Adverse weather conditions, including fog, heavy rain, and snow, pose significant challenges for vision-based systems in road safety and autonomous driving. This review synthesizes recent advances in deep learning for road scene analysis under degraded weather conditions, presenting a taxonomy organized by task (classification, detection, segmentation), data sources (real versus synthetic), and model architectures (CNNs, GANs, transformers). The methodology integrates both real-world and synthetic datasets, assesses robustness using standard benchmarks such as mIoU, PSNR, and FID, and compares architectures for accuracy, inference efficiency, and generalization capacity. Findings indicate that GANs significantly enhance training diversity through synthetic augmentation, transformers improve global context modeling, and hybrid designs offer promising robustness, albeit at the expense of computational efficiency. Despite these advances, current approaches remain limited by the availability of meteorologically diverse datasets and weak geographic transferability. Future research should prioritize the development of multimodal and geographically balanced benchmarks, lightweight hybrid architectures, and sensor fusion strategies that ensure real-time adaptability. This review contributes both a comparative synthesis and a roadmap toward resilient perception systems capable of reliable operation in adverse weather.