An Interpretable Transformer-Based Model for Post-Disaster Infrastructure Damage Assessment Using High-Resolution Remote Sensing Imagery

Abstract

Climate change impacts have raised the frequency of natural disasters throughout the world. The impacts of these natural events, like drought, floods, cyclones, fire, hurricanes, and others, are complex for both developing and developed countries. Specifically, post-disaster recovery and disaster risk management become highly significant to ensure strong development. With the emerging accessibility of high-resolution satellite images, AI applications like machine learning and deep learning (DL) are employed to systematize the process of interpretation. This study introduces a Post-Disaster Structural Damage Analysis using Remote Sensing Images (PDSDA-RSI) technique to design an explainable deep learning framework for detecting and assessing post-disaster structural damage from high-resolution remote sensing images. Soft voting-based classification integrated with a stacked sparse denoising autoencoder, a bidirectional gated recurrent network, and a Siamese neural network is performed to classify the post-disaster damages into multiple classes. Furthermore, rectified Adam is applied for stabilizing the model training. At last, Grad-CAM++ is employed to visualize and interpret model predictions by highlighting key regions in the image. To exhibit the improved performance of the PDSDA-RSI model, extensive simulations take place the outcomes are investigated under numerous measures. The comparison analysis reported the betterment of the PDSDA-RSI model under several metrics.