3D Dynamic Radio Map Prediction Using Vision Transformers for Low-Altitude Wireless Networks

Nguyen Duc Minh Quang, Chang Liu, Huy-Trung Nguyen, Shuangyang Li, Derrick Wing Kwan Ng, Wei XiangarXiv 2025 · DOI · arXiv

Introduction

Low-altitude wireless networks (LAWNs) are increasingly vital for drone-based logistics and emergency response. However, maintaining reliable connectivity is challenging due to the highly non-stationary nature of the environment. Drones move rapidly in 3D space, and ground base stations (BSs) frequently adjust transmit power to meet fluctuating demands.

To manage these dynamics, engineers utilize Radio Maps (RMs)—spatial visualizations of signal strength. Traditional RM tools often fall short because they are typically static, restricted to 2D, or computationally too slow for real-time applications. This paper introduces 3D-DRM, a Transformer-based framework designed to reconstruct and forecast 3D radio signal evolution in real-time.

The 3D-DRM Framework

The proposed 3D-DRM architecture leverages an end-to-end Transformer model to handle the dual task of spatial reconstruction from sparse measurements and temporal prediction of future states.

The 3D-DRM architecture featuring a ViT encoder, temporal transformer, and cross-attention decoder.

The framework consists of several key components:

  • Vision Transformer (ViT) Encoder: Processes irregular, sparse drone measurements by embedding signal power, 3D coordinates, and timestamps into high-dimensional tokens.
  • Temporal Transformer Module: Learns sequential dependencies to model how the radio field evolves over time.
  • Cross-Attention Decoder: Uses "voxel queries" to reconstruct the full 3D grid, performing cross-attention between queries and latent features to generate the predicted map.
  • Hybrid Loss Function: Combines Voxel-wise Fidelity Loss for numerical accuracy with Temporal Gradient Loss to ensure smooth, physically plausible transitions between frames.

Experimental Setup

The model was evaluated using a synthetic dataset of 120 sequences (200 frames each) representing a 500m x 500m x 250m area. The simulation involved 4 base stations and 40–80 UAVs moving at 5–10 m/s. The task required predicting the next 5 frames of the 3D Radio Map based on the previous 15 frames of sparse measurements.

Key Results and Performance

The 3D-DRM model demonstrated superior performance compared to standard benchmarks like RadioUNet and ConvLSTM.

Performance comparison of 3D-DRM against ConvLSTM and RadioUNet across varying UAV densities.

Key findings include:

  • Enhanced Accuracy: 3D-DRM achieved an average reduction in Root Mean Squared Error (RMSE) of approximately 35% compared to ConvLSTM and 25% compared to RadioUNet.
  • Temporal Stability: The model exhibited significantly lower temporal gradient error, indicating smoother transitions without erratic "flickering" between predicted time steps.
  • Robustness to Density: While high drone density (over 60 UAVs) increases signal complexity and slightly raises RMSE, 3D-DRM remains more effective than baseline models in balancing data density with environmental complexity.

Conclusion

By unifying spatial reconstruction and temporal prediction within a Transformer architecture, 3D-DRM provides a scalable, accurate solution for real-time radio environment awareness in drone networks. Future work will explore multi-layer RM prediction for different frequency bands and distributed learning approaches to reduce centralized data overhead.

Summary updated

BibTeX
@misc{quang2025d,
  author = {Nguyen Duc Minh Quang and Chang Liu and Huy-Trung Nguyen and Shuangyang Li and Derrick Wing Kwan Ng and Wei Xiang},
  title = {{3D Dynamic Radio Map Prediction Using Vision Transformers for Low-Altitude Wireless Networks}},
  year = {2025},
  eprint = {2511.19019},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2511.19019},
}
Figure