An object oriented approach to detect earthquake damage in urban area from VHR optical imagery

When an earthquake occurs a rapid and accurate damage assessment of the hit urban area is essential. In this contest, valuable information can be obtained from Earth Observation (EO) data at Very High Resolution (VHR) which provide detailed information on single structures present in the scene, allowing to produce damage maps at single building level.In this work we propose an object-oriented image analysis approach to detect damaged buildings from a pair of VHR optical images acquired before and after a seism. We have tested our procedure using images taken by the QuickBird satellite before and after the earthquake that hit L'Aquila city (Italy) on April 6, 2009. Two layers of polygons reporting the damage ground survey of the whole urban area of L'Aquila, coming from two different sources, are used for validation purposes. The first one is related to the survey performed by the National Institute of Geophysics and Volcanology (INGV) Macroiseismic team, while the second one refers to ground survey carried out by the Italian Civil Protection Department (DPC). The geolocated version of the data set has been provided by the Construction Technologies Institute of the Italian National Research Council (ITC-CNR). As usual in any object oriented image analysis approach, the proposed methodology comprises two steps: the image segmentation, for obtaining a buildings map, and the object classification, for detecting damaged buildings. In this work the segmentation of the pre-event optical image in objects corresponding to the buildings in the scene is performed using a pre-existing building map provided as GIS layer, but it can also be obtained by means of suitable segmentation/classification algorithms applied to a pre-event VHR optical image. Once the objects are identified, a feature extraction step is carried out in order to build the classifier input space. Several features can be computed within the objects in order to identify those changed due to the earthquake. Here we consider: a) change metrics derived from the Information theory, such as Kullback-Leibler divergence and the Mutual Information; b) changes in the textural parameters derived from the grey level co-occurrence matrix; c) changes in the colour space, i.e. differences in the Hue, Saturation and Intensity parameters. Instead of using a moving window with fix size, all the aforementioned change features are evaluated at object level, by considering only the pixels within the building footprints. The discrimination between collapsed or heavy damaged buildings and less damaged or undamaged buildings is performed in the Bayesian framework. A non parametric approach is used as the statistics of the features cannot be easily predicted and a leave-one-out validation approach has been carried out to test the classification accuracy against ground truth. Although the classification performances using EO data are not excellent, similar uncertainties were observed between the ground datasets. Compared to the highly costly and time consuming ground surveys, the EO still appear a valid tool for a rapid response after an earthquake. The work has been funded by the EC-FP7 APhoRISM project (Research, Technological Development and Demonstration Activities, grant agreement n. 606738).