Hyperspectral Image Classification using Machine Learning Approaches

Classification of hyperspectral remote sensing images (HSI) is one of the research problems of the remote sensing community due to the high dimensionality of the input data and fewer samples in it. This paper aims at comparing the methodologies for the hyperspectral classification. Initially, the spectral information from the hyperspectral cube is extracted through the principal component analysis and probabilistic principal component analysis via expectation and maximization and the non-linear edge-preserving filter for spatial information extraction. The reduced bands from both of these methods are then given to a non-linear support vector machine for classification. The proposed approach is tested with a different number of bands such as 5, 10, 15 and 25 extracted through the feature extraction techniques. The different performance metrics are computed with the proposed methods and a comparative analysis is made. Also, the results are tested by selecting different bands from the feature extraction methods. The overall accuracy, average accuracy and kappa coefficient over the university of Pavia data set are 95.87, 97.33 and 96.42 respectively which shows the robustness of the algorithm over the other two methods. The process can be used in agriculture, food technology, forestry applications for the extraction of different indices and detection of quality of the food items.