Deep Learning Image-Based Classification for Post-Earthquake Damage Level Prediction Using UAVs

Unmanned Aerial Vehicles (UAVs) integrated with lightweight deep learning models represent an effective solution for image-based rapid post-earthquake damage assessment. UAVs, equipped with cameras, capture high-resolution aerial imagery of disaster-stricken areas, providing essential data for evaluating structural damage. When paired with light eight Convolutional Neural Network (CNN) models, these UAVs can process the captured images onboard, enabling real-time, accurate damage level predictions that might with potential interest to orient efficiently the efforts of the Search and Rescue (SAR) teams. This study investigates the use of the MobileNetV3-Small lightweight CNN model for real-time post-earthquake damage level prediction using UAV-captured imagery. The model is trained to classify three levels of post-earthquake damage, ranging from no damage to severe damage. Experimental results show that the adapted MobileNetV3-Small model achieves the lowest FLOPs, with a significant reduction of 58.8% compared to the ShuffleNetv2 model. Fine-tuning the last five layers resulted in a slight increase of approximately 0.2% in FLOPs, but significantly improved accuracy and robustness, yielding a 4.5% performance boost over the baseline. The model achieved a weighted average F-score of 0.93 on a merged dataset composed of three post-earthquake damage level datasets. It was successfully deployed and tested on a Raspberry Pi 5, demonstrating its feasibility for edge-device applications. This deployment highlighted the model's efficiency and real-time performance in resource-constrained environments.