Satellite Image Segmentation and Classification

The resolution of remote sensing images increases every day, raising the level of detail and the heterogeneity of the scenes. Most of the existing geographic information systems classification tools have used the same methods for years. With these new high resolution images basic classification methods do not provide satisfactory results. In this study we developed a region-based classification method, consisting in two steps: a segmentation and a classification. The segmentation uses a Markov model to divide the image into several homogenous regions. Then follows the region-based classification performed either with the Mahalanobis distance or by the support vector machine classifier. This method was validated and a comparison between pixel-based and region-based classification was performed. We demonstrated that this method provides better results comparing to the existing remote sensing classification tools, even if some work should be done to prove its robustness. We also proved that the prior segmentation significantly improves the results of classification, both from the quantitative and qualitative points of view.