Spectral-spatial based super pixel remote sensing image classification

Apart from the rich spectral information provided by multispectral or hyperspectral sensors, the spatial information has been paid more and more attention in remote sensing classification, especially for high spatial resolution images. Pixel-wise spatial features can be generated by applying Gray Level Co-occurrence Matrix (GLCM) locally to describe an image's texture properties. Morphological filtering provides spatial structure enhancement and watershed processing aims at contextual boundary identification. In this paper, the advantages and disadvantages of these spatial treatments are investigated. A combined procedure is developed to maximize spatial information extraction. Texture feature selection is emphasized for class separability enhancement. Morphological filtering is introduced as a preprocessing for watershed segmentation in order to reduce false alarm on contextual boundaries. Super pixels are formed for the objects defined from the watershed segmentation. The experimental results show that the combined spatial treatment is effective and, by integrating it with spectral information, an object oriented classification map can be obtained with significantly reduced `salt and pepper' noise.