PaperlayerView exact revision on arXiv

Source-linked AI summary

AI-Enabled Wireless Propagation Modeling and Radio Environment Maps for 5G Aerial Wireless Networks

Gautham Reddy, Kürşat Tekbıyık, Bryton Petersen, Antoine Lesage-Landry, Gunes Karabulut Kurt, Ismail Güvenç

arXiv:2609.27083v1eess.SP

TL;DR

Aerial connectivity from terrestrial networks is challenged by LoS interference, sidelobe-related coverage holes, localized fading, and fast UAV dynamics. The paper proposes a cascaded REM that separates large-scale spatial structure from fast fading using a radial sequence formulation. It reports signal-strength prediction errors near 3 dB and REM similarity indices above 0.75, and examines UAV link-quality prediction.

  • Problem

    Aerial users encounter LoS links and terrestrial base-station sidelobe interference, complicating reliable three-dimensional connectivity.

    Source: §I · p. 1

  • Method

    A cascaded REM framework decouples large-scale spatial radio footprints from localized fast-fading dynamics.

    Source: §V · p. 13

  • Results

    Signal-strength predictions have errors near 3 dB, and cross-dataset REM similarity indexes exceed 0.75.

    Source: §V · p. 14

  • Takeaways & Limitations

    The generated REM is used to predict UAV link quality, a capability the paper identifies as relevant to coordinated aerial network operations and trajectory planning.

    Source: §V · p. 14

Abstract

from arXiv · show

With the gaining prominence of aerial mobility applications, their success depends on the seamless integration of terrestrial and non-terrestrial network connectivity. However, providing reliable connectivity from terrestrial telecommunication networks remains challenging due to multi-cell interference from base stations (BSs) under line-of-sight (LoS) conditions to unmanned aerial vehicles (UAVs), coverage holes caused by antenna sidelobe degradation, localized multipath fading effects, and the high-speed dynamics of aerial users. To model such complexities, often exacerbated by sparse real-world data, this work proposes a dual-stage radio environment map (REM) framework. Our approach physically decouples the channel modeling, where a spatial Transformer first anchors the deterministic, large-scale path loss geometry, while a gated recurrent unit (GRU) subsequently extrapolates the stochastic, localized fast- fading deviations. By reformulating 3D spatial interpolation as a 1D radial sequence prediction task, the framework inherently aligns with the physics of propagation. We evaluate the proposed framework against state-of-the-art baselines, including 3D Kriging, UNet, Mamba, and Inception, using empirical 5G datasets. The results demonstrate improved intra-site generalization across diverse altitudes, user dynamics, and reference signal received power (RSRP) datasets, achieving signal-strength predictions with errors near 3 dB and REM spatial-similarity indices exceeding 0.75. Finally, we examine the influence of REMs and channel rank conditions on UAV channel quality, underscoring the necessity of reliable channel modeling for robust aerial connectivity.

I. INTRODUCTION

Aerial networks require reliable three-dimensional connectivity despite LoS links and interference from terrestrial base-station sidelobes. The paper replaces grid-based map reconstruction with radial sequence modeling and a physically decoupled REM architecture, evaluated on real-world 5G datasets.

I. INTRODUCTION

  • Aerial connectivity challenges include high-probability LoS links and interference from terrestrial base-station sidelobes in dynamic three-dimensional environments.The introduction frames reliable connectivity as a planning challenge for infrastructure designed primarily for terrestrial users.Source: §I · p. 1

A. Related Works

  • Prior REM methods use deterministic propagation or statistical interpolation, while image-based approaches treat mapping as grid inpainting that is challenging for dynamic aerial networks.Ray tracing requires detailed environmental data and substantial computation; Kriging can scale cubically –.Source: A. Related Works · p. 2
  • The framework reformulates propagation from the base station to the UAV as a continuous one-dimensional radial sequence rather than a discretized map.This addresses the stated mismatch between image-based mapping and aerial networks with limited LoS obstruction, narrow sidelobes, and rapid mobility.Source: A. Related Works · p. 2
  • A spatial Transformer models macroscopic channel effects, while a GRU estimates unobservable micro-environmental fast fading from sparse, temporally varying measurements.The architecture separates large-scale spatial structure from localized stochastic variation.Source: A. Related Works · p. 3

B. Contributions

  • The paper identifies strong one-dimensional radial spatial correlation and sequential temporal fast-fading variation as physical properties motivating its architecture.These properties are reported from analysis of three-dimensional signal-strength data.Source: B. Contributions · p. 3
  • The proposed framework outperforms 3D Kriging, UNet, Mamba, and Inception for REM reconstruction and sparse-measurement RSRP prediction across altitudes and UAV dynamics.Evaluation uses PawPrints, Nemo, and Quectel datasets collected in the same AERPAW environment –.Source: B. Contributions · p. 3

C. Outline

This section presents a ground-to-air two-ray propagation model that separates deterministic LoS path loss from stochastic fading. It also outlines how spatial and fast-fading components motivate a staged learning strategy.

