Data-and-Semantic Dual-Driven Spectrum Map Construction for 6G Spectrum Management

Fuhui Zhou; Jiayu Liu; Lu Yuan; Qihui Wu; Rui Ding; XiaoDong Liu

REVIEW 4 major objections 4 minor 16 references

Data-and-Semantic Dual-Driven Spectrum Map Construction for 6G Spectrum Management

T0 review · 4 major / 4 minor · reviewed 2026-08-10 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read A UNet that reads binary city maps reconstructs spectrum maps and infers unsampled frequencies.

desk verdict A plausible incremental UNet extension for multi-frequency spectrum map construction, but the missing simulator specification and thin empirical evidence make the reported gains hard to trust. read the letter →

arxiv 2501.12853 v1 pith:3VOFMUVE submitted 2025-01-22 cs.LG

Jiayu Liu , Fuhui Zhou , Xiaodong Liu , Rui Ding , Lu Yuan , Qihui Wu This is my paper · ORCID

classification cs.LG

keywords spectrummapmanagementdeeplearningUNetbinarycityfrequency-spacereasoningradioenvironment6G

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper proposes DSD-UNet, a method for building spectrum maps—grids showing radio signal strength across an urban area—from a few measurement points. It argues that adding two cheap pieces of semantic knowledge, a binary map of where buildings are and a binary map of where sensors sit, lets a UNet learn signal propagation better than raw measurements alone. It further claims that stacking the spatial maps for several frequencies into one 3D tensor lets the network infer the spectrum map at a frequency that was never sampled. In simulations, the method beats a semantic-free UNet and kriging by about 10.72% and 11.93% RMSE in single- and multi-transmitter scenarios, with roughly 1 dB and 1.43 dB lower error at low sampling density. If it holds in real environments, spectrum managers could monitor whole frequency bands with sparse, cheap sensor deployments.

What carries the argument

The central object is the joint frequency-space three-dimensional spectrum map model: an $N\times N\times(K+1)$ tensor formed by stacking the two-dimensional incomplete spectrum maps of $K$ sampled frequencies in ascending order, with the missing frequency's layer blank. This tensor is the input to DSD-UNet, a UNet with encoder-decoder convolutions and skip connections. The load-bearing mechanism is semantic mixing: binary city map and binary sampling location map, each also stacked along frequency, are concatenated with the spectral data so that the network's filters can associate spatial structure (building blockage, sensor placement) with received power. The third skip connection duplicates the raw incomplete spectrum map into the decoder path, preserving the original measurements. This machinery lets one network do two jobs at once: interpolate each sampled frequency's map and extrapolate the unsampled frequency's map from cross-frequency correlations.

What would settle it

On a real multi-band urban measurement campaign, or in a simulator with frequency-independent shadow fading, hold out one carrier frequency, train DSD-UNet on the rest, and compare its predicted map at the held-out frequency with a measured map; if the RMSE is no better than a semantic-free UNet or than using sampled frequencies alone, the claimed cross-frequency and semantic gains are falsified.

Watch

Extended reading notes

Core claim

DSD-UNet treats spectrum map construction as learning a mapping $f(\cdot)$ from an incomplete tensor $\mathcal{S}\in\mathbb{R}^{N\times N\times(K+1)}$—built by stacking the two-dimensional incomplete maps at $K$ sampled frequencies plus a blank layer for the missing frequency—to a complete tensor $\mathcal{E}$. The network is a UNet whose encoder receives three parallel inputs: $\mathcal{S}$, a binary city map $Z\in\mathbb{R}^{N\times N}$ (1 where buildings block propagation, 0 elsewhere), and a binary sampling location map $M\in\mathbb{R}^{N\times N}$ (1 where a receiver sits). These semantic maps are stacked along the frequency dimension and concatenated with the spectral data; a third skip connection re-injects the raw incomplete spectrum map into the decoder. Training minimizes the Frobenius norm $\|\mathcal{E}-\mathcal{P}\|_F^2$ against the ground-truth tensor $\mathcal{P}$. The paper claims this design lowers RMSE and speeds convergence relative to a semantic-free UNet and to kriging, and that it reconstructs the spectrum map at the unsampled frequency $f_0=1800$ MHz using data from 900, 1500, and 2100 MHz.

