Class-Guided Building Extraction from Ikonos Imagery*

DEM (digital elevation model) is available. A study on the use Abstract of stereo Ikonos imagery shows that 5to 10-m contour lines can Recent high-resolution satellite images provide a valuable new be derived with the highest topographic standard (Toutin et al., data source for geospatial information acquisition. This paper 2001). As for geospatial feature extraction for topographic mapaddresses building extraction from Ikonos images in urban ping, Baltsavias et al. (2001a) and Fraser et al. (2001; 2002) presareas. The proposed approach uses the classification results ent results on building extraction from Ikonos stereo images. A of Ikonos multispectral images to provide approximate locacomparative study with the results obtained from aerial phototion and shape for candidate building objects. Their fine graphs concludes that about 15 percent of the building areas, as extraction is then carried out in the corresponding panchromeasured in aerial images, cannot be modeled using Ikonos matic image through segmentation and squaring. The ECHO images. An assessment based on 19 GPS surveyed check points classifier is used for supervised classification while the ISODATA at roof corners suggests that the Ikonos-derived building model algorithm is used for unsupervised classification and subsecan reach an accuracy of better than 1 meter both in planimetry quent image segmentation. The classification performance is and elevation (Baltsavias et al., 2001a). Sohn and Dowman evaluated using the classification confusion matrix, while the (2001) used a local Fourier transformation to analyze the domifinal building extraction results are assessed based on the nant orientation angle in a building cluster and extract rectilinmanually delineated results. A building squaring approach ear building outlines from Ikonos imagery based on a binarybased on the Hough transformation is developed that detects space partitioning tree. Dial et al. (2001) present an investigaand forms the rectilinear building boundaries. A number of tion on automated road extraction in wide suburban roads. sample results are presented to illustrate the approach and Regarding the fundamental methodology in geospatial feademonstrate its efficiency. It is shown that about 64.4 percent ture extraction from aerial and space images, abundant experiof the buildings can be detected, extracted, and accurately ence has been gained in the past few years. A collection of stateformed through this process. Remaining difficulties are high of-the-art articles can be found in the periodical proceedings percentage false alarm errors caused by the misclassification edited by Grün et al. (1995), Grün et al. (1997), and Baltsavias et of road and building classes as well as occlusion and shadows al. (2001b). Mayer (1999) presents a comprehensive survey on that may mislead the extraction process. the techniques used for image-based building extraction. These collections and surveys suggest a common understanding that Introduction Recent successfully launched satellites (Ikonos, EROS, Quickmultiple data sources and external knowledge or models about the buildings are needed to reach an automated solution. ReBird) provide high resolution (around 1 meter or better) images that can be used for the extraction of geospatial features for topported methodologies include, only naming a few as representative, the use of additional information, such as topographic ographic mapping applications (Petrie, 2001). Both the potential and methodology for this objective have been studied over maps (Haala, 1999; Förstner and Pluemer, 1997), laser ranging or elevation data collected from lidar or stereo pairs (Baillard the last two years. Li et al. (2000) and Muller et al. (2001) used simulated Ikonos imagery to investigate the obtainable accuand Maitre, 1999; Hofmann, 2001; Kim and Muller, 2001; Weidner and Förstner, 1995), multispectral and hyperspectral racy from precision photogrammetric processing. Dial (2000) reported on the investigation on the potential of Ikonos images images (Haala, 1999; Huertas et al., 1999; Huertas et al., 2000; McKeown et al., 1999), as well as various building models and conclude that the accuracy requirement for large scale (1:4,800) topographic maps can be met with Ikonos images. Bal(Braun et al., 1995; Förstner and Pluemer, 1997; Förstner, 1999). This paper is focused on automated building extraction tsavias et al. (2001a) present a comprehensive evaluation of the radiometric and geometric property and accuracy of the standguided by classification results using both multispectral and panchromatic Ikonos images. The approach developed consists ard Ikonos product. Fraser et al. (2001; 2002) used different forms of geometric transformations to further rectify the standof the following steps. First, a supervised classification is undertaken for the four-band multispectral Ikonos images with ard Ikonos image product and attempt to reach a sub-meter accuracy. Toutin (2001) used 13 panchromatic and multispecfocus on building detection. After the classification results are vectorized, the location and extent of the building objects are tral Ikonos Geo-product images over seven study sites to study the error propagation in the bundle adjustment and ortho-rectiused to define a working window over the corresponding panchromatic image for precision delineation and extraction. In fication process. It was found that a 2to 4-meter accuracy in the ortho-rectified Ikonos image can be achieved if a proper order to avoid the difficulty in determining the threshold for