Domain shift in deep learning occurs when labeled source data and unlabeled target data follow different distributions, degrading model performance. Unsupervised Domain Adaptation (UDA) tackles this by learning transferable features from labeled source data. While marginal distribution alignment is well-studied, aligning the feature distribution conditioned on class labels remains challenging. We propose RAW-Mix, a novel UDA framework combining region-aware data mixing, adversarial training, and self-supervised learning. RAW-Mix consists of two key modules: Marginal Domain Alignment (MDA), which aligns feature distributions via adversarial learning, and Discriminative Feature Alignment (DFA), which extracts target-discriminative regions using self-attention and kernel density estimation, blending them with source patches for augmentation. This enhances conditional alignment through self-supervised learning. We demonstrate the effectiveness of RAW-Mix, showing significant performance improvements on various domain adaptation datasets, including Cityscapes, Foggy Cityscapes, Sim10K, and KITTI.