A multi-scale and attention-enhanced deep learning approach for tunnel leakage identification

Purpose This study aims to develop a lightweight, dual-modal deep learning model for accurate classification of water leakage in cable tunnels. By integrating visible and infrared image fusion with a multi-scale global feature extraction module and a Tiny Spatial Attention (Tiny spatial attention [TSA]) mechanism, the model addresses challenges such as background noise, structural occlusion and irregular leakage patterns. The goal is to enhance recognition accuracy, reduce misdiagnosis and support real-time deployment in complex tunnel environments. Design/methodology/approach A dual-modal data set combining visible and infrared images of cable tunnels was constructed to capture comprehensive leakage features. A deep neural network based on ResNet-50 was enhanced with a Multi-Scale Global Feature Extraction Module and a TSA mechanism to improve spatial sensitivity and classification accuracy. The model was trained and evaluated on 10,000 labeled images using cross-entropy loss and the Adam optimizer. Performance was compared with traditional segmentation methods and other deep learning baselines through accuracy, error rate and inference time analyses. Findings The proposed dual-modal deep learning model significantly outperforms traditional image segmentation methods and baseline deep learning models in classifying cable tunnel leakage. It achieved an average accuracy of 86.1% on the test set and 85.1% on an external data set, demonstrating strong generalization. The model reduced error rates by up to 88.3% and inference time by over 50% compared to classical approaches. Ablation studies confirmed the effectiveness of both the multi-scale feature module and the TSA mechanism in enhancing classification performance. Originality/value This study introduces a novel dual-modal deep learning framework that integrates visible–infrared image fusion, a multi-scale global feature extraction module and a TSA mechanism for cable tunnel leakage detection. Unlike traditional or single-modal approaches, the proposed method enhances spatial feature representation while maintaining a lightweight design suitable for real-time deployment. It addresses challenges such as occlusion, background noise and irregular leakage patterns. The model offers high accuracy and generalization, providing a practical and scalable solution for intelligent tunnel inspection and fault diagnosis in power infrastructure.