Automated detection of buildings from single VHR multispectral images using shadow information and graph cuts

In this study, we propose a novel methodology for automated detection of buildings from single very-high-resolution (VHR) multispectral images. The methodology uses the principal evidence of buildings: the shadows that they cast. We model the directional spatial relationship between buildings and their shadows using a recently proposed probabilistic landscape approach. An effective shadow post-processing step is developed to focus on landscapes that belong to building regions. The building regions are detected using an original two-level graph theory approach. In the first level, each shadow region is addressed separately, and building regions are identified via iterative graph cuts designed in two-label partitioning. The final building regions are characterised in a second level in which the previously labelled building regions are subjected to a single-step multi-label graph optimisation performed over the entire image domain. Numerical assessments performed on 16 VHR GeoEye-1 images demonstrate that the proposed approach is highly robust and reliable. A distinctive specialty of the proposed approach is its applicability to buildings with diverse characteristics as well as to VHR images with significantly different illumination properties.