arXiv:2609.05793v1 [cs.NI] 5 Sep 2026
HalifaxDT: A Wireless Digital Twin from Open Geospatial Data Alexis Delplace
Xinyi Li
Faculty of Computer Science, Dalhousie University Halifax, Canada LISN, Université Paris-Saclay Gif-sur-Yvette, France [email protected]
Faculty of Computer Science, Dalhousie University Halifax, Canada [email protected]
Samer Lahoud
Nur Zincir-Heywood
Faculty of Computer Science, Dalhousie University Halifax, Canada [email protected]
Faculty of Computer Science, Dalhousie University Halifax, Canada [email protected]
Abstract
CCS Concepts
Wireless digital twins can support site-specific analysis, planning, and experimentation for future wireless networks. However, constructing them at city scale remains challenging when accurate 3D city models are unavailable. This paper presents HalifaxDT, a wireless digital twin of the Halifax Peninsula in Nova Scotia, Canada, constructed from heterogeneous open geospatial and spectrum data. HalifaxDT combines building footprints, LiDAR elevation products, building metadata, and spectrum licensing records through a workflow that reconciles multiple sources of building-height information while preserving the provenance of geometry decisions. The resulting terrain, buildings, and gateway metadata are integrated into Sionna RT for wireless simulation. We evaluate HalifaxDT through two use cases. The first compares its coverage predictions with a reference derived from field measurements and with an analytical baseline. The second uses the digital twin to predict the received signal under normal operating conditions and detect interference. The coverage results show that HalifaxDT better preserves the spatial structure of the measured radio map than the analytical baseline. The interference study shows that deviations from the predicted reference can reveal interference that is difficult to detect from received power alone. We also identify current fidelity limitations, including simplified material representation, missing vegetation, and the need for broader RF calibration and synchronization with live measurements.
• Networks → Wireless access networks; Network simulations; Network measurement.
This work is licensed under a Creative Commons Attribution 4.0 International License. DTwin ’26, Austin, TX, USA © 2026 Copyright held by the owner/author(s). ACM ISBN 979-8-4007-2957-7/2026/10 https://doi.org/10.1145/3842434.3844958
Keywords digital twins, wireless networks, ray tracing, Sionna RT, interference detection, coverage mapping ACM Reference Format: Alexis Delplace, Xinyi Li, Samer Lahoud, and Nur Zincir-Heywood. 2026. HalifaxDT: A Wireless Digital Twin from Open Geospatial Data. In ACM Workshop on Digital Twins over NextG Wireless Networks (DTwin ’26), October 26–30, 2026, Austin, TX, USA. ACM, New York, NY, USA, 9 pages. https://doi.org/10.1145/3842434.3844958
1
Introduction
Wireless digital twins are becoming important tools for sitespecific simulation, deployment planning, synthetic data generation, and what-if analysis in future wireless networks. Their usefulness, however, depends on the availability of an environment model that is sufficiently accurate for propagation analysis. In dense urban areas, radio performance is shaped by local geometry, blockage, reflection, diffraction, terrain elevation, antenna placement, and material properties. A wireless digital twin therefore cannot be treated only as a network model; it must also represent the physical environment in which the network operates. Prior work spans simulation platforms, calibrated propagation models, and city-scale scene instantiations. Sionna RT [8] provides the differentiable ray-tracing engine used here, while measurement-calibrated digital twins show how field observations correct systematic RSSI offsets and improve fidelity [16]. City-scale systems such as BostonTwin [17] demonstrate detailed urban geometry when such geometry is already available. HalifaxDT complements these efforts
DTwin ’26, October 26–30, 2026, Austin, TX, USA
by combining open-data scene reconstruction with sparse LoRaWAN field-measurement evaluation. However, many cities lack detailed 3D models that can be directly converted into wireless simulation scenes. HalifaxDT therefore focuses on traceable reconstruction from open or institutional layers, including building footprints, LiDAR elevation products, volunteered geographic information, municipal records, and spectrum-licence databases. This is especially important for building-height estimation, where elevation rasters, explicit height attributes, floor-count estimates, and external topographic records differ in coverage, temporal validity, spatial alignment, and height definition. We present HalifaxDT, a ray-tracing-ready wireless digital twin of the Halifax Peninsula in Nova Scotia, Canada. HalifaxDT is constructed from publicly available geospatial and spectrum data rather than from an existing 3D city model. It integrates municipal building footprints, LiDARderived elevation products, OpenStreetMap metadata, municipal building attributes, national topographic building records, and spectrum-licence information, and exports the resulting scene in a format compatible with Sionna RT. The generated scene assets, processed inputs, and reconstruction scripts are made available online for inspection and reuse [2]. The contributions are threefold. First, we present a reusable workflow for constructing ray-tracing-ready wireless digital twins from heterogeneous open geospatial data. Second, we introduce a building-height inference process that combines elevation products, building attributes, floor-count estimates, and external records while preserving provenance. Third, we instantiate the workflow as HalifaxDT and evaluate it through coverage mapping and interference detection, showing how an open-data wireless digital twin can support both propagation analysis and jammer detection without labeled jammer data.
