Segmentation based building detection in high resolution satellite images

We demonstrate an integrated strategy for identifying buildings in very high resolution satellite imagery of urban areas. Buildings are extracted using structural, contextual, and spectral information. We perform multi-resolution and spectral difference segmentation to obtain a proper object segmentation. First, we use One-Class support vector machine (SVM) in order to determine the man-made structures (buildings, roads, etc.). Next, we proceed with texture segmentation approach using a conditional threshold value to extract the buildings. And then, we use geodesic opening and closing operations to extract bright foreground objects. After this, shadows and vegetation regions are detected in these segments based on their spectral properties. We then remove noise, vegetation and shadows from the candidate building regions. And finally, we classify the buildings by checking for the presence of shadows along the buildings opposite to the sun's azimuth direction to distinguish buildings from other bright regions. Performance evaluation of the proposed algorithm is performed on data acquired using WorldView satellite imagery over Abu Dhabi, United Arab Emirates.