Abstract: Traditional dike seepage inspections rely on manual labor, making it difficult to achieve round-the-clock, comprehensive coverage. Existing automated inspection systems face two core challenges: first, the conflict between limited onboard computing power and real-time processing demands; second, the difficulty in accurately extracting faint seepage features against complex backgrounds, resulting in failure to meet stringent low false alarm rate requirements. To address this, we constructed a multi-altitude, multi-temporal, multi-scenario dike leakage image dataset. We innovatively proposed a dual-model collaborative architecture combining lightweight YOLOv8n and high-precision YOLOv8x. Leveraging fine-grained annotation and data augmentation techniques, the system performs preliminary screening of leakage areas on the drone side and high-precision identification on the server side. This approach ensures real-time performance while significantly enhancing detection accuracy. Through parameter tuning-particularly optimizing confidence thresholds and intersection-over-union ratios during model inference-performance saw significant improvement. The system's false negative rate and false positive rate were reduced to 2.7% and 3.2%, respectively. The optimized algorithm was deployed on our proprietary intelligent inspection platform, integrated with drone pods to enable round-the-clock safety inspections. Using flood-season leakage detection on flood control embankments in Xiangyin County and Miluo City, Hunan Province as a case study, the system instantly identifies leaks through real-time drone monitoring and data processing, triggering alerts to enable rapid response by embankment safety personnel and effectively prevent escalation of hazards.