3D Radio Environment Map (REM)

• 3D REM is a high-dimensional function mapping 3D coordinates to radio field metrics such as path-loss, delay, and angular information for precise network modeling.
• It integrates real-world measurements with environmental priors and employs statistical, physics-informed, and deep learning methods for robust construction.
• 3D REMs enhance wireless network operations by enabling improved resource allocation, mobility management, and coverage forecasting in complex urban and aerial scenarios.

A 3D Radio Environment Map (REM) is a structured, high-dimensional representation that quantifies radio field characteristics—including received signal strength, path-loss, delay, and angular information—at every point in a three-dimensional spatial domain. As wireless systems evolve to support complex scenarios such as UAV connectivity, 6G vehicular networks, and environment-aware spectrum management, volumetric REMs have become a critical tool for coverage forecasting, interference analysis, and resource allocation. Unlike traditional 2D REMs, 3D REMs address the challenges posed by vertical spatial variation, non-uniform antenna patterns, building morphology, and the mobility of aerial or multi-story users.

1. Mathematical Representation and Physical Modeling

At the core, a 3D REM is formalized as a continuous or discretized function that maps spatial coordinates to channel observables. The canonical form is:

$$R : \mathbb{R}^3 \rightarrow \mathbb{R}^K,\quad R(\mathbf{p}) = \{L(\mathbf{p}), \theta(\mathbf{p}), \tau(\mathbf{p}), \cdots\}$$

where $L(\mathbf{p})$ is path-loss or RSRP, $\theta(\mathbf{p})$ encapsulates geometric quantities such as direction of arrival, and $\tau(\mathbf{p})$ is time-of-arrival for dominant multipath. Typically, the domain is sampled onto a regular voxel grid, rendering $R$ as a tensor $$\mathcal{R} \in \mathbb{R}^{H \times W \times D \times K}$$ (Wang et al., 16 Jul 2025). Models incorporate deterministic propagation (e.g., FSPL, log-distance, shadowing, diffraction, and scattering), site-specific antenna patterns, and channel stochasticity (zero-mean Gaussian shadowing with spatial correlation) (Reddy et al., 23 Jan 2026, Chen et al., 2024).

2. Data Acquisition and Preprocessing

3D REM construction is predicated on a combination of real-world measurements and a priori environmental knowledge. Measurement campaigns utilize RTK-equipped UAVs for outdoor scenarios (Chen et al., 2024), drone swarms for indoor mapping (Mendes, 2021), or multimodal sensor rigs integrating LiDAR and signal probes for scene geometry and RF data (Milosheski et al., 1 Nov 2025). Each sample associates a 3D coordinate with the observed metric (e.g., RSRP, RSSI), optionally annotated with transmitter/receiver position, antenna configuration, and environment labels.

Preprocessing entails (i) geometric coordinate transformation (e.g., (x, y, z) to (ρ, φ, θ)), (ii) binning into angular or cubic spatial cells, (iii) filtering out low-SNR/noise entries (thresholding RSRP/RSSI), and (iv) integrating environmental priors (building masks, heights, LoS metrics, materials) (Reddy et al., 23 Jan 2026, Sallouha et al., 2023). In indoor mapping, registration and alignment of point clouds and measurement grids are crucial for spatial consistency (Milosheski et al., 1 Nov 2025).

3. Modeling Approaches: Statistical, Physics-Inspired, and Learning-Based

Three principal paradigms dominate current REM estimation methodology:

3.1 Statistical Interpolation and Kriging

Classical Kriging and Gaussian Process Regression (GPR) interpolate sparse measurements based on spatial covariance, leveraging kernel functions (e.g., Matérn) and often incorporating deterministic path-loss as a mean function. In high-dimensional REMs, GPR is favored for its uncertainty quantification but is challenged by cubic complexity in sample volume and over-smoothing in regions with abrupt transitions (e.g., at building edges) (Chen et al., 2024, Reddy et al., 23 Jan 2026).

3.2 Physics-Informed and Hybrid Models

Physics-inspired models embed environment semantics into the REM inversion process. These include virtual obstacle grids—parameterizations of scene blockage as cellwise "virtual heights"—and multi-path-aware propagation decompositions (LoS/NLoS, diffraction, scattering), supporting environment-aware map construction and inference based on geometric relationships between TX/RX paths and obstacles (Liu et al., 2021, Chen et al., 2024). Simultaneous recovery of the radio map and environment (joint channel and obstacle learning) is achieved by alternating blockwise optimization and regression.

3.3 Deep Learning and Generative Models

Recent advances deploy U-Net and Transformer architectures, as well as generative diffusion models (DDPMs), in the REM space. Fully-convolutional 3D U-Nets map multi-channel inputs (building height, LoS, TX map) to REM target volumes, exploiting skip connections and data augmentation for runtime efficiency and improved generalization (Sallouha et al., 2023, Wang et al., 16 Jul 2025). Transformer-based approaches, such as TransfoREM, recast REM interpolation as a masked sequence translation task—mapping spherical coordinates to RSRP sequences—enabling efficient pretraining and fine-tuning strategies (Reddy et al., 23 Jan 2026).