Load-bearing premise

The method assumes that the simulated radio environment, whose propagation model is never specified, has a learnable frequency-space correlation that also exists in real cities, and that a binary map of building footprints is enough to capture the geometry that shapes signal blockage and multipath.

Editorial extensions

If this is right

If the claims hold, a single DSD-UNet trained on $K$ frequencies can output complete maps for all $K+1$ frequencies, removing the need to deploy sensors tuned to every band.
At low sampling density (5% of grid cells), the semantic maps provide enough geometric prior to keep construction error roughly 1 dB below a data-only UNet and 1.43 dB below kriging, so sparse sensing becomes viable.
The binary city map and sampling map are cheap to store and compute, so the accuracy gain does not require expensive high-fidelity city models.
The network converges in about 5 epochs with semantics versus 10 without, meaning training cost drops for each new area.
Since any frequency in $F$ can be chosen as the target, the method is a template for full-band spectrum map construction rather than a fixed-frequency estimator.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

A testable extension is to check whether the binary city map still helps when real urban geometry includes terrain height, building materials, and vegetation, all of which the binary map discards.
The frequency-space reasoning suggests a transfer-learning recipe: train once on a city with dense sampling, then fine-tune on a new city with only sparse samples, with the semantic maps providing a shared geometric prior.
The claim implies that the simulator's frequency correlation is strong enough for cross-frequency inference; if real propagation is dominated by frequency-selective fading independent of location, the blank-frequency layer would be harder to fill, so the method's main test is real multi-band measurement data.
A natural extension is to add time as a fourth axis, allowing the network to predict future spectrum occupancy rather than only missing frequencies.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

The genuinely new thing here is stacking binary city and sampling-location maps as extra semantic channels in a 3D UNet over frequency, so the network can jointly complete spatial gaps and infer a missing frequency slice. That combination is not in RadioUNet or the authors' own prior frequency-spatial work, and the idea is sensible for reducing sensing cost. Credit where due: the ablation against a UNet without semantics is the right control, and the frequency-space framing is cleanly posed in Section III. If the reported 10–12% RMSE gains hold, that is a useful increment for the spectrum-map subfield.

The soft spots are real and mostly about evidence. The simulator's propagation model is never specified: no path-loss, shadowing, or multipath equation, and no statement of how the binary city map enters the label generation. If the ground truth is generated by attenuating power according to that same binary mask, then feeding the mask to the network is handing it the answer key. That is the load-bearing concern, and the paper does nothing to rule it out. The empirical support is also thin: one synthetic simulation, no error bars or multi-seed runs, no comparison to the most relevant baselines ([7] and [16]), and a single missing-frequency case (f0 = 1800 MHz) stretched into a claim about completing the whole band. There is also a small notational inconsistency where f0 is defined as the target after F is described as ascending from f0, then set to 1800 MHz in the simulation. None of this clearly breaks the central direction; the architecture and the semantic-input idea are plausible. But the current evidence does not establish that the gains transfer to real urban geometry, because the coupling between the semantic channels and the label generation is unknown.

Who gets value: readers working on radio map estimation or environment-aware spectrum management will find the frequency-space extension worth a look, especially the 3D stacking trick and the convergence-speed observation. The paper is short, it is basically a workshop/conference contribution, and it reads like an incremental result that could be made solid with code, a specified forward model, and one ablation over which frequency is missing. I would not cite it yet in its current form, but I would send a revised, reproducible version to a serious referee. On that basis: engage with it, ask for the simulator and the missing baselines, and treat the current quantitative claims as unverified until then.

