Physics-Driven Lightweight Weather Image Classification via Ultra-Lightweight Optical Attribute Modulation

Abstract

Weather image classification is essential for autonomous driving and outdoor visual perception. However, existing deep learning methods face two critical challenges: fine-grained visual confusion between sunny and cloudy conditions, and domain shift between training and deployment data. This paper reveals a systematic blind spot in standard convolutional neural networks—they cannot explicitly perceive global scene-level physical statistics that serve as critical cues for distinguishing weather conditions. We propose RawOAM (Raw Optical Attribute Modulation), an ultra-lightweight module that extracts six groups of physical scene statistics from raw RGB images in a strictly parameter-free manner, and maps these physical priors to channel-wise calibration weights via a tiny two-layer MLP that directly modulates the backbone’s feature maps. Built upon the FastViT-T12 architecture, our method achieves an F1-macro score of 0.9282 on a public weather dataset, outperforming six representative lightweight baselines, while reducing the validation-test generalization gap from 6.7% to 6.1%. Experiments demonstrate that physics-guided modulation not only improves absolute accuracy but also alleviates the classification degradation on cloudy and sunny images caused by traditional feature concatenation methods.