Deep learning-based classification of earthquake-impacted buildings using textual damage descriptions

Abstract

To make informed decisions and have a well-coordinated response to a major hazard event, a rapid assessment of the spatial distribution and severity of building damage is needed. To this end, the long short-term memory (LSTM) deep learning method is applied to classify building damage based on textual descriptions of damage. The damaged state of an individual building is classified using the ATC-20 tags (red, yellow and green). The application of the LSTM approach is demonstrated using building damage descriptions recorded following the 2014 South Napa, California earthquake. The dataset, which consists of 3423 buildings (1552 green tagged, 1674 yellow tagged, 197 red tagged), is randomly divided into training and testing subsets. A predictive model is established using the training set and the performance of this model is evaluated with the test set. An overall accuracy of 86% is achieved when the deep learning model is used to identify the ATC-20 tags for the test set. Despite some noted limitations, the study highlights the overall potential of using the LSTM method to rapidly assess building cluster damage using textual information, which can be generated by building professionals, stakeholders or social media platforms.