Enhancing Post-Disaster Damage Detection and Recovery Monitoring by Addressing Class Imbalance in Satellite Imagery Using Enhanced Super-Resolution GANs (ESRGAN)

Abstract. Access to very high-resolution (HR) satellite imagery is often limited, delayed, or cost-prohibitive, restricting accurate and timely post-disaster damage detection and recovery monitoring (PDDRM). Additionally, class imbalance in disaster classification datasets further complicates deep learning (DL)-based assessments. This study addresses these challenges by leveraging ESRGAN to enhance low-resolution (LR) satellite imagery, thereby improving damage classification accuracy and the ability to monitor post-disaster recovery over time with three state-of-the-art DL models: Vision Transformer (ViT), ConvNeXt, and MaxViT for PDDRM classification across four key recovery states: Not Damaged, Not Recovered, Recovered, and New Buildings. To generate super-resolution (SR) images, LR images were first paired with HR images to train ESRGAN. Numerical evaluations using Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index Measure (SSIM) between SR and HR images confirm that ESRGAN effectively reconstructs high-resolution features, with Not Damaged (PSNR: 29.2, SSIM: 0.78) and New Buildings (PSNR: 30.3, SSIM: 0.81) exhibiting the highest reconstruction quality. ESRGAN-generated SR images were then compared against LR images in terms of classification accuracy and reliability. The results demonstrate that SR improves classification accuracy and precision, particularly for ViT and ConvNeXt, with ViT achieving an accuracy of 84% and ConvNeXt 82% on SR images, compared to 79% and 78% on LR images. We also employed Grad-CAM++ visualizations to interpret model predictions, which highlighted reliability improvements in certain classes. This study demonstrates that SR is a scalable and cost-effective alternative to very high-resolution satellite imagery, reducing dependency on expensive data sources while improving classification accuracy for PDDRM.