ON THE CLASSIFIER PERFORMANCE FOR SIMULATION BASED DEBRIS DETECTION IN SAR IMAGERY

Abstract

Abstract. Urban areas struck by disasters such as earthquakes are in need of a fast damage detection assessment. A post-event SAR image often is the first available image, most likely with no matching pre-event image to perform change detection. In previous work we have introduced a debris detection algorithm for this scenario that is trained exclusively with synthetically generated training data. A classification step is employed to separate debris from similar textures such as vegetation. In order to verify the use of a random forest classifier for this context, we conduct a performance comparison with two alternative popular classifiers, a support vector machine and a convolutional neural network. With the direct comparison revealing the random forest classifier to be best suited, the effective performance on the prospect of debris detection is investigated for the post-earthquake Christchurch scene. Results show a good separation of debris from vegetation and gravel, thus reducing the false alarm rate in the damage detection operation considerably.