Abstract: Landslide cracks are key surface manifestations of slope deformation,and their early identification is essential for effective landslide early warning. This study selected the Heifangtai terrace in Gansu Province as the research area and constructed a comprehensive dataset of loess landslide cracks with diverse background conditions. Using this dataset,we systematically evaluated the performance of three deep learning models—Faster R-CNN,Cascade R-CNN,and YOLO v8—in identifying loess landslide cracks. The main findings are as follows: (1)An integrated acquisition scheme combining low-altitude UAV photography and close-range camera imaging was established. Using both geometric transformation and pixel-level data augmentation methods,a loess landslide crack dataset containing 2856 samples was constructed. (2)The one-stage detection model YOLO v8 demonstrated superior performance in recall,real-time inference,training efficiency,and model lightweightness. It achieved a mean average precision of 0.989,a frame rate exceeding 140 FPS,and a model size of only 6.116 MB,significantly outperforming the two-stage detection models. (3)The quality and scale of the dataset were found to constrain the crack identification performance of deep learning models. Further improvements are needed to enhance the detection of subtle and concealed cracks under complex environmental conditions. The proposed deep learning-based automatic identification method for loess landslide cracks offers valuable technical support for landslide risk mitigation in loess regions.