Improving Hyperspectral Image Classification with Watershed Segmentation-Based Texture Features

Hyperspectral images offer a good separability among land cover classes, owing to their advantages to provide spectral bands with narrow wavelength intervals. It is also possible to further increase classification accuracy of hyperspectral images by integrating some auxiliary data to classification process. The aim of this study was to increase the classification accuracy of AVIRIS (Airborne Visible Infrared Imaging Spectrometer) and ROSIS (Reflective Optics System Imaging Spectrometer) hyperspectral data by integrating the Watershed Segmentation (WS)-based texture features to the SVM (Support Vector Machines) and RF (Random Forest) classification process. Since the used hyperspectral images contained a high number of bands, the PCA (Principal Component Analysis) technique was used to reduce the dimensionality of texture features and avoid redundant texture characteristics. It was found that the used procedure increased the RF classification accuracy of the AVIRIS and ROSIS data by 13.89% and 23.93%, respectively. It was also concluded that the SVM classification accuracy of the AVIRIS and ROSIS data was increased by 18.89% and 30.42%, respectively.