A Benchmark and Investigation of Deep-Learning-Based Techniques for Detecting Natural Disasters in Aerial Images

Rapid emergency response and early detection of hazards caused by natural disasters are critical to preserving the lives of those in danger. Deep learning can aid emergency response authorities by automating UAV-based real-time disaster recognition. In this work, we provide an extended dataset for aerial disaster recognition and present a comprehensive investigation of popular Convolutional Neural Network models using transfer learning. In addition, we propose a new lightweight model, referred to as DiRecNet, that provides the best trade-off between accuracy and inference speed. We introduce a tunable metric that combines speed and accuracy to choose the best model based on application requirements. Lastly, we used the Grad-CAM explainability algorithm to investigate which models focus on human-aligned features. The experimental results show that the proposed model achieves a weighted F1-Score of 96.15% on four classes in the test set. When utilizing metrics that consider both inference time and accuracy, our model surpasses other pre-trained CNNs, offering a more efficient and precise solution for disaster recognition. This research provides a foundation for developing more specialized models within the computer vision community.