[论文解读] TransfoREM: Transformer aided 3D Radio Environment Mapping

Gautham Reddy, Ismail Guvenc+4 more|arXiv (Cornell University)|Jan 23, 2026

UAV Applications and Optimization被引用 0

一句话总结

TransfoREM 将3D无线电环境映射视为在球坐标中的序列预测任务，借助 transformer 实现 REM 重建优于 Kriging 与其他 ML 方法；采用两阶段训练（基于模型的预训练 + 数据驱动的微调）以及真实世界数据微调。

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.

研究动机与目标

• Motivate 3D REM for UAV connectivity beyond ground-level coverage.
• Propose a physics-inspired, sequence-based REM construction using a transformer.
• Combine model-based propagation with real-world data via a two-stage pre-training and fine-tuning scheme.
• Demonstrate improved interpolation/extrapolation over Kriging and other ML methods.
• Show integration potential of TransfoREM at base stations for dynamic network management.

提出的方法

• Model radio propagation as a sequence prediction task in spherical coordinates around a base station.
• Use an encoder-only transformer to predict RSRP sequences from a multi-feature input Gamma_i.
• Two-stage training: (Stage-1) model-based pre-training using FSPL and deterministic gains to generate synthetic sequences; (Stage-2) data-driven fine-tuning with masked inputs and Smooth L1 loss on real data.
• Represent each spatial point i with a Gamma_i feature vector including log(delta), angles (theta, phi), and x,y,z coordinates across radial bins.
• Mask-based training mirrors language modeling: random masking during pre-training; targeted masking during fine-tuning to align with site-specific data.
• Compare against Kriging and a Triple-Layer ML baseline; evaluate using RMSE, MAE, and R^2 on real-world datasets.

Figure 1: The spherical coordinate system representation of a UAV position and its associated received signal strength indicator (RSSI) sequences in space.

实验结果

研究问题

• RQ1 Can a transformer-based REM generate accurate 3D radio maps from limited real-world measurements?
• RQ2 Does a two-stage training (model-based pre-training plus data-driven fine-tuning) improve REM accuracy over stage-1 alone?
• RQ3 How does TransfoREM perform in interpolation and extrapolation across altitudes compared with Kriging and TripLeveL ML baselines?
• RQ4 Is the approach viable for deployment at individual BS sites for proactive network management?

主要发现

Dataset	REM Method	RMSE	MAE	R^2
AERPAW	TransfoREM Stage 1	7.49 dB	6.20 dB	0.33
dataset [8]	TransfoREM Stage 2	4.57 dB	3.13 dB	0.77
dataset [8]	TransfoREM Stage 1	5.47 dB	3.43 dB	-0.27
dataset [2]	TransfoREM Stage 2	1.29 dB	0.78 dB	0.93
Case Study	TripleLayerML Stage 1	4.07 dB	3.04 dB	0.90
dataset [2]	TripleLayerML Stage 2	1.27 dB	0.82 dB	0.95
dataset [2]	TripleLayerML Stage 3	1.12 dB	0.69 dB	0.95

• Stage-1 results show reasonable REM predictions using FSPL-based synthetic data.
• Stage-2 fine-tuning yields substantial performance gains (RMSE/MAE reductions, R^2 improvement).
• On the AERPAW dataset, Stage-2 achieves RMSE 4.57 dB, MAE 3.13 dB, R^2 0.77 (versus Stage-1 7.49 dB, 6.20 dB, 0.33).
• On dataset [2], Stage-2 achieves RMSE 1.29 dB, MAE 0.78 dB, R^2 0.93 (Stage-1: 5.47 dB, 3.43 dB, -0.27).
• TransfoREM matches or surpasses TripleLayerML performance after Stage-2 (RMSE/MAE/R^2 comparable; e.g., 1.29 dB / 0.78 dB / 0.93 vs 1.27 dB / 0.82 dB / 0.95).
• TransfoREM demonstrates improved interpolation/extrapolation across altitudes better than Kriging by about 1.5 dB on median AB tests.

Figure 2: Radial distance up to 100 m have correlation values greater than 0.5.