PaperlayerView exact revision on arXiv

Source-linked AI summary

RadioDiff-3D: A 3D$\times$3D Radio Map Dataset and Generative Diffusion Based Benchmark for 6G Environment-Aware Communication

Xiucheng Wang, Qiming Zhang, Nan Cheng, Junting Chen, Zezhong Zhang, Zan Li, Shuguang Cui, Xuemin Shen

arXiv:2507.12166v1cs.LGcs.SYeess.SY

TL;DR

Existing radio-map methods mainly predict fixed-height 2D pathloss, leaving DoA, ToA, and vertical variation underrepresented. The paper introduces a ray-traced 3D dataset and 3D-convolutional benchmarks, including RadioDiff-3D; the authors report viable data-driven 3D generation and strong baselines for future work.

  • Problem

    Most existing radio-map predictors focus on fixed-height 2D pathloss, while key 3D descriptors such as DoA and ToA are largely absent.

    Source: A. RM Construction Methods · p. 4

  • Method

    The paper constructs UrbanRadio3D and evaluates a 3D-UNet and 3D-convolutional RadioDiff-3D across channel modalities and transmitter-aware and transmitter-agnostic settings.

    Source: §I · pp. 2–3, §VI · p. 15

  • Results

    The authors report viable data-driven 3D radio-map generation and strong baselines for future research.

    Source: §VI · p. 15

  • Takeaways & Limitations

    The dataset and framework provide a foundational toolset for advancing environment-aware wireless communication.

    Source: §VI · p. 15

Abstract

from arXiv · show

Radio maps (RMs) serve as a critical foundation for enabling environment-aware wireless communication, as they provide the spatial distribution of wireless channel characteristics. Despite recent progress in RM construction using data-driven approaches, most existing methods focus solely on pathloss prediction in a fixed 2D plane, neglecting key parameters such as direction of arrival (DoA), time of arrival (ToA), and vertical spatial variations. Such a limitation is primarily due to the reliance on static learning paradigms, which hinder generalization beyond the training data distribution. To address these challenges, we propose UrbanRadio3D, a large-scale, high-resolution 3D RM dataset constructed via ray tracing in realistic urban environments. UrbanRadio3D is over 37$\times$3 larger than previous datasets across a 3D space with 3 metrics as pathloss, DoA, and ToA, forming a novel 3D$\times$33D dataset with 7$\times$3 more height layers than prior state-of-the-art (SOTA) dataset. To benchmark 3D RM construction, a UNet with 3D convolutional operators is proposed. Moreover, we further introduce RadioDiff-3D, a diffusion-model-based generative framework utilizing the 3D convolutional architecture. RadioDiff-3D supports both radiation-aware scenarios with known transmitter locations and radiation-unaware settings based on sparse spatial observations. Extensive evaluations on UrbanRadio3D validate that RadioDiff-3D achieves superior performance in constructing rich, high-dimensional radio maps under diverse environmental dynamics. This work provides a foundational dataset and benchmark for future research in 3D environment-aware communication. The dataset is available at https://github.com/UNIC-Lab/UrbanRadio3D.

I. INTRODUCTION

Radio maps can support environment-aware 6G, but existing learning-based methods largely predict fixed-height 2D pathloss and lack public, densely sampled 3D data. The paper introduces UrbanRadio3D and 3D-convolutional benchmarks for volumetric maps with multiple channel features.

I. INTRODUCTION

  • RadioDiff-3D supports transmitter-aware construction from known transmitter information and transmitter-unaware estimation from sparse observations and environmental priors.The latter setting addresses interference from non-cooperative transmitters.Source: §I · p. 2
  • UrbanRadio3D represents pathloss, DoA, and ToA across 3D space in a one-meter voxel format spanning receiver altitudes.Its format is designed to integrate structured simulation data with volumetric learning models.Source: §I · p. 2
  • RadioDiff-3D uses a diffusion model with 3D convolutions to synthesize volumetric maps, alongside a 3D-convolutional UNet benchmark.The models target dense, multi-feature channel representations rather than fixed-height prediction.Source: §I · p. 3

A. RM Construction Methods

  • Sampling-based approaches use sparse path-loss measurements and interpolation, but become impractical where measurements are difficult and can degrade in sparsely sampled NLOS regions.Kriging also incurs cubic matrix-inversion cost as sample counts grow.Source: A. RM Construction Methods · p. 3
  • Sampling-free approaches use environmental priors, including raster-based neural models and diffusion, but existing approaches retain limits in spatial scope, measurement dependence, or scalability.Examples include 2D pathloss predictors and generative models that still require sparse measurements.Source: A. RM Construction Methods · pp. 3–4
  • Most radio-map work remains fixed-height and pathloss-focused, while scarce high-resolution 3D data and 2D architectures limit learning of volumetric propagation.The paper identifies native 3D modeling and high-precision 3D channel datasets as two bottlenecks.Source: §I · p. 2, A. RM Construction Methods · p. 4

B. Data Acquisition of RM

The section surveys radio-map data-generation approaches and their trade-offs, then describes diffusion models and DDIM sampling as a faster alternative to standard DDPM inference.