2
HalifaxDT Construction Methodology
The construction of a city-scale digital twin for wireless simulation can be organized as a reusable four-stage workflow: (i) identifying the geospatial data needed to describe the environment, (ii) aligning heterogeneous sources in a common spatial frame, (iii) converting the aligned layers into a ray-tracing-ready 3D scene, and (iv) integrating the resulting scene with a wireless simulator. HalifaxDT instantiates this workflow for the Halifax Peninsula, but the same procedure can be adapted to other cities where similar open or institutional geospatial sources are available.
2.1 Geospatial Inputs and Source Discovery The first step is to identify the data layers required for the intended wireless simulation. If a georeferenced 3D city model is already available, the construction problem is mainly one
Alexis Delplace, Xinyi Li, Samer Lahoud, and Nur Zincir-Heywood
of conversion, coordinate normalization, and radio-material annotation. Otherwise, the scene must be reconstructed from 2D and 2.5D geospatial sources. In practice, the required inputs can be grouped into a few main categories: existing 3D models, building footprints, elevation products, and building metadata. Table 1 summarizes the role of each layer, typical sources, and the corresponding HalifaxDT data. The selected geospatial layers are integrated by spatially aligning them and transforming them into a common metric coordinate system for scene generation.
Table 1: Core geospatial inputs for constructing a cityscale wireless digital twin. Input
Typical sources
HalifaxDT example
3D city model
CityGML, OBJ/PLY, existing city models OSM, cadastral data, municipal/provincial GIS LiDAR, photogrammetry, satellite DSM LiDAR, national elevation datasets, terrain APIs OSM tags, municipal building layers, building registries, topographic databases
Not available; reconstructed from layers HRM Buildings; Shapefile
Building footprints DSM DEM/DTM Building metadata
HRM LiDAR DSM; GeoTIFF HRM LiDAR DEM; GeoTIFF OSM/Overpass; GeoJSON, HRM Buildings CSV, NSTDB; Shapefile/GPKG
Two elevation layers are especially useful when explicit building heights are missing. A Digital Surface Model (DSM) describes the visible surface, including buildings, vegetation, and other above-ground objects. A Digital Elevation Model (DEM), or Digital Terrain Model (DTM), describes the bareearth terrain. When a DSM and DEM are co-registered, their difference provides height above the local terrain, which can be used to estimate building heights. This DSM–DEM strategy can be generalized to other urban digital twins when co-registered surface and terrain elevation products are available. In practice, the main challenges are the availability and resolution of the elevation layers, their temporal consistency with the building footprints, and the quality of the spatial alignment between datasets. HalifaxDT uses building footprints from Halifax Regional Municipality (HRM) as the base geometry, HRM LiDAR products for elevation, and OpenStreetMap (OSM), the HRM Buildings CSV, and Nova Scotia Topographic Database (NSTDB) records as complementary sources of building attributes and height evidence [4–6, 10–12]. OSM provides volunteered height, floor/level, material, and roof tags when available. The HRM Buildings CSV provides municipal building attributes, including explicit height values and floor-count metadata, and NSTDB provides an external building-height reference. The coverage and complementarity of these sources are analyzed below.
HalifaxDT: A Wireless Digital Twin from Open Geospatial Data
2.2
DTwin ’26, October 26–30, 2026, Austin, TX, USA
Preprocessing and Height Inference
After source discovery, the vector and raster layers are normalized into a common metric frame. In HalifaxDT, this step crops the HRM building footprints, OSM/NSTDB building records, and HRM DSM/DEM rasters to the Halifax Peninsula, reprojects them to the same local coordinate system, and links each footprint to the raster samples and metadata records used for height inference. This normalization is required because the sources do not share common geometries or identifiers: municipal footprints, OSM polygons, and NSTDB records may describe the same building using different outlines, coordinate systems, and attributes. The specific projection and spatial extent are implementationdependent. The key requirement is to place all geometry and raster evidence in a single local coordinate frame before scene generation. The most important preprocessing task in HalifaxDT is building-height inference. For a building footprint 𝑏, the DSM-derived height candidate is computed from the distribution of normalized DSM–DEM samples intersecting the footprint: ℎ𝑏DSM = 𝑄 0.95 ({𝑧 DSM (𝑢) − 𝑧 DEM (𝑢) | 𝑢 ∈ 𝑏, 𝑢 valid}) ,
(1)
where 𝑄 0.95 denotes the 95th percentile of valid raster samples. This approximates the upper roof surface while being less sensitive than the maximum to isolated vegetation, noise, missing raster values, or single-pixel outliers. In practice, this estimate must still be robust to footprint misalignment, roofshape variability, multi-volume buildings, and differences between height definitions used by external datasets. Each data source can provide one or more candidate height estimates. Candidate height estimates can be derived from explicit height fields; floor or level counts combined with building-type-dependent floor-height priors; roof-height metadata; external elevation records corrected for local terrain elevation; or metadata such as building type, feature class, footprint area, and construction type. In HalifaxDT, these non-DSM candidates are obtained from OSM tags, the HRM Buildings CSV, and NSTDB records. OSM and HRM contribute explicit-height candidates when available, falling back to level-count-derived candidates when level counts are provided. The source coverage observed in HalifaxDT motivates this multi-source design. As shown in Figure 1, OSM provides height-related metadata for only a limited fraction of buildings, with few explicit heights. DSM-derived candidates are available for all footprints, while NSTDB records and the HRM Buildings CSV add elevation-derived, explicit-height, and floor-count-derived evidence. Because these sources differ in coverage, derivation method, temporal validity, and spatial matching quality, combining them reduces generic
Figure 1: Availability of building-height evidence across the sources integrated in HalifaxDT.
