Rapid and accurate windstorm damage assessment is critical for effective disaster response. This assessment traditionally involves manual analysis of visual data collected by experts after the disaster, which is time-consuming and cannot meet the requirements for rapid evaluation. Besides expert-collected data, images capturing the windstorm aftermath are readily available on the Internet, presenting a valuable but underutilized asset for damage assessment. However, the use of these Internet images for assessment is challenging, because most of them are collected by nonexperts, exhibit complex scenes, and are often scattered among a vast number of irrelevant images. To achieve rapid assessment with these readily available resources, a deep learning-based framework is proposed in this study. This framework includes an automated data scraping module for collecting Internet images and advanced deep learning models to simultaneously sort out irrelevant images and classify damage severity in a single step. Multiple deep learning models vision transformers (ViTs) and deep convolutional neural networks (DCNNs) are trained, evaluated, and compared on a windstorm damage dataset collected from the Internet via the developed scraping module. Furthermore, several pretraining strategies, including supervised and self-supervised masked autoencoder (MAE) pretraining, are investigated to enhance the performance of deep learning models. Results show that ViTs, when pretrained using the MAE strategy, consistently outperform DCNNs in accuracy. The best-performing model, FasterViT-2 pretrained with the MAE strategy using ImageNet, achieved 92.77% accuracy, a macro F1-score of 91.09%, and a throughput of 310.2 images per second, offering a strong balance between performance and efficiency. In particular, MAE pretraining using ImageNet significantly reduces misclassification rates between damage categories with complex and overlapping visual features, leading to more accurate predictions. These findings highlight the potential of the proposed framework to improve the speed and accuracy of damage assessment using publicly accessible Internet images, leading to rapid and effective disaster response.
Canchila et al. (Fri,) studied this question.