C. Outline

  • The paper's outline assigns propagation modeling, data collection and analysis to Section II, model development to Section III, experiments to Section IV, and conclusions to Section V.Section IV includes a case study on UAV channel-quality prediction using REMs.Source: C. Outline · p. 3

II. RADIO PROPAGATION MODEL

  • The two-ray ground-to-air model separates a deterministic LoS component from a stochastic fading factor.The channel geometry couples a direct LoS ray with a ground-reflected component.Source: §II · p. 3
  • The geometry uses base-station and UAV locations to derive their three-dimensional separation and the LoS elevation and ground-reflection angles.The passage defines d3D as LoS separation, θl as LoS elevation, and θr as the reflection angle with respect to the ground.Source: §II · p. 3
  • The model expresses path loss in dB as LoS path loss plus a fast-fading factor dependent on receiver gain and multipath coupling.The fast-fading term can account for additional multipath components across different regions in three-dimensional space.Source: §II · p. 4

A. 5G Measurement Datasets

The study uses three receiver configurations on the AERPAW rural testbed, with datasets collected over shared UAV trajectories. Measured RSRP follows large-scale LoS trends but also contains device-specific variation not captured by the deterministic model.

A. 5G Measurement Datasets

  • PawPrints, Nemo, and Quectel receivers collect 5G NR measurements on the n77 band at 3.4 GHz with 100 MHz bandwidth.Their reporting intervals are 1.0 s, 0.5 s, and 0.25 s, respectively; Quectel measurements are limited to 30 m altitude, while PawPrints and Nemo include 30 m and 50 m.Source: A. 5G Measurement Datasets · pp. 4–5
  • SS-RSRP is modeled as the sum of normalized transmit power per synchronization-signal resource element and path loss.The normalization uses total transmit power, physical resource blocks, and subcarriers per resource block; w accounts for additional shadowing and noise effects.Source: A. 5G Measurement Datasets · p. 5
  • Measured RSRP exhibits device-specific variation, while the deterministic LoS model captures large-scale spatial trends but misses scene-specific shadowing and small-scale fading.The reported variation may stem from hardware-specific signal calculations, filtering, and measurement averaging.Source: A. 5G Measurement Datasets · p. 5

B. Spatial Signal Correlation Properties

The section distinguishes correlation patterns in the LoS and fast-fading components of aerial RSRP.

B. Spatial Signal Correlation Properties

  • The LoS component has strong spherical-domain correlation driven by radial propagation and the base-station antenna gain pattern.The section quantifies correlation properties across three experimental datasets.Source: B. Spatial Signal Correlation Properties · p. 5
  • Fast-fading values are correlated locally across spatially and temporally adjacent samples.This contrasts with the LoS component's stronger spherical-domain correlation.Source: B. Spatial Signal Correlation Properties · p. 5
  • The section analyzes the two components separately: large-scale LoS correlation and localized fast-fading correlation.The stated analysis covers the three experimental datasets.Source: B. Spatial Signal Correlation Properties · p. 5

1) RSRP correlation in the spherical coordinate system

The section measures RSRP correlation along spherical radial directions and fast-fading correlation over recent spatially separated samples. The datasets show radial correlation useful for large-scale path-loss mapping and dataset-dependent fast-fading correlation over short intervals.

1) RSRP correlation in the spherical coordinate system

  • The spatial analysis groups measurements into 0.05-radian azimuth and elevation bins, then averages pairwise correlations by 5 m radial-separation bins.The resulting spherical radial profile summarizes correlations across angular bins.Source: 1) RSRP correlation in the spherical… · p. 5
  • PawPrints RSRP correlation remains close to 0.5 through 50 m, while Nemo decays faster and Quectel has a narrower range; all three show radial correlation.Quectel measurements cover only the 30 m altitude profile, contributing to spikier correlation behavior.Source: 1) RSRP correlation in the spherical… · p. 6

2) Fast-fading factor correlation in spatio-temporal domain

  • Fast-fading correlation is computed against the five preceding measurements after removing LoS path loss and transmit power, with samples grouped in 3 m separation bins.This rolling window reflects UAV speeds up to 10 m/s and reporting intervals up to 1 s.Source: 2) Fast-fading factor correlation in… · p. 6
  • Nemo fast-fading values are most well-correlated; PawPrints and Quectel's smoothed, integer-quantized reports do not capture fine-grained fast fading.Shorter Quectel reporting intervals better preserve correlation; localized analysis spans about 50 m for PawPrints, 25 m for Nemo, and 13 m for Quectel.Source: 2) Fast-fading factor correlation in… · p. 6

C. Implications for Deep Learning Architecture

The paper identifies sparse-data difficulty in jointly learning large-scale LoS decay and stochastic multipath deviations, despite radial spatial correlation and sequential fast-fading continuity. It responds with a radial sequence design and a spatial Transformer followed by a GRU.

