Performance evaluation of automated approaches to building detection in multi-source aerial data

Automated approaches to building detection in multi-source aerial data are important in many applications, including map updating, city modeling, urban growth analysis and monitoring of informal settlements. This paper presents a comparative analysis of different methods for automated building detection in aerial images and laser data at different spatial resolutions. Five methods are tested in two study areas using features extracted at both pixel level and object level, but with the strong prerequisite of using the same training set for all methods. The evaluation of the methods is based on error measures obtained by superimposing the results on a manually generated reference map of each area. The results in both study areas show a better performance of the Dempster-Shafer and the AdaBoost methods, although these two methods also yield a number of unclassified pixels. The method of thresholding a normalized DSM performs well in terms of the detection rate and reliability in the less vegetated Mannheim study area, but also yields a high rate of false positive errors. The Bayesian methods perform better in the Memmingen study area where buildings have more or less the same heights.