Unsupervised domain adaptation (UDA) for semantic segmentation has achieved significant progress, yet most existing approaches emphasize high-level feature alignment while overlooking critical image-level priors such as illumination, color, and texture. These low-level discrepancies often blur semantic boundaries and degrade structural consistency in the target domain, limiting generalization. To address this issue, we propose a UDA framework for semantic segmentation that leverages image-level priors to enhance cross-domain adaptation. At the image level, we propose a Spatial-Frequency Joint Alignment (SFJA) module, which extracts informative priors by spatial-domain structures and frequency-domain characteristics of the input images, thereby mitigating domain shifts caused by lighting variations and texture artifacts. To further enhance semantic representation, at the feature level, we design a Multi-Scale Cross-axis Interaction (MSCI) module to capture long-range semantic dependencies and fine-grained structural cues across multiple scales, thereby improving boundary delineation and contextual understanding. The SFJA module injects structurally coherent image-level priors into the MSCI module, guiding feature-level adaptation with informative, appearance-aware representations. Extensive experiments conducted on multiple synthetic-to-real benchmarks demonstrate that the proposed method achieves performance comparable to state-of-the-art approaches in both quantitative metrics and qualitative assessments.