Environment-Aware Channel Modeling

• Environment-aware channel modeling is a site-specific method that represents channel behavior as a function of physical environment, node location, and context rather than coarse statistics.
• It integrates explicit geometry, map-based priors, and multimodal generative inference to construct accurate, scalable channel models that overcome limitations of stochastic approaches.
• Applications span UAV networks, urban intersections, and THz sensing, improving channel estimation, interference management, and real-time CSI acquisition.

Environment-aware channel modeling framework denotes a site-specific methodology in which channel behavior is represented as a function of physical environment, node location, and, when needed, time and frequency, rather than only through coarse scenario labels or distance-based statistics. In this view, the channel may be expressed as a mapping such as $$\mathcal{W}: \mathbf{q}\in\mathcal{Q}\mapsto\mathbf{r}$$, a location-conditioned prior $$\mathcal{M}:\mathbf{q}\rightarrow P_{h|q}(\mathbf{h}\mid\mathbf{q})$$, or a broader environment-to-channel relation $\mathbf{H}=\mathcal{F}(\mathcal{E},\mathbf{s},t)$, depending on whether the framework stores channel knowledge, a conditional distribution, or a direct inference model from multimodal observations (Zeng et al., 2020, Qiu et al., 8 Jul 2025, He et al., 14 Jan 2026). Across recent work, the common objective is to exploit explicit environmental structure—geometry, materials, land cover, clutter, blockage, or sensed scene context—to produce channel descriptions that are site-specific, mobility-aware, and more informative than purely stochastic abstractions (Vinogradov et al., 19 Nov 2025, Arun et al., 14 May 2026).

1. Conceptual foundations and recurrent representations

The central premise is that the propagation channel is not treated as a realization of a generic scenario model alone, but as a location-indexed or environment-conditioned object. In the CKM formulation, a site-specific database stores “whatever channel-related information useful to enhance environmental-awareness and facilitate or even obviate sophisticated real-time CSI acquisition,” and may take B2X or X2X forms, optionally indexed by time $t$ and frequency $f$ (Zeng et al., 2020). In the CGM specialization, the stored quantity is a scalar large-scale channel gain in dB, commonly written as $r=\beta+10\alpha\log_{10}\|\mathbf{q}\|+S$, whereas other CKM variants store path-level or angle-delay knowledge (Li et al., 2021).

Three recurrent representational choices appear in the literature. One is explicit geometry, where buildings, terrain, and surfaces are modeled directly and LOS/NLOS or reflection structure is derived from that geometry. A second is map-based prior modeling, where historical measurements or ray-tracing outputs are compressed into parametric or learned site-specific maps, such as CKMs, CADMs, CSFMs, or VOMs. A third is multimodal generative inference, where environmental observations such as images, LiDAR, height maps, and coordinates are mapped directly to channel realizations or channel fields by conditional generative models (Liang et al., 4 Dec 2025, Arun et al., 14 May 2026).

These formulations differ in what they store and what they predict, but they converge on the same design principle: environmental information is treated as a first-class input to channel construction. This suggests that “environment-aware” is best understood as a family of methods rather than a single model class. Some frameworks target quasi-stationary quantities such as path loss, shadowing, or LOS regions; others target complete MPC sets, near-field channels, or pilot-free CSI inference (Zeng et al., 2020, Guo et al., 5 Apr 2026).

2. Geometry-driven and semi-deterministic constructions

A prominent strand is geometry-driven large-scale modeling. In the semi-deterministic A2G framework of “Spatially Consistent Air-to-Ground Channel Modeling and Simulation via 3D Shadow Projections,” the large-scale loss at user location $\mathbf{r}_i$ is modeled as

$$\Lambda(\mathbf{r}_{i}) = \Lambda_0 + \Lambda_{\mathrm{ex}}(\mathbf{r}_{i}) + \xi(\mathbf{r}_{i}) \quad [\mathrm{dB}],$$

