A Hierarchical Building Detection Method for Very High Resolution Remotely Sensed Images Combined with DSM Using Graph Cut Optimization

Detecting buildings in remotely sensed data plays an important role for urban analysis and geographical information systems. This study proposes a hierarchical approach for extracting buildings from very high resolution (9 cm GSD (Ground Sampling Distance)), multi-spectral aerial images and matched DSMs (Digital Surface Models). There are three steps in the proposed method: first, shadows are detected with a morphological index, and corrected for NDVI (Normalized Difference Vegetation Index) computation; second, the NDVI is incorporated using a top-hat reconstruction of the DSM to obtain the initial building mask; finally, a graph cut optimization based on modified superpixel segmentation is carried out to consolidate building segments with high probability and thus eliminates segments that have low probability to be buildings. Experiments were performed over the whole Vaihingen dataset, covering 3.4 km2 with around 3000 buildings. The proposed algorithm effectively extracted 94 percent of the buildings with 87 percent correctness. This demonstrates that the proposed method achieved satisfactory results over a large dataset and has the potential for many practical applications. Introduction The identification and localization of buildings in an urban area is very important for planning, building analysis, automatic 3D reconstruction of building models and change detection (Qin and Gruen, 2014). The development of very high resolution (VHR) remote sensing images (Qin et al., 2013) creates a possible avenue to sense individual buildings in an urban scenario, e.g., Ikonos with 1-meter resolution, or Worldview with 0.5-meter resolution. Sensors with even higher resolution are in the planning stages (e.g., Geoeye-2 and Worldview-3 with 0.3-meter resolution). However, this increasing level of detail does not necessarily facilitate building detection with an improved accuracy (Huang and Zhang, 2011). Indeed, more detailed image contents actually increase spectral ambiguities in remotely sensed images, such as symbol patterns on the road, and big vehicles. Therefore, researchers have devoted a lot of effort toward using multisource data and designing better detection strategies to increase the building detection rate. Multispectral images provide shadow information as primitives for building locations. Furthermore, shadow information are especially effective in single image based methods (Huang and Zhang, 2012; Ok, 2013; Ok et al., 2013). Meanwhile, NDVI data extracted from a multispectral image can be used as vegetation indicators to eliminate trees. Vector features such as parallel lines and corner junctions reveal the characteristics of rectangular buildings, which have been investigated and used to develop single-image based methods for building detection (Lin and Nevatia, 1998; Sirmacek and Unsalan, 2011; Sirmacek and Unsalan, 2010; Sirmaçek and Unsalan, 2009). Lidar (Light Detection and Ranging) point clouds provide height information for a ground scene and are used for building detection. By subtracting the DTM (Digital Terrain Model) from the DSM (Digital Surface Model), a nDSM (normalized DSM) can be computed to obtain off-terrain points for building detection (Weidner and Förstner, 1995). In addition, the multi-return characteristics of lidar provide useful information to eliminate the vegetation for point clouds based methods (Ekhtari et al., 2008; Meng et al., 2009), to increase the accuracy of building detection. Both multispectral image and lidar point clouds have their advantages and deficiencies. Complex algorithms based on a single image usually have assumptions concerning building distribution and sometimes are only able to detect certain types of buildings. For example, methods based on feature point extraction from a single image are only able to detect isolated buildings with regular patterns, and methods relying on parallel lines are not able to detect dome roofs. As compared to multi-spectral images, lidar point clouds provide accurate height information, but less accurate boundaries. There are also null values for lidar point clouds due to occlusion and specular reflection from water surfaces on the roofs. Therefore, integration of both sources is a possible direction for improving building detection accuracy as well as robustness. There has been a spate of integrated methods proposed in the literature. Rottensteiner et al. (2007) and Rottensteiner et al. (2005) proposed a supervised classification-based building Rongjun Qin is with the Singapore ETH Center, Future Cities Laboratory, ETH, Zurich. 1 CREATE Way, #06-01 CREATE Tower, Singapore 138602 (rqin@student.ethz.ch). Wei Fang is with the Singapore ETH Center, Future Cities Laboratory, ETH, Zurich. 1 CREATE Way, #06-01 CREATE Tower, Singapore 138602, and the State Key Laboratory of Information Engineering in Surveying, Mapping and Remote Sensing (LIESMARS), Wuhan University, China. #129, Luoyu Road, Wuchang District, LIESMARS, Wuhan University, Wuhan, P. R. China, 430079. Photogrammetric Engineering & Remote Sensing Vol. 80, No. 8, September 2014, pp. 000–000. 0099-1112/14/8009–000 © 2014 American Society for Photogrammetry and Remote Sensing doi: 10.14358/PERS.80.9.000 PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING September 2014 37 This is the pre-print verison, technical content is appeared as it is, with only a few minor differences due to language editing