LeMoRe-MTKD: A Multi-Teacher–Student Knowledge Distillation Deep Learning Framework for Landslide Segmentation Using UAV Imagery

Landslides (LSs) are a natural phenomenon, typically triggered by heavy rainfall and earthquakes, and pose a global danger to human lives and property. LSs frequently occur on steep slopes and in loose soil conditions, and damaged sites can be hard to reach thereafter. Thus, remote approaches would be relevant and practical for investigations, and the use of an Unmanned Aerial Vehicle (UAV) for imaging, along with robust methods to identify and quantify LS-affected areas, is recommended. Therefore, this study proposed a multi-teacher–student knowledge distillation (KD) deep learning framework for landslide semantic segmentation. The lightweight LeMoRe model is selected for the learner model (student), and SegNet, U-Net, DeepLabV3+, and LinkNet are used as the teacher networks. To establish the UAV training dataset, we conducted UAV imaging for four LS sites, including Mogo, Buri, Oejeong, and Sangneung (testing site) villages, in Sancheong County, South Korea, in 2025. First, all plain networks are trained on the training dataset, and then KD is applied for single-teacher and multi-teacher KD. Subsequently, ablation studies are implemented to identify the optimal KD component set and the best student–teacher combination (with equal weights) for subsequent experiments. In addition, the equal-weight combination is tuned for the optimal teacher weight proportion to maximize the teacher knowledge transfer to the student. The best model weight is saved for inference with the testing site of Sangneung to map the LS-damaged areas. The experiment results reveal an improvement in accuracy (6.3% F1, 8.2% IoU with equal weight and 12.3% F1 and around 16.8% IoU with the tuned weight). The inference results for the Sangneung testing site were qualitatively and quantitatively evaluated through visual comparison with the original UAV imagery with three zoom-in areas presented for detailed examination. Finally, efficiency, the trade-off between accuracy and efficiency, limitations, and practical KD are discussed.