Post-Disaster Damage Assessment from High-Resolution Satellite Imagery using Dilated LinkNet Hierarchical Radon Transformer

Abstract

Effective post-disaster damage assessment is critical for supporting timely disaster response and recovery efforts. To address the challenges of large-scale damage analysis, in this study, we propose a novel deep learning model, the Dilated LinkNet Hierarchical Radon Transformer (DL-HRT) network. The DL-HRT network integrates feature extraction, change detection, and damage classification within a unified framework, enabling accurate and efficient satellite imagery analysis. This model was trained on datasets including the Haiti Earthquake dataset and the xBD dataset from the xView2 Challenge, achieving precise classification of damage severity across four levels, from minor damage (level 1) to destroyed (level 4). Experimental results demonstrate that the DL-HRT network outperforms state-of-the-art models, such as DAHiTrA and BiTransUNet, in terms of accuracy, computational efficiency, and robustness. Such a key innovation lies in its incorporation of a dilation module within the U-Net architecture, preserving low-level features during down-sampling. This capability enhances the network’s ability to effectively capture fine-grained details and global contextual information. Additionally, the inclusion of hierarchically structured difference blocks, leveraging positional embeddings derived from Radon image transformations, improve both classification speed and accuracy. By providing a comprehensive and efficient solution for post-disaster damage assessment, the DL-HRT network has the potential to significantly enhance disaster management operations, aiding rapid and informed decision-making during response and recovery efforts.