UrbanRadio3D: 3D Urban Radio Mapping

• UrbanRadio3D is a framework that creates 3D urban radio maps using ray tracing and sparse measurements, capturing key channel descriptors like pathloss, DoA, and ToA.
• It employs diverse mathematical formulations—from scalar fields to tensor and radiation-field models—to accurately represent complex radio propagation in urban environments.
• The framework supports applications such as UAV connectivity, wireless digital twins, and spectrum-aware planning by integrating geometry-aware reconstructions and active sensing techniques.

UrbanRadio3D denotes, in current literature, both a specific dataset and a broader research direction for constructing three-dimensional urban radio maps. In the narrower sense, it is a “large-scale, high-resolution, 3D RM dataset constructed via ray tracing in realistic urban environments,” organized as volumetric radio fields over urban blocks and annotated with pathloss, direction of arrival, and time of arrival (Wang et al., 16 Jul 2025). In the broader sense used across related works, it refers to systems that recover or generate urban 3D radio fields from geometry, sparse measurements, or both, for applications such as low-altitude networking, UAV connectivity, wireless digital twins, spectrum-aware planning, and integrated sensing and communication (Wen et al., 27 Jan 2026, Chen et al., 2024, Huang et al., 26 Apr 2026).

1. Conceptual scope and mathematical formulations

In the supplied literature, UrbanRadio3D appears in two closely related senses. One is a named benchmark dataset with volumetric labels for multiple channel descriptors (Wang et al., 16 Jul 2025). The other is a system-level abstraction in which an urban environment is mapped into a 3D radio representation that can be queried at arbitrary transmitter–receiver configurations (Wen et al., 27 Jan 2026).

Several mathematical formulations coexist. A minimal 3D radio map is a scalar field over space, such as the RSRP mapping

$\gamma : \mathbb{R}^3 \rightarrow \mathbb{R},$

where $\gamma(\mathbf{p})$ gives received power at location $\mathbf{p}=[p_x,p_y,p_z]^\mathrm{T}$ (Chen et al., 2024). A richer volumetric representation is a tensor

$$\mathcal{R} \in \mathbb{R}^{H \times W \times D \times C},$$

with spatial axes $H,W,D$ and modality dimension $C$, where modalities include pathloss, DoA, and ToA (Wang et al., 16 Jul 2025). A still more general view is the radiation-field formulation

$$\mathcal{R} : (\mathbf{x}, \boldsymbol{\omega}, f, \text{Tx}, \text{Rx}) \mapsto \text{signal quantity},$$

which treats optical radiance and RF behavior as two renderings of a shared physical field (Wen et al., 27 Jan 2026).

The literature also contains a full-dimensional formulation in which the map depends on both transmitter and receiver positions. In that case the link state is a 6D variable $$\mathbf{p}=(\mathbf{p}_{\mathrm{u}},\mathbf{p}_{\mathrm{d}})\in\mathbb{R}^6$$, and the radio map becomes $g(\mathbf{p}):\mathbb{R}^6\to\mathbb{R}$, returning large-scale channel gain for arbitrary 3D user–3D UAV pairs (Liu et al., 2021). This suggests that UrbanRadio3D is not restricted to receiver-side volumetric coverage alone; it can also be understood as a Tx–Rx relational field.

2. Radio quantities and environmental representations

A common misconception is that UrbanRadio3D concerns only 3D pathloss images. The literature is broader. Depending on the paper, the mapped quantity may be RSRP, RSSI, pathloss, free-space path loss, received gain, spatial spectrum, multipath components, direction of arrival, angle of departure, or time of arrival (Chen et al., 2024, Wen et al., 27 Jan 2026, Wang et al., 16 Jul 2025).

The UrbanRadio3D dataset introduced with RadioDiff-3D stores pathloss, DoA in azimuth and elevation, and ToA on a $256\times256\times20$ grid, where receiver heights run from 1 m to 20 m in 1 m steps (Wang et al., 16 Jul 2025). URF-GS uses a more general RF side based on spatial spectrum and MPC-derived attributes, including delay, AoD, AoA, and power, and reorganizes measured MPCs into spatial spectrum images via equirectangular projection (Wen et al., 27 Jan 2026). SpectrumNet, while not itself called UrbanRadio3D, extends the representation along other axes—frequency, terrain, and climate—by modeling

$\gamma(\mathbf{p})$0

over 3 heights, 5 frequency bands, 11 terrain scenarios, and 3 climate scenarios (Zhang et al., 2024).