Table 2: Final building-height source decisions in HalifaxDT. Selected height decision
Count
Share
DSM-P95 DSM-P75 for high-tail profiles External or consensus source FCODE fallback
16,163 1,747 1,511 246
82.2% 8.9% 7.7% 1.3%
fallback heights and provides cross-checks where independent candidates exist. Full DSM-derived candidate availability does not imply full reliable-height coverage: DSM-P95 may be affected by footprint alignment, roof shape, vegetation, temporal mismatch with the 2018 LiDAR acquisition, and height-definition differences. For high-elevation tails or mixed roof levels, the selector uses DSM-P75 rather than DSM-P95 to avoid letting small rooftop structures or adjacent objects determine the height of the whole building. Figure 1 reports source availability, whereas Table 2 reports the resulting height-selection decisions. Most buildings retain DSM-derived heights, while multi-source evidence mainly affects multi-volume, externally confirmed, or invalid DSM cases. In 14.5% of buildings, the final selected height differs from DSM-P95 by more than 2 m. The height-inference protocol treats each possible height as a candidate supported by one or more sources and assigns a deterministic rule-based confidence score. Initial scores depend on source and derivation method, while adjustments use DSM pixel count, inner-footprint availability, roof-profile stability, OSM/HRM overlap, OSM polygon ambiguity, NSTDB overlap or matching distance, temporal consistency with the 2018 LiDAR acquisition date, and physical plausibility. Scores only rank competing candidates for the same building.
DTwin ’26, October 26–30, 2026, Austin, TX, USA
A DSM-derived value is informative when recent and well aligned, but less reliable for buildings constructed or modified after the DSM acquisition date. This matters in HalifaxDT because the DSM and DEM come from 2018 LiDAR data and the city has continued to grow. When construction or update metadata indicates that a building postdates the DSM, the DSM-derived candidate is retained but assigned lower confidence; buildings predating the DSM with good raster coverage usually retain strong DSM evidence. After scoring, the pipeline selects the final height. Physically plausible estimates that agree within a predefined tolerance are combined with a weighted median, which is robust to outliers while giving more influence to higher-scored candidates. The tolerance accounts for source-specific height definitions (e.g., roof base, maximum roof point, elevationderived height, or level-derived height), so small discrepancies are expected. If no strong consensus exists, the highestscored plausible candidate is selected; if direct evidence is missing or unreliable, the pipeline falls back to metadatabased inference. The output of this stage includes both a height value and a provenance record: the selected source or consensus, the decision strategy, the confidence level, any post-DSM building flag, the candidates considered, and the reason for the decision. This provenance supports manual inspection, later correction, and adaptation of the same construction workflow to other cities with different data availability.
2.3
Scene Generation for Wireless Simulation
The third stage converts the preprocessed geospatial layers into a ray-tracing-ready scene. This requires choosing a terrain representation, generating a ground mesh, extruding building footprints, assigning material groups, and creating the scene description used by the ray tracer. For terrain modeling, the main trade-off is between spatial resolution and the complexity of the resulting ground mesh. HalifaxDT uses the 5 m DEM for terrain because terrain elevation varies more smoothly than building geometry at the scale of the study area. A 1 m terrain surface would substantially increase mesh complexity with limited benefit for the intended city-scale simulations. This choice illustrates a general rule: high-resolution data should be preserved where it affects propagation-relevant structure, but simplified where it mainly increases computational cost. Buildings are generated by extruding the height-enriched footprints. During extrusion, building bases are sampled from the DEM so that walls follow local terrain elevation. Footprint edges can be densified before extrusion to better capture terrain variation along long building edges, and roofs are generated at the selected roof elevation. These steps avoid
Alexis Delplace, Xinyi Li, Samer Lahoud, and Nur Zincir-Heywood
treating buildings as flat objects placed on a flat plane, while remaining simpler than full architectural 3D reconstruction. The scene is also annotated with approximate radio-frequency materials. Detailed facade materials are rarely available in open city-scale datasets. HalifaxDT therefore maps available semantic attributes to a small set of approximate RF material classes. HalifaxDT uses explicit OSM material tags when available and otherwise infers coarse material labels from building type, feature class, height, footprint geometry, and spatial priors. These labels are then mapped to the material models available in the ray tracer. Terrain and common material classes such as brick, concrete, metal, and wood can be assigned to their corresponding or closest available RF material proxies. Because city-scale material labels are sparse and coarse, these annotations are treated as optional scene metadata. The coverage-map evaluation below uses a single building material to avoid introducing an additional source of uncertainty from unvalidated material-label differences. For larger cities, the same protocol can be applied per spatial tile while preserving the mapping between local tile coordinates and the global coordinate reference system. This makes the construction workflow scalable without changing the underlying data model.