Referee Report

4 major / 4 minor

Summary. The paper proposes DSD-UNet, a UNet-based method for urban spectrum map construction that augments incomplete spectrum maps with two additional inputs: a binary city map and a binary sampling location map, which the authors call semantic knowledge. The method stacks incomplete spectrum maps across four frequencies and trains the network to reconstruct the full frequency-space volume, including one frequency that is entirely absent from the measurements. Experiments on a synthetic 256 m by 256 m grid with 64 by 64 cells, with sampling densities of 5%, 20%, 35%, and 50%, compare DSD-UNet against a UNet without semantic inputs and against kriging. The paper reports roughly 10.72% and 11.93% RMSE reductions in single- and multi-transmitter scenarios, about 1 dB and 1.43 dB lower error at low sampling density, faster convergence, and successful inference of a missing frequency map.

Significance. The high-level idea of combining environment semantics with joint frequency-space reasoning is relevant and potentially useful for 6G spectrum management, where sampling is sparse and many frequency bands are unmonitored. The paper has a clear problem formulation, a sensible ablation baseline (UNet without semantics), and a straightforward architecture description. However, the significance is currently limited by the fact that all evidence comes from one unspecified synthetic simulator. If the propagation model is specified and the method is shown to work across multiple missing frequencies, independent simulator configurations, or real measurements, the contribution would be solid. As it stands, the central empirical claims are not yet supported beyond the particular, under-specified simulation setting.

major comments (4)

[Section IV-A] The simulation's forward model for the ground-truth spectrum maps is never specified. Section IV-A gives the area, grid size, frequency set F={900,1500,1800,2100} MHz, sampling densities, and states that transmitter parameters are randomly set, but it provides no equation for P(fk, xi,j), no path-loss, shadowing, or multipath model, and no description of how the binary city map Z enters the generation of the labels. This is load-bearing because DSD-UNet receives Z as an input channel; if the simulator computes received power by attenuating according to exactly the same binary mask, the network is handed the answer key. The reported 10.72% and 11.93% RMSE improvements and the faster convergence could then be artifacts of this coupling rather than evidence that semantic knowledge transfers to real urban geometry. Please specify the complete propagation model, state explicitly how Z affects the generated labels, and ideally validate on an independent simulator or on real-world measurement data to rule out the shortcut.
[Section IV-A and Section III-C] The claim of constructing the 'complete frequency band' or 'entire frequency set' is supported by only one missing frequency. The experiment uses F={900,1500,1800,2100} MHz and sets f0=1800 MHz as the only target inference frequency. A single missing-frequency test with three sampled frequencies does not demonstrate general complete-band construction, particularly because the four discrete frequencies are not a continuous band. Please vary the target frequency across all elements of F, report per-frequency results, and discuss how the method scales to a denser frequency grid; otherwise the central missing-frequency inference claim is under-supported.
[Section IV-B, Figures 5 and 6] The empirical comparison lacks error bars and the nearest frequency-spatial baseline. The ablation against a UNet without semantics is appropriate, but the claim that DSD-UNet outperforms 'benchmark schemes' would be much stronger with a comparison to a city-map-assisted method such as RadioUNet [13] or the authors' earlier frequency-spatial reasoning approach in [7] and [16]. In addition, the reported RMSE and dB improvements appear to be single-run numbers with no standard deviations across random seeds, transmitter configurations, or train/test splits, so the reader cannot assess statistical significance. Please report mean and standard deviation over at least several seeds and add the missing baseline.
[Section IV-B, Figure 6] The faster-convergence claim is based on training loss curves, not validation or test loss. Figure 6 shows that the semantic-augmented network reaches low training loss in 5 epochs versus 10 for the baseline, but this does not establish faster generalization unless the test RMSE or validation loss at those epochs is reported. Please show convergence on held-out data or clarify that the claim concerns training loss only.

minor comments (4)

