Automatic Classification of Collapsed Buildings Using Object and Image Space Features

Abstract

We present a method based on two kinds of image-extracted features comparing stereo pairs of aerial images before and after an earthquake. The study area is a part of the city of Bam, Iran which was hit strongly by an earthquake on December 26, 2003. In order to classify damages caused by earthquakes, we have explored the use of two kinds of extracted features: volumes (defined in object space) and edges (defined in image space). For this purpose, digital surface models (DSM) were created automatically from pre- and post-earthquake aerial images. Then the volumes of the buildings were calculated. In addition, a criterion for edge existence - in accordance with pre-event building polygon lines — from post-event images is proposed. A simple clustering algorithm, based on the nearest neighbor rule was implemented using these two features simultaneously. Based on visual inspection of the stereo images, three levels of damage (total collapse, partial collapse, no damage) were considered. By comparing pre- and post-earthquake data the results have been evaluated. The overall success rate — total number of correctly classified divided by the total number of samples — was found to be 71.4%. With respect to the totally collapsed buildings we obtained a success rate of 86.5% and 90.4% in terms of producer’s and user’s accuracies respectively, which is quite encouraging. The results of the analysis show that using multiple features can be useful to classify damages automatically and with high success rate. This can give first very valuable hints to rescue teams.