An information fusion method for multispectral image classification postprocessing

Remote-sensing image classification is one of the most important techniques in understanding the dynamics of the Earth's ecosystems. Various approaches have been proposed for performing this classification task. Obtained classification results are generally shown as a thematic (or class) map in which each pixel is assigned a class label. Due to sensor noise and algorithm limitations, obtained thematic maps are very noisy. The noise has a "salt-and-pepper" appearance in homogeneous regions and produces weakly defined interregion borders. In this paper, a new postprocessing approach aiming to produce thematic maps with sharp interregion boundaries and homogeneous regions is presented. This approach is conducted in two steps: (1) relevant features derived from the original multispectral image (edge maps) as well as from the thematic map, the Smoothed Thematic Map (STM), are determined and (2) a region-growing algorithm is applied over the thematic map. This algorithm grows until reaching an edge (from the edge maps) or a class change in the STM. The proposed approach fills the requirements of being independent of the used classification algorithm and not knowledge-based (in the sense that no a priori information concerning the contents of the considered image is needed). Tests have been conducted on a Landsat image covering mainly agricultural areas.