Region Based Maximum Likelihood Estimation for Small Geospatial Object Extraction

The availability of high resolution aerial imagery makes it possible to identify small geospatial objects in dense urban areas. The challenge lies in the overlapping problem in the feature space of the object classes. In this paper, we propose a region-based maximum likelihood (RML) method for geospatial object extraction in urban areas. For accurate extraction of the object region, the target objects are segmented by SVM classification. And the colour distribution is subsequently compared against the SVM training data via a Mahalanobis colour distance and a ML approach is developed to distinguish between regions that might be overlapped in their feature space. A quantitative measure which evaluates the resulting extractions is presented. The experimental results show that the proposed approach yields intuitively correct as well as accurate extraction of objects in urban aerial images.