The seismic product proposed by APHORISM is a Likelihood Index Damage Map (LIDaM). LIDaM is generated at two different scales: single buildings scale by using very high resolution imagery, and groups of buildings scale with medium resolution satellite data. Examples of LiDAM have been generated for two events: Port-au-Prince January 2010 and Christchurch February 2011 earthquakes.
Damage map at city block scale
The damage map at city block scale consists in a classification of Collapse Ratio damage index, which is defined as the ratio between the number of collapsed buildings and the total number of buildings within a city block. The city blocks for the Port-au-Prince case study have been extracted by using the online database of Open Street Map service.
Two unsupervised algorithms have been tested: a standard K-means procedure and a new one, called Features Stepwise Thresholding - FST. This last is an iterative algorithm, based on the knowledge of the trend (increasing or decreasing) of each satellite feature with respect to the damage grade. In order to assign the final Collapse Ratio class to a city block, the majority of the damage class is considered after n iterations.
As far as the satellite features is concerned, we exploited for this classification the Normalised Difference Index (NDI), the Kullback-Leibler Divergence (KLD) and the Mutual Information (MI), for the optical data, and the Intensity Correlation Difference (ICD and KLD for the SAR data. NDI, KLD and MI have been calculated from GeoEye-1 pre- and post- seismic optical images, at 2m resolution per pixel, while for the SAR features three images (two pre-event, and one post event) acquired by the German mission TerraSAR-X have been used. The resulting damage map is shown below.
Collapse Ratio map from FST algorithm.
Damage map at single building scale
A damage map at single building scale was generated from Satellite imagery and Structural Vulnerability data, considering the earthquake that hit Christchurch city on February 2011.
A Naïve Bayes classification approach was applied in order to associate a probability of damage to each building in the area of interest based on change features extracted from a pair of optical images collected by the WorldView-2 satellite sensor before and after the earthquake. Probability Density Functions conditioned to the class were estimated from training data selected by visual inspection of the satellite images. For producing the final damage map, the features subset giving the best classification performances on the training set, measured trough the Cohen’s Kappa, was selected. It was composed by: Mutual Information from Pansharpened data, change in Contrast, Energy and Homogeneity from Panchromatic (PAN) data, changes in Hue and Saturation from pansharpened (PSH) data.
The structural vulnerability assessment of Christchurch’s buildings and the evaluation of the expected damage scenarios were performed according to two vulnerability models, namely: a macroseismic model and a mechanical one. The latter was introduced in the Aphorism procedure, in consideration of the high demand for vulnerability estimation of the final users that persuaded the team to make an extra effort in this direction.
Finally, change detection products from Optical data and Structural module were integrated according a Bayesian data fusion approach. Based on the posterior probabilities resulting from the data integration, buildings were labelled as damaged or undamaged generating the final damage product. In particular, a building was classified as damaged if its corresponding probability of collapse was greater than 0.9, which is equivalent to imposing a collapse prior probability of 0.1.
Christchurch building scale damage map from Optical and Structural data (close up).