PaperlayerView exact revision on arXiv

Source-linked AI summary

Channel Knowledge Maps for FR3 UAV Base Stations: Dataset and ML-based Modeling

Guillem Moreno~Garcia, Sergi Abadal, Evgenii Vinogradov

arXiv:2610.06356v1eess.SP

TL;DR

Existing CKM datasets lack FR3 channels and UxNB-specific geometries, leaving limited dataset support for studying these deployments. This paper introduces an FR3 dataset and a height-aware attenuation model with separate calibrated LOS and NLOS branches. The model reports 1.93 dB overall test RMSE with supplied visibility masks.

  • Problem

    Existing CKM datasets lack FR3 channels and UxNB-specific geometries, while UxNB-relevant channel maps require urban coverage and heights reaching up to 500 m.

    Source: §I · p. 1

  • Method

    The paper introduces an FR3 CKM dataset and a height-aware framework that calibrates separate LOS and NLOS propagation models using CKM training data.

    Source: §I · p. 2, A. Line-of-Sight Branch · p. 4, B. Non-Line-of-Sight Branch · p. 4

  • Results

    1.93 dB overall test RMSE with supplied visibility masks; the simulator’s median CPU runtime is 0.06 s per map.

    Source: §V · p. 6

  • Takeaways & Limitations

    The dataset provides FR3 attenuation and other channel data, while the calibrated model generates attenuation maps and generalizes to unseen city layouts.

    Source: §V · p. 6

    Reported accuracy assumes supplied visibility masks, and generating the masks is excluded from the runtime measurement.

    Source: §V · p. 6

Abstract

from arXiv · show

This paper introduces a dense upper mid-band Frequency Range (FR3) Channel Knowledge Map (CKM) dataset for Uncrewed Aerial Vehicles (UAVs) with an on-board base station (UxNB). It contains 16,180 CKM samples from 3,236 urban locations in 66 cities at 7.125 GHz and UxNB heights of up to 500 meters. Every 513x513 sample stores the building geometry, the line-of-sight (LOS) state, attenuation, delay spread, angular spread, and variable transmitter height. As a focused first use-case, we propose an interpretable height-aware attenuation model. Its LOS branch combines free-space loss, two-ray term, and compact data-based height/range calibration. Its non-LOS (NLOS) branch combines a COST 231-Hata basis with local environment-specific linear calibration. The proposed model reaches 1.93 dB overall root mean square error, with 1.74 dB in LOS and 3.54 dB in NLOS on over 500 unique locations in 14 unseen cities (2,590 samples in total). The result provides attenuation data for FR3 UxNB deployments and a physical anchor for future FR3 modeling efforts.

I. INTRODUCTION

The introduction identifies a lack of public FR3 UxNB channel-map data and limitations in existing datasets and learning-based models. It presents a new FR3 dataset and a height-aware framework that separates LOS and NLOS propagation.

I. INTRODUCTION

  • No publicly available UxNB CKM datasets in FR3 are identified, despite FR3’s potential for UxNB coverage and capacity.The introduction describes FR3 as balancing sub-6 GHz propagation reliability and millimeter-wave bandwidth, while requiring reliable NLOS operation.Source: §I · p. 1
  • Existing CKM datasets lack FR3 channels, UxNB-specific geometries, and accurate 3D city models.The introduction also notes that available UrbanRadio3D receiver heights reach only 20 m, while UxNBs may operate up to 500 m.Source: §I · p. 1
  • Existing learning-based channel-map methods use heavy image-to-image networks, assume fixed terrestrial geometries, and often require local fine-tuning for unseen cities.The stated limitations include rediscovering propagation physics independently for each environment and difficulty representing variable transmitter altitude.Source: §I · p. 2
  • The paper introduces a 7.125 GHz FR3 CKM dataset with 16,180 samples across 3,236 city layouts in 66 cities and UxNB heights up to 500 m.The dataset uses complete, global-scale 3D building data,.Source: §I · p. 2
  • The proposed height-aware framework separates LOS and NLOS propagation and calibrates deterministic baseline models with lightweight regression.The introduction names free-space, two-ray, and COST 231-Hata models as examples, and reports zero-shot generalization to unseen cities.Source: §I · p. 2

II. CKM: DEFINITION AND DATASET DESCRIPTION

A CKM sample combines spatial environmental features, channel parameters, and operational metadata. CKM estimation is framed as mapping the environment and metadata to an estimated channel tensor.

A. CKM Definition and Mapping Formulation

  • A CKM sample combines environmental features, channel parameters, and operational metadata.The sample represents the environment and channel on spatial grids and stores operational information in a metadata vector.Source: A. CKM Definition and Mapping… · p. 2
  • The environmental tensor represents the propagation environment, while the channel tensor stores receiver-grid parameters such as attenuation, delay spread, and angular parameters.The metadata vector can include carrier frequency and UxNB location information.Source: A. CKM Definition and Mapping… · p. 2
  • CKM estimation is formulated as learning a mapping from the environmental tensor and metadata to an estimated channel tensor.The estimated tensor is optimized to approximate the ground-truth target channel tensor.Source: A. CKM Definition and Mapping… · p. 2

B. The FR3 UxNB CKM Dataset

The dataset uses ray tracing over diverse urban layouts and elevated UxNB positions to produce spatially aligned environmental and channel maps. Its samples span 7.125 GHz and transmitter heights from 10 m to 500 m.

