A Shadow based trainable method for building detection in satellite images

Dikmen, Mehmet
The purpose of this thesis is to develop a supervised building detection and extraction algorithm with a shadow based learning method for high-resolution satellite images. First, shadow segments are identified on an over-segmented image, and then neighboring shadow segments are merged by assuming that they are cast by a single building. Next, these shadow regions are used to detect the candidate regions where buildings most likely occur. Together with this information, distance to shadows towards illumination direction and spectral properties of segments are used to classify them as belonging to a building or not. Then, a resegmentation is performed to extract building patches by merging only the neighboring segments, which are classified as building. Next, a postprocessing step is implemented to eliminate some false building patches. Finally, a one class modeling approach was introduced to refine extracted building patches. The approach was tested on several Google Earth images of varying characteristics in order to examine the effects of the change in illumination direction, shadow amount and building variety (size, shape, density, etc.). The results were examined by both pixel and object based performance evaluation methods. Best results were obtained on images having relatively shorter shadows and captured almost at the nadir. Best quality for the extracted patches and the least false detections were also observed in the same case.
Citation Formats
M. Dikmen, “A Shadow based trainable method for building detection in satellite images,” Ph.D. - Doctoral Program, Middle East Technical University, 2014.