2.4
HalifaxDT Instance and Sionna Integration
The final stage packages the generated terrain, buildings, and RF material annotations into a scene that can be loaded by Sionna RT. The exported scene is organized around a Mitsuba-compatible XML scene file, which is the scene format expected by Sionna RT. This XML file declares the ITU radio-material definitions and references the mesh objects used by the simulator. The terrain and building geometries are exported as PLY mesh files and linked from the XML scene file, with material references assigned to each mesh group. The scene description therefore acts as the central entry point connecting the generated geometry with the RF material annotations. Sionna RT can then load the XML scene and compute radio maps or propagation paths using scenario-defined enddevices and gateways. In this way, the same HalifaxDT scene can support multiple wireless analyses while keeping the environment representation separate from experiment-specific radio configuration.
3 Use Cases 3.1 Coverage Mapping Coverage maps describe how radio-signal quality varies across a geographical area. They support network planning, gateway placement, coverage-hole identification, and the assessment of service availability before or after deployment.
HalifaxDT: A Wireless Digital Twin from Open Geospatial Data
DTwin ’26, October 26–30, 2026, Austin, TX, USA
Conventional coverage maps can be produced from field measurements or analytical propagation models, but measurement campaigns are costly and analytical models provide only a coarse representation of a specific urban environment. A wireless digital twin provides an alternative for generating such maps: once a city and its radio infrastructure are represented in a propagation-ready virtual environment, radio conditions can be estimated at locations where no measurement is available and updated when the deployment configuration changes. 3.1.1 Coverage-Map Construction. All maps use the same 25 m outdoor grid and geographical extent. The reference map is derived from Long Range Wide Area Network (LoRaWAN) transmissions collected on the Halifax Peninsula [1]. To construct a grid-level reference map, transmission attempts, including received packets and non-receptions, are assigned to grid cells. Non-receptions are treated as censored observations at −140 dBm, the minimum gateway sensitivity observed in the dataset. Irregular measurements are then converted into a continuous map using regression kriging [7], which decomposes the estimated RSSI field into a large-scale trend and a kriged residual: ˆ 𝑟ˆ (𝑠) = 𝑚(𝑠) + 𝜀ˆ(𝑠),
(2)
ˆ where 𝑚(𝑠) is the regression trend and 𝜀ˆ(𝑠) is the kriged residual field. In HalifaxDT, the trend is modeled as ˆ 𝑚(𝑠) = 𝛽 0 + 𝛽 1 log10 (𝑑 (𝑠)) + 𝛽 2𝑑 (𝑠) + 𝛽 3𝑥 (𝑠) + 𝛽 4𝑦 (𝑠), (3) where 𝑑 (𝑠) is the distance to the gateway, and 𝑥 (𝑠) and 𝑦 (𝑠) are the local projected coordinates of the measurement grid cell. The residual field is estimated at each prediction location 𝑠 0 using local ordinary kriging: 𝑛 ∑︁ ˆ 𝑖 )) , 𝜀ˆ(𝑠 0 ) = 𝜆𝑖 (𝑟 (𝑠𝑖 ) − 𝑚(𝑠 (4)
reciprocal path gain. The end-device antenna is placed at terrain height plus 2.5 m to match the measurement setup, while the gateway antenna is 12 m above local ground level, corresponding to approximately 60 m above sea level over the Halifax Peninsula. The simulation uses 904.6 MHz, 25 dBm enddevice EIRP, 6 dBi gateway receive gain, isotropic vertically polarized antennas, max_depth=16, samples_per_src=20M, max_num_paths_per_src=500k, and seed 42. Line-of-sight, diffuse reflection, and transmission/refraction are enabled; specular reflection and diffraction are disabled. Buildings use a single concrete material and terrain uses medium dry ground, keeping the comparison focused on geometry, terrain, antenna placement, and ray-traced spatial structure rather than coarse material-label assumptions. The pointwise RSSI raster is floored at −140 dBm and smoothed with a 3by-3 median filter followed by Gaussian smoothing with 𝜎 = 62.5 m. This post-processing is applied only to HalifaxDT because the analytical map is already spatially smooth. Simulations use sionna-rt 2.0.1 with the cuda_ad_mono_polarized Mitsuba variant. For comparison, we retain the urban Okumura–Hata model, a well-established analytical baseline for large urban environments around 900 MHz. It is evaluated at every outdoor grid position using the same carrier frequency, fixed-gateway configuration, end-device height, antenna gains, and −140 dBm floor. We first evaluate the maps without RSSI calibration. As reported in Table 3, HalifaxDT achieves lower MAE and better spatial metrics than Okumura–Hata. This raw comparison includes global link-budget mismatches, such as cable or connector losses, antenna-gain mismatch, and campaign-specific RSSI reporting effects that can penalize a map independently of its spatial structure, so we also evaluate a limited-data calibration setting.