B. Data Acquisition of RM

  • Ray-tracing datasets offer environmental precision and scale, while analytical path-loss maps are computationally attractive but omit location-specific propagation effects.Ray tracing can yield richer channel descriptors at higher computational cost; analytical-model discrepancies from field measurements can exceed 10 dB.Source: B. Data Acquisition of RM · pp. 4–5
  • RadioGAT, RMDirectionalBerlin, RadioMapSeer, CKMImageNet, and SpectrumNet vary in urban coverage and modalities, but retain limits in 3D, channel detail, or spatial granularity.RadioGAT and RadioMapSeer use 2D layouts; RMDirectionalBerlin lacks DoA and ToA; CKMImageNet lacks explicit DoA/DoD or ToA in many scenes; SpectrumNet has coarse, discrete heights and scalar pathloss.Source: B. Data Acquisition of RM · p. 5

C. Generative Diffusion Model

  • DDPM generates data through forward noise addition and a learned reverse denoising process, with a simplified variance choice reported not to degrade performance.The forward process adds Gaussian noise; the reverse process learns the transition distribution. Using β_t instead of the theoretically correct variance scaling simplifies implementation.Source: C. Generative Diffusion Model · p. 5
  • DDIM redefines reverse sampling for faster, deterministic generation, and η = 0 yields consistent samples with fewer steps.DDIM retains the DDPM training objective and is presented as useful for time-sensitive radio-map construction.Source: C. Generative Diffusion Model · p. 6

III. PROPERTIES OF URBANRADIO3D

UrbanRadio3D comprises high-resolution urban propagation simulations across 20 receiver heights, with multiple channel modalities and auxiliary spatial maps. The simulations use uniform default material properties across building facades.

A. 3D RadioMap Dataset Construction

  • The dataset contains 701 urban regions, 200 transmitter locations per region, and 20 receiver heights, yielding 2.84 million simulation instances at 1 m spatial resolution.Regions cover 256 × 256 m² and represent cities including Kara, Berlin, Glasgow, Ljubljana, London, and Tel Aviv.Source: A. 3D RadioMap Dataset Construction · p. 6
  • Each 3D map stacks 20 horizontal slices simulated at fixed receiver heights, with pathloss, DoA, ToA, building masks, height maps, and transmitter-location maps.The simulations use 1 m resolution within each slice and the Dominant Path Model (DPM).Source: A. 3D RadioMap Dataset Construction · p. 6
  • All building facades use the same default material parameters, assuming uniform electromagnetic behavior across surfaces.The paper says this simplifies large-scale simulation while maintaining reasonable physical accuracy.Source: A. 3D RadioMap Dataset Construction · p. 6
  • The dataset presents three scenarios with both 2D and 3D views, alongside 3D ray-tracing views.Source: Figure 1 · p. 6, Figure 2 · p. 6

B. Visualization of the Dataset

The dataset visualizations show how elevation and urban structures relate to radio-map patterns, while the naming convention records modality, height, building, and transmitter location. Channel parameters are normalized using global thresholds.

B. Visualization of the Dataset

  • Visualizations cover pathloss at receiver heights from 1 to 20 m, ToA, DoA azimuth and elevation, propagation paths, and building or transmitter maps.The structural maps include building segmentation, building heights, and transmitter locations.Source: B. Visualization of the Dataset · p. 7, Figure 3 · p. 7
  • Above surrounding building heights, structures may visually disappear from radio maps because the receiver has clear line of sight to the transmitter.With obstructions removed, attenuation or reflection changes and spatial signal patterns are significantly altered.Source: B. Visualization of the Dataset · p. 7

C. Dataset Naming Conventions

  • Files are grouped by modality and height; the filename 103_62X_125Y.png encodes building ID 103 and transmitter coordinates X=62, Y=125.That image shows the 1 m height layer and path-loss distribution around the transmitter and nearby urban structures.Source: C. Dataset Naming Conventions · pp. 7–8, Figure 4 · p. 7

D. Data Normalization and Heatmap Generation

  • Each channel parameter is normalized separately using fixed global minimum and maximum thresholds, with normalized values optionally mapped to [0, 255].The channels include pathloss, ToA, DoA azimuth, and DoA elevation.Source: D. Data Normalization and Heatmap… · p. 8

IV. DIFFUSION-BASED 3D RM GENERATION

The section frames 3D radio-map construction as multimodal volumetric prediction and presents RadioDiff-3D as a conditional diffusion model for generating these maps.