Generative models, notably conditional diffusion and 3D U-Net-based DDPMs, enable high-fidelity REM synthesis even with limited ground-truth data, outperforming VAEs, GANs, and Kriging in MSE and structural similarity (Cao et al., 27 Dec 2025, Wang et al., 16 Jul 2025). Unified radiation field representations, as instantiated in URF-GS, couple optical and wireless inverse rendering on 3D Gaussian primitives, bridging geometry, material, and multipath physics while achieving superior sample efficiency and generalization (Wen et al., 27 Jan 2026).

4. Sparse Sampling, Computational Complexity, and Performance

A major challenge is the high acquisition and computational cost of dense 3D REMs. Accordingly, optimized sampling strategies—offline clustering, online uncertainty-based MAP selection, and E-optimality designs—are employed to minimize the number of UAV sorties or measurement points for a target accuracy. Hierarchical Bayesian learning further exploits structure in the propagation field, yielding state-of-the-art reconstruction with low sampling rates (Jie et al., 2024, Chen et al., 2024).

Scale is addressed by dimensionality reduction (PCA on path-loss matrices), inducing-point GPs, or generative paradigms tractable under GPU acceleration (e.g., batched self-attention, DDIM samplers). Performance benchmarks consistently report RMSE in the 1–5 dB range for advanced methods, with deep learning and diffusion outperforming classical interpolation at fractional measurement budgets (Reddy et al., 23 Jan 2026, Chen et al., 2024, Wang et al., 16 Jul 2025, Cao et al., 27 Dec 2025).

5. Integration with Wireless Network Operations

3D REMs serve as the backbone for numerous network primitives, including:

• Resource Scheduling: Enhanced base station resource allocation, dynamic spectrum sharing, and spatial bin utilization (Reddy et al., 23 Jan 2026).
• Mobility Management: Proactive handover thresholding, 3D beam-tilt optimization, UAV/DRONE corridor planning (Reddy et al., 23 Jan 2026, Liu et al., 2021, Wang et al., 16 Jul 2025).
• Energy Optimization: Efficient selection of cell-free APs for user service via REM-guided on/off switching strategies (MPL-ASO) (Sallouha et al., 2023).
• Relay Placement: UAV-aided relay deployment with REM-informed site selection yielding substantial throughput gains over traditional heuristics (Liu et al., 2021, Chen et al., 2024).
• Robustness to Environmental Dynamics: Multi-modal REMs incorporating LiDAR, 3D visual inputs, and human presence modulation (Milosheski et al., 1 Nov 2025, Wen et al., 27 Jan 2026).

6. Evaluations, Benchmarks, and Limitations

Quantitative evaluation relies on test-set RMSE, MAE, SSIM, and coverage probability metrics, validated on large-scale real-world datasets (e.g., UrbanRadio3D, AERPAW, custom indoor 3D LiDAR/WiFi scans). Notable results include <1.5 dB RMSE on fine-tuned transformer and UNet-diffusion models and 2–3× error reduction relative to IDW/KNN/Kriging baselines, even at sub-5% sampling density (Reddy et al., 23 Jan 2026, Chen et al., 2024, Wang et al., 16 Jul 2025).

Model limitations span computational cost (especially for GPR/kriging and high-resolution generative models), sensitivity to environmental dynamics (static scene assumptions common), reduced accuracy in severe multipath regimes, and the need for improved uncertainty quantification in generative prediction outside the support of training data. Extensions under active exploration include spatio-temporal REMs, frequency/band-aware mapping, dynamic scene adaptation, physics-regularized generative models, and transformer-based 3D architectures with attention locality control (Reddy et al., 23 Jan 2026, Wang et al., 16 Jul 2025, Wen et al., 27 Jan 2026).

---

References:

• "TransfoREM: Transformer aided 3D Radio Environment Mapping" (Reddy et al., 23 Jan 2026)
• "Bridging Visual and Wireless Sensing: A Unified Radiation Field for 3D Radio Map Construction" (Wen et al., 27 Jan 2026)
• "High-Efficiency Urban 3D Radio Map Estimation Based on Sparse Measurements" (Chen et al., 2024)
• "RadioDiff-3D: A 3D$\times$3D Radio Map Dataset and Generative Diffusion Based Benchmark for 6G Environment-Aware Communication" (Wang et al., 16 Jul 2025)
• "Sparse Bayesian Learning-Based Hierarchical Construction for 3D Radio Environment Maps Incorporating Channel Shadowing" (Jie et al., 2024)
• "Diffraction and Scattering Aware Radio Map and Environment Reconstruction using Geometry Model-Assisted Deep Learning" (Chen et al., 2024)
• "REM-U-net: Deep Learning Based Agile REM Prediction with Energy-Efficient Cell-Free Use Case" (Sallouha et al., 2023)
• "A Multimodal Dataset for Indoor Radio Mapping with 3D Point Clouds and RSSI" (Milosheski et al., 1 Nov 2025)
• "UAV-aided Radio Map Construction Exploiting Environment Semantics" (Liu et al., 2021)
• "Research Project 2: Drone-supported AI-based Generation of 3D Maps of Indoor Radio Environments" (Mendes, 2021)
• "A Lightweight Coordinate-Conditioned Diffusion Approach for 6G C-V2X Radio Environment Maps" (Cao et al., 27 Dec 2025)