Verification of Tsunami Reconstruction Projects by Object-oriented Building Extraction from High Resolution Satellite Imagery

A powerful earthquake (Richter magnitude 9.2) occurred near to the northwest shore of Sumatra, on 26 December, 2004. It triggered a giant tsunami that devastated Banda Aceh, Indonesia. Many donors provided recovery aid, without coordination or proper auditing. This may have led to waste, fraud and corruption. This study investigated an application of remote sensing to enhance financial accountability and transparency in managing reconstruction projects following natural disasters, by automatically identifying buildings constructed as a result of the disaster response, using Banda Aceh as a test area. The increasing availability of high-resolution satellite images, such as the KOMPSAT-2 used in this study, together with aerial orthophotos, makes such a procedure potentially a practical part of a disaster recovery audit.The segmentation algorithm of eCognition was used to generate image segments. These segments were then classified as "building" and "background" by using a rule-base decision tree based on ancillary information: texture, contextual and semantic properties of objects. Building footprints were extracted to a GIS.Accuracy was assessed by four methods. First was the traditional approach of generating random points and computing an error matrix. This gave accuracies of 98.6% (user's) and 63.4% (producer's) for the "building" class. The second method was based on the overlaying of geometric centres of extracted and manual-identified buildings, with a threshold based on building size. This method gave accuracies at the optimal threshold of 81.0% (correctness) and 84.7% (completeness). The third method applied a bounding box to the extracted and reference data, to take both shape and size into account. The ratio of length to width was defined as the shape condition, and the ratio of areas as the size condition; these were then averaged, giving accuracies of 82.3% (correctness) and 82.5% (completeness). The fourth method combined the second and the third methods, giving the highest accuracy. None of the object-based assessments accounted for "one to many" and "many to one" relationships between extracted and reference data.New Buildings were separated from old by overlaying the extracted footprints with a pre-reconstruction building map, taking those with common areas less than 50% as new buildings. These are then ready for audit.Building footprints were successfully extracted from high-resolution images by object-oriented classification. Remaining problems include identification of multi-faceted roofs and connected buildings, and correction for these.