𝑖=1
ˆ 𝑖 ) is the residual at measurement grid cell where 𝑟 (𝑠𝑖 ) − 𝑚(𝑠 𝑠𝑖 . The weights 𝜆𝑖 come from ordinary kriging with up to 18 neighboring measurement cells within 300 m. A Gaussian smoothing step with standard deviation 1.2 grid cells suppresses isolated interpolation artifacts. The resulting map is a measurement-derived interpolation product, not an independent ground-truth RSSI field; accordingly, we report measurement-support and kriging-uncertainty diagnostics over the evaluation mask. For the regression-kriging reference, the standard deviation over this mask has median, 90th percentile, 95th percentile, and maximum values of 5.51, 8.52, 9.10, and 10.27 dB, respectively. The HalifaxDT coverage map is generated with Sionna RT on the HalifaxDT scene. We use PathSolver point-bypoint on the 25 m outdoor grid rather than RadioMapSolver. For each valid grid point, the solver computes uplink paths from a LoRaWAN end-device to the fixed gateway, relying on
3.1.2 Limited-Data Calibration. We consider a lightweight calibration scenario that corrects global RSSI offset and scale errors from a small subset of the measurement campaign. Because these records come from the same campaign used to construct the measurement-derived reference map, this is limited-data calibration rather than independent validation. To assess sampling variability, we repeat the procedure over 50 independent random draws; each draw selects thirty real transmission records, including received RSSI values and non-receptions, to calibrate both models. We use a leftcensored Tobit model with latent RSSI, ∗ 𝑦𝑖,𝑚 = 𝛼𝑚 + 𝛽𝑚b 𝑟𝑚 (s𝑖 ) + 𝜀𝑖 ,
2 𝜀𝑖 ∼ N (0, 𝜎𝑚 ),
(5)
where b 𝑟𝑚 (s𝑖 ) is the uncalibrated prediction of model 𝑚. Parameter 𝛼𝑚 represents a global offset, while 𝛽𝑚 adjusts the RSSI scale. For received transmissions, the likelihood uses the observed RSSI. For non-receptions, it uses the probability that
DTwin ’26, October 26–30, 2026, Austin, TX, USA
Alexis Delplace, Xinyi Li, Samer Lahoud, and Nur Zincir-Heywood
Table 3: Global and spatial coverage-map metrics. Arrows indicate the preferred direction, and Gain denotes the relative improvement of HalifaxDT over Okumura– Hata. For the 30-point calibration rows, each entry reports the central value over 50 independent random calibration draws as mean ± standard deviation. MAE↓ (dB)
Cal.