[Section II, Eq. (2)] The minimization in Eq. (2) is written as min over the grid indices, but the optimization variable should be the network parameters; please write min_θ or define θ explicitly.
[Section III-B, Eqs. (3) and (4)] There are typos in the text around Eqs. (3) and (4): 'gird' should be 'grid', and the sentences beginning 'If there exists an obstacle...' and 'If the grid point I contains...' are missing closing conditions. Also, the phrase 'the set of the measured frequency band monitored by the sampling receiver' in Section II is awkward and should be rephrased.
[Section II] The problem formulation is ambiguous about whether the same transmitter locations and powers are used for every frequency in F. Please clarify whether each frequency has its own propagation scenario or whether the network learns to exploit a shared spatial layout across frequencies.
[Section III-B] Calling the binary sampling location map 'semantic knowledge' is potentially misleading: it is a mask derived directly from the measured data, not a semantic description of the environment. Please clarify the distinction between environmental semantics (city map) and data availability information (sampling mask).

Circularity Check

0 steps flagged · score 2.0 of 10

No circular derivation found: the claims rest on supervised held-out evaluation and an architecture ablation; self-citations are background only.

full rationale

The central empirical claim (Section IV-B) is supported by training on 20,000 synthetic samples and testing on 1,000 held-out samples (Section IV-A), minimizing the squared error against ground-truth maps via Eq. (2). The reported accuracy gains are computed against a UNet without semantic inputs and against kriging, so they are the outcome of a controlled comparison, not a fitted parameter renamed as a prediction. The joint frequency-space model in Section III-A is simply the stacking of incomplete maps along the frequency axis; no equation in the paper defines the target spectrum map P in terms of the model's inputs Z and M, so there is no self-definitional reduction. The self-citations to prior frequency-space reasoning ([7],[16]) are used for background and for the Frobenius-norm notation, not as a load-bearing uniqueness theorem; removing them would not alter the architecture or the training objective. The unspecified simulator's forward model is a legitimate external-validity and reproducibility concern, but it is not circularity: even under a hypothetical coupling between the binary city map and label generation, the network must still generalize to held-out labels from sparse samples, and the semantics-free UNet baseline controls for the architecture. Accordingly, no Eq. X = Eq. Y by construction or fitted-input-as-prediction step can be exhibited, and the paper is scored 2 only for minor, non-load-bearing self-citation.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The central claim rests on the simulator's realism, the sufficiency of binary maps as semantic priors, and the presence of learnable frequency-space correlation. None of these are validated against real measurements or released artifacts.

free parameters (4)

Learning rate = 0.0003
Chosen training hyperparameter; no sensitivity analysis reported (Section IV-A).
Training epochs = 30
Training length; the claimed 5-epoch convergence result depends on this budget (Section IV-B).
Batch size = 4
Chosen training hyperparameter (Section IV-A).
Target missing frequency = 1800 MHz
The simulation sets f0=1800 MHz as the frequency to infer; the result may depend on this choice, and no ablation over other missing frequencies is reported (Section IV-A).

assumptions (4)

domain assumption Signal strength is approximately constant within each 4 m grid cell
Section II states 'the signal strength within the same grid exhibits only fluctuates slightly. Hence... remains constant.' This discretization is standard but limits fidelity.
domain assumption Binary city maps and binary sampling maps contain the semantic information needed to infer propagation
Section III-B argues buildings and sensor locations shape received signal strength; the binary maps are the only environment information provided to the network.
domain assumption Spectrum maps at different frequencies are correlated in a learnable way
Section III-A builds a 3D frequency-space model on this assumption; no real multi-frequency dataset validates it.
ad hoc to paper The simulated propagation environment represents real urban deployments
Section IV describes random transmitter and sensor placement but omits the channel and building model, so generalization to real electromagnetic environments is assumed.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Data-and-Semantic Dual-Driven Spectrum Map Construction for 6G Spectrum Management." pith.science (2026). https://pith.science/paper/3VOFMUVE

