Abstract

Accurate weather classification is essential for the safety and adaptability of autonomous vehicles (AVs). This paper proposes a deep learning (DL) detection framework leveraging transfer learning to categorize outdoor weather conditions in both normal and adverse scenarios. We evaluated the performance of three specific CNN architectures SqueezeNet, ResNet-50, and EfficientNet utilizing Nvidia GPU hardware. The study employed a combined dataset from DAWN2020 and MCWRD2018 to train the models on six weather classes: cloudy, rainy, snowy, sandy, shine, and sunrise. Experimental results identified ResNet-50 as the superior model, achieving detection accuracy, precision, and sensitivity of 98.48%, 98.51%, and 98.41%, respectively. Additionally, the model exhibited a low inference time of 5 ms. When compared to current state-of-the-art methods, our approach improved classification accuracy by a factor of 0.5–21%. Consequently, this framework is well-suited for real-time implementation, providing AVs with on-demand, high-precision decision support.