Source-linked AI summary
KODAMA: Multimodal Digital Twin Reconstruction for Urban RF Propagation Modelling
Maximiliano Wardle, A. Ryo Koblitz
TL;DR
RF digital twins need to reproduce measured site-specific channels, but existing approaches trade scene fidelity, scale, and acquisition effort. KODAMA reconstructs ray-tracing-ready city-scale twins from existing geospatial and street-level imagery. Across three environments, its uncalibrated predictions improve on automated pipelines and approach a measurement-calibrated hand-built twin.
Problem
No automated method reconstructs high-fidelity city-scale RF scenes from pre-existing data alone, despite RF twins being used to predict site-specific channels.
Source: §3 · p. 3Method
KODAMA uses aerial imagery, photogrammetry, and LiDAR for terrain and building volumes, then street-level imagery to recover facade relief, materials, and clutter.
Source: §1 · p. 2Results
Up to 5.35 dB point-to-point RMSE improvement over automated pipelines and within 0.22 dB of a measurement-calibrated hand-built twin.
Source: §7 · p. 8Takeaways & Limitations
RF propagation provides a strict, falsifiable test of 3D reconstruction by judging whether signals propagate as they do at the real site.
Source: §7 · p. 8KODAMA depends on unevenly available geospatial products, uses a 2.5D representation that struggles with tunnels and bridges, and models outdoor environments only.
Source: §6 · p. 8
Abstract
from arXiv · showhide
3D reconstruction typically strives for geometric fidelity or visual plausibility. Radio frequency digital twins (RFDT) are instead judged by whether communication channels behave in them as they do in the real world. RFDTs promise site-specific channel prediction but current practice forces a choice between coarse automated scenes and hand-built, measurement-calibrated models that take weeks to construct per-site. We present KODAMA, an automated pipeline that reconstructs ray tracing-ready RFDTs at city scale from off-the-shelf geospatial data alone: aerial imagery, LiDAR, and photogrammetry yield terrain and watertight building meshes, while exposure-weighted multi-view fusion of street-level imagery recovers fa\c{c}ade relief, electromagnetic materials, and clutter---all without site visits or calibration. Across three sites spanning 3.6 to 28 GHz, KODAMA's uncalibrated predictions achieve single-digit RMSE, reducing point-to-point error by up to 5.35 dB over automated baselines and coming within 0.22 dB of a measurement-calibrated, hand-built RFDT.
1. Introduction
RF digital twins are judged by whether simulated channels match measurements at the real site, not by visual resemblance alone. KODAMA reconstructs city-scale, ray-tracing-ready twins from existing geospatial data and is validated against real-world measurements.
1. Introduction
- An RFDT is evaluated by how closely simulated channels match measurements at the site, since visual or geometric quality can miss propagation-relevant errors.RFDT channel predictions can inform network tasks and training data for signal-processing algorithms.Source: §1 · p. 2
- KODAMA combines aerial imagery, photogrammetry, and LiDAR with street-level imagery to reconstruct terrain, building volumes, facade relief, materials, and clutter.The paper says kilometre-scale, ray-tracing-ready scenes can be produced in hours instead of weeks.Source: §1 · p. 2
- The paper identifies RF-specific reconstruction requirements that conventional geometric and perceptual measures do not guarantee.KODAMA is presented as an automated pipeline using widely available spatial data to meet those requirements.Source: §1 · p. 2
- KODAMA improves on existing automated RFDT approaches, and its uncalibrated predictions rival hand-built twins calibrated against on-site measurements.The validation covers three sites.Source: §1 · p. 2
2. Radio-frequency ray tracing desiderata
RF ray tracing depends on explicit, physically meaningful scene properties that conventional visual or geometric criteria may not ensure. The required scene detail also varies with the radio frequency being modelled.