@misc{pith2026250112853,
  author       = {Pith},
  title        = {Pith review of: Data-and-Semantic Dual-Driven Spectrum Map Construction for 6G Spectrum Management},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/3VOFMUVE}},
  note         = {Machine review of arXiv:2501.12853}
}

read the original abstract

Spectrum maps reflect the utilization and distribution of spectrum resources in the electromagnetic environment, serving as an effective approach to support spectrum management. However, the construction of spectrum maps in urban environments is challenging because of high-density connection and complex terrain. Moreover, the existing spectrum map construction methods are typically applied to a fixed frequency, which cannot cover the entire frequency band. To address the aforementioned challenges, a UNet-based data-and-semantic dual-driven method is proposed by introducing the semantic knowledge of binary city maps and binary sampling location maps to enhance the accuracy of spectrum map construction in complex urban environments with dense communications. Moreover, a joint frequency-space reasoning model is exploited to capture the correlation of spectrum data in terms of space and frequency, enabling the realization of complete spectrum map construction without sampling all frequencies of spectrum data. The simulation results demonstrate that the proposed method can infer the spectrum utilization status of missing frequencies and improve the completeness of the spectrum map construction. Furthermore, the accuracy of spectrum map construction achieved by the proposed data-and-semantic dual-driven method outperforms the benchmark schemes, especially in scenarios with low sampling density.

Figures

Figures reproduced from arXiv: 2501.12853 by the authors.

**Figure 2.** Illustration of the semantic knowledge in the proposed [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

**Figure 3.** The scheme of our proposed DSD-UNet. a value of 1. Conversely, if no obstacle hinders the grid point I, Zi,j is assigned a value of 0. Based on this setting, a binary city map can be generated as, Zi,j = 1, if gird I exists obstacles 0, else, (3) where Zi,j represents the value of binary city map Z at the grid point I = (i, j). Similarly, a binary sampling location map is defined as a two-dimensional tensor M. If … view at source ↗

Figures from the paper (2 more)

**Figure 4.** Figure 4: The visualization results of the estimated spectrum maps under different methods. [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]

**Figure 5.** Figure 5: Performance comparison under different sampling [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

16 extracted references · 16 canonical work pages

[7]

Accurate spectrum map construction for spectrum management through intelligent frequency-spatial reasoning,

F. Zhou, C. Wang, G. Wu, Y . Wu, Q. Wu, and N. Al-Dhahir, “Accurate spectrum map construction for spectrum management through intelligent frequency-spatial reasoning,” IEEE Trans. Commun , vol.71, no.7, pp. 3932-3945, 2023

work page 2023
[16]

Ac- curate spectrum map construction using an intelligent frequencyspatial reasoning approach,

C. Wang, Y . Wu, F. Zhou, Q. Wu, C. Dong, and K. K. Wong, “Ac- curate spectrum map construction using an intelligent frequencyspatial reasoning approach,” in Proc. IEEE GLOBECOM, 2022, pp. 3460-3465

work page 2022
[13]

RadioUNet: Fast radio map estimation with convolutional neural networks,

R. Levie, C ¸ . Yapar, G. Kutyniok, and G. Caire, “RadioUNet: Fast radio map estimation with convolutional neural networks,” IEEE Trans. Wireless Commun., vol. 20, no. 6, pp. 4001-4015, 2021

work page 2021
[1]

On the road to 6G: Visions, requirements, key technologies, and testbeds,

C.-X. Wang, X. You, X. Gao, X. Zhu, Z. Li, C. Zhang, H. Wang, Y . Huang, Y . Chen, H. Haas, J. S. Thompson, E. G. Larsson, M. D. Renzo, W. Tong, P. Zhu, X. Shen, H. V . Po, and L. Hanzo, “On the road to 6G: Visions, requirements, key technologies, and testbeds,” IEEE Commun. Surveys. Tuts., vol. 25, no. 2, pp. 905-974, 2023

work page 2023
[2]

