<p>RGB-T semantic segmentation is a fundamental task in computer vision. It plays a significant role in various applications, such as autonomous driving, nighttime surveillance, and human–computer interaction. Recognizing that image features can be characterized in both the spatial and frequency domains, and that prior research has often focused on one while neglecting the potential of their synergy, this paper aims to achieve an effective integration of the two. To this end, we propose FTA-Net, a novel cross-modal fusion network. It employs a dual-branch architecture to explore the complementary relationship between RGB and thermal infrared features from both the spatial and frequency domains. Specifically, the top-k attention-guided differential fusion (TAD) branch leverages difference convolution and a top-k attention mechanism to model local salient regions, thereby achieving fine-grained inter-modal interaction and noise suppression. Concurrently, the Fourier cross-modal fusion (FCF) branch captures multi-scale contextual information in the frequency domain to augment the global feature modeling capability. By fusing the features from these two branches, FTA-Net achieves a synergistic optimization of local precision and global consistency. Extensive experiments on the public datasets MFNet and PST900 demonstrate its superior performance, achieving mIoU scores of 59.3% and 87.58%, respectively. These results significantly surpass those of current state-of-the-art methods, validating the effectiveness and superiority of the proposed dual-domain fusion paradigm. The experimental results and code of our network can be accessed at the following URL: <a href="https://github.com/shen2325/FTANet.git">https://github.com/shen2325/FTANet.git</a></p>