TransfoREM: Transformer aided 3D Radio Environment Mapping

TL;DR

TransfoREM reframes 3D Radio Environment Mapping as a sequence-prediction problem in spherical coordinates ($ ho$, $ ext{phi}$, $ ext{theta}$) and uses a Transformer-based generator to interpolate and extrapolate radio strength for UAV connectivity. It combines a model-based pre-training stage that captures basic propagation (via $P_{b,k_{ ext{dB}}} = P_{b_{ ext{dB}}} + G_{b,k_{ ext{dB}}} + N$ with $G_{b,k_{ ext{dB}}} = ext{PL}_{ ext{dB}} + ext{SF}_{ ext{dB}} + A_{ ext{dB}}$) with a data-driven fine-tuning stage that leverages real measurements through masked inputs and Smooth L1 loss. The approach demonstrates improved interpolation and altitude extrapolation over Kriging baselines and compares favorably to a cascaded TripleLayerML method, achieving similar accuracy with reduced complexity and enabling online, BS-level REM updates. These results suggest TransfoREM can support proactive resource allocation, interference management, and spatial spectrum utilization for high-altitude UAV communications in dynamic networks. The work offers a practical pathway to deploy physics-informed REM learning directly at base stations for real-time network optimization.

Abstract

Providing reliable cellular connectivity to Unmanned Aerial Vehicles (UAV) is a key challenge, as existing terrestrial networks are deployed mainly for ground-level coverage. The cellular network coverage may be available for a limited range from the antenna side lobes, with poor connectivity further exacerbated by UAV flight dynamics. In this work, we propose TransfoREM, a 3D Radio Environment Map (REM) generation method that combines deterministic channel models and real-world data to map terrestrial network coverage at higher altitudes. At the core of our solution is a transformer model that translates radio propagation mapping into a sequence prediction task to construct REMs. Our results demonstrate that TransfoREM offers improved interpolation capability on real-world data compared against conventional Kriging and other machine learning (ML) techniques. Furthermore, TransfoREM is designed for holistic integration into cellular networks at the base station (BS) level, where it can build REMs, which can then be leveraged for enhanced resource allocation, interference management, and spatial spectrum utilization.

Figures (6)

• Figure 1: The spherical coordinate system representation of a position and its associated received signal strength indicator (RSSI) sequences in space.
• Figure 2: Radial distance up to 100 m have correlation values greater than 0.5.
• Figure 3: The masked input feature $\Gamma_i$ conveys the position information of the $i^\mathrm{th}$ spatial point, and the transformer encoder is trained to predict its $\Omega_i$.
• Figure 4: The feature mask from Fig. \ref{['fig:TransformerModel']} is varied by training stage. The pretraining stage leverages complete sequence information by using random masks at each training iteration, and the fine-tuning stage corrects predictions using masks to align with the limited real-world data.
• Figure 5: a) and b) are the test datasets at the lowest and highest altitude. c) and d) are the stage 1 test set predictions learned from FSPL data. Finally, e) and f) are the best fit test predictions after stage 2 fine-tuning.