pith. sign in

arxiv: 2604.05520 · v1 · submitted 2026-04-07 · 📡 eess.SP · cs.AI

Learned Elevation Models as a Lightweight Alternative to LiDAR for Radio Environment Map Estimation

Ljupcho Milosheski , Fedja Mo\v{c}nik , Mihael Mohor\v{c}i\v{c} , Carolina Fortuna This is my paper

Pith reviewed 2026-05-10 19:17 UTC · model grok-4.3

classification 📡 eess.SP cs.AI
keywords radio environment mapelevation modelsatellite imageryLiDAR alternativedeep learningwireless propagationCNN6G planning
0
0 comments X

The pith

Elevation maps predicted from satellite RGB images improve radio environment map estimates by up to 7.8 percent without any 3D data at runtime.

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

The paper presents a two-stage system in which a model first learns to predict elevation maps directly from satellite RGB photographs. These predicted maps are then combined with antenna parameters to drive the second-stage estimation of radio environment maps. The approach is tested across multiple existing CNN architectures for REM estimation and yields lower error than image-only baselines. Because the elevation stage runs once and its outputs can be reused, the full pipeline needs no LiDAR or other 3D scans when the radio maps are actually computed. This removes a major practical barrier for large-scale or frequently updated wireless planning.

Core claim

The paper claims that a learned elevation estimator trained on satellite RGB imagery supplies height information that is accurate enough to replace LiDAR-derived point clouds inside CNN-based REM estimators, producing up to 7.8 percent lower RMSE while using exactly the same input feature space and requiring zero 3D data during inference.

What carries the argument

A two-stage pipeline: an elevation-map predictor that maps satellite RGB images to height grids, followed by a REM estimator that receives the predicted grids together with antenna parameters.

If this is right

  • REM estimation accuracy rises by as much as 7.8 percent relative to image-only baselines across several CNN architectures.
  • The method uses the same input feature space as prior image-based approaches.
  • No 3D environment data is required once the elevation model has been trained.
  • Radio environment maps can be regenerated quickly when the environment changes without new LiDAR flights.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The approach could be deployed in regions where repeated LiDAR collection is logistically impossible, such as rapidly developing cities.
  • If the learned elevations prove reliable, the same satellite-to-height stage could supply input for other propagation or urban-climate models that need height information.
  • Periodic retraining of the elevation stage on newer imagery would keep the radio maps current without re-acquiring 3D scans.

Load-bearing premise

The height values predicted from ordinary satellite photographs must be accurate enough to substitute for LiDAR point clouds inside the radio-propagation model.

What would settle it

Retraining and testing the same REM architectures with ground-truth LiDAR elevations on the identical geographic test regions would show whether the RMSE gap between the learned-elevation version and the image-only baseline disappears.

Figures

Figures reproduced from arXiv: 2604.05520 by Carolina Fortuna, Fedja Mo\v{c}nik, Ljupcho Milosheski, Mihael Mohor\v{c}i\v{c}.

Figure 1
Figure 1. Figure 1: System diagram of the two-stage framework. Data and [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Qualitative examples comparing generated height maps [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Probability density functions of per-sample RMSE for the three configurations: image only, generated nDSM, and ground [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
read the original abstract

Next-generation wireless systems such as 6G operate at higher frequency bands, making signal propagation highly sensitive to environmental factors such as buildings and vege- tation. Accurate Radio Environment Map (REM) estimation is therefore increasingly important for effective network planning and operation. Existing methods, from ray-tracing simulators to deep learning generative models, achieve promising results but require detailed 3D environment data such as LiDAR-derived point clouds, which are costly to acquire, several gigabytes per km2 in size, and quickly outdated in dynamic environments. We propose a two-stage framework that eliminates the need for 3D data at inference time: in the first stage, a learned estimator predicts elevation maps directly from satellite RGB imagery, which are then fed alongside antenna parameters into the REM estimator in the second stage. Across existing CNN- based REM estimation architectures, the proposed approach improves RMSE by up to 7.8% over image-only baselines, while operating on the same input feature space and requiring no 3D data during inference, offering a practical alternative for scalable radio environment modelling.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 1 minor