Environmental representation is equally heterogeneous. URF-GS attaches geometry, optical appearance, and RF material attributes to the same 3D Gaussian primitives,

$\gamma(\mathbf{p})$1

so that geometry, normals, color, albedo, metallicity, and roughness are all part of a single field (Wen et al., 27 Jan 2026). R$\gamma(\mathbf{p})$2Net instead embeds height directly into 2D images through

$\gamma(\mathbf{p})$3

then predicts a multi-channel tensor whose channels correspond to receiver heights (Rao et al., 18 May 2026). GeoUQ-GFNet uses rasterized geometry priors $\gamma(\mathbf{p})$4, namely occupancy, height, BS-relative coordinates, distance, and a LoS proxy (Zeng et al., 7 Apr 2026). Older semantically structured work represents the environment through a multi-class virtual obstacle map $\gamma(\mathbf{p})$5, learned from RSS alone and intended to encode radio semantics rather than literal building geometry (Liu et al., 2021).

3. Datasets, benchmarks, and empirical regimes

UrbanRadio3D research is strongly dataset-driven. Representative resources differ in whether they are ray-traced or measured, whether they are volumetric or slice-based, and whether they model a single metric or a multimodal channel state.

Resource	Scale and geometry	Radio quantities
UrbanRadio3D	701 urban regions, $\gamma(\mathbf{p})$6, 20 height layers, 200 Tx per map, $\gamma(\mathbf{p})$7 simulation instances, 11.2M samples	Pathloss, DoA_Azi, DoA_Ele, ToA
SpectrumNet	15,300 areas, $\gamma(\mathbf{p})$8 grid over $\gamma(\mathbf{p})$9, 3 heights, 5 bands, 11 terrains, 3 climates	RSS
Nanchang UAV dataset	$\mathbf{p}=[p_x,p_y,p_z]^\mathrm{T}$0, 4274 UAV points	RSRP
UrbanRT-RM	$\mathbf{p}=[p_x,p_y,p_z]^\mathrm{T}$1, 7 urban scenes, 8 BS deployments, 3 sampling modes	Large-scale gain map

UrbanRadio3D is explicitly described as “over 37× larger than previous datasets” and as having “7× more height layers than prior state-of-the-art dataset” (Wang et al., 16 Jul 2025). It uses real cities including Kara, Berlin, Glasgow, Ljubljana, London, and Tel Aviv, with building heights ranging from 6.6 m to 19.8 m and simulations performed with WinProp and FEKO under the Dominant Path Model at 5.9 GHz (Wang et al., 16 Jul 2025). SpectrumNet broadens the scenario space rather than the vertical depth, providing over 300,000 radio map images with terrain and climate effects, including dense urban, urban, suburban, rural, mountainous, forest, lake, and ocean cases (Zhang et al., 2024).

Measurement-based urban 3D mapping remains central. The Nanchang campus dataset covers a “densely-built urban” region with many buildings 6–38.6 m tall, obstacles including cars, walls, trees, and vegetation, and 4274 RSRP measurements collected by a DJI Matrice 30 RTK UAV at 2.645 GHz (Chen et al., 2024). AERPAW-based 3D REM studies contribute C-Band UAV measurements at altitudes from 30 m to 110 m and supply empirical 3D correlation and semi-variogram models for Kriging (Maeng et al., 2023). UrbanRT-RM provides a controllable 3D-ray-traced but 2D-observation benchmark with road-oriented and non-road urban layouts, including canyon, crossroad, T-junction, and sparse-building scenes (Zeng et al., 7 Apr 2026).

This diversity suggests that UrbanRadio3D is not a single benchmark culture. It spans dense urban field measurement, synthetic volumetric ray tracing, and hybrid benchmarks that project 3D propagation onto selected slices or patches.

4. Construction methodologies

Methodologically, the literature divides into four broad families: residual geostatistical cartography, sequence models, geometry-aware neural reconstruction, and unified physics-based rendering.

A classic hybrid formulation decomposes measured power into simulated prior, residual, and noise: $\mathbf{p}=[p_x,p_y,p_z]^\mathrm{T}$2 The Nanchang work then models $\mathbf{p}=[p_x,p_y,p_z]^\mathrm{T}$3 as a Gaussian process over the augmented feature

$\mathbf{p}=[p_x,p_y,p_z]^\mathrm{T}$4

and uses a composite kernel $\mathbf{p}=[p_x,p_y,p_z]^\mathrm{T}$5, for which “constant × Matérn + white noise” achieves RMSE $\mathbf{p}=[p_x,p_y,p_z]^\mathrm{T}$6 dB (Chen et al., 2024). Closely related 3D Kriging work uses empirical air-to-ground correlation $\mathbf{p}=[p_x,p_y,p_z]^\mathrm{T}$7 and the induced semi-variogram $\mathbf{p}=[p_x,p_y,p_z]^\mathrm{T}$8 to interpolate 3D REMs from UAV measurements (Maeng et al., 2023).

