Contextual information for the classification of high resolution remotely sensed images

The use of remote sensed images in many applications of environmental monitoring, change detection, risks analysis, damage prevention, etc. is continuously growing. Classification of remote sensed images, exploited for the production of land cover maps, involves continuous efforts in the refinement of the employed methodologies. The pixel- wise approach, which considers the spectral information associated to each pixel in the image, is the standard classification methodology. The continuous improving of spatial resolution in remote sensors requires the focus on what is around a single pixel with the integration of "contextual" information. In order to produce more reliable land cover maps from the classification of high resolution images, this paper analyzes the effectiveness of the integration of contextual information comparing two different pixel-wise techniques for its extraction: 1) the post-classification filtering with a Majority filter applied to the map produced by the standard Maximum Likelihood algorithm; 2) the segmentation algorithm SMAP. The results were compared. A GeoEye-1 image, exploited in the framework of the Asi-Morfeo project, was considered.