A Model Based Framework For Land Cover Classification In High Resolution Satellite Images

This paper presents a parametric framework for the solution of land cover classification problem in high resolution IKONOS and QuickBird satellite images. In the first step, we merge the multispectral images of low spatial resolution with a monospectral image of higher spatial resolution in order to improve the resolution of multispectral images. This image fusion is based on the search of an optimal 3D hybrid colour space. The second step consists of the parametric segmentation of these high resolution images. This segmentation is based on the 2D multichannel QPAR linear prediction model. A multivariate Gaussian mixture model approximation is used to model the probability distribution of the multichannel linear prediction errors for an initial class label field estimation. These coarsely classified results are then improved by using a spatial regularization algorithm. Experimental results show the pertinence and the robustness of the approach. A comparison of the land cover classification results in different colour spaces is also presented.