Transformer-based methods convert propagation into sequence prediction. TransfoREM discretizes radial distance from a BS up to $\mathbf{p}=[p_x,p_y,p_z]^\mathrm{T}$9 m and learns per-direction RSRP sequences from a six-feature representation that combines $$\mathcal{R} \in \mathbb{R}^{H \times W \times D \times C},$$0, spherical angles, and Cartesian coordinates. It uses an encoder-only transformer with embedding dimension 64, 6 encoder layers, and 8 heads, trained first on FSPL plus antenna gain and then fine-tuned on real data with Smooth L1 loss (Reddy et al., 23 Jan 2026). PILOT also uses autoregression, but replaces raster scan with a wavefront sequence expanding outward from the transmitter and conditions each prediction on aligned environment tokens from a ViT encoder. The same model handles 2D and 3D path-loss generation through height-slice stacking and a 3D gradient loss (Huang et al., 26 Apr 2026).

Geometry-aware neural models usually trade explicit physics for efficient volumetric inference. R$$\mathcal{R} \in \mathbb{R}^{H \times W \times D \times C},$$1Net reduces 3D radio map estimation to a 2D residual network with height embedding, using R$$\mathcal{R} \in \mathbb{R}^{H \times W \times D \times C},$$2Net-In for penetration-dominated indoor settings and R$$\mathcal{R} \in \mathbb{R}^{H \times W \times D \times C},$$3Net-Out for diffraction-dominated outdoor settings (Rao et al., 18 May 2026). GeoUQ-GFNet fuses structural priors, relative coordinates, sparse observations, and accessibility masks in a lightweight Ghost/Grid-KAN encoder with an FPN decoder, and predicts both dense gain and heteroscedastic uncertainty (Zeng et al., 7 Apr 2026). RadioDiff-3D uses a 3D U-Net within a latent diffusion framework to generate volumetric pathloss maps in either radiation-aware or radiation-unaware settings (Wang et al., 16 Jul 2025).

URF-GS is the most explicitly physics-grounded formulation in the set. It reconstructs a shared radio–optical field with 3D Gaussian Splatting, trains an optical stage from RGB, monocular depth, and monocular normals, then freezes geometry and learns RF material attributes from spatial spectrum supervision. RF propagation is rendered through differentiable path tracing with a PBR-inspired BRDF and explicit FSPL: $$\mathcal{R} \in \mathbb{R}^{H \times W \times D \times C},$$4 This approach makes UrbanRadio3D a geometry- and material-aware inverse-rendering problem rather than a pure regression task (Wen et al., 27 Jan 2026).

5. Sparse sensing, uncertainty, and active perception

Sparse sampling is not an incidental constraint in UrbanRadio3D; it is often the primary design condition. The Nanchang GPR framework introduces both an online MAP-based selection rule,

$$\mathcal{R} \in \mathbb{R}^{H \times W \times D \times C},$$5

and an offline KMeans-based design in the augmented feature space $$\mathcal{R} \in \mathbb{R}^{H \times W \times D \times C},$$6. Both reach RMSE $$\mathcal{R} \in \mathbb{R}^{H \times W \times D \times C},$$7 dB using only 2% of the sampling points, whereas IDW, KNN, and Kriging require more than 14% to reach the same error level (Chen et al., 2024).

Uncertainty estimation increasingly functions as both a reliability signal and an acquisition function. GeoUQ-GFNet predicts a log-variance map $$\mathcal{R} \in \mathbb{R}^{H \times W \times D \times C},$$8 and optimizes a heteroscedastic Gaussian NLL combined with $$\mathcal{R} \in \mathbb{R}^{H \times W \times D \times C},$$9, gradient, and variance-regularization terms. Additional measurements are then chosen by top-$H,W,D$0 uncertainty over accessible and unobserved locations. Under a +4% additional budget, uncertainty-guided querying reduces RMSE to 3.1002 dB, compared with 5.9144 dB for random sampling (Zeng et al., 7 Apr 2026).