C. Implications for Deep Learning Architecture

  • Strong radial spatial correlation and sequential continuity in temporal fast fading support replacing generic grid-based spatial models with a physically decoupled radial-sequence design.The paper notes that monolithic spatial models cannot easily learn LoS decay and stochastic multipath deviations together from sparse real-world data.Source: C. Implications for Deep Learning… · p. 6

III. REM SYSTEM DESIGN

  • The spatial Transformer learns the LoS path-loss profile first, and a GRU head then estimates sequential fast-fading behavior from recent RSRP data.The 3D REM is formulated as a 1D RSRP sequence prediction task using measurements along radial directions.Source: §III · p. 6, Figure 5 · p. 6

A. Transformer-based Spatial Modeling

The spatial model samples physical channel parameters along radial directions and uses a Transformer encoder to predict continuous RSRP sequences. Its inputs bound large-scale attenuation but do not capture fast fading.

A. Transformer-based Spatial Modeling

  • The model samples channel parameters at discrete 1 m radial steps to a maximum range, then maps them to an RSRP sequence with quantized radial resolution.The spherical formulation is motivated by dominant LoS propagation and strong radial signal-strength correlation.Source: A. Transformer-based Spatial Modeling · p. 7
  • The input feature matrix Γi is an 8 × Rmax array of radial features for each measured 3D position and direction.Its features include transmit power and channel parameters such as frequency, distance, angles, and Cartesian coordinates.Source: A. Transformer-based Spatial Modeling · p. 7
  • The input is restricted to observable macroscopic parameters sufficient to bound large-scale attenuation but insufficient to capture fast fading.An encoder-only Transformer uses embeddings, positional encodings, and six layers with eight-head attention to predict the RSRP profile.Source: A. Transformer-based Spatial Modeling · p. 7

B. GRU-based Fast-Fading Factor Estimation

The GRU estimates localized fast fading from recent trajectory samples, complementing the spatial encoder’s global REM. The framework separates spatial-map learning from sequential residual estimation and uses staged training to support sparse 3D reconstruction.

B. GRU-based Fast-Fading Factor Estimation

  • The GRU extrapolates hidden multipath effects from recent signal continuity because fast fading cannot be inferred from spatial coordinates alone.It combines the spatial RSRP estimate with a predicted fast-fading value, using five previous factors and their relative separation distances.Source: B. GRU-based Fast-Fading Factor… · pp. 7–8

C. Three-Stage Training Technique

  • Stagewise training pre-trains the spatial encoder on synthetic LoS data, fine-tunes it on real measurements, then freezes it while training the GRU on fast-fading residuals.The stages respectively target geometric correlations, regional shadowing and antenna patterns, and residual fast fading.Source: C. Three-Stage Training Technique · p. 8
  • The radial sequence Transformer uses global attention and synthetic pre-training as structural priors to reconstruct continuous 3D REMs from sparse 2D flight trajectories.A single observed coordinate can update attention-based spatial representations across the propagation ray, while pre-training bounds extrapolation using deterministic path-loss geometry.Source: C. Three-Stage Training Technique · p. 8

D. Baseline Methods for Comparison

  • The benchmark set spans convolutional, state-space, multiscale convolutional, and statistical approaches, with the same preprocessing, loss, and optimizer across models.Only the network architecture differs across experiments.Source: D. Baseline Methods for Comparison · p. 8

1) UNet

  • The UNet baseline adapts RadioUNet’s encoder-decoder design to 1D sparse measurements by replacing its 2D convolutions, pooling, and upsampling with 1D operations.The radial-sequence formulation avoids reliance on dense static city maps and the computational overhead of 3D convolutions under sparse aerial measurements.Source: 1) UNet · p. 9

2) Mamba

The section describes Mamba, InceptionTime, and 3D Ordinary Kriging as benchmark alternatives for one-dimensional REM reconstruction and volumetric interpolation.

