Post-Disaster Damage Analysis Using Deep Learning Techniques and Convolutional Neural Networks (CNNs)

In disaster response and recovery, post disaster damage assessment is an important piece that should be done. Damage assessment methods of the traditional type tend to be laborious, time consuming and prone to human error. This thesis investigates recent advancements in deep learning forums — such as Convolutional Neural Networks (CNNs) — in using automated and improved accuracy to automatically and optimize the damage analysis for postdisaster. This work presents a comprehensive framework of post disaster building damage detection through CNN’s trained on high resolution satellite and aerial imagery. Using the state of the art deep learning architectures including multi scale feature extraction and attention mechanism, our proposed approach is able to obtain the increased classification accuracy and increased segmentation precision on detection of damaged structures. CNN’s have proven to be effective in disaster situation scenarios including earthquakes, floods, and hurricanes (disaster caused) already using them to assess structural damage levels [3]. By combining the multi-temporal analysis and explainable AI (XAI) techniques, we improve prior methodologies by bringing model interpretability and decision making [19]. We demonstrate that our model is bereft with errors, valid for a wide range of disasters, and surpasses other methods in terms of accuracy as well as adaptability to different types of disasters obtained on benchmark disaster datasets. Our findings support the effort to enhance AI-driven decision support systems for quick disaster response and asset allocation. We then integrate real time satellite feed and improve the transfer learning strategies for adapting to unseen disaster events domain.