Advanced Fusion Techniques for Automated Detection of Settlement Areas

Data fusion exists in different forms in different scientific communities. The term is used by the image community to address the problem of sensor fusion, where images from different sensors are combined. The term is also used by the database community for parts of the interoperability problem. The logic community uses the term for knowledge fusion. Usually, they can be classified into three levels: pixel level (ikonic), feature level (symbolic) and knowledge or decision level (Pohl and van Genderen, 1998). In this paper, we focus on the development of a decision based fusion. For automated detection of settlement areas, we developed a hierarchical decision tree which is based on homogeneous image segments rather than on image pixels. Within an integrated GIS/remote sensing environment registered multisensor image and GIS datasets were used to facilitate an automated settlement detection. The datasets included remote sensing images from SPOT 5 (5 m GSD), LANDSAT 7 ETM (30 m GSD), KOMPSAT 1 (6,6 m GSD), and ASTER (15 m GSD) satellite programs as well as GIS data from the ATKIS landscape and digital elevation models. Our method for multisensor decision-based data fusion was initially developed and tested in selected urban and suburban areas first with the fusion of SPOT and LANDSAT data and later on extended to KOMPSAT and ASTER images. The methodology is based mainly on an adapted texture and segment oriented hierarchical classification approach: based on panchromatic high resolution image data as primary input. Segments at three different scales (levels) are the basis for a hierarchical decision-based classification procedure. Beginning with large segments, texture and shape parameters were calculated for each single segment. In addition, we used the normalized vegetation index (NDVI) calculated from the multispectral lower resolution satellite image data to distinguish between vegetation and non-vegetation areas. Using adapted threshold parameters, candidate regions for settlement areas were identified. In the second level with medium-sized segments, texture and shape parameters were calculated again using different restriction thresholds. This procedure was only performed within the selected settlement candidates of the previous step. This procedure was repeated for fine segments and high threshold values to isolate the actual settlement areas. Finally, the settlement segments were merged and cleaned by automated filter procedures to eliminate small remaining agriculture segments and to include urban parks and lakes in the settlement areas. By checking the results with actual GIS base data and photo interpretation results, the automated procedure achieved a total accuracy of more than 90% for the settlement class. Furthermore, we applied the same technique to fuse KOMPSAT and ASTER image data. Because of the lower spatial resolution of KOMPSAT compared to SPOT 5, only the threshold values for the texture parameters had to be readjusted. Again, the achieved accuracy exceeded 90%. Future research will include the extension of this method to differentiate between residential and commercial areas and the detection of abandoned mining areas as well as the analysis of their actual state.