Spectral discrimination and classification of sugarcane varieties using EO-1 hyperion hyperspectral imagery

The genetic variety of sugarcane is a major factor in many aspects of sugarcane production. It can control growth rates, yield, sugar content, and resistance or susceptibility to pest and diseases. Thus, reliable auditing of the varieties grown in different areas is necessary for profitable sugarcane cropping. The specific objectives of this study were: a) to assess the spectral separability of different sugarcane varieties, b) to determine which plant attributes will provide the potential discriminating features, and c) to assess the accuracy of image classification to map sugarcane varieties. Using an atmospherically corrected EO-1 Hyperion image acquired over Mackay, Queensland, Australia, the apparent reflectance signatures from sample areas of sugarcane varieties were analysed using discriminant analysis (DA) to explore spectral separability and to determine the optimum bands and indices. Five and eight cane varieties were separately used for each DA run. Then, image classification was implemented for eight cane varieties using four selected classification algorithms. These were independently performed for: a) 152 individual Hyperion bands, and b) a selection of 20 spectral vegetation indices. From the discriminant analysis, the classification results indicate a high discrimination between cane varieties, i.e. 97% accuracy for five varieties and 74% for eight varieties. The best indices for discrimination were OSAVI, TCARI, Ratio 770/550, Pigment Specific Simple Ratio (Chlorophyll b) and Simple Ratio 800/550, indicating that pigments and the leaf internal structure were relevant in the discrimination. However, for the classification of the entire image, the results were not encouraging; the highest classification accuracy was only 46%. The low accuracy can be attributed to the high number of classes used (i.e. eight cane varieties) and the many confounding factors pertaining to crops, management regime, growth stage, and background features such as soils. Thus, future approaches should consider the integration of non-image information (e.g. soil information, crop calendar, etc.) and/or exploring the usefulness of other measurable plant attributes (e.g. leaf/canopy geometry)