Natural disaster and wars wreak havoc not only on individuals and critical infrastructure, but also leave behind ruined residential buildings and housings. The size, type and location of damaged houses are essential data sources for the post-disaster reconstruction process.
Field damage assessment is time consuming and requires trained personnel, whilst remote sensing techniques can be used to provide rescue teams, reconstruction and rehabilitation authorities with damage related information and features in a timely manner.
In this work, a novel autonomous building damage detection technique that relies on both pre- and post-war aerial images is proposed. To the best of our knowledge, building damage detection due to war activities has not been discussed in the literature. Pre- and post-war images may be captured in different conditions, camera type, angle, and capturing conditions. Thus applying affine transformation as a pre-processing step is firstly used to correct for geometric distortions or deformations that occur with non-ideal camera angles.
Next, our novel building detection algorithm is applied on the pre-war image. Detected buildings positions will be projected on the post-image, resulting in a set of damaged buildings candidates. Then, thorough damage analysis is done using three main features: (i) Shadow, (ii) Correlation, and (iii) Uniformity.
Shadow can play an important role in analyzing buildings state (damaged or not). Following airstrikes, bombing explosions or other military activities, building’s structure will be partially ruined and/or totally demolished, thus, changing the shadow area and orientation. Uniformity, a statistical metric from the Gray-Level Co-Occurrence Matrix (GLCM), is used to measure a Region of Interest (ROI) homogeneity. Uniformity is a useful feature in this scope since damaged buildings are expected to show lower homogeneity in post-war Images. Correlation, lastly, measures the linear dependency of grey levels of neighboring pixels within ROI. Correlation is affected by the destruction and thus relevant towards damage estimation.
Those three features are compared in both pre- and post-war images. The validity of all three hypotheses is needed to label a specific building under study as damaged. Otherwise, the building is classified as a non-damaged residential unit.
The literature does not provide for the best of our knowledge a relevant dataset to benchmark against, and thus we focused in this work on images taken between 2014 and 2016 during the Syrian civil war. Accuracy assessment applied over several regions including different affected war zones reveals the high performance of our novel approach.
Future work includes the extension of the proposed approach to classify damaged buildings into several states based on destruction severity. In addition, we plan to assemble and publish a dataset to be used by researchers in the field.