A tutorial on environment-aware communications via channel knowledge map for 6G,

Y . Zeng, J. Chen, J. Xu, D. Wu, X. Xu, S. Jin, X. Gao, D. Gesbert, S. Cui, and R.Zhang, “A tutorial on environment-aware communications via channel knowledge map for 6G,” IEEE Commun. Surveys. Tuts. , early access, 2024

work page 2024
[3]

Deep stacked autoencoder-based long-term spectrum prediction using real- world data,

G. Pan, Q. Wu, G. Ding, W. Wang, J. Li, F. Xu, and B.Zhou, “Deep stacked autoencoder-based long-term spectrum prediction using real- world data,” IEEE Trans. Cogn. Commun. Netw. , vol. 9, no. 3, pp. 534- 548, 2023. Fig. 6: Comparison of training loss under different sampling densities

work page 2023
[4]

Compressed wideband spectrum mapping in 3D spectrum-heterogeneous environment,

F. Shen, G. Ding, Q. Wu, and Z. Wang, “Compressed wideband spectrum mapping in 3D spectrum-heterogeneous environment,” IEEE Trans. V eh. Technol., vol. 72, no. 4, pp. 4875-4886, 2023

work page 2023
[5]

Deep-learning-based radio map reconstruction for V2X communications,

S. Roger, M. Brambilla, B. C. Tedeschini, C. Botella-Mascarell, M. Cobos, and M. Nicoli, “Deep-learning-based radio map reconstruction for V2X communications,”IEEE Trans. V eh. Technol., vol. 73, no. 3, pp. 3863-3871, 2024

work page 2024

Show all 16 references

[6]

Blockage-aware radio map construction via exploiting the diffraction and obstruction structure,

W. Chen, and J. Chen, “Blockage-aware radio map construction via exploiting the diffraction and obstruction structure,” in Proc. IEEE GLOBECOM, 2023, pp. 1920-1925

2023
[8]

Crowdsourced indoor Wi-Fi REMs: Does the spatial interpolation method matter?

Z. El-friakh, A. M. V oicu, S. Shabani, L. Simi ´c, and P. M ¨ah¨onen, “Crowdsourced indoor Wi-Fi REMs: Does the spatial interpolation method matter?” in Proc. IEEE DySPAN , 2018, pp. 1-10

2018
[9]

K-nearest neighbors gaussian process regression for urban radio map reconstruction,

Y . Zhang and S. Wang, “K-nearest neighbors gaussian process regression for urban radio map reconstruction,” IEEE Commun. Lett. , vol. 26, no. 12, pp. 3049-3053, 2022

2022
[10]

Ordinary Kriging interpolation for indoor 3D REM,

P. Maiti and D. Mitra, “Ordinary Kriging interpolation for indoor 3D REM,” J. Ambient Intell. Hum. Comput. , vol. 14, no. 10, pp. 13285- 13299, 2023

2023
[11]

Real-time outdoor localization using radio maps: A deep learning approach,

C ¸ . Yapar, R. Levie, G. Kutyniok, and G. Caire, “Real-time outdoor localization using radio maps: A deep learning approach,” IEEE Trans. Wireless Commun., vol. 22, no. 12, pp. 9703-9717, 2023

2023
[12]

Deep completion autoencoders for radio map estimation,

Y . Teganya and D. Romero, “Deep completion autoencoders for radio map estimation,” IEEE Trans. Wireless Commun. , vol. 21, no. 3, pp. 1710-1724, 2022

2022
[14]

RME-GAN: A learning framework for radio map estimation based on conditional generative adversarial network,

S. Zhang, A. Wijesinghe, and Z. Ding, “RME-GAN: A learning framework for radio map estimation based on conditional generative adversarial network,” IEEE Internet Things J., vol. 10, no. 20, pp. 18016- 18027, 2023

2023
[15]

Model and transfer spatial-temporal knowledge for fine-grained radio map reconstruction,

K. Li, C. Li, B. Yu, Z. Shen, Q. Zhang, S. He, and J.Chen, “Model and transfer spatial-temporal knowledge for fine-grained radio map reconstruction,” IEEE Trans. Cogn. Commun. Netw. , vol. 8, no. 2, pp. 828–841, 2022

2022

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.

