Earthquake building damage classification based on full suite of Sentinel-1 features

Abstract This study investigates the capabilities of the proposed approach that fully utilizes the intensity, phase, and polarimetry features of Sentinel-1 for multi-class and binary classification of building damage, aiming to address the limitations of existing studies, which typically only utilize partial Sentinel-1 features or integrate them with an ancillary dataset for building damage classification. Using the Puerto Rico earthquake and Turkey earthquake as case studies, multi-class classification achieved F 1 score of 0.50/0.58, an overall accuracy of 0.52/0.60, and receiver operating characteristic – area under the ROC curve (AUC) value of 0.59/0.69, respectively. For binary classification under varying proportions of collapsed buildings and no damaged buildings, overall accuracy ranged from 0.631 to 0.813 and F 1 score from 0.630 to 0.806 across the two earthquakes. The performance for both classification tasks is comparable to previous studies that employed other synthetic aperture radar data or integrated one or two Sentinel-1 features with ancillary dataset, and the performance is better than that of studies using only one or two Sentinel-1 features for the destroyed vs undamaged classification. These quantitative comparisons highlight that our proposed approach, which solely rely on Sentinel-1, exploiting all features of Sentinel-1 data and using a readily applied machine learning model, is suitable for the timely, rapid, and microregion-level building damage detection in the local region after the earthquake.