Improving Rooftop Detection in Aerial Images Through Machine Learning

Abstract : In this paper, we examine the use of machine learning to improve a rooftop detection process, which is one step in a vision system that recognizes buildings in overhead imagery. We review the problem of analyzing aerial images and describe an existing vision system that automates the recognition of buildings in such images. After this, we briefly review two well known learning algorithms, representing different inductive biases, that we selected to improve rooftop detection. An important aspect of this problem is that the data sets are highly skewed and the cost of mistakes differs for the two classes, so we evaluate the algorithms under varying misclassification costs using ROC analysis. We report three sets of experiments designed to illuminate facets of applying machine learning to the image analysis task. One set of studies focuses on within image learning, in which both training and testing data are derived from the same image. Another addresses between image learning, in which training and testing sets come from different images. A final set investigates learning using all available images in an effort to determine the best performing method. Experimental results demonstrate that useful generalization occurs when training and testing on data derived from images that differ in location and in aspect. Furthermore, they demonstrate that, under most conditions and across a range of misclassification costs, a trained naive Bayesian classifier exceeded, by as much as a factor of two, the predictive accuracy of nearest neighbor and a handcrafted linear classifier, the solution currently used in the building detection system. Analysis of learning curves reveals that naive Bayes achieved superiority using as little as 6% of the available training data.