基于深度学习的滑坡目标检测模型对比研究

Comparative Study of Deep Learning Models for Landslide Object Detection

摘要: 针对滑坡目标识别中人工解译效率低、复杂场景下目标检测效果不稳定等问题，本文基于公开滑坡目标检测数据集，选取YOLOv8、YOLOv11和RT-DETR三种目标检测模型开展对比实验，对不同模型在滑坡识别任务中的性能进行分析。实验采用统一数据集和训练条件，从模型训练过程、验证集定量指标以及可视化检测结果三个方面进行评价。结果表明，三种模型均能够较好地完成滑坡目标检测任务，训练过程中损失函数整体下降，精确率(Precision, P)、召回率(Recall, R)和平均精度均值(mean Average Precision, mAP)等指标持续提升并趋于稳定，表现出较好的收敛性。其中，RT-DETR的综合检测性能最好，其Precision、Recall、mAP@0.5和mAP@0.5:0.95分别达到0.954、0.950、0.964和0.815，均优于YOLOv8和YOLOv11，说明其在滑坡目标识别和边界框定位方面具有更高精度。可视化结果进一步表明，RT-DETR的检测框完整性和稳定性最好，YOLOv11次之，YOLOv8在个别复杂场景下存在重复检测现象。研究表明，三种模型均适用于滑坡目标检测任务，其中RT-DETR更适合精度要求较高的场景，YOLOv11更适合兼顾检测效果与实时性的应用需求。

Abstract: To address the problems of low efficiency in manual interpretation and unstable detection performance in complex landslide scenes, this study conducted comparative experiments using three object detection models, namely YOLOv8, YOLOv11, and RT-DETR, based on a public landslide object detection dataset, and analyzed their performance in landslide recognition tasks. Under the same dataset and training conditions, the models were evaluated from three aspects: the training process, quantitative validation metrics, and visualization of detection results. The results show that all three models can effectively accomplish landslide object detection. During training, the loss functions generally decreased, while indicators such as Precision (P), Recall (R), and mean Average Precision (mAP) continuously improved and gradually stabilized, indicating good convergence. Among them, RT-DETR achieved the best overall detection performance, with Precision, Recall, mAP@0.5, and mAP@0.5:0.95 reaching 0.954, 0.950, 0.964, and 0.815, respectively, all higher than those of YOLOv8 and YOLOv11, demonstrating higher accuracy in landslide recognition and bounding box localization. The visualization results further show that RT-DETR achieved the best completeness and stability of detection boxes, followed by YOLOv11, while YOLOv8 showed repeated detections in some complex scenes. The study indicates that all three models are suitable for landslide object detection tasks. RT-DETR is more suitable for scenarios requiring higher accuracy, while YOLOv11 is more appropriate for applications that need to balance detection performance and real-time efficiency.