Integrated Local Features to Detect Building Locations in High-Resolution Satellite Imagery

A building is one of the most important man-made objects forming part of a class of interest to photogrammetry and remote-sensing researchers that many studies in recent years have focused on extracting the locations of buildings, especially from high-resolution images, which are widely used to detect geospatial targets such as buildings. Conventional building-detection methods apply various image-processing techniques based on pattern recognition, classification and neural network methods, all of which have some problems, such as eliminating small buildings, needing a training sample, detecting trees as a building and mixing buildings with other classes. Local features are significant algorithm invariants to viewpoint and illumination changes that have been made over the past decade, to find local images of interesting structures. Each local feature algorithm has distinct characteristics that can help in building detection. In this study, a novel, integrated method is presented for detecting building locations from high-resolution satellite imagery, based on local features. In the proposed method, three types of local features, i.e. point features, circular blob-like features and regional elliptical features, are applied in order to describe building areas on high-resolution images. Considering the different characteristics of local feature extraction algorithms, such as differences in the numbers, densities, sizes and shapes of extracted features, a weighted integrated approach based on a multi-kernel probability density function has been developed. The proposed method does not require image information in different bands, as it works on black-and-white images. A blunder-detection method, based on an orientation histogram, is also used to refine the building extraction process. The proposed method is successfully applied to the detection of buildings on various satellite images, and the results demonstrate its capability in terms of recall, precision and F-score.