the measurements derived from confusion matrix, such as true positive rate, true negative rate, positive predictive value and F1 score. The SVM setting is also discussed with respect to prediction accuracy of validation.