Model
Raw Raw Gain
HalifaxDT 10.92 Okumura–Hata 20.66 +47.2%
V-RMSE↓ (dB2 )
FSS↑
IoU↑
BDist↓ (m)
11.24 16.10 +30.2%
0.840 0.679 +23.8%
0.664 0.503 +31.9%
220.4 469.5 +53.1%
30-pt HalifaxDT 5.58 ± 0.31 4.18 ± 1.42 0.876 ± 0.008 0.705 ± 0.014 152.5 ± 5.0 30-pt Okumura–Hata 6.61 ± 0.35 14.68 ± 0.17 0.802 ± 0.018 0.621 ± 0.021 339.2 ± 31.9 Gain +15.6% +71.5% +9.2% +13.6% +55.0%
Figure 2: Outdoor coverage map of Halifax: measurement-derived reference map (left) and Tobit-calibrated maps from HalifaxDT (center) and Okumura–Hata (right). the latent RSSI is below the sensitivity floor 𝑐 = −140 dBm: 𝑟𝑖 − 𝜇𝑖,𝑚 1 𝜙 , 𝑖 ∈ R, 𝜎 𝜎𝑚 𝜇𝑖,𝑚 = 𝛼𝑚 +𝛽𝑚b 𝑟𝑚 (s𝑖 ), 𝐿𝑖,𝑚 = 𝑚 𝑐 − 𝜇𝑖,𝑚 , 𝑖 ∈ C, Φ 𝜎𝑚 (6) where R and C denote received and censored non-received transmissions, respectively. This avoids treating every nonreception as an exact RSSI measurement at the sensitivity floor. With only 30 observations, an unconstrained affine calibration can overfit. We therefore estimate 𝜽𝑚 = (𝛼𝑚 , 𝛽𝑚 , 𝜎𝑚 ), with 𝜎𝑚 > 0, by minimizing a ridge-regularized Tobit negative log-likelihood, " # 30 𝛼 2 1 ∑︁ 𝑚 2 b 𝜽𝑚 = arg min − log 𝐿𝑖,𝑚 + 𝜆𝛼 + 𝜆𝛽 (𝛽𝑚 − 1) , 𝜽𝑚 30 𝑖=1 10 (7) with 𝜆𝛼 = 0.1 and 𝜆𝛽 = 1. These penalties shrink the offset toward zero and RSSI scale toward one, keeping calibration close to the unmodified map unless supported by the data. The calibrated raster reports the fitted latent RSSI mean after affine calibration, while the likelihood accounts for left censoring at −140 dBm. The calibrated comparison therefore emphasizes how well each map preserves the spatial organization of the measurement-derived reference. 3.1.3 Spatial Interpretation. Figure 2 provides a spatial view of the differences summarized in Table 3. The Okumura–Hata map follows a smooth radial attenuation trend but cannot reproduce site-specific shadowing and irregular coverage boundaries. In contrast, the HalifaxDT map exhibits spatial
variations aligned with those visible in the measurementderived map. We quantify this observation using global and spatial-comparison metrics in Table 3. The comparison also shows how scene fidelity affects the radio map. In the real southern park area, dense vegetation partly obstructs propagation, whereas HalifaxDT represents this area as relatively flat and open. The model therefore allows stronger direct and reflected paths between end-devices and the gateway, leading to RSSI overestimation. Two localized building-adjacent overestimation areas likely reflect similar missing vegetation losses. These errors reflect environment-model limitations and show where additional scene detail would most improve the digital twin. 3.1.4 Quantitative Comparison. All metrics in Table 3 use the same outdoor evaluation mask, which keeps grid cells within 300 m of at least one raw field-measurement record used to construct the reference map. For the evaluation subset, the mask contains 28,852 grid cells over 18.03 km2 , with 4,958 raw records and 2,903 occupied measurement cells, corresponding to 274.9 records/km2 , 161.0 occupied measurement cells/km2 , and direct support in 10.1% of evaluation cells. Across all available campaign records, the same region contains 7,435 records, including 4,926 received packets and 2,509 non-receptions. The mask therefore limits comparison to locally supported areas, ensuring that metrics are not dominated by regions where the reference map relies mainly on interpolation. MAE measures global pixelwise RSSI accuracy: HalifaxDT reduces raw MAE by 47.2% relative to Okumura–Hata, while after 50 repeated 30-point Tobit calibrations it remains 15.6% better and both models reach similar city-scale average error. This motivates the complementary spatial metrics below. As a non-interpolated sanity check, we evaluate receivedonly errors at the raw measurement locations used for the map comparison. Using 3,445 received records, calibrated
HalifaxDT: A Wireless Digital Twin from Open Geospatial Data
HalifaxDT obtains MAE/RMSE of 7.96/9.84 dB, while calibrated Okumura–Hata obtains 7.57/9.38 dB. Thus, HalifaxDT’s advantage in Table 3 should be interpreted as improved measurement-supported coverage-map structure, not uniformly better packet-level RSSI prediction. Okumura– Hata captures the dominant city-scale attenuation trend, but its smooth radial structure cannot reproduce local coverage boundaries and shadowed regions as well as HalifaxDT. Nonreceptions are excluded from this MAE/RMSE calculation and handled through censored Tobit calibration. To evaluate spatial organization beyond average RSSI error, we use four complementary metrics. Variogram RMSE (V-RMSE) compares predicted and reference empirical semivariograms over the same lag distances [9]; after calibration, HalifaxDT is 71.5% lower. Fractions Skill Score (FSS) [14] compares local covered-area fractions while tolerating small spatial displacements and is 9.2% higher. Intersection over Union (IoU) measures covered-area overlap [13] and is 13.6% higher. Boundary distance compares predicted and reference coverage contours through symmetric nearest-contour distance and is 55.0% lower. These metrics show that HalifaxDT better preserves the site-specific structure of the measured radio map. Without HalifaxDT smoothing, the uncalibrated map still outperforms Okumura–Hata in MAE, FSS, IoU, and boundary distance (12.30 dB, 0.845, 0.633, and 238.5 m), but V-RMSE increases to 63.17 dB2 , confirming its sensitivity to cell-scale spatial variation. Overall, HalifaxDT is markedly more accurate without RSSI calibration, showing that the digital-twin map requires substantially less empirical adjustment than the analytical baseline. Across repeated 30-observation calibrations, regularized censored calibration improves both models, yet HalifaxDT retains a consistent advantage across the spatial coverage metrics. The coverage-map use case therefore shows that HalifaxDT can provide measurement-supported spatial coverage predictions beyond what is captured by a calibrated analytical baseline.