B. The FR3 UxNB CKM Dataset

  • The environmental maps use global-scale 3D building datasets that provide consistent building-height information for UxNB scenarios.The paper contrasts this with optional OSM building-height tags available in fewer than 10% of cases.Source: 1) 3D urban environment modeling (E) · p. 3
  • The dataset covers 3,236 urban locations in 66 cities with 513 × 513 environmental maps at 1 m resolution.The two environmental layers encode building height profiles and binary building-occupancy masks.Source: 1) 3D urban environment modeling (E) · p. 3
  • The 16,180 simulated samples pair 3,236 layouts with five UxNB heights each, spanning 10 m to 500 m at 7.125 GHz.Receivers are placed at ground level in non-building coordinates, and the UxNB is positioned at the layout center.Source: 2) Ray Tracing channel simulation (C) · p. 3
  • Each channel tensor stores LOS/NLOS visibility, attenuation, RMS delay spread, and RMS angular spread.In-building elements receive zero channel values and should be masked during training and evaluation.Source: 2) Ray Tracing channel simulation (C) · p. 3
  • The dataset assigns transmitter-height and ITU-topology metadata, with topology selected by comparing relative differences in coverage, building density, and mean building height.The topology distance uses equal-weight log-ratios so the scale of building density does not dominate.Source: 3) Metadata (m) · p. 3

III. HEIGHT-AWARE CKM MODELING

The attenuation simulator models LOS and NLOS regions separately because their propagation mechanisms differ and vary with UxNB altitude. Its reported accuracy uses a dataset visibility mask to identify the regions.

III. HEIGHT-AWARE CKM MODELING

  • The simulator combines distinct LOS and NLOS attenuation models and uses the dataset visibility mask to assign propagation regions.The reported accuracy assumes the mask; generating it via ray tracing or shadow projection is excluded from timing.Source: §III · pp. 3–4
  • LOS propagation is modeled with free-space loss and coherent two-ray ground reflection, while NLOS uses a CKM-data-corrected COST 231-Hata model.The NLOS branch also accounts for ITU topology characteristics.Source: §III · p. 4
  • The model precomputes horizontal and 3D distances, ground-reflected path length, and elevation angle across the 513 × 513 grid.These geometry matrices are indexed by receiver location for the propagation models.Source: §III · p. 4

A. Line-of-Sight Branch

The LOS branch combines free-space path loss and coherent ground reflection, then calibrates the model using transmitter-height and distance-dependent corrections.

A. Line-of-Sight Branch

  • The LOS attenuation model combines direct free-space path loss with a coherent ground-reflection term to represent two-ray propagation.The free-space term follows Friis distance dependence, while the coherent correction retains the direct/reflected-path phase difference.Source: A. Line-of-Sight Branch · p. 4
  • Height-dependent bias ζ(htx) and residual correction r(d2D, htx) calibrate the LOS prediction for transmitter height and horizontal distance.The residual term is computed as the mean residual error after choosing calibration parameters that minimize RMSE in each height bin.Source: A. Line-of-Sight Branch · p. 4
  • LOS calibration uses a lookup table of height- and distance-dependent ρ, ϕ, ζ, and r, with inference interpolating neighboring entries.The candidate reflection amplitude and phase are evaluated within 5 m height bins, and the minimum-RMSE combination is retained.Source: A. Line-of-Sight Branch · p. 4

B. Non-Line-of-Sight Branch

The NLOS branch augments COST 231-Hata with local environment features and calibrated weights specific to topology and transmitter-height regime.

B. Non-Line-of-Sight Branch

  • NLOS attenuation is predicted from a COST 231-Hata baseline augmented with local topographical features and a height- and topology-dependent weight vector.The feature vector has 14 components, and weights are calibrated offline for nine topology-by-altitude regimes.Source: B. Non-Line-of-Sight Branch · p. 4
  • Thirteen contextual features summarize building density, normalized height, shadowed-ground fraction, path elevation, and distance at local and regional scales.The local and regional windows use sizes 15 and 41, respectively; the feature set also includes interaction cross-terms.Source: B. Non-Line-of-Sight Branch · p. 4
  • Designed for sub-3 GHz terrestrial links, the COST 231-Hata formulation is used as a regression feature, with its coefficients fitted on training data.Source: B. Non-Line-of-Sight Branch · p. 4
  • A Vancouver example at 50.30 m has 2.18 dB overall RMSE, with 2.10 dB LOS RMSE and 2.52 dB NLOS RMSE.The example contains 119,180 LOS and 25,607 NLOS channels.Source: Figure 3 · p. 5

A. Experimental Setup and Evaluation Protocol

Evaluation separates cities into training, validation, and unseen test partitions, then reports test coverage and regime-focused assessments.