A. Problem Formulation of 3D RM Construction

  • The radio map is a 4D tensor over three spatial dimensions and channel modalities: pathloss, DoA azimuth and elevation, and ToA, with building occupancy represented separately.The formulation uses a 3D occupancy grid to describe environmental obstructions.Source: A. Problem Formulation of 3D RM… · p. 8

1) RM Construction with Known Base Station and Environmental Information:

  • With environmental geometry and base-station configuration known, the task is to predict the full radio-map tensor from these inputs.The configuration includes transmitter location, elevation, and radiation characteristics.Source: 1) RM Construction with Known Base… · p. 8

2) RM Construction with Sparse Sampling and Environmental Information:

  • When the transmitter is unknown or inaccessible, sparse 3D channel observations and environmental priors are used to reconstruct the complete radio-map tensor.The observations may include pathloss, DoA, or ToA; the setting is relevant to spectrum sensing, interference detection, and coverage estimation.Source: 2) RM Construction with Sparse… · p. 8

B. Diffusion Model with 3D Operator

1) Latent Diffusion Modeling:

The method learns to denoise a progressively perturbed latent representation, then conditions generation on either known transmitter information or sparse observations. The two settings support full-volume map generation from different available inputs.

1) Latent Diffusion Modeling:

  • The latent diffusion process progressively adds Gaussian noise, then trains a network to predict that noise by minimizing expected denoising loss.The process is described over timesteps with conditioning variables supplied to the denoising network.Source: 1) Latent Diffusion Modeling · p. 8

2) Conditional Generation for Two Application Settings:

  • With the environment and base-station configuration known, conditioning is injected into denoising U-Net residual blocks to generate radio maps aligned with propagation laws and urban topology.Figure 5 visualizes pathloss, ToA spread, and DoA distributions at heights from 3 m to 15 m.Source: 2) Conditional Generation for Two… · p. 9, Figure 5 · p. 9
  • When the base station is unknown, the model conditions on the environment and sparse pathloss, DoA, and ToA samples, using reconstruction-guided sampling to cohere generated and observed regions.The guidance uses an interpolated sparse map and a time-dependent weight; Figure 6 shows the channel maps on a common voxel grid.Source: 2) Conditional Generation for Two… · pp. 9–10, Figure 6 · p. 10

3) 3D U-Net with Conditioning:

RadioDiff-3D uses a conditioned 3D U-Net and autoregressive height-wise generation. The supplied experimental passages describe a preliminary feasibility check, not full-scale outcome values.

3) 3D U-Net with Conditioning:

  • The 3D convolutional U-Net takes a noisy latent tensor and predicts noise or a denoised latent, integrating conditioning through cross-attention and using spectral refinement for high-frequency multipath details.Skip connections bridge encoder and decoder layers.Source: 3) 3D U-Net with Conditioning · p. 10

4) Autoregressive Height-wise Generation:

  • Autoregressive generation conditions each new height slice on the preceding slice, supporting vertical continuity and inference across varying building heights and altitudes.The approach is intended to avoid high memory or compute cost for large 3D tensors.Source: 4) Autoregressive Height-wise Generation · p. 10

V. EXPERIMENTAL PROCESS

  • The experiments use standardized settings and control data preprocessing and model training to support reproducibility and comparability.The passage introduces the subsequent experimental procedures and configurations.Source: §V · p. 10

A. Preliminary Verification Experiment

  • The preliminary feasibility check trained on 10% of the training set and evaluated 100 test samples without sampling, tracking RMSE, NMSE, SSIM, and PSNR.Figure 8 presents generated and compared frames for this preliminary stage.Source: A. Preliminary Verification Experiment · p. 10

B. Dataset and Input Preparation

The experiments prepare low-height pathloss volumes with environmental conditioning, split samples randomly at file level, and also test sparse sampling strategies.

B. Dataset and Input Preparation

  • The experiments use pathloss maps from 1–4 m heights as output and condition the model on building segmentation, building height, and transmitter-location maps.For sampling experiments, a sampling map is added to the three environmental feature maps.Source: B. Dataset and Input Preparation · pp. 10–11
  • The dataset is randomly split by file, with 90% assigned to training and 10% to testing; the subsets have no spatial or semantic overlap.Each sample has unique spatial coordinates and beamforming configuration.Source: B. Dataset and Input Preparation · p. 11

C. Sampling Strategies

  • Uniform and random sampling each select 10% of radio-map data points, approximately 6,553 voxels per height level for a 256×256 slice.The stated sampling density balances computational tractability and generalization capability.Source: C. Sampling Strategies · p. 11

1) Uniform Sampling:

The section describes two ways to create sparse frame tensors by retaining sampled pixels and zeroing the rest, then reports the model’s training configuration.

