A Classification Approach to Finding Buildings in Large Scale Aerial Photographs

Automatic building extraction remains an open research problem in digital photogrammetry. While many algorithms are proposed for building extraction, none of these solve the problem completely. One of their limitations is in the initial detection of the presence or absence of a building in the image region. One approach to the initial detection of buildings is to cast the problem as one of classification, where the image is divided into patches that either contain or do not contain a building. Support Vector Machines (SVMs) are a relatively new classification tool that appear well suited to this task. They are closely related to other machine learning techniques such as neural networks but have a stronger base in statistical theory and produce a generalised solution to the classification problem, using the principles of structural risk minimisation. They have been used successfully in other image classification and object recognition problems. Due to the high resolution of digital aerial photographs, compression and characterization of the image content is an essential part of the process. While many methods are available for this, an over-sampled, multi-resolution form of the Haar wavelet is a simple and convenient method for establishing coefficients that can be used in the machine learning phase. A Support Vector Machine (SVM) was trained on a large sample of building and non-building examples and achieved very high training accuracy. The trained SVM was then used to classify previously unseen image patches from the same photography and achieved an accuracy of more than 80%. Images from a variety of other sources were also tested and the classification accuracy was again consistently high. The relative simplicity of the method and the high success rates suggest that these techniques are quite promising for use in the initial detection of buildings in aerial images and may be a useful adjunct to the suite of algorithmic tools employed for building recognition and extraction.