COMPARATIVE ANALYSIS OF SVM, ANN AND CNN FOR CLASSIFYING VEGETATION SPECIES USING HYPERSPECTRAL THERMAL INFRARED DATA
Vegetation includes a significant class of terrestrial ecosystem. Information on tree species categorization is important for environmentalists, foresters, agriculturist, urban managers, landscape architects and biodiversity conservationist. The traditional methods of measuring and identifying tree species (i.e., through field-based survey) are time taking, laborious and costly. Remote sensing data provides an opportunity to identify and classify vegetation species over a large spatial extent. Hyperspectral remote sensing can detect the sublet spectral details among species classes and thus make it possible to differentiate vegetation species based on these subtle variations. This research examines the thermal infrared (2.5 to 14.0 μm) hyperspectral emissivity spectra (comprised of 3456 spectral bands) for the classification of thirteen different plant species. The use of thermal infrared hyperspectral emissivity spectra for the identification of vegetation species is very rare. Three different machine learning methods including support vector machine (SVM), artificial neural network (ANN) and convolutional neural network (CNN) are used to classify thirteen vegetation species and their performance is assessed based on their overall accuracy. The accuracy obtained by CNN, ANN and SVM is 99%, 94% and 91%, respectively. Each classifier was also tested for the advantage associated with increase in training samples or object segmentation size. Increase in the training samples improved the performance of SVM. In a nutshell, all comparative machine learning methods provide very high classification accuracy and CNN outperformed the comparative methods. This study concludes that thermal infrared hyperspectral emissivity data has the potential to discern vegetation species using state of the art machine learning and deep learning methods.