Rapid post-earthquake damage assessment of building portfolios through deep learning-based component-level image recognition

Abstract

Frequent seismic events significantly heighten the likelihood of the building structure being impacted throughout its operational lifecycle. Seismic effects on these structures result in a notable reduction in structural safety and serviceability. Enabling the post-earthquake recovery of building structures necessitates a precise and swift assessment of earthquake-induced damage. The conventional method for assessing building damage involves collecting data through close manual observation or contact inspection. Although it can obtain relatively accurate results by combining evaluation specifications, it is time-consuming and labour-intensive. Deep learning (DL) is data-driven and employs computational methods for data processing. The image classification function of convolutional neural network (CNN) in DL brings great convenience to image data processing. It has been applied to various aspects of structural health monitoring/post-disaster assessment. However, most of the current studies focus on the component level and lack of a comprehensive perspective on the whole structure, which is not conducive to the judgment of the overall condition. Consequently, this paper proposes a rapid assessment method of the overall damage level of post-earthquake buildings based on component images and DL. Two component-level image classification models, component type and damage level, are firstly trained. Then, the images of the buildings to be evaluated after the earthquake are collected and classified. Combined with the weights of different component types and damage levels proposed, the overall condition scores and grades of the structures can be obtained by weighted calculation. The proposed method realizes an overall evaluation of the structural damage condition after seismic events, and further extends to the building portfolios, providing actionable guidance for subsequent personnel and fund allocation, rescue, and maintenance measures. Due to limitations in the dataset, such as the potential biases in the training dataset, the trained model is not perfect and faces challenges in distinguishing between minor and moderate damage. In the future, the dataset should update and ideally cover a wide range of building types, component types, and damage levels to ensure the model's robustness and applicability to various real-world scenarios.