Tunnels, as a core component of modern transportation infrastructure, play a critical role in ensuring the safety of transportation operations and public safety. Water leakage, one of the most common issues in tunnels, can lead to severe consequences such as structural corrosion, strength degradation, or even collapse if not detected and ad-dressed promptly. However, the complexity of the tunnel environment and the varia-bility of water leakage patterns pose significant challenges to the accuracy and effec-tiveness of target detection results. To address these issues, this paper proposes an in-telligent water leakage detection method for tunnels based on an improved YOLOv10, LPM-YOLO. First, this paper proposes a Local-Global Attention Fusion Module (LGAFM) to achieve adaptive weighted fusion of multi-scale features. Second, by in-troducing the PTSSA module based on Token-Statistical Self-Attention, the traditional quadratic complexity of self-attention calculations is eliminated, and the efficiency of modeling long-range dependencies is effectively improved. Finally, based on the Mamba Vision hybrid architecture concept, the C2f_MV module is introduced to en-hance the model's ability to capture long-range spatial dependencies. Experiments demonstrate that LPM-YOLO achieves 75.6% AP₅₀ and 41.7% AP_50:95 while maintaining its lightweight characteristics, representing improvements of 6.7% and 2.5% over the original YOLOv10-s, respectively.