泥石流灾害频发且破坏性强，传统监测手段实时性差、误报率高，亟需高效精准的智能检测方法提升预警能力。基于此，提出基于YOLOv8框架的改进模型YOLOv8m-GCSlide，通过集成全局注意力机制特征提取模块（GCNet）强化泥石流动态特征感知，设计滑动损失函数优化分类边界，结合知识蒸馏技术生成轻量化模型YOLOv8n-GCSlide。构建多源视频数据集，采用0.25 s每帧提取策略平衡训练效率，融合数据增强与负样本提升泛化能力。实验表明，改进模型检测精确率达94.6%（较基线+1.2%），召回率88.0%（+0.7%），平均精度95.9%（+2.0%），推理速度244.1 FPS，参数量减少88.1%，性能优于主流轻量模型。实际案例测试中，模型在复杂地形下召回率为82.3%，误报率为4.2%，帧率为240.6 FPS。结果表明，全局注意力机制特征提取模块与损失函数能有效捕捉泥石流运动特征，模型压缩技术兼顾精度与效率，可为地质灾害预警系统提供高精度、低延时的端侧部署技术支持。

Debris flow disasters, known for their frequent occurrence and high destructiveness, are difficult to monitor effectively due to the limited real-time performance and high false-alarm rates of conventional monitoring methods. This critical limitation underscores the urgent need to develop highly efficient and precise intelligent detection techniques to substantially enhance early warning capabilities. To address the challenges of poor real-time performance and high false alarm rates in traditional debris flow monitoring systems, this study proposes an enhanced YOLOv8m-GCSlide model based on the YOLOv8 framework. The GlobalContext Network (GCNet) is integrated into the backbone network to improve spatial dependency modeling of dynamic fluid boundaries in complex terrains, while a Sliding Loss function (SlideLoss) is designed to dynamically adjust classification thresholds and mitigate sample imbalance. Knowledge distillation is applied to compress the model, resulting in a lightweight variant (YOLOv8n-GCSlide) with reduced computational complexity. A multi-source video dataset was constructed using publicly available resources, with frames extracted at 0.25-second intervals to balance feature retention and training efficiency. Data augmentation techniques, including random cropping, rotation, scaling, Gaussian blur, and color jittering, were used to enhance generalization, supplemented with negative samples (e.g., dry riverbeds and landslides) to reduce false positives. Experimental results show that the optimized model achieves 94.6% (+2.0%) detection accuracy, 88.0% recall, 95.9% mean Average Precision (mAP), and an inference speed of 244.1 FPS, outperforming mainstream lightweight models such as SwinTransformer and MobileNet variants. After compression, the model parameters were reduced by 88.1%, with the distilled version retaining 94.6% (+1.2%) accuracy and 88.0% (+0.7%) recall while maintaining an inference speed of 244.1 FPS. Field validation conducted in Sedongpu Gully, a high-risk debris flow region, confirmed the model’s practical applicability. Under complex environmental interference, the model achieved 82.3% recall, 4.2% false positive rate, and a processing speed of 240.6 FPS. The integration of global attention mechanisms and task-specific loss functions effectively captures dynamic motion features and suppress environmental noise. Additionally, model compression techniques help balance accuracy and computational efficiency, enabling edge deployment for real-time disaster warnings. This approach provides a robust technical foundation for intelligent geological hazard monitoring systems, emphasizing high precision, low latency, and adaptability to resource-constrained scenarios.