Hyperspectral Image Classification by Means of Suprepixel Representation with KNN

In real-world application, especially in remote sensing based on image processing hyperspectral imaging (HSI) shows promising results. Superpixel-based image segmentation is the powerful tool in hyperspectral image processing. Series of neighboring pixels composes superpixel which may belong to different classes but can be regarded as homogenous region. Extraction of more representative feature is considered to be most important thing in hyperspectral imaging. Training and testing samples that are more representative are found by proposing a new method for selecting two k values for representing optimal superpixels. This paper starts with superpixel shifting as first step and followed by KNN classifier. Which is performed by pixels with minimal spectral features in HSI are clustered together in the same superpixel. Followed by spatial-spectral feature is extraction by a domain transformation from spectral to spatial. For each superpixel, training and test samples are selected to eliminate classification within the same class. An average distance between test and training samples are used for determining class label. Finally, by the results from most common hyperspectral images Indian pines, Salinas, Pavia show that this method shows a better classification performance.