Fast inline tobacco classification by near-infrared hyperspectral imaging and support vector machine-discriminant analysis

Abstract

Classification systems are frequently used in tobacco Green Leaf Threshing (GLT) facilities to assess the chemical characteristics and quality of tobacco leaves. This classification is usually performed by a trained specialist who has to analyse and classify the leaf, in a short time, based on experience and visual information. However, this approach is not robust and may be biased due to its highly subjective nature. This work proposes the use of infrared hyperspectral imaging and chemometric tools for discriminant analysis. A fast and real-time method to classify flue-cured Virginia and air-cured Burley tobacco was developed by classifying the leaf based on three criteria: (a) stalk position, (b) leaf colour, and (c) leaf quality; this occurs within 5 seconds. The applicability of the method was evaluated by analysing standard tobacco leaf bundles using near-infrared imaging and Support Vector Machine-Discriminant Analysis (SVM-DA). The models classified the stalk position with a global prediction accuracy of 80.4% for flue-cured Virginia and 88.1% for air-cured Burley. The models for targeting leaf classification by colour presented a global prediction accuracy of 95.9% for flue-cured Virginia and 96.5% for air-cured Burley. For leaf quality, the prediction accuracy ranged from 61.5% to 100.0% for flue-cured Virginia and 78.8% to 100.0% for air-cured Burley. The system was implemented in a tobacco GLT facility for one year to cover an entire crop season and extended to include tobacco bales and leaf bundles. For the most complex classification parameter, leaf quality, flue-cured Virginia models presented a global accuracy of 50.2% to 71.6%. Considering the high complexity of the system and short period of analysis, the obtained accuracy was accepted as suitable when compared to traditional human classification.