A lightweight, robust, real-time disaster detection network

As natural disasters become more frequent, low-cost, low-power edge devices are critical for monitoring in harsh environments. To address resource-constrained disaster scenarios, we propose EdgeDisNet, an ultra-lightweight neural network based on EfficientNet architecture. It enhances classification performance through efficient feature modeling. EdgeDisNet achieves 0.09 M parameters and 66.72 MFLOPs, with F1-scores of 0.946 and 0.960 on AIDER and AIDERv2 datasets. Hardware validation on NVIDIA Jetson Orin confirms an end-to-end latency of 5.37 ms per image, corresponding to a throughput of 186.2 FPS, providing robust support for efficient disaster management.