<p>RGB-T salient object detection aims to fuse information from visible and thermal infrared modalities to accurately localize salient regions in complex scenes. However, existing methods primarily rely on spatial-domain features for cross-modal fusion and fail to fully exploit the potential of frequency-domain information. Additionally, inherent differences in spatial structures and frequency distributions often constrain cross-domain feature fusion, leading to suboptimal fusion results. To address these issues, we propose a dual-branch spatial-frequency integration network (SFINet), designed to effectively achieve complementary fusion and cross-domain alignment of bimodal information. Specifically, we introduce a spatial-domain cross-modal fusion module and a frequency-domain cross-modal perception module, which capture long-range dependencies and global structural information, respectively. Furthermore, to overcome semantic inconsistencies between cross-domain features, we propose a dual-domain semantic interaction module that enhances fusion through semantic alignment and feature interaction. Additionally, a feature decoupling module is designed to extract both the main and edge details of the fused features, further improving localization accuracy and boundary clarity through a custom loss function. Experimental results show that SFINet outperforms state-of-the-art methods across multiple benchmark datasets, demonstrating its superior robustness and broad generalization capability. The code is available at <a href="https://github.com/icehire/SFINet.">https://github.com/icehire/SFINet.</a></p>