Tornadonet-Lite: a Lightweight Deep Learning Framework for Real-Time Tornado Video Detection on Mobile Devices

Tornadoes are highly destructive weather phenomena characterized by their short duration and small spatial scale, making them difficult to capture using conventional meteorological observation systems, which are primarily designed for long-duration, large-scale weather events. The rise of short videos on social media offers a novel opportunity for tornado observation, enabling real-time detection and reporting when equipped with tornado detection algorithms on mobile devices. However, the computational limitations of mobile platforms make it challenging to deploy conventional detection algorithms in real-time applications. To address this issue, we propose TornadoNet-Lite, a lightweight video detection model specifically designed for mobile deployment to enable rapid and efficient tornado detection. First, we build a high-capacity teacher model, TornadoNet, which achieves reliable tornado video detection at the expense of a large number of parameters. To reduce model complexity, we then apply offline knowledge distillation, using the teacher model to guide the training of a lightweight student model. During inference, we employ structural re-parameterization techniques, particularly merging multi-branch depthwise convolutions in the inverted residual blocks, to minimize computational and memory overhead. Experimental results demonstrate that TornadoNet-Lite achieves an 85.3 % F1 score, while reducing model parameters by 41 % and inference memory usage by 23 % compared to the classical lightweight model MobileNetV2. In real-world deployment tests within the StormEye mobile application, TornadoNet-Lite exhibited strong real-time performance and practical feasibility, validating its effectiveness for on-device tornado video detection.