3D-URAM extends this logic to closed-loop active mapping in uncharted air–ground environments. Stage I is a Bayesian UNet trained with dual masking over sparse radio and partial geometry; Stage II is a dynamic probabilistic roadmap plus transformer waypoint policy trained with PPO to maximize normalized uncertainty reduction under travel budget. The reward is tied to the decrease in global uncertainty

$H,W,D$1

and the policy selects among candidate 3D waypoints whose features include position, predicted mean radio map, and uncertainty (Lu et al., 11 Jun 2026). The paper reports that 3D-URAM reduces reconstruction error by over 50% compared to representative baselines and validates the approach in a real-world $H,W,D$2 space (Lu et al., 11 Jun 2026).

This suggests that active perception is becoming a structural component of UrbanRadio3D rather than a downstream add-on. The map is increasingly treated as a latent field whose uncertainty must guide where the next measurement is taken.

6. Applications, empirical patterns, and recurrent limitations

UrbanRadio3D methods are repeatedly connected to network planning, AP and BS deployment, blockage prediction, interference management, robot and UAV path planning, radio dynamic zones, wireless digital twins, and immersive visualization (Wen et al., 27 Jan 2026, Reddy et al., 23 Jan 2026, Maeng et al., 2023). URF-GS explicitly frames 3D radio maps as tools for “spectrum-aware planning” and “environment-aware sensing,” including AP placement and robot or drone navigation with connectivity guarantees (Wen et al., 27 Jan 2026). TransfoREM is designed for per-BS deployment so that REMs can support “enhanced resource allocation, interference management, and spatial spectrum utilization” (Reddy et al., 23 Jan 2026). The 6D virtual-obstacle framework shows that a radio-map-assisted UAV placement strategy can achieve more than 50% capacity gain in a relay scenario (Liu et al., 2021).

Across benchmarks, several performance patterns recur. URF-GS reports up to a 24.7% improvement in spatial spectrum prediction accuracy and a 10x increase in sample efficiency relative to NeRF-based methods (Wen et al., 27 Jan 2026). PILOT reports NMSE 0.0120 in volumetric mode and 0.0071 in height-conditioned mode on UrbanRadio3D 1–4 m slices, reducing NMSE by 78% relative to the diffusion baseline while being roughly $H,W,D$3 faster at inference (Huang et al., 26 Apr 2026). TransfoREM improves from RMSE 7.49 dB to 4.57 dB on AERPAW between pre-training and fine-tuning, and reaches RMSE 1.29 dB on a separate LTE UAV dataset, while outperforming Kriging in altitude extrapolation by about 1.5 dB in median error (Reddy et al., 23 Jan 2026). R$H,W,D$4Net demonstrates that height-embedded 2D architectures can approximate 3D estimation efficiently, reaching NMSE 0.0268 on 3DiRM3200 and 0.0046 on RadioMapSeer (Rao et al., 18 May 2026).

The literature also converges on several limitations. Static-scene assumptions are common: URF-GS notes that dynamic blockage from vehicles and pedestrians is not captured, and 3D-URAM likewise assumes time-invariant maps (Wen et al., 27 Jan 2026, Lu et al., 11 Jun 2026). Frequency generalization remains limited in many studies: URF-GS experiments are fixed at 60 GHz, UrbanRT-RM at 3.5 GHz, and 3D-URAM at 2.4 GHz (Wen et al., 27 Jan 2026, Zeng et al., 7 Apr 2026, Lu et al., 11 Jun 2026). Some benchmarks remain effectively 2.5D even when the underlying propagation is fully 3D; UrbanRT-RM learns a ground-plane gain slice, while several 3D-URAM experiments operate on altitude-indexed 2D grids rather than full voxel CNNs (Zeng et al., 7 Apr 2026, Lu et al., 11 Jun 2026). Sim-to-real calibration is another persistent issue, especially for ray-traced datasets such as SpectrumNet and UrbanRadio3D, whose authors explicitly discuss calibration and domain-transfer needs (Zhang et al., 2024, Wang et al., 16 Jul 2025).

Taken together, these works indicate that UrbanRadio3D has evolved from single-plane pathloss interpolation into a technically distinct program of volumetric radio-environment reconstruction. Its defining features are explicit height dependence, geometric conditioning, sparse-measurement efficiency, and a shift from scalar coverage prediction toward richer channel descriptors such as DoA, ToA, and spatial spectrum. The outstanding research tension is between physical fidelity, scalable inference, and real-world adaptability.