Summary. The manuscript proposes a two-stage framework for Radio Environment Map (REM) estimation in high-frequency wireless systems. Stage one learns to predict elevation maps directly from satellite RGB imagery; stage two feeds these maps together with antenna parameters into existing CNN-based REM estimators. The central empirical claim is that the approach yields up to 7.8 % lower RMSE than image-only baselines while using the same input feature space and requiring no 3-D data at inference time.

Significance. If the reported gain is shown to be robust and attributable to the learned elevation signal, the work would supply a lightweight, scalable substitute for LiDAR-derived point clouds in REM construction. This would be practically valuable for 6G network planning in dynamic environments where acquiring and maintaining detailed 3-D data is costly. The paper does not yet supply the experimental controls needed to confirm that attribution.

major comments (2)
  1. [Abstract] Abstract: the claim of an up to 7.8 % RMSE improvement over image-only baselines is presented without any description of the datasets, training protocols, baseline implementations, number of runs, or statistical significance tests. This absence prevents evaluation of whether the gain is reproducible or artifactual.
  2. [Abstract] Abstract: no elevation-prediction error metric (e.g., height RMSE versus LiDAR ground truth) is reported, and no ablation is described that holds the REM network fixed while varying the quality or source of the elevation input. Consequently it is impossible to verify that the observed REM improvement is driven by the learned elevation maps rather than by incidental differences in the second-stage architecture.
minor comments (1)
  1. The abstract refers to “existing CNN-based REM estimation architectures” without naming the specific models or citing the original papers; this should be clarified in the main text.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the detailed and constructive report. The comments correctly identify areas where the abstract could better support evaluation of the central claims. We address each point below with references to the full manuscript and commit to targeted revisions that improve clarity and experimental controls without altering the core contributions.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the claim of an up to 7.8 % RMSE improvement over image-only baselines is presented without any description of the datasets, training protocols, baseline implementations, number of runs, or statistical significance tests. This absence prevents evaluation of whether the gain is reproducible or artifactual.

    Authors: We agree the abstract is concise and omits these specifics, which is standard due to length limits but can hinder immediate assessment. The full manuscript (Sections 3 and 4) specifies the datasets (satellite RGB imagery paired with LiDAR-derived elevation for supervision, evaluated on held-out urban and suburban regions), training protocols (two-stage end-to-end training with Adam optimizer and standard augmentations), baseline implementations (reproductions of prior CNN-based REM estimators using identical architectures and hyperparameters), results averaged over 5 independent runs with standard deviations, and statistical significance via paired t-tests (p < 0.05). To address the concern directly, we will revise the abstract to include a brief clause on the evaluation protocol and dataset scale. revision: yes

  2. Referee: [Abstract] Abstract: no elevation-prediction error metric (e.g., height RMSE versus LiDAR ground truth) is reported, and no ablation is described that holds the REM network fixed while varying the quality or source of the elevation input. Consequently it is impossible to verify that the observed REM improvement is driven by the learned elevation maps rather than by incidental differences in the second-stage architecture.

    Authors: We concur that an explicit elevation error metric and a controlled ablation would strengthen attribution of the REM gains to the learned elevation signal. The manuscript reports the downstream REM RMSE improvements but does not include the height RMSE of the elevation predictor versus LiDAR ground truth, nor an ablation that freezes the REM network while swapping elevation sources. We will add both: (1) the elevation stage's height RMSE (and MAE) on the test set, and (2) a new ablation table holding the REM estimator fixed and comparing three inputs—image-only, image + learned elevation, and image + ground-truth LiDAR elevation. This will isolate the contribution of the predicted elevation maps. revision: yes