3.2
Interference Detection
HalifaxDT can also provide a model-based reference for interference detection. Using the modeled environment and ray-traced channels, HalifaxDT estimates the signal that a sensing unit should receive from legitimate transmitters under normal operating conditions. This reference can be compared with real-world measurements [15]; deviations may indicate unexpected interference from malfunctioning equipment, unauthorized transmitters, or jammers. 3.2.1 Detection Method. We consider a scenario with multiple legitimate transmitters and a jammer. The received signal
DTwin ’26, October 26–30, 2026, Austin, TX, USA
at the sensing unit is modeled as 𝑦PT =
𝐾 ∑︁
𝐻𝑘 𝑥𝑘 + 𝐻 jam𝑥 jam + 𝑛
𝑘=1
where 𝐻𝑘 is the channel frequency response (CFR) between the 𝑘-th legitimate transmitter and the sensing unit, 𝑥𝑘 is the transmitted OFDM signal, and 𝑛 is additive white Gaussian noise. The digital twin predicts the interference-free received Í𝐾 signal as 𝑦DT = 𝑘=1 𝐻𝑘 𝑥𝑘 + 𝑛. Interference is detected by computing the residual power 𝑃𝜀 = ∥𝑦PT − 𝑦DT ∥ 2 . As shown in Fig. 3, the DT requires transmitter and sensingunit locations, transmitted signals, and transmit powers to predict the interference-free signal. These inputs can be obtained through 5G positioning and network-side transmitter information, but may be noisy in practice. Here, they are assumed exact to isolate detection performance under an idealized DT. A jammer is declared present when 𝑃𝜀 exceeds a predefined threshold.
Figure 3: System overview of jammer detection using digital twin. We compare the DT-based detector with conventional energy detection (ED), which uses received power 𝑃 𝑦 = ∥𝑦PT ∥ 2 without channel or environment knowledge. Both methods are training-free and require no labeled jammer data; legitimate samples are used only to calibrate the threshold. 3.2.2 Experimental Setup. We consider a 500 × 500 m subset of HalifaxDT with three fixed street-level legitimate transmitters and one rooftop sensing unit at (2990, 3190) m. Five jammer locations are randomly sampled and evaluated independently, yielding five test scenarios shown in Fig. 4. The carrier frequency is 3.7 GHz with 100 MHz bandwidth. OFDM signals use 64 subcarriers and 14 symbols per frame, with 0 dBm transmit power for both legitimate transmitters and the jammer. We consider eight jammer waveforms: continuous-wave jamming with amplitude variation (CWJ-A), continuous-wave jamming with frequency variation (CWJ-W), amplitude modulation jamming (AMJ), noise amplitude modulation jamming (NAMJ), narrowband noise jamming (NBNJ), multitone jamming (MTJ), linear frequency modulation jamming (LFMJ), and periodic pulse noise jamming (PPNJ), following the definitions in [3].
DTwin ’26, October 26–30, 2026, Austin, TX, USA
Alexis Delplace, Xinyi Li, Samer Lahoud, and Nur Zincir-Heywood
Scenario layout on HalifaxDT 3300
Building Transmitter (TX) Sensing Unit (SU) Jammer (J)
The DT-based detector outperforms energy detection at all five jammer positions. At the highest JNR (J2), both methods reach 100% TPR because the jammer signal is strong. At moderate JNR (J0, J1, J3), the DT-based method remains above 99% TPR, while energy detection ranges from 70.8% to 95.2%. At the lowest JNR (J4), the DT-based detector reaches 48.9% TPR, compared with 4.5% for energy detection. Its advantage comes from predicting the legitimate aggregate signal and testing the residual, whereas energy detection relies only on total received power and cannot separate legitimate and interference components.
TX1 SU
3200
y [m]
J1
3100
3000
J2
TX0 J3 J0
2900
TX2
J4
4
2800 2800
2900
3000
3100
3200
x [m]
Figure 4: Simulation scenario layout.
