To address the inefficiency of traditional methods and the insufficient utilization of texture features in existing deep learning models for tunnel leakage detection, this paper proposes a leakage image segmentation method based on longitudinal texture convolution and learnable channel attention mechanism. First, considering the longitudinal extension characteristics of leakage regions, we design a longitudinal texture convolution module consisting of three cascaded asymmetric convolutional kernels with different scales, which enhances multi-scale receptive fields to capture longitudinal texture features. Second, we introduce a learnable parameter matrix into the channel attention mechanism to dynamically optimize feature channel weights, thereby focusing on critical regions. Based on the SegNet encoder-decoder architecture, we construct a lightweight segmentation network by incorporating Sobel operator preprocessing and pooling indices upsampling technique. Experimental validation on a public dataset containing 4 555 leakage images demonstrates that compared with U-Net and DeepLabv3+, our method achieves improvements of 0.72%~2.82% in intersection over union (IoU), reaching 80.64%, with accuracy and F1-score increased to 94.68% and 89.30%, respectively. Ablation studies further verify the effectiveness of the longitudinal texture convolution (LTC) and learnable channel attention (LCA) modules, whose combination improves IoU by 3.96%. By integrating multi-scale longitudinal texture features with adaptive channel weighting mechanism, the proposed method significantly enhances segmentation precision for leakage regions, providing an efficient solution for infrastructure safety monitoring.