where $\Lambda_0$ is a deterministic reference FSPL term, $\Lambda_{\mathrm{ex}}$ is deterministic excess path loss selected by geometry-derived LOS/NLOS state, and $$\mathcal{M}:\mathbf{q}\rightarrow P_{h|q}(\mathbf{h}\mid\mathbf{q})$$0 is spatially correlated shadow fading (Vinogradov et al., 19 Nov 2025). The framework is built around three key pillars: explicit 3D geometry of the environment, deterministic LOS/NLOS and path loss tied to that geometry, and spatially correlated stochastic shadow fading. LOS regions are obtained by projecting each building roof vertex onto the ground plane, building per-building shadow polygons, and defining

$$\mathcal{M}:\mathbf{q}\rightarrow P_{h|q}(\mathbf{h}\mid\mathbf{q})$$1

so that LOS/NLOS transitions along a route are binary and entirely determined by geometry. The resulting LOS mapping is computed once per environment and ABS placement, reused across routes, and yields complexity dominated by $$\mathcal{M}:\mathbf{q}\rightarrow P_{h|q}(\mathbf{h}\mid\mathbf{q})$$2, rather than per-sample ray–building testing (Vinogradov et al., 19 Nov 2025).

A related but more statistical geometry-to-channel mapping appears in urban street-canyon intersections, where a composite environmental factor

$$\mathcal{M}:\mathbf{q}\rightarrow P_{h|q}(\mathbf{h}\mid\mathbf{q})$$3

compresses average building height, building height dispersion, and building density into a single scalar descriptor (Chen et al., 2 Apr 2026). That factor is then embedded into 3GPP-style LOS/NLOS path-loss formulas and linked to small-scale channel parameters such as multipath power, delay, angle spreads, and cluster counts. In the reported measurements, this environment-related path-loss model reduces RMSE by about 8 dB in LOS and about 3 dB in NLOS relative to the original 3GPP UMi model, while the small-scale parameterization yields DS, ASA, and ESA distributions that match those of a validation intersection (Chen et al., 2 Apr 2026).

At 300 GHz, monostatic THz sensing work makes the geometry–material link still more explicit by mapping reflector quantity, surface roughness, reflector geometry, and material type to cluster counts, cluster shapes, specular and diffuse power, reflection loss, and intra-cluster dispersion (Lyu et al., 2 Sep 2025). There, SAGE extracts MPC parameters, connected component labeling groups them in delay–angle space, and reflection-loss statistics are then tied to material classes such as metal, glass, cement, tile, and polymer. The framework therefore treats the channel not only as a function of geometry, but also as a function of physically interpretable surface properties (Lyu et al., 2 Sep 2025).

Taken together, these models replace probabilistic LOS rules or scenario-average coefficients with explicit environment-conditioned channel construction. They are especially effective when blockage geometry, path segmentation, or spatial continuity is essential, such as urban A2G links, street-canyon V2X, or THz sensing.

3. Channel Knowledge Maps and map construction

CKM-based frameworks treat channel knowledge as a location-indexed database or function. In its general form, a CKM maps location to a channel-related quantity useful for communications, sensing, or both, and may store large-scale gain, shadowing, path-level features, or complete path statistics (Zeng et al., 2020). The appeal of CKM is strongest for channels for yet-to-reach locations, channels for non-cooperative nodes, channels with large dimensions, and channels with severe hardware or processing limitations, because location and environment structure can replace part of the burden of instantaneous CSI acquisition (Zeng et al., 2020).

The EM-based CKM construction framework develops this idea into a parametric estimation problem. Given finite measurements $$\mathcal{M}:\mathbf{q}\rightarrow P_{h|q}(\mathbf{h}\mid\mathbf{q})$$4, it assumes a mixture of $$\mathcal{M}:\mathbf{q}\rightarrow P_{h|q}(\mathbf{h}\mid\mathbf{q})$$5 propagation regimes, each with parameters $$\mathcal{M}:\mathbf{q}\rightarrow P_{h|q}(\mathbf{h}\mid\mathbf{q})$$6, and introduces latent responsibilities

