Spatially Consistent Air-to-Ground Channel Modeling via Generative Neural Networks

TL;DR

This work addresses the challenge of generating spatially coherent air-to-ground channel representations for UAV mobility by modeling the conditional distribution of RSS sequences given UAV–gNB distances and gNB identity. It proposes a Transformer-based conditional GAN with distance and gNB conditioning, trained with LS adversarial and categorical losses to produce RSS sequences that mirror real spatial statistics. Validation uses first- and second-order statistics, notably the Correlation Matrix Distance, showing convergence of generated sequences to real data, with data augmentation further improving fidelity. The approach, demonstrated on ray-traced 28 GHz urban scenarios, provides a scalable workhorse for mobility-aware system-level evaluations and could be extended to full multipath modeling and cross-environment generalization.

Abstract

This article proposes a generative neural network architecture for spatially consistent air-to-ground channel modeling. The approach considers the trajectories of uncrewed aerial vehicles along typical urban paths, capturing spatial dependencies within received signal strength (RSS) sequences from multiple cellular base stations (gNBs). Through the incorporation of conditioning data, the model accurately discriminates between gNBs and drives the correlation matrix distance between real and generated sequences to minimal values. This enables evaluating performance and mobility management metrics with spatially (and by extension temporally) consistent RSS values, rather than independent snapshots. For some tasks underpinned by these metrics, say handovers, consistency is essential.

Figures (6)

• Figure 1: Simulation area showing an example of a UAV trajectory (red) between buildings (green) and the locations of three gNBs ($\triangle$).
• Figure 2: Proposed generative channel model architecture.
• Figure 3: CMD computed on the test set vs. number of training iterations for the case of a single gNB, with and without data augmentation.
• Figure 4: Evolution of the correlation matrix for generated RSS sequences as compared to the one of real RSS sequences.
• Figure 5: Left-hand side: CDF of the real and generated RSS values. Right-hand side: mean (solid and dashed lines) plus/minus standard deviation (shaded areas) of the real and generated RSS vs. distance and respective log-distance least-squares fits. Generated sequences are obtained by training with data augmentation.