Swinlitekd: Optimizing Landslide Recognition Using Swin-Transformer Networks with Knowledge Distillation

Renxiang Huang, Tao Chen

Abstract: Recognizing landslides effectively is crucial for disaster prevention and post-disaster rescue operations. Our SwinLiteKD, an innovative knowledge distillation network based on Swin-Transformer, addresses challenges related to model runtimes and inefficiencies in current deep learning approaches. Validated in a landslide-prone region in Zigui County, Hubei Province, China, our proposed method incorporates landslide influencing factors to significantly enhance model performance. Compared to ResNet50, Swin-Transformer, and DeiT, SwinLiteKD achieves superior Overall Accuracy (OA: 97.0000%), Precision (97.1698%), Recall (96.9999%), F1 (97.0848%), and Kappa (93.99%). With the lowest number of FLOPs, our model ensures crucial computational efficiency in landslide recognition after geological disasters, requiring only 4.4987 GFLOPs. This is 0.1553 GFLOPs, 0.0723 GFLOPs, and 0.3597 GFLOPs less than ResNet, Swin, and DeiT, respectively. SwinLiteKD demonstrates excellent adaptability in scenarios demanding swift landslide recognition post-geological disasters.