1) Uniform Sampling:

  • Uniform sampling retains a fixed number of uniformly selected pixels from each frame in a sparse tensor, with unsampled pixels set to zero.The input tensor has height, width, frame, and channel dimensions.Source: 1) Uniform Sampling · pp. 11–12

2) Random Sampling:

  • Random sampling forms a sparse tensor by selecting a random subset of pixels from each frame.The selected pixel indices are represented by Irandom.Source: 2) Random Sampling · p. 12

D. Experimental Parameter Configuration

  • Training used batch size 2, learning rate 5e-5, EMA decay 0.995, AMP, 10-iteration EMA updates, L1 loss, and an NVIDIA 4090.The authors state these settings were chosen to balance training speed and stability.Source: D. Experimental Parameter Configuration · p. 12

E. Performance Metrics

The section evaluates radio-map reconstruction with image-quality metrics and describes how MSE-related and SSIM measures assess prediction accuracy and structure.

E. Performance Metrics

  • RMSE, NMSE, SSIM, and PSNR are used to assess reconstruction quality, with qualitative comparisons across altitudes shown in Fig. 9.The metrics are summarized in Table V, whose values are not included in the supplied passages.Source: E. Performance Metrics · p. 12, Figure 9 · p. 13

1) MSE:

  • MSE averages squared pixel-intensity differences; NMSE scales MSE, and RMSE is the square root of MSE.The stated definitions concern comparison between the ground-truth and predicted images.Source: 1) MSE · pp. 12–13

2) SSIM:

  • SSIM evaluates brightness, contrast, and structural information, reflecting the paper’s emphasis on signal brightness, surrounding contrast, and map accuracy.The section motivates SSIM for radio maps with high-frequency texture details.Source: 2) SSIM · p. 13

3) PSNR:

The section reports inference latency across DDIM sampling steps, defines PSNR as a reconstruction-quality measure, and outlines the 3D-UNet design.

3) PSNR:

  • PSNR measures the ratio of maximum signal power to interfering noise power, is expressed in dB, and generally indicates better image quality at higher values.For radio maps, it also assesses edge-detail quality.Source: 3) PSNR · p. 13

F. Inference Time

  • 2.43 seconds at 20 DDIM steps and 24.35 seconds at 200 steps show inference time rising approximately linearly with sampling steps.The paper identifies a trade-off between reconstruction quality and computational cost for real-time or resource-constrained deployment.Source: F. Inference Time · pp. 13–14, Table VII · p. 14

G. 3D-UNet Architecture Overview

  • 3D-UNet replaces 2D convolutions and pooling with 3D operations to capture volumetric spatial context, while its encoder-decoder and skip connections integrate high- and low-level features.The design is used for dense prediction over three-dimensional space.Source: G. 3D-UNet Architecture Overview · p. 14

H. Application of 3D-UNet in Our Experiment

The conclusion presents UrbanRadio3D and evaluates 3D-UNet across channel characteristics on four altitude layers, with the stated goal of supporting future 3D radio-map research.

H. Application of 3D-UNet in Our Experiment

  • The 3D-UNet input stacks building segmentation, building height, and transmitter-location maps; it predicts ToA, DoA azimuth, and DoA elevation at four altitude levels.The four levels are 1 m, 2 m, 3 m, and 4 m; each output has the same spatial and altitude resolution as the input.Source: H. Application of 3D-UNet in Our… · pp. 14–15

I. 3D-UNet Performance Metrics

  • The 3D-UNet was evaluated on 1,000 test samples disjoint from training using RMSE, NMSE, SSIM, and PSNR, with a visual comparison in Fig. 10.The supplied passage identifies Table VIII as the summary of results for ToA, DoA azimuth, and DoA elevation, but gives no metric values.Source: §I · p. 15, Table VIII · p. 15

VI. CONCLUSION

  • UrbanRadio3D and the RadioDiff-3D and 3D-UNet benchmarks were evaluated across pathloss, DoA, and ToA, in transmitter-aware and transmitter-agnostic settings.The authors describe the results as demonstrating the viability of data-driven 3D radio-map generation and providing baselines for future work.Source: §VI · p. 15
Loading 2507.12166v1…