Digital twin-enabled post-disaster damage and recovery monitoring with deep learning: leveraging transfer learning, attention mechanisms, and explainable AI

This study presents a novel approach to combine Digital Twins (DTs) with advanced deep learning (DL) models to deliver accurate, timely and actionable insights. Very high-resolution multi-temporal satellite images from various platforms were employed, covering pre-disaster, immediate post-disaster, and recovery phases. The dataset used consists of 1,303 labeled samples, distributed across three classes: ‘not damaged’, ‘recovered’, and ‘not recovered’. By leveraging transfer learning and spatial and multi-head attention mechanisms with state-of-the-art DL models, we significantly enhance model performance and reduce training time to address the (near) real time processing needs of DTs. Multi-head Attention ShallowNet (MAS), specifically developed for this task, achieved an accuracy of 83% with a training time of just 26 min. RegNetX002 demonstrated the highest accuracy (88%) but required significantly more time for training (519 min). Additionally, this study integrates Explainable AI methods, including Saliency Maps and Grad-CAM, to provide transparency, reliability checks, and detailed insights into the models’ decision-making processes. The results indicate that MAS demonstrates reliable performance with a balanced focus across pre-disaster, event, and post-disaster timeframes, effectively identifying critical regions for damage and recovery assessment. This comprehensive approach lays the foundation for integrated solutions in real-time response and recovery monitoring with DTs.