PHASE-CORRELATION AND SPARSE MODELLING FOR DAMAGE DETECTION DUE TO THE 2018 SULAWESI-EARTHQUAKE TSUNAMI

Moya L.; Mas E.; Koshimura S. · World Conference on Earthquake Engineering proceedings (2021)

Resumen

On September 28, 2018, a magnitude Mw 7.5 earthquake occurred at 75 km north from the city of Palu, Central Sulawesi, Indonesia. The earthquake induced catastrophic effects, such as a series of tsunamis, landslides, and liquefaction. It has been reported more than 70,000 damaged houses. In this event, significant permanent ground deformation was observed. In fact, the ruptured fault crosses the city of Palu. In this kind of disaster events, one of the earliest needs is the identification of collapsed buildings, which may contain survivors trapped inside. An automatic procedure for such purpose can be performed from a pair of satellite images taken before and after the event. The fundamental basis lies in the detection of changes between both images. However, before detecting changes, the pair of images must be perfectly aligned. In ordinary situations, aligning a pair of images, a pre-process termed image co-registration, is a straightforward procedure. In general lines, first, some ground control points are identified, manually or automatically, in both images and, second, use them for mapping one image by some linear or polynomial transformation. Unfortunately, because image co-registration assumes smooth transformation, it did not work well for the case of the 2018 Sulawesi earthquake-tsunami. The complex ground deformations due to the fault line and the liquefaction could not be removed by linear/polynomial transformation. In this paper we report a novel procedure to identify the areas affected by the earthquake and tsunami. The method does not require image co-registration. It is based on the application of phase-correlation and sparse modeling. Consider the phase-correlation a function whose input is the two images and its output is a bi-dimensional array. On ideal conditions, phase-correlation exhibit a distinctive peak in areas without damage (no-change); otherwise, there is no such peak. In real scenarios, though, such a trend is difficult to identify in medium/low-resolution images. Therefore, statistical learning with sparsity, more precisely sparse logistic regression based on l1-regularization, is employed here to identify the significant components of the phase-correlation for the identification of collapsed buildings. The proposed framework was applied to optical and microwave images and the results achieved high accuracy.

Instituciones: Universidad Nacional de Ingeniería