A hybrid image segmentation method for building extraction from high-resolution RGB images

Buildings are the most striking artificial features, and extracting buildings becomes critical for many service-providing agencies. Although improvements have been achieved, building detection from remotely sensed images is still challenging. In the past few years, many building extraction methods have been put forward by researchers, such as line- or edge-based, template matching, knowledge- and auxiliary data-based, machine learning, morphological operations-based, and GEographic Object-Based Image Analysis (GEOBIA) -based. GEOBIA is a paradigm for analyzing high-resolution images; however, GEOBIA-based building extraction methods encounter problems in the segmentation and classification stage. Thus, the accuracy of those methods is lesser than other building detection approaches. This research introduced several modifications to the previously proposed hybrid segmentation methods, such as using the reference polygon to identify optimal parameters, a donut filling technique to reduce over-segmentation caused by roof elements, and illumination differences to restrict merging with shadow. The proposed methodology was tested on a UAV image with visible bands only. Better results were achieved using this approach when compared to the multiresolution proposed by Baatz and Schäpe (2000) and the other two-hybrid methods proposed by Wang et al. (2018a) and Yang et al. (2017). This hybrid segmentation method was also applied to subsets of the Wuhan University buildings dataset and produced similar results. One of the great strengths of the proposed method was that there were no parameter tuning and user interaction at running time. In addition, it was able to segment both small and large buildings without using any scale or object size parameters.