Landslides, as one of the most frequent natural disasters, pose significant threats to human life and property, making accurate occurred landslide extraction crucial for disaster prevention and control. Current research predominantly focuses on optical imagery, overlooking the potential of Synthetic Aperture Radar (SAR) data, while limited landslide samples constrain model generalization capabilities. To address these limitations, a novel deep learning network model integrating optical and SAR data was developed for landslide identification. The model employs SwinUnet as its backbone architecture. Unsupervised pretraining via contrastive learning was adopted to improve landslide identification accuracy under limited sample conditions. Experimental validation in the Loess Plateau of western Shanxi demonstrated superior performance compared to conventional landslide identification methods. The proposed approach achieved high-precision landslide identification with limited training samples, showing significant improvement in accuracy following contrastive learning pretraining. These results highlight the model's substantial potential for advancing landslide identification research and applications.