$$\mathcal{M}:\mathbf{q}\rightarrow P_{h|q}(\mathbf{h}\mid\mathbf{q})$$7

For CGMs, the component model is Gaussian in dB with parameters $$\mathcal{M}:\mathbf{q}\rightarrow P_{h|q}(\mathbf{h}\mid\mathbf{q})$$8, and the M-step reduces to weighted regression with closed-form updates (Li et al., 2021). In the numerical example, the true environment contains five propagation regimes—LoS, two blocked NLoS groups, and two indoor groups. With $$\mathcal{M}:\mathbf{q}\rightarrow P_{h|q}(\mathbf{h}\mid\mathbf{q})$$9 measurements and $\mathbf{H}=\mathcal{F}(\mathcal{E},\mathbf{s},t)$0, the reconstructed CGM captures the highly non-symmetric ground-truth structure, whereas single-model fitting with $\mathbf{H}=\mathcal{F}(\mathcal{E},\mathbf{s},t)$1 produces concentric contours and persistently high NRMSE (Li et al., 2021).

This line of work emphasizes compactness as well as prediction accuracy. Rather than storing a dense grid of channel values, the CKM stores a small number of parameter vectors, mixing coefficients, and optionally responsibilities. A plausible implication is that CKM frameworks are best understood as structured surrogates for radio propagation fields: they retain environment sensitivity while reducing storage and query-time cost. This is also why DNN-based CKMs were proposed early for high-dimensional X2X maps, where explicit tables can become impractical (Zeng et al., 2020).

CKM-based representations are not restricted to scalar gain. The same logic extends to angle-delay maps, score-function maps, path maps, and virtual-object maps. What changes is the definition of the channel knowledge vector, not the underlying principle that the map is site-specific, location-indexed, and designed to expose environment-conditioned priors.

4. Estimation, sensing, and environment-conditioned priors

Environment-aware frameworks increasingly use environmental priors directly inside channel estimation algorithms. In XL-MIMO, the CSFM framework defines a location-specific channel prior

$\mathbf{H}=\mathcal{F}(\mathcal{E},\mathbf{s},t)$2

discretizes space into grids, and stores for each grid a denoiser $\mathbf{H}=\mathcal{F}(\mathcal{E},\mathbf{s},t)$3 trained on historical channels (Qiu et al., 8 Jul 2025). The estimator solves a regularized MAP problem,

$\mathbf{H}=\mathcal{F}(\mathcal{E},\mathbf{s},t)$4

and then uses plug-and-play splitting plus Tweedie’s formula to implement the prior through the denoiser. For grid size $\mathbf{H}=\mathcal{F}(\mathcal{E},\mathbf{s},t)$5 m, single-step denoising yields average improvement of 7.84 dB over noisy channels, and CSFM-PnP outperforms LS/ML, LMMSE, and nearest-neighbor baselines under limited pilots, abundant pilots, and even $\mathbf{H}=\mathcal{F}(\mathcal{E},\mathbf{s},t)$6 pilot-free generation (Qiu et al., 8 Jul 2025).

A related but architecturally different approach appears in pilot-constrained upper mid-band MIMO, where an enhanced U-Net with cross-attention fuses LS channel estimates with EM-derived RSS maps (Javid et al., 8 Nov 2025). The framework is physics-informed through the loss

$\mathbf{H}=\mathcal{F}(\mathcal{E},\mathbf{s},t)$7

where $\mathbf{H}=\mathcal{F}(\mathcal{E},\mathbf{s},t)$8 enforces consistency between EM-based RSS and channel-derived power. In realistic urban ray-tracing data, this PINN achieves over 5 dB gain in NMSE compared to state-of-the-art methods, maintains practical computational complexity, and transfers across frequencies and environments with only minimal fine-tuning (Javid et al., 8 Nov 2025).

In sensing-oriented CKM reuse, CADM treats the target as a virtual UE and transforms communication angle-delay priors into sensing priors. For a location $\mathbf{H}=\mathcal{F}(\mathcal{E},\mathbf{s},t)$9, the sensing likelihood is constructed from Gaussian angle-delay distributions learned in the communication CKM, yielding a maximum-likelihood localization problem over $t$0 (Wu et al., 4 Jul 2025). The same map thus supports environment-aware communication and NLoS sensing, with CRLB and simulation results showing that the full $t$1 non-reciprocal path model outperforms conventional reciprocal-path formulations in pure NLoS conditions (Wu et al., 4 Jul 2025).

