Towards Efficient Edge Computing: Optimizing Lightweight Models for Weather Classification

In edge computing environments, weather classification plays a vital role in various applications, such as optimizing crop growth, solar energy production, autonomous vehicle routing, and video surveillance. Traditional weather classification models, primarily based on ResNet, have recently been surpassed by Vision Transformers (ViT), which offer high accuracy. However, most ViT models are designed for server environments, leading to limitations in data transmission delays, operational costs, and application-specific optimizations. This paper proposes a method to generate and optimize lightweight weather classification models for edge computing environments by applying post-training quantization to ViT models. By converting model weights and activation values to low-bit representations, we significantly enhance computational speed and reduce memory usage. Performance analysis on the Nvidia Jetson AGX Orin board demonstrates that our lightweight models maintain high accuracy while greatly improving processing speed and reducing model size compared to non-quantized models.