2) Mamba

  • Mamba uses stacked selective state-space blocks with local feature mixing to model short-term variations and long-range propagation effects.Each of its four blocks includes pre-normalization, a gated linear unit, depthwise convolution, and a selective state-space scan.Source: 2) Mamba · p. 9

3) Inception

  • InceptionTime applies parallel 1D convolutions at multiple kernel sizes to time-series REM prediction, with stacked blocks intended to capture multiscale signal patterns.Its modules concatenate and normalize convolution and pooling branches before a pointwise convolution produces a REM value.Source: 3) Inception · p. 9

4) 3D Kriging

  • 3D Ordinary Kriging interpolates aerial measurements across spatial coordinates using an empirical semivariogram to represent volumetric correlation and distance-dependent decay.It is presented as a classical statistical baseline for reconstructing continuous radio maps across flight altitudes and spatial planes.Source: 4) 3D Kriging · pp. 9–10

IV. EVALUATION AND RESULTS

The evaluation compares REM models on three real-world datasets using continuous-time test splits and accuracy metrics. The Transformer+GRU improves RSRP predictions across datasets, while heatmap comparison assesses how learned spatial maps align with a deterministic reference.

IV. EVALUATION AND RESULTS

  • Testing uses separate continuous-time train and test intervals; Quectel trains on yaw 45° data and tests on yaw 315° data to assess orientation generalization.Nemo and PawPrints measurements cover 30 m and 50 m altitudes along the same UAV trajectory relative to the base station.Source: §IV · p. 10

A. RSRP Prediction Performance

  • Near 3 dB RMSE and R2 above 0.60 across datasets, Transformer+GRU improves RSRP prediction, with diminishing returns on smoothed PawPrints data.Baselines outperform the proposed models only on PawPrints; Table 2 reports the cross-dataset model comparison.Source: A. RSRP Prediction Performance · pp. 10–11

B. Spatial REM Visualization

  • PawPrints’ learned heatmap closely aligns with the deterministic LoS reference, although discernible deviations may reflect real-world propagation effects absent from that idealized map.The comparison uses SSIM for global RSRP distribution and FSIM for antenna-lobe patterns and coverage holes; Table 3 summarizes the results.Source: B. Spatial REM Visualization · p. 11, Table 3 · p. 11

C. Case Study: UAV Channel Quality Prediction

The Quectel REM-derived CQI map achieves above 80% Top-3 accuracy against observed channel quality. Under cross-altitude testing, the decoupled Transformer+GRU is reported as the most robust evaluated model, including a 4.19 dB RMSE result for Nemo.

C. Case Study: UAV Channel Quality Prediction

  • Above 80% Top-3 accuracy is achieved for REM-derived CQI predictions on Quectel data, with success defined as an absolute error within three indices.The SINR-to-CQI mapping assumes constant interference-plus-noise of −115 dBm per NR resource element and predominantly rank 2 channels.Source: C. Case Study: UAV Channel Quality… · p. 12, Table 4 · p. 12

D. Cross-Altitude Generalization

  • 4.19 dB RMSE is achieved on Nemo when training at 30 m and testing at 50 m, outperforming 3D Kriging and Mamba in that cross-altitude split.Across evaluated baselines, the decoupled Transformer+GRU is reported as the most robust to altitude shifts; the reverse direction is harder for models to infer lower-altitude antenna sidelobes.Source: D. Cross-Altitude Generalization · pp. 12–13, Table 5 · p. 13

E. Computational Complexity and System Trade-offs

The framework reports near-3 dB signal-strength errors and REM similarity above 0.75, while evaluating inference latency and accuracy trade-offs for real-time use. Its cascaded design separates large-scale radio footprints from localized fast-fading dynamics.

E. Computational Complexity and System Trade-offs

  • The optimal 6-layer Transformer+GRU uses 1,697k parameters versus Mamba’s 602k, while RSRP prediction RMSE remains under 3 dB across encoder-layer counts.Transformer parameter count depends on network depth and hidden dimensions, not spatial sequence length.Source: E. Computational Complexity and System… · p. 13
  • Increasing the radial step from 1 m to 5 m cuts T4 inference time from 1.76 ms to 0.57 ms, with a marginal accuracy penalty.The reported 5 m and 10 m step sizes have identical inference times; Mamba scales as O(N), while 3D Kriging scales as O(N^3).Source: E. Computational Complexity and System… · p. 13, Table 6 · p. 14

V. CONCLUSION

  • The cascaded REM framework separates learning large-scale spatial radio footprints from localized fast-fading dynamics.Source: §V · p. 13

E. Computational Complexity and System Trade-offs

V. CONCLUSION

  • Errors near 3 dB and REM similarity indexes above 0.75 accompany cross-dataset generalization in the benchmarked aerial predictions.The conclusion also describes the generated REM as useful for predicting UAV link quality.Source: §V · p. 14
Loading 2609.27083v1…