<p>Gas leakage poses a significant threat to workplace safety in industrial settings, and as the features of the leaked gas are subtle, they are difficult to detect effectively. Methods based on the fusion of RGB and thermal images have demonstrated considerable potential in gas leakage detection tasks, owing to the complementary information fusion advantages they afford. However, most RGB-Thermal (RGB-T) image fusion models struggle to achieve high-precision segmentation results while maintaining low parameter counts and are therefore unable to meet the requirements for lightweight design and real-time gas leakage detection. To address this, we propose LLGSeg, a lightweight network for segmenting leaked gases based on RGB-T feature fusion. LLGSeg employs a specially designed cross-modal gated feature alignment and fusion module, using the GBConv and gating mechanism to achieve lightweight alignment and fusion of cross-modal features. Additionally, a local variance-based significance enhancement module has been designed. This module divides the input features equally across the channel dimension, performs distinct feature extraction operations for each partition, and uses local variance to enhance feature significance, thereby achieving feature extraction and enhancement with a low number of parameters. Furthermore, a cross-scale information integration decoder is employed to generate segmentation results, achieving a balance between the accuracy of gas leakage segmentation and the scale of parameters. On the public dataset Gas-DB, LLGSeg achieved Acc, IoU and F2 scores of 77.71%, 55.63% and 75.10%, respectively, significantly outperforming state-of-the-art methods. LLGSeg achieved 26.72 FPS on the Gas-DB dataset, meeting the requirements for real-time segmentation.</p>