For both methods, the threshold is calibrated with 400 legitimate samples to target a 1% false positive rate (FPR). At each jammer location, the test set contains 500 legitimate samples and 500 jammer-contaminated samples for each of the eight jammer types, i.e., 4,500 samples per scenario: 500 legitimate and 4,000 jammer-contaminated. The physical environment is simulated with the full HalifaxDT ray-tracing scene: LoS and specular reflection are enabled, diffuse reflection and refraction are disabled, maximum interaction depth is 5, and a single vertically polarized isotropic antenna is used. The same scene generates both observations and the DT reference, so this setting is an upper bound: 𝑃𝜀 reflects detection under exact knowledge of 𝐻𝑘 and does not account for DT-model mismatch. The DT reference 𝑦DT also includes an independently drawn noise term 𝑛, so the residual reflects both jammer contribution and independent noise realizations. 3.2.3 Results. The detection performance is shown in Table 4. Table 4: Detection performance. Jammer
JNR (dB)
TPRDT
TNRDT
TPRED
TNRED
J0 J1 J2 J3 J4
7.9 1.0 17.2 4.0 −8.1
99.8% 100.0% 100.0% 100.0% 48.9%
97.4% 98.6% 99.0% 99.0% 98.8%
76.0% 70.8% 100.0% 95.2% 4.5%
99.2% 100.0% 98.8% 99.2% 98.0%
Mean
–
89.7%
98.6%
69.3%
99.0%
Limitations and Future Work
HalifaxDT has several limitations that inform future work. Some scene inputs are approximate or incomplete: the coveragemap evaluation uses a single concrete building material because city-scale material labels are sparse and coarse, and the 2018 LiDAR elevation data may miss recently constructed or modified buildings. Vegetation is also omitted, which can affect propagation accuracy, as observed in park areas. Finally, the current evaluation does not include independent RF measurement validation of the digital twin itself. In the interference-detection use case, the ground-truth received signal is also generated using the same ray-tracing scene, meaning the digital twin is assumed to represent the real environment. Validation against independent RF measurements and under DT-model mismatch is therefore an important direction for future work.
5
Conclusion
This paper presented HalifaxDT, a propagation-oriented wireless digital twin of the Halifax Peninsula built from open geospatial and spectrum data. Its reproducible workflow transforms heterogeneous 2D and elevation sources into a ray-tracing-ready urban scene with inferred building heights, terrain elevation, optional coarse material metadata, and gateway metadata. Through coverage mapping and interference detection, HalifaxDT shows how an environmentaware digital twin can provide spatial and signal-level information not captured by coarse analytical propagation models or received-power-only detection. The results also show that performance depends on scene fidelity: missing vegetation, simplified material representation, and limited RF calibration can create systematic errors, motivating measurementdriven calibration, richer environmental modeling, and tighter synchronization with live wireless observations.
Acknowledgments The authors gratefully acknowledge Noah Jehanno and Nathan Chappe for their contributions to the data collection and prototyping activities that supported this work.
HalifaxDT: A Wireless Digital Twin from Open Geospatial Data
References [1] Alexis Delplace, Samer Lahoud, and Kinda Khawam. 2025. Exploring LR-FHSS Modulation for Enhanced IoT Connectivity: A Measurement Campaign. In 2025 IEEE 102nd Vehicular Technology Conference (VTC2025-Fall). 1–7. doi:10.1109/VTC2025-Fall65116.2025.11309920 [2] Alexis Delplace, Xinyi Li, Samer Lahoud, and Nur Zincir-Heywood. 2026. HalifaxDT. https://github.com/AlexisDel/HalifaxDT. Online repository. [3] Xuhui Ding, Qi Lu, Yue Zhang, Gaoyang Li, Xiaozheng Gao, Neng Ye, Dusit Niyato, and Kai Yang. 2026. Few-Shot Recognition and Classification Framework for Jamming Signal: A CGAN-Based Fusion CNN Approach. IEEE Transactions on Vehicular Technology 75, 8 (2026), 17494–17508. doi:10.1109/TVT.2026.3678290 [4] Halifax Regional Municipality. n.d.. Buildings. Building polygons and associated building information. https://data-hrm.hub.arcgis.com/ datasets/HRM::buildings-1/about?layer=0 [5] Halifax Regional Municipality. n.d.. LiDAR DEM 5m 2018. Digital elevation model. https://data-hrm.hub.arcgis.com/documents/HRM:: lidar-dem-5m-2018/about [6] Halifax Regional Municipality. n.d.. LiDAR DSM 1m 2018. Digital surface model. https://data-hrm.hub.arcgis.com/documents/HRM:: lidar-dsm-1m-2018/about [7] Tomislav Hengl, Gerard B. M. Heuvelink, and David G. Rossiter. 2007. About Regression-Kriging: From Equations to Case Studies. Computers & Geosciences 33, 10 (2007), 1301–1315. doi:10.1016/j.cageo.2007.05.001 [8] Jakob Hoydis, Fayçal Aït Aoudia, Sebastian Cammerer, Merlin NimierDavid, Nikolaus Binder, Guillermo Marcus, and Alexander Keller. 2023. Sionna RT: Differentiable Ray Tracing for Radio Propagation Modeling. arXiv:2303.11103 