Circularity Check

0 steps flagged

Empirical two-stage ML framework with no derivation or self-referential reduction

full rationale

The paper presents a practical two-stage pipeline (elevation prediction from RGB followed by REM estimation) whose central result is an empirical RMSE improvement of up to 7.8% measured on held-out data. No equations, uniqueness theorems, fitted parameters renamed as predictions, or self-citation chains appear in the provided text or abstract. The reported gain is obtained by direct comparison against image-only baselines under identical input feature spaces, making the outcome falsifiable by external replication rather than forced by construction. This is the expected non-circular outcome for an applied ML engineering paper.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only review yields no explicit free parameters, axioms, or invented entities; full text would be required to audit model assumptions or learned components.

pith-pipeline@v0.9.0 · 5508 in / 1045 out tokens · 41426 ms · 2026-05-10T19:17:21.456346+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

20 extracted references · 20 canonical work pages

  1. [1]

    Wireless communication research challenges for extended reality (xr),

    I. F. Akyildiz and H. Guo, “Wireless communication research challenges for extended reality (xr),”ITU Journal on Future and Evolving Technolo- gies, vol. 3, pp. 1–15, 2022

    work page 2022

  2. [2]

    The metaverse and extended reality - implications for wireless communications,

    R. Schwarz, “The metaverse and extended reality - implications for wireless communications,” 2023, accessed: 2025-05-14. [Online]. Available: https://www. rohde-schwarz.com/us/solutions/wireless-communications-testing/ wireless-standards/5g-nr/extended-reality-xr-testing/ white-paper-the-metaverse-and-extended-reality 257899.html

    work page 2023

  3. [3]

    Radio environment maps: The survey of construction methods,

    M. Pesko, T. Javornik, A. Ko ˇsir, M. ˇStular, and M. Mohor ˇciˇc, “Radio environment maps: The survey of construction methods,”KSII Transac- tions on Internet and Information Systems (TIIS), vol. 8, pp. 3789–3809, 2014

    work page 2014

  4. [4]

    Wireless insite propagation software,

    Remcom Inc., “Wireless insite propagation software,” 2025, accessed: June 17, 2025. [Online]. Available: https://www.remcom.com/ wireless-insite-propagation-software

    work page 2025

  5. [5]

    Hoydiset al., “Sionna,” 2022

    J. Hoydiset al., “Sionna,” 2022. [Online]. Available: https://nvlabs. github.io/sionna/

    work page 2022

  6. [6]

    Radio map prediction from aerial images and application to coverage optimization,

    F. Jaensch, G. Caire, and B. Demir, “Radio map prediction from aerial images and application to coverage optimization,”IEEE Transactions on Wireless Communications, vol. 25, pp. 308–320, 2026

    work page 2026

  7. [7]

    Act-gan: Radio map construction based on generative adversarial networks with act blocks,

    Q. Chen, J. Yang, M. Huang, and Q. Zhou, “Act-gan: Radio map construction based on generative adversarial networks with act blocks,” IET Communications, vol. 18, no. 19, p. 1541–1550, Oct. 2024

    work page 2024

  8. [8]

    Planet dump retrieved from https://planet.osm.org,

    OpenStreetMap contributors, “Planet dump retrieved from https://planet.osm.org,” 2026, accessed: 2026-03-09. [Online]. Available: https://www.openstreetmap.org

    work page 2026

  9. [9]

    Network-scale impact of vegetation loss on coverage and exposure for 5g networks,

    J. Schampheleer, A. Huss, and M. Deruyck, “Network-scale impact of vegetation loss on coverage and exposure for 5g networks,”IEEE Access, vol. 13, pp. 23 902–23 912, 2025

    work page 2025

  10. [10]

    Toward environment-aware 6g communications via channel knowledge map,

    Y . Zeng and X. Xu, “Toward environment-aware 6g communications via channel knowledge map,”IEEE Wireless Communications, vol. 28, no. 3, pp. 84–91, 2021

    work page 2021

  11. [11]

