[Objectives] Water leakage is one of the most prevalent and critical pathological defects in water conveyance tunnels. Its long-term evolution inevitably leads to the deterioration of lining structures, fundamentally threatening the overall engineering safety and operational longevity of hydraulic systems. Consequently, leveraging high-definition inspection imagery captured during maintenance periods to achieve precise defect identification and facilitate targeted, specialized repair efforts has emerged as a crucial strategy for ensuring the safe operation of water conveyance networks. However, current automated method development faces significant hurdles because leakage images in tunnel environments are characterized by highly irregular morphology, multi-scale variations, and inherently blurred boundaries between water and concrete. Furthermore, uneven and insufficient illumination inside the tunnels drastically limits the segmentation accuracy of conventional models. To address these challenging imaging conditions, this study proposes a lightweight semantic segmentation method specifically optimized for water conveyance tunnel leakage, based on an improved DeepLabV3+ architecture. [Methods] The proposed model architecture incorporates three specific modifications to achieve a balance between accuracy and computational efficiency. First, to mitigate high computational costs, the model adopts MobileNetV3 as its lightweight backbone network. This choice replaces traditional, parameter-heavy large-scale feature extraction networks, thereby effectively compressing the number of model parameters and reducing computational overhead to facilitate future deployment on resource-constrained devices. Second, to tackle the issues of varying scales and boundaries, a novel cascaded multi-scale context aggregation module is embedded directly into the encoder stage. By utilizing a specific cascaded structure of multi-level atrous convolutions, this module significantly enhances the network’s multi-scale contextual representation capabilities, specifically improving the extraction and description of features pertaining to multi-scale and slender leakage regions. Third, during the decoding phase, the model integrates the Concurrent Spatial and Channel Squeeze & Excitation (scSE) attention mechanism. The scSE module performs parallel recalibration on the fused features along both the channel and spatial dimensions. This dual-axis recalibration achieves adaptive feature enhancement for target leakage regions by strengthening relevant foreground responses while simultaneously suppressing non-relevant background noise and interference caused by uneven lighting. [Results] The performance of the proposed improved DeepLabV3+ model was rigorously validated on a comprehensive self-built dataset consisting of water conveyance tunnel leakage images. Validation results demonstrate that the model achieves highly satisfactory performance across key evaluation metrics. Specifically, the model obtained a Mean Intersection over Union (MIoU) of 91.98%, a specific Leakage Intersection over Union (Leak-IoU) of 86.48%, and a mean Precision (mPrecision) of 96.41%. Crucially, despite maintaining high segmentation accuracy, the model parameters are reduced to only 5.68 M,The actual FPS measured for the edge device is 12.20 f∙s^-1. In comparative analysis with the original DeepLabV3+ base model and several other mainstream semantic segmentation models, the quantitative data indicates that the proposed method significantly lowers model complexity (in terms of parameters and floating-point operations) while simultaneously achieving or exceeding the high-precision segmentation performance of much larger networks.[Conclusions] In conclusion, the experimental results demonstrate that the improved lightweight DeepLabV3+ semantic segmentation model striking an optimal balance between accuracy and efficiency. By integrating MobileNetV3, the cascaded context aggregation module, and the scSE attention mechanism, the model overcomes common environmental challenges in tunnels to provide robust segmentation of irregular and blurred leakage areas. The specific lightweight design of this model aligns directly with the practical requirements for field-based edge computing deployment during tunnel inspection routines. Therefore, this research offers a reliable, advanced, and practical technical support system for the intelligent identification of water leakage and the digitized engineering operation and maintenance (O&M) of water conveyance tunnels.