As frequent and highly destructive geologic hazards, debris flows necessitate effective monitoring alongside rapid and accurate detection to support disaster prevention and mitigate losses of life and property. Current detection technologies, however, are often limited by high false alarm rates, insufficient accuracy, considerable model complexity that complicates deployment, and the elevated costs of contact-based detection. To overcome these limitations, this paper introduces a lightweight real-time debris flow detection model based on the Roboflow Detection Transformer (RF-DETR). First, we constructed a dataset of realistic debris flow scenarios including debris flow disaster events worldwide. Based on this, the lightweight vision transformer model EfficientFormerV2 is adopted as the backbone. By employing a dimension-consistent architecture, the model avoids the frequent switching between 4D and 3D features found in traditional Vision Transformers (ViTs), thereby reducing a significant number of inefficient operations. Additionally, we optimized the multiscale projection layer by removing downsampling and feature aggregation operations, which reduces redundant computations and improves feature extraction efficiency. Furthermore, the introduction of the Efficient Intersection over Union (EIoU) loss function achieves faster convergence and improved detection accuracy for debris flows. Ablation studies performed on our debris flow dataset demonstrate that the improved model reduces parameters by 54.2% and computational load by 36.4% while ensuring acceptable losses in detection accuracy and latency. This significant reduction in model size and complexity achieves effective lightweighting, fulfilling the requirements for deployment on edge devices and enabling real-time debris flow detection.