Environment-aware sensing also changes the channel model itself when the target modifies the environment. In a 105 GHz indoor factory scenario, the target–environment coupling model augments a 3GPP-style GBSM with the Blockage-Region Coupling Factor $t$2 and the Forward-Scattering Coupling Factor $t$3, so that blocked background paths are removed and new forward-scattered target paths are generated from them (Liu et al., 31 May 2025). The coupled model improves similarity indices in both LoS and NLoS cases, for example raising joint angle-delay similarity from 48.99% to 79.02% in the LoS case (Liu et al., 31 May 2025).

Near-field ISAC pushes this logic further. The VOM framework stores a library $t$4 of virtual environment objects, maps each UE location $t$5 to a dominant subset $t$6, and uses monostatic sensing plus QR/SVD-based clutter suppression to extract a dynamic subspace (Guo et al., 5 Apr 2026). The downlink channel is then approximated as

$t$7

which yields a joint VOM- and sensing-aided regularized LS estimator that outperforms schemes without VOM priors and/or dynamic sensing in both channel estimation accuracy and achievable rate (Guo et al., 5 Apr 2026).

5. Generative and multimodal environment-to-channel models

A newer direction replaces explicit per-link ray tracing at inference with conditional generative modeling. CITYMPC learns the complete per-link MPC parameter set from environmental observations using a conditional VAE (Arun et al., 14 May 2026). The input $t$8 combines TX and RX point-of-view stacks, a global height map, and TX/RX coordinates; the output is a complete channel realization with up to $t$9 paths, each represented by normalized complex gain, excess delay, AoD and AoA unit vectors, and a presence mask. Trained on 427,397 ray-traced links across five cities, CITYMPC achieves received power MAE as low as $f$0 dB and $f$1 MAE of $f$2 ns in Austin, while reproducing the empirical distributions of received power, delay, path count, and angle statistics (Arun et al., 14 May 2026). At the same time, cross-city transfer degrades markedly, which indicates that the learned distribution remains strongly site-dependent (Arun et al., 14 May 2026).

Cross-modal flow matching generalizes this idea to pilot-free CSI inference from camera images, LiDAR point clouds, and GPS coordinates (Liang et al., 4 Dec 2025). The model learns the conditional distribution $f$3 by transporting a multimodal latent distribution in channel space toward the angular-domain channel distribution with conditional flow matching and a modality-alignment loss. In the reported ablation, CFM+MA achieves NMSE $f$4 dB and cosine similarity $f$5, and runtime on RTX 4090 is about 1.25 ms for $f$6 integration steps and about 2.5 ms for $f$7, which is below the 10 ms frame duration used in the system model (Liang et al., 4 Dec 2025).

The AI-empowered channel inference paradigm casts the channel directly as $f$8, with multimodal environment acquisition, physics-aware feature fusion, AI-native or AI-hybrid inference, and transfer-learning and XAI modules (He et al., 14 Jan 2026). Reported results include average path loss prediction RMSE of $f$9 dB and training time reduction by 60%–75%. This suggests a broader shift from scenario-based parameter fitting toward direct environment-to-channel inference, especially when link-level and area-level predictions must coexist (He et al., 14 Jan 2026).

Generative diffusion has also been applied to large-scale A2G attenuation fields. In the scalable air-to-ground framework, DEM, slope, aspect, land cover embeddings, and satellite geometry condition a dual-head U-Net trained with v-parameterization and physics-informed inpainting (Tian et al., 24 Mar 2026). RT-derived obstruction and reflection profiles provide sparse anchors, while the diffusion model reconstructs complete channel loss maps. For a 256×256 tile, RT computation takes about 144,000 seconds and diffusion inference about 300 seconds; with 4% sampling in mountainous Region C, MAE drops to 1.23 dB (Tian et al., 24 Mar 2026). The same environment-aware physics layer yields high agreement with measured Starlink, OneWeb, and LTE dynamics, including sign agreement 0.734 for latency-aligned Starlink traces and Pearson correlations 0.906 and 0.932 for OneWeb and LTE respectively (Tian et al., 24 Mar 2026).

6. Applications, validation, and open problems

Environment-aware channel modeling frameworks support a wide range of tasks because they preserve spatial structure that purely stochastic models typically smooth away. In UAV-assisted and non-terrestrial A2G networks, they enable mobility-aware performance evaluation, user-route segmentation into LOS and NLOS regions, link planning, outage prediction, and radio-map generation (Vinogradov et al., 19 Nov 2025). In V2X street-canyon intersections, they tie building morphology to both large-scale loss and small-scale dispersion, supporting channel simulation and performance evaluation for collision warning and intersection movement assist scenarios (Chen et al., 2 Apr 2026). In ISAC, they allow the same CKM to serve communication and NLoS sensing, and they make target–environment coupling or material-aware THz reflection behavior explicit rather than treating them as residual clutter (Wu et al., 4 Jul 2025, Lyu et al., 2 Sep 2025, Liu et al., 31 May 2025).

Several recurring limitations appear across the literature. Geometry-driven A2G work assumes a regular Manhattan grid with axis-aligned square buildings, fixed carrier frequency $r=\beta+10\alpha\log_{10}\|\mathbf{q}\|+S$0 GHz, and a static ABS, even though the shadow-projection formulation itself is generic (Vinogradov et al., 19 Nov 2025). Intersection models based on a single scalar $r=\beta+10\alpha\log_{10}\|\mathbf{q}\|+S$1 do not fully capture corner asymmetry or all geometric details (Chen et al., 2 Apr 2026). CKM- and CSFM-based methods assume sufficiently accurate location information and a suitable grid size; too small a grid yields too little data for stable deep denoiser training, while too large a grid mixes heterogeneous channel statistics (Qiu et al., 8 Jul 2025). Near-field VOM methods remain narrowband, single-user, and quasi-static in their current formulation (Guo et al., 5 Apr 2026).

Learning-based frameworks introduce a different set of concerns. CITYMPC is trained and validated only on ray-traced data and shows strong cross-city distribution shift (Arun et al., 14 May 2026). Cross-modal flow depends on multimodal sensing quality and has not yet established generalization across different sensing pipelines or cities (Liang et al., 4 Dec 2025). Physics-informed RSS-assisted estimation demonstrates robustness with minimal fine-tuning, but still relies on realistic ray-tracing data rather than field measurements (Javid et al., 8 Nov 2025). This suggests that environment awareness is not synonymous with universal generalization; in many cases it means stronger site specificity, which improves local fidelity while increasing sensitivity to domain shift.

A common misconception is that environment-aware channel modeling is equivalent to full ray tracing. The surveyed work does not support that equivalence. Some frameworks explicitly position themselves between purely stochastic models and full RT through semi-deterministic geometry and correlated random fields; others use RT only offline to train surrogates or to construct priors; still others rely primarily on measurements and map learning (Vinogradov et al., 19 Nov 2025, Li et al., 2021, Arun et al., 14 May 2026). The broader trend is therefore not a replacement of every model by a single paradigm, but a redistribution of modeling effort: explicit environment knowledge is moved into geometry processing, map construction, sensing-derived priors, or conditional generative learning, depending on the intended operating regime.

Future directions identified in the literature include spatially and frequency-consistent small-scale fading on top of large-scale CKMs, GIS-based generalization beyond regular grids, validation against dedicated measurement campaigns, ML-based fast generation of CKMs, explicit UAV mobility, adaptive grid design, online CKM updating, multi-BS cooperative maps, score-based generative models, and richer multimodal environment representations (Vinogradov et al., 19 Nov 2025, Qiu et al., 8 Jul 2025, Guo et al., 5 Apr 2026, He et al., 14 Jan 2026). Taken together, these directions indicate that the field is moving toward channel models that are simultaneously site-specific, computationally scalable, and tightly coupled to environmental sensing or environmental databases, rather than anchored to a single universal stochastic law.