Comparative analysis of automatic approaches to building detection from multi-source aerial data

Automatic building detection has been a hot topic since the early 1990’s. Early approaches were based on a single aerial image. Detecting buildings is a difficult task so it can be more effective when multiple sources of information are obtained and fused. The objective of this paper is to provide a comparative analysis of automatic approaches to building detection from multi-source aerial images. We analysed data related to both urban and suburban areas and took into consideration both object based and pixel-based methods. Although many of these methods perform full data classification, we focused only on the detection of building regions. Three measures were used for the evaluation of the performance of each method: number of detected buildings to their total number (detection rate), number of objects wrongly detected as buildings (false positive) and number of missed buildings (false negative) to the number of detected buildings. The data sets we used were RGB and colour infrared (CIR) orthoimages and Digital Surface Models (DSMs) obtained by an airborne laser scanner, which provides a first pulse DSM and a last pulse DSM. In addition, we derived from these data and used other four sources of information: a Digital Terrain Model (DTM) obtained from a filtered version of the last pulse DSM, the height difference between the last pulse and the DTM, the height difference between the first and the last pulse and the Normalized Difference Vegetation Index (NVDI) derived from the red and infrared channels.We analysed results coming from three classification algorithms, namely Bayesian, Dempster-Shafer and AdaBoost, applied to the features extracted both at pixel level and at object level. To obtain a very realistic comparison we used the same training set for all methods, either pixel-based or object-based. Results obtained are interesting and can be synthesised in the need of fusing (the results of) more approaches to yield the best results.