Landcover classification of satellite imagery with tesselated spatial structure model

In this paper, a tesselated spatial structure model is proposed for unsupervised land-cover classification. The model can manage some fundamental problems such as existence of mixed pixels and class parameter estimation of finite mixture distribution in a systematic manner. Some areas in a Landsat TM image are checked if they fit in the model by their appearance and statistics. Based on the proposed model, spatial segmentation by pyramid linking and clustering by K-means are applied to the satellite image. The image is filtered by using spatial median operation of IDL in order to avoid the effect of mixed pixels on segment value. It is shown that the median filtering is effective not only for rural area classification but also for urban area classification.