s: 【Objective】 Electrical Resistivity Tomography (ERT) enables non-destructive acquisition of internal resistivity distribution in dam bodies, yet traditional manual interpretation suffers from low efficiency and high subjectivity. 【Methods】 A multimodal data augmentation strategy was constructed using techniques such as geometric transformation, Mosaic augmentation, and local occlusion, expanding the original 813 images to 2, 000 images and effectively mitigating the risk of overfitting. The ASF multi-scale fusion module and MCAttn dynamic attention mechanism were introduced into the YOLOv8 neck network, enhancing the model’s ability to resolve multi-scale features and fuzzy boundaries of leakage areas through cross-layer feature interaction and adaptive weight allocation. A multi-task joint optimization framework integrating VFL classification loss, DFL distribution focal loss, and an improved CIoU geometric loss was designed to achieve task alignment between classification confidence and localization accuracy. 【Results】 The results show that the improved model achieves a mean average precision (mAP@0.5) of 77.0%, which is 5.6% higher than that of the original YOLOv8, along with an 8.0% increase in recall and an inference speed of 42.4 FPS. The algorithm was applied to leakage detection projects at Maoshan earth-rockfill dam and Fengcheng geomembrane core-wall dam, where the segmentation results exhibited over 95% spatial overlap with manual interpretation. 【Conclusion】 This verifies the effectiveness and robustness of the proposed model in practical engineering, providing a new intelligent technical approach for dam leakage identification based on high-density electrical method.