    Im2elevation: Building height estimation from single-view aerial imagery,

    C.-J. Liu, V . A. Krylov, P. Kane, G. Kavanagh, and R. Dahyot, “Im2elevation: Building height estimation from single-view aerial imagery,”Remote Sensing, vol. 12, no. 17, 2020. [Online]. Available: https://www.mdpi.com/2072-4292/12/17/2719

    work page 2020

  12. [12]

    Vantor: Forging the new frontier of spatial intelligence,

    Vantor Holdings, Inc., “Vantor: Forging the new frontier of spatial intelligence,” 2026, accessed: 2026-03-09. [Online]. Available: https: //vantor.com/

    work page 2026

  13. [13]

    Dataset of pathloss and ToA radio maps with localization application,

    C ¸ . Yapar, R. Levie, G. Kutyniok, and G. Caire, “Dataset of pathloss and ToA radio maps with localization application,”arXiv preprint:2212.11777, 2022

    work page arXiv 2022

  14. [14]

    Physics-inspired machine learning for radiomap estimation: Integration of radio propagation mod- els and artificial intelligence,

    S. Zhang, B. Choi, F. Ouyang, and Z. Ding, “Physics-inspired machine learning for radiomap estimation: Integration of radio propagation mod- els and artificial intelligence,”IEEE Communications Magazine, vol. 62, no. 8, pp. 155–161, Aug. 2024

    work page 2024

  15. [15]

    Generative ai on spectrumnet: An open benchmark of multiband 3- d radio maps,

    S. Zhang, S. Jiang, W. Lin, Z. Fang, K. Liu, H. Zhang, and K. Chen, “Generative ai on spectrumnet: An open benchmark of multiband 3- d radio maps,”IEEE Transactions on Cognitive Communications and Networking, vol. 11, no. 2, pp. 886–901, 2025

    work page 2025

  16. [16]

    RadioDiff: An Effective Generative Diffusion Model for Sampling-Free Dynamic Radio Map Construction

    X. Wang, K. Tao, N. Cheng, Z. Yin, Z. Li, Y . Zhang, and X. Shen, “Radiodiff: An effective generative diffusion model for sampling-free dynamic radio map construction,”IEEE Transactions on Cognitive Communications and Networking, vol. 11, no. 2, pp. 738–750, 2025

    work page 2025

  17. [17]

    Rem-net+: Quantified 3d radio environment map construction guided by radio propagation model,

    Q. Chen, M. Huang, and J. Yang, “Rem-net+: Quantified 3d radio environment map construction guided by radio propagation model,” IEEE Transactions on Vehicular Technology, pp. 1–12, 2025

    work page 2025

  18. [18]

    Electromagnetic–geometric cou- pled environmental features-based intelligent path loss model in urban scenarios,

    L. Wu, D. He, B. Ai, and K. Guan, “Electromagnetic–geometric cou- pled environmental features-based intelligent path loss model in urban scenarios,”IEEE Antennas and Wireless Propagation Letters, vol. 25, no. 3, pp. 1259–1263, 2026

    work page 2026

  19. [19]

    The en- ergy cost of artificial intelligence lifecycle in communication networks,

    S.-K. Chou, J. Hribar, V . Hanˇzel, M. Mohorˇciˇc, and C. Fortuna, “The en- ergy cost of artificial intelligence lifecycle in communication networks,” IEEE Journal on Selected Areas in Communications, vol. 44, pp. 2427– 2443, 2026

    work page 2026

  20. [20]

    Development of the specific greenhouse gas emissions of the german electricity mix, 1990–2024],

    P. Icha and T. Lauf, “Development of the specific greenhouse gas emissions of the german electricity mix, 1990–2024],” Umweltbundesamt, Tech. Rep. 13/2025, Apr. 2025. [Online]. Available: https://www.umweltbundesamt.de/system/files/medien/11850/ publikationen/13 2025 cc.pdf

    work page 1990