A stereo image matching method to improve the DSM accuracy inside building boundaries

In a digital surface model derived from high-resolution stereo images, the accuracies of building rooftop elevations are usually much lower than that of open areas without tall objects. This problem makes building height estimation difficult from the stereo images. The inaccuracy of building rooftop elevation is caused by many factors including shadows, occlusions, smoothing constraints in the matching algorithms, and the mismatch on extremely high buildings. To improve the accuracy of building rooftop elevations, this study enhances the existing image matching methods by adding building footprint maps as a constraint. The proposed image matching method consists of three steps. First, a building footprint was used to identify the corresponding building rooftop locations in the stereo images. Second, the left and right stereo images were matched according to color and shape. Third, the left and right sub-images of the building were further matched at pixel level to generate detailed roof elevations. Validation using surveyed rooftop elevations demonstrated that the proposed method can estimate building rooftop elevation with a one-third error using the current commercial software. In addition, the errors for low-rise and high-rise buildings were consistent in the proposed method.