A. Experimental Setup and Evaluation Protocol

  • The city-disjoint split uses 10,840 training samples, 2,750 validation samples, and 2,590 test samples from 14 unseen cities.Training uses 37 cities and validation uses 15; validation supports development checks and ablation, while final performance uses the test partition.Source: A. Experimental Setup and Evaluation… · p. 5
  • The test evaluates over 4.23×10^8 ground receivers, of which 7.4% are NLOS; 424 test maps have more than 25% NLOS receivers.The paper reports separate regime-specific evaluations because the overall element-weighted score is LOS-dominated.Source: A. Experimental Setup and Evaluation… · p. 5
  • The evaluation includes RMSE by ITU topology and a comparison with a data-driven baseline.The supplied table titles identify these assessment scopes but do not provide table values.Source: Table III · p. 5, Table IV · p. 5

B. Prediction Accuracy and Height Generalization

The proposed simulator achieves 1.93 dB overall RMSE on unseen test cities, with performance varying by propagation regime, topology, and UxNB altitude. Against a pure end-to-end neural model, it has lower global and LOS errors, while the neural model has lower NLOS error.

B. Prediction Accuracy and Height Generalization

  • 1.93 dB overall RMSE, with 1.74 dB for LOS and 3.54 dB for NLOS, on unseen test cities; the unweighted 14-city mean is 1.98 dB.Suburban topology has the lowest total error at 1.70 dB, while Dense Urban NLOS error is 4.13 dB; per-city RMSE ranges from 1.61 dB in Halifax to 2.59 dB in Osaka.Source: B. Prediction Accuracy and Height… · p. 5
  • Overall RMSE decreases monotonically from 2.14 dB at 10 m to 1.71 dB at 500 m UxNB altitude.The passage relates this trend to observations that higher-altitude channels are usually more stable and predictable,.Source: B. Prediction Accuracy and Height… · p. 5
  • Against GMM U-Net, the proposed model has lower global RMSE (1.93 dB vs. 4.36 dB) and LOS error (1.74 dB vs. 4.43 dB), while GMM U-Net has lower NLOS error (3.34 dB).Both models were evaluated on the same 2,590 test maps and mask.Source: B. Prediction Accuracy and Height… · p. 5
  • The 1.93 dB error is presented as strong accuracy for larger areas with variable UxNB transmitter altitudes, but published models are not directly rankable under their different conditions.The cited context includes RadioUNet at 1.60 dB on fixed-height 256 × 256 grids, TSRDiff at 3.62 dB under sparse target sampling, and RO-HMTGP at 4.19–6.00 dB on empirical RSRP data.Source: B. Prediction Accuracy and Height… · pp. 5–6

C. Ablation and Computational Complexity Studies

Validation ablations show substantial gains from calibrated propagation terms, while a compact feature set retains most of the complete model’s improvement. The computational context specifies same-CPU median comparisons and inference-run setup, but gives no runtime values in these passages.

C. Ablation and Computational Complexity Studies

  • On the 2,750-sample validation set, adding coherent ground reflection reduces LOS RMSE from 3.79 dB to 1.7412 dB; calibrated COST 231-Hata reduces NLOS RMSE from 25 dB to 3.20 dB.The NLOS calibration includes fitted regression coefficients and topology- and height-specific calibration.Source: C. Ablation and Computational… · p. 6
  • A compact model retaining eight of 14 features and no interaction cross-terms preserves over 90% of the complete model’s improvement over COST 231-Hata.The retained configuration uses only l = 41.Source: C. Ablation and Computational… · p. 6
  • The runtime comparison uses same-CPU medians on an AMD Ryzen 5 5600X with GPU disabled, with 10 untimed and 50 timed RadioUNet/PMNet runs and 100 proposed-model samples.The proposed-model samples use supplied masks and exclude file I/O.Source: C. Ablation and Computational… · p. 6

V. CONCLUSIONS

The conclusion presents a dense FR3 UAV CKM dataset and a computationally efficient, height-aware attenuation framework that generalizes to unseen city layouts. With supplied visibility masks, the simulator achieves 1.93 dB overall test RMSE and a median CPU runtime of 0.06 s per map.

V. CONCLUSIONS

  • The dataset provides path loss, delay spread, angular spread, and geometric visibility across 16,180 simulated samples from 66 global cities at 7.125 GHz.The paper proposes a computationally efficient, height-aware attenuation framework based on conventional models calibrated on the dataset; it reports generalization to unseen city layouts.Source: §V · p. 6
  • With supplied visibility masks, the proposed simulator achieves 1.93 dB overall test RMSE and a median CPU runtime of 0.06 s per map.Source: §V · p. 6
Loading 2610.06356v1…