Reported works on cyber-physical disaster systems (CPDS) deal with the assessment of loss and damage aftermath of a large-scale disaster such as earthquake, wildfire, and cyclone, etc. involves collecting data from the IoT devices sent to the cloud data centers for analysis, often causes high bandwidth usage with substantial delay. In our work, we have shown to eliminate bandwidth cost and reducing latency substantially suitable for post-disaster response for rescue operations. We propose a low-latency and energy-efficient CPDS applying cloud-IoT-edge by bringing intelligence and infer-encing proximity to the disaster site to detect the disaster events in real-time and inform to the rescue teams. The edge computing model of CPDS uses convolutional neural network (CNN) with MobileNetV2 lightweight model and gradient weighted class activation mapping (Grad-CAM++) to locate and quantify degree of the damage into classes- severe, mild, and no damage. We implemented CPDS on a real-world laboratory testbed that comprises resource-constrained edge devices (Raspberry Pi, smartphones, and PCs) and docker-based containerization of deep learning models and analyzed the computational complexity. With the rigorous experiments of the proposed approach, we evaluated the performance in terms of classification accuracy, energy saving, and end-to-end (E2E) delay comparing with the current state-of-the-art approaches.