Terrain segmentation of high resolution satellite images using multi-class adaboost algorithm

Terrain segmentation is still a challenging issue in pattern recognition, especially in the application of high resolution satellite images. Among the various segmentation approaches are those based on graph partitioning, which present some drawbacks such as high processing time, low accuracy on detection of targets on the large scaled images such as high resolution satellite images. In this paper, we focus on the computational intelligence approach to classify and detect building, foliage, grass, bare-ground, and road of land cover. We propose a method, which has a high accuracy on classification and object detection by using multi-class AdaBoost algorithm based on a combination of two extracted features, which are cooccurrence and Haar-like features. With all features, multi-class Adaboost selects only critical features and performs as an extremely efficient classifier. Experimental results show that the classification accuracy is over 91% with a high resolution satellite image.