Edge-Accelerated Hail Damage Detection Using Lightweight Neural Models and On-Site Weather Sensor Integration

The increasing frequency and severity of hailstorms due to climate change pose significant risks to agriculture, infrastructure, and public safety. Traditional methods for hail damage assessment are often slow, labor-intensive, and subjective, leading to delayed response and financial losses. While deep learning models offer a promising alternative, their high computational cost and latency make them unsuitable for real-time, on-site deployment. This paper proposes a novel framework for edge-accelerated hail damage detection that integrates a pruned and quantized YOLOv8 object detection model with real-time data from on-site weather sensors. Our methodology focuses on creating a highly efficient model through a structured pruning and post-training integer quantization pipeline, reducing its size by 92% and inference time by 78% compared to the baseline, with only a marginal 2.1% drop in mean Average Precision (mAP). The system is further refined by a sensor-fusion gating logic, which activates the visual analysis only when specific meteorological thresholds (e.g., hail kinetic energy, precipitation rate) are exceeded, thereby conserving edge resources. Experimental results on a custom dataset of vehicle and rooftop hail damage demonstrate that our optimized model achieves an inference speed of 18 ms per image on a Jetson Nano, making it suitable for real-time applications. This research validates the feasibility of deploying robust AI models on resource-constrained edge devices, paving the way for rapid, automated hail damage assessment in field deployments.