Matching Spectral with Spatial Analysis to Improve Building Damage Recognition in VHR Images Mattia Stasolla, Paolo Gamba and Fabio Dell’Acqua Department of Electronics University of Pavia Pavia, Italy Ellen M. Rathje Department of Civil, Architectural, and Environmental Engineering University of Texas at Austin Austin, Texas USA Abstract — VHR satellites can potentially provide important information for hazard management, especially for rapid damage assessment. This work aims at improving building damage recognition after natural disasters by matching spectral with spatial features in VHR optical images. The proposed procedure exploits morphological operators and a change detection algorithm to create an accurate damaged building inventory, according to standardized scales. The test area is the city of Bam, Iran, hit by an earthquake in 2003. Keywords: earthquake; damage assessment; VHR images. I. INTRODUCTION Disaster management by means of pre- and post-event very high resolution (VHR) satellite or airborne images is nowadays an acquired practice, as testified by a number of internationally funded bodies and projects [1-3]. What is still needed, however, are procedures, or at least series of processing steps, which may provide satisfying scene interpretation results in many (possibly all) situations. To this aim, refined algorithms should be devised, and implemented into software able to extract information from the scenery in a semi-automatic way. This work is devoted to the definition of such a processing chain, whose goal is to match the spectral discrimination capability of a classification methodology with the spatial analysis performed via segmentation. The procedure, though simple, pave the way to more complex improvements, based on further refinements of the input feature, or they adaptability to different VHR sensors with basically different behaviors, like optical and SAR ones. The general outline can be summarized into two basic steps (they will be discussed in the next section). 1. Pre- and post-images are classified exploiting spectral and geometrical information. The goal of this first step is to improve common spectral classifications through a segmentation approach. By segmenting an image, in fact, the classification gets at a higher level of analysis, since images are no longer just pixels, but they are made by different homogeneous regions having certain properties. 2. A change detection algorithm is applied to first step output land cover maps in order to provide an assessment index for the creation of a damaged building inventory. The next section describes in detail the whole algorithm, whose results are discussed in section III. Moreover, since the main part of preprocessing makes use of the morphological image theory, an appendix, extracted from [5], is dedicated to provide some useful background. II. DATASET AND PROCESSING CHAIN The original dataset is formed of a couple of Quickbird images (multi-spectral at 2.4 m resolution and panchromatic at 0.6 m) of Bam city, acquired on 30 th September 2003 and 3 rd January 2006, respectively before and after the earthquake occurred (26 th December 2003). To exploit the high spatial resolution of the PAN data and high spectral resolution of the MS data, two pan-sharpened images have been generated and co-registered. As said in the introduction, the proposed algorithm consists of two main steps: Classification and Damage Extent Evaluation. The former is separately applied to both pre- and post-images, the latter involves a joint analysis of the first level outputs. Figure 1. 1 st level scheme