Abstract: Influenced by extreme weather events, especially rainstorms, and anthropogenic engineering activities, the Loess Plateau region in China is subjected to frequent landslides, which pose a significant threat to the safety of human life and property. This study proposes a hierarchical constraint-based method for identifying slow-moving loess landslides. First, a lightweight network model is constructed through knowledge distillation in deep learning to rapidly identify deformation areas from interferometric synthetic aperture radar (InSAR) data. Compared to the original teacher model, the distilled model enhances the identification accuracy by 3%, increases the F1-score by 2%, and reduces the model size by 50.1%. Given that the identified deformation areas from InSAR data contain a large quantity of non-landslide deformations such as land subsidence and mining-induced deformations, two constraint approaches, namely mining-induced negative sample filtering and slope unit-based constraint, are introduced to further exclude non-landslide deformations. The proposed method enables the accurate identification of slow-moving landslides in loess regions, holding great importance for the monitoring and early warning of geological hazards.