Seismic damage assessment in Sarpole-Zahab town (Iran) using synthetic aperture radar (SAR) images and texture analysis

Abstract

The synthetic aperture radar SAR system with the capability of imaging during the night, day, and the all-weather conditions has a high potential in change detection on the ground surface. In this research, we used three SAR images of ALOS-2 satellite over Sarpole-Zahab town in the west of Iran that had an earthquake with the magnitude of 7.3 on November 12, 2017. The effects of speckle noise on the accuracy of the results were assessed based on noise reduction filters. Correlation coefficient, difference of intensity (in five window sizes), and difference of coherence and texture (in six window sizes) of the pre- and post-event images were calculated, and the output parameters were extracted. Then, the damage assessment was carried out based on four machine learning classifiers, containing the random forest (RDF), the support vector machine, the naive Bayes classifier, and K-nearest neighbor. The RDF showed an overall accuracy of 86.3%. Seventy percent of the dataset was used for training, and 30% of it was used for the prediction purpose (~ 300 buildings). Based on the training dataset, the total number of structures in the study area was predicted (approximately 9200 buildings). Finally, a discriminant analysis was carried out among the damaged and undamaged buildings.