2. Radio-frequency ray tracing desiderata
- At higher frequencies and shorter wavelengths, smaller scene features have greater effects on propagation, so the necessary level of detail is not fixed.Radio wavelengths are orders of magnitude larger than light wavelengths, requiring a higher-order approximation to Maxwell’s equations.Source: §2 · p. 2
- RF ray tracing requires watertight meshes, semantically separable objects, accurate terrain and building volumes, detailed observed facades, and per-surface electromagnetic materials.Material properties determine reflection strength and energy lost at each bounce.Source: §2 · p. 2
- Clutter such as vegetation, furniture, signage, and cars can block, weaken, or redirect signals, while georeferencing positions objects in real-world coordinates.Correct placement also lets diverse data fuse cleanly and predictions transfer to the site.Source: §2 · p. 3
3. Related work
Related work spans coarse automated RF scenes, labor-intensive site-specific twins, and broader urban reconstruction methods that do not jointly provide the required fidelity, scale, and low capture burden. KODAMA is positioned against these gaps in both RF-specific and general urban-scene reconstruction.
3. Related work
- Existing approaches satisfy different subsets of RF scene requirements, and no automated method reconstructs city-scale environments for RF ray tracing while fulfilling them all.This is the related-work section’s central gap.Source: §3 · p. 3
3.1. RF digital twin reconstruction
- Automated RFDT pipelines commonly extrude OSM footprints over coarse SRTM terrain, yielding approximate prisms without facade detail or clutter that depart from measurements.OSM buildings are often misshapen, displaced, or missing, and fewer than 8% include height data.Source: §3.1 · p. 3
- A hand-refined Brooklyn RFDT matched measurements within 4 dB but required rangefinder surveys, LiDAR, photogrammetry, and labor-intensive manual refinement [76].KODAMA is evaluated on the same site using open geospatial data and existing street imagery.Source: §3.1 · p. 3
- Government CityGML, LiDAR, or photogrammetry models offer richer geometry but describe individual cities; vision-based RFDT methods require dedicated capture and remain small-scale.The vision-based methods cited are demonstrated in small, bounded, mostly indoor environments.Source: §3.1 · p. 3
3.2. Urban scene reconstruction beyond RF
- Neural city reconstructions target photorealism or generic geometry, while satellite-generated scenes may invent facade details that create RF paths absent from the real site.KODAMA instead adds detail supported by real observations and retains evidence for each geometric operation.Source: §3.2 · p. 3
- Overhead-driven mesh methods leave facades weakly observed; VoxCity remains too coarse for the considered RF frequencies and lacks detailed geometry, materials, and clutter.Other feature-refinement approaches require accurate, well-structured geometric priors that may not exist for a target site.Source: §3.2 · pp. 3–4
4. Method
KODAMA turns geospatial height data into watertight building and terrain geometry, then uses registered street-level imagery to add façade form, electromagnetic materials, clutter, and foliage. Its detail-recovery stages combine image evidence with exposure-aware multi-view fusion and, where available, source-mesh information.
4. Method
- KODAMA converts geospatial inputs into shared height fields, builds watertight buildings from dense height data, and uses street imagery to add façade geometry, materials, and clutter.It accepts orthophotos with LiDAR or surface models, optionally bare-earth terrain data, or georeferenced photogrammetric models.Source: §4 · p. 4
4.1. Reconstructing terrain and buildings
- Iterative terrain estimation separates ground from above-ground structure, while cleaned roof surfaces are extruded to terrain and remeshed as watertight building solids.Tightening the terrain tolerance across passes recovers finer variation without absorbing elevated structures; the building meshes preserve roof shapes and features such as chimneys and solar panels.Source: §4.1 · p. 4
4.2. Recovering detail from street-level imagery
- Ground-level data leave façade geometry, electromagnetic materials, and clutter unresolved, so KODAMA registers street-level views and applies exposure-aware multi-view fusion.Registration uses provider metadata because pose refinement is unreliable when metadata quality varies and landmarks may be missing from the scene.Source: §4.2 · pp. 4–5
4.2.1. Parsing street-level imagery
- SAM3 generates aligned architectural and clutter masks that guide façade and material recovery and prevent foreground objects from contaminating building surfaces.The masks also support adding 3D clutter objects to the scene.Source: §4.2.1 · p. 5
4.2.2. Reconstructing façade geometry
- For façade relief, semantic and geometric cues classify regions as recessed, neutral, or protruding, then exposure-weighted image votes and source-mesh discrepancy inform per-texel displacement posteriors.Thresholding and connected-component processing select candidate regions for Boolean mesh edits; source geometry modulates rather than overrides image evidence.Source: §4.2.2 · pp. 5–6, Figure 2 · p. 5
4.2.3. Assigning electromagnetic materials.
- Material assignment favors broad electromagnetic coverage: image predictions are refined with architectural and clutter masks, fused across views, and mapped to standardized material parameters.Unobserved areas may inherit the dominant material of the same surface, with direct evidence distinguished from extrapolated assignments.Source: §4.2.3 · p. 6
4.2.4. Placing RF clutter and foliage.
- Clutter instances are depth-calibrated against scene geometry, reconciled across detections, and placed using class-specific assets; large canopies combine absorbing volumes with planar reflection and scattering surfaces.The hybrid foliage representation captures principal propagation effects while remaining tractable at city scale.Source: §4.2.4 · pp. 6–7
5. Evaluation and results
KODAMA was evaluated at three sites against independent measurements, with automated OSM and 3GPP baselines and a calibrated NYURay comparison in Brooklyn. Its performance varied by setting, from substantial Espoo gains to a close match with OSM in Boulder and near-parity with NYURay in Brooklyn.
Evaluation setup
- The evaluation compares ray-traced predictions with independent site measurements, using OSM and 3GPP baselines and Brooklyn’s calibrated NYURay twin.Solver settings were held fixed per scene so prediction differences could be attributed to reconstruction.Source: §5 · p. 7
Evaluation metrics
- Point-to-point comparisons reveal local propagation effects, such as around-corner propagation from street-clutter diffraction, that PD and CI trend-line scores can mask.P2P evaluations are less frequently reported in the literature.Source: §5 · p. 7
Espoo
- 5.35 dB P2P RMSE and 2.66 dB PD RMSE reductions over OSM were reported in Espoo.P2P RMSE fell from 13.11 dB to 7.76 dB; the gap arose primarily in NLOS conditions.Source: Table 1 · p. 8, §5 · p. 8
Brooklyn
- 0.22 dB separated KODAMA from the measurement-calibrated NYURay twin in Brooklyn CI, while KODAMA beat OSM in all three measures and 3GPP in PD.The evaluation used measurements at 6.75 GHz and 16.95 GHz.Source: Table 1 · p. 8, §5 · p. 8
Boulder
- KODAMA and OSM were within around 0.1 dB in Boulder, and both outperformed 3GPP PD predictions.Measured pairs were almost entirely line-of-sight, so direct-path loss dominated regardless of scene modelling.Source: §5 · p. 8
6. Limitations and future work
KODAMA depends on available geospatial products and its 2.5D representation and material classes constrain what it can model. The reconstruction stages have only been evaluated in combination; the authors identify component-level ablation and measurement-informed training as future work.
Scope and representation
- Geodata coverage is uneven and concentrated in urban centres and the Global North, limiting where KODAMA’s inputs are available.Its 2.5D height fields also make larger openings and overhangs, including tunnels and bridges, challenging.Source: §6 · p. 8
- ITU-R P.2040 material classes leave some electromagnetic behaviour and surface roughness unmodelled, and KODAMA models outdoor environments only.Outdoor-indoor convergence is identified as a natural extension.Source: §6 · p. 8
Evaluation and future work
- Reconstruction stages have been evaluated only in combination, so their individual contributions remain unisolated.A full factorial ablation is identified as the immediate next step.Source: §6 · p. 8
- Because the propagation model is differentiable, measurements could serve as a training signal as well as a benchmark.The proposed longer-term direction is to close the loop between scene generation and channel prediction.Source: §6 · p. 8
7. Conclusions
KODAMA reconstructs city-scale RF digital twins automatically from widely available geospatial data, without site visits, bespoke capture, or calibration. Across three environments and 3.6–28 GHz, its uncalibrated predictions improve on automated pipelines by up to 5.35 dB in point-to-point RMSE and come within 0.22 dB of a measurement-calibrated hand-built twin.
Conclusion
- Across three environments and 3.6–28 GHz, KODAMA improves on automated pipelines by up to 5.35 dB point-to-point RMSE and comes within 0.22 dB of a calibrated hand-built twin.The conclusion frames RF propagation as a falsifiable test of reconstruction, judging scenes by whether signals propagate as they do at the real site.Source: §7 · p. 8