[1] [7]

Accurate spectrum map construction for spectrum management through intelligent frequency-spatial reasoning,

F. Zhou, C. Wang, G. Wu, Y . Wu, Q. Wu, and N. Al-Dhahir, “Accurate spectrum map construction for spectrum management through intelligent frequency-spatial reasoning,” IEEE Trans. Commun , vol.71, no.7, pp. 3932-3945, 2023

work page 2023

[2] [16]

Ac- curate spectrum map construction using an intelligent frequencyspatial reasoning approach,

C. Wang, Y . Wu, F. Zhou, Q. Wu, C. Dong, and K. K. Wong, “Ac- curate spectrum map construction using an intelligent frequencyspatial reasoning approach,” in Proc. IEEE GLOBECOM, 2022, pp. 3460-3465

work page 2022

[3] [13]

RadioUNet: Fast radio map estimation with convolutional neural networks,

R. Levie, C ¸ . Yapar, G. Kutyniok, and G. Caire, “RadioUNet: Fast radio map estimation with convolutional neural networks,” IEEE Trans. Wireless Commun., vol. 20, no. 6, pp. 4001-4015, 2021

work page 2021

[4] [1]

On the road to 6G: Visions, requirements, key technologies, and testbeds,

C.-X. Wang, X. You, X. Gao, X. Zhu, Z. Li, C. Zhang, H. Wang, Y . Huang, Y . Chen, H. Haas, J. S. Thompson, E. G. Larsson, M. D. Renzo, W. Tong, P. Zhu, X. Shen, H. V . Po, and L. Hanzo, “On the road to 6G: Visions, requirements, key technologies, and testbeds,” IEEE Commun. Surveys. Tuts., vol. 25, no. 2, pp. 905-974, 2023

work page 2023

[5] [2]

A tutorial on environment-aware communications via channel knowledge map for 6G,

Y . Zeng, J. Chen, J. Xu, D. Wu, X. Xu, S. Jin, X. Gao, D. Gesbert, S. Cui, and R.Zhang, “A tutorial on environment-aware communications via channel knowledge map for 6G,” IEEE Commun. Surveys. Tuts. , early access, 2024

work page 2024

[6] [3]

Deep stacked autoencoder-based long-term spectrum prediction using real- world data,

G. Pan, Q. Wu, G. Ding, W. Wang, J. Li, F. Xu, and B.Zhou, “Deep stacked autoencoder-based long-term spectrum prediction using real- world data,” IEEE Trans. Cogn. Commun. Netw. , vol. 9, no. 3, pp. 534- 548, 2023. Fig. 6: Comparison of training loss under different sampling densities

work page 2023

[7] [4]

Compressed wideband spectrum mapping in 3D spectrum-heterogeneous environment,

F. Shen, G. Ding, Q. Wu, and Z. Wang, “Compressed wideband spectrum mapping in 3D spectrum-heterogeneous environment,” IEEE Trans. V eh. Technol., vol. 72, no. 4, pp. 4875-4886, 2023

work page 2023

[8] [5]

Deep-learning-based radio map reconstruction for V2X communications,

S. Roger, M. Brambilla, B. C. Tedeschini, C. Botella-Mascarell, M. Cobos, and M. Nicoli, “Deep-learning-based radio map reconstruction for V2X communications,”IEEE Trans. V eh. Technol., vol. 73, no. 3, pp. 3863-3871, 2024

work page 2024