Accurately identifying and monitoring loess landslide cracks is crucial for understanding slope failure mechanisms and predicting their evolution. However, the complex background conditions of the Loess Plateau-including variations in light intensity, visual clutter, irregular crack morphologies, and heterogeneous scales-severely limit extraction accuracy and impede the detection of subtle crack evolution. To overcome these limitations, this study proposes FCS-YOLO, a novel, high-precision deep learning model for automatic identification of loess landslide cracks from UAV photogrammetry data. In the proposed model, three elaborate modules-Feature Enhanced Gated Convolutional Attention (FGCA), Cooperative Channel-Spatial-Pixel Attention (CCSP), and Scale-Aware Feature Enhancement and Fusion (SAEF)-are integrated into YOLOv8. These modules significantly enhance crack-specific feature extraction, reducing the overall error rate by similar to 23.1% relative to the baseline and enabling precise crack delineation. Subsequently, we introduce the primary crack identification method, Hazard-Response Crack Significance Index (HRCSI), a novel metric designed to identify cracks with the highest potential to evolve into slip surfaces. Finally, we resolve two long-standing questions-where cracks are most likely to develop and which crack parameter best correlates with their evolution process-through a coupling analysis of crack dynamics and surface deformation. Our code and data are publicly available at: https://github.com/Zpx517/FCS-YOLO.