earticle

영어

Good segmentation of satellite images plays a significant role in monitoring of urban areas, as well as of natural terrain. The analysis of two segmented observations can provide new information relating to land use, shifting cultivation, deforestation, or environmental changes. This paper introduces a combination of textural features that can achieve very good results for terrain segmentation. We implement BPNN (Back Propagation neural network) and Adaboost algorithms for the classification of an urban area in terms of a combination of several textural features. Using high resolution IKONOS satellite images of the Boston area, we carry out experiments on terrain classification. Experimental results show that a combination of co-occurrence and Harr-like features can be used to obtain high accuracy of terrain classification of 89.8-94.5% with the Adaboost classifier; this new method outperforms other implementations. To verify the efficiency of the proposed classification method, change detection using temporal images is also tested via experiment. The resulting change map shows that a newly developed area can be successfully detected.