Spectral-spatial approaches for hyperspectral data classification

Classification of hyperspectral data is very challenging and mapping of land cover is one of its applications. Improving the classification accuracy and computation time of hyperspectral data were achieved incorporating contextual information in combination with spectral information for correcting classification errors along class boundaries and within class. In the proposed method, the original hyperspectral image was first classified using the Support Vector Machine (SVM) classifier, followed by the Markov Random Field (MRF) approach applied to the boundary areas and Unsupervised Extraction and Classification of Homogeneous Objects (UnECHO) classifier used for the interior parts of regions to produce the final classification map. In this study two agricultural (Hyperion and AVIRIS) and one urban (ROSIS) datasets were used. Investigations of the spectral and various contextual approaches including feature reduction show that the SVM-MRF method with grid search works best for all of the datasets. The highest overall accuracy of 97.35% was achieved for the urban dataset.