Automatic building change detection through linear feature fusion and difference of Gaussian classification

Many applications in infrastructure planning and maintenance are currently aided by the collection of aerial image data and manual examination by human analysts. The increasing availability and quality of this image data presents an opportunity for computer vision and machine learning techniques to aid in infrastructure planning and maintenance. Due to the immense effort required for human analysts to view and organize the data, there is great demand for computer automation of these tasks. A strategy for detecting changes in known building regions in multitemporal visible and near-infrared imagery based on a linear combination of independent features and a difference of Gaussian based classification approach is proposed. Initial building candidates are discovered using a linear combination of features representing vegetation intensity, image texture, shadow intensity and distance from known road areas. The resulting building candidates are classified by shape using a unique difference of Gaussians technique and a standard Support Vector Machine classifier. Building regions reported in the reference data set from the prior observation time are revisited using the same classification approach to minimize the number of false positive detections from the feature fusion strategy. The effectiveness of the proposed technique is evaluated on five wide area real-world images. Ground truths for the building regions in all five images are manually created and used to measure the accuracy of the building detection and change detection results. Detection statistics and visualized results of the proposed algorithm are presented, and it is observed that the results are promising compared to the manually created ground truth. As a possible continuation of this research, a brief discussion on parameter estimation for building change detection based on image characteristics is included