Weather Detection Systems (WDS) are of utmost importance in providing vital information for the decision-making processes of autonomous vehicles, especially when faced with adverse weather conditions. Deep learning techniques provide a reliable and effective solution for enabling autonomous vehicles to accurately perceive and interpret outdoor weather conditions. This capability plays a crucial role in facilitating adaptive decision-making in diverse and dynamic environments. This research paper introduces an innovative detection framework leveraging Deep Learning (DL) techniques, tailored to accurately classify weather conditions encountered by autonomous vehicles in both regular and challenging scenarios. The developed framework utilizes transfer learning methodologies and capitalizes on the computational capabilities of the Nvidia GPU to assess the effectiveness of three specific deep Convolutional Neural Networks (CNNs), namely SqueezeNet, ResNet-50, and EfficientNet. The evaluation process entails the utilization of two modern weather imaging datasets, specifically DAWN2020 and MCWRD2018. These datasets encompass a total of six distinct weather classes, namely rainy, sandy, cloudy, snowy, sunny, and sunrise. The experimental findings provide evidence of the exceptional classification capabilities of all the models under consideration. Notably, the ResNet-50 Convolutional Neural Network (CNN) model showcases outstanding performance metrics, attaining precision, accuracy, and sensitivity rates of 98.51%, 98.48% and 98.41%, respectively. Moreover, it is worth noting that the ResNet-50 Convolutional Neural Network (CNN) model demonstrates an exceptionally brief duration for detection, with an average inference time of 5 milliseconds when leveraging the Graphics Processing Unit (GPU) component. The results of our comparative analysis, when compared pre-trained models, demonstrate the superior accuracy of our proposed model. We observed significant improvements in classification accuracy across the six weather conditions classifiers, ranging from 0.5% to 21%. Therefore, the aforementioned framework presents itself as a viable and efficient solution for the timely execution of tasks, offering autonomous vehicles the ability to detect objects accurately and promptly. This, in turn, improves the vehicles' decision-making process in complex and ever-changing surroundings.

Index Terms- Autonomous vehicle, weather condition, Transfer learning, Deep learning, SqueezNet, CNN, ResNet 50, and Efficient-b0.