Segmentation of high resolution satellite imagery using texture and spectral information

Multispectral data of Earth resource satellites has been widely used for automated land-cover-land-use classification since the launch of Landsat MSS. There are many classification algorithms, most of them are implemented by per-pixel classifier based on spectral response. It is sufficient to use only the spectral data to perform pixel-based classification for low resolution satellite imagery. However with the improvement of the spatial resolution of satellite image, e.g., 20 m and 10 m of SPOT, the detail of the image has become more complicated than that of low resolution image. It is apparent that the per-pixel approach with spectral information alone is inadequate for classifying high resolution data. This study describes a multiple level segmentation method which uses texture as well as spectral information for classification. The basic idea of this method is that at the first level of the segmentation moving window operation is performed for the whole image, then the spectral statistics of the window is compared with a lookup table of a training set. The comparison will statistically determine the class of the window. If the class can be separated into more detail classes according to spatial information, then segmentation continues to the second level. The spatial measurement is then used to perform comparison between moving window and spatial lookup table. Accordingly, the segmentation technique and additional spatial information should be able to classify the image to more detailed level. The testing results indicate that the proposed multi-level segmentation approach and the use of spatial information is feasible and useful in classifying high resolution satellite imagery