Cross-Modal Knowledge Distillation for Path Loss Prediction in Urban mmWave UAV-assisted Networks

Accurate path loss (PL) prediction is an essential step for the design and deployment of mmWave UAV-assisted communication networks, especially in urban environments. In this work, we propose a voxel-based 3D Convolutional Neural Network (3D-CNN) that directly processes high-resolution 3D models of the urban environment surrounding the ground receiver to predict PL. To further improve prediction accuracy, we introduce a Cross-Modal Knowledge Distillation (CMKD) framework that transfers structural information from a geometrical model to the 3D map–based model. The teacher–student architecture employs output-level and intermediate-layer distillation strategies. Experimental results demonstrate that the proposed CMKD frameworks clearly outperform the baseline method, with the intermediate-layer distillation achieving a Root Mean Square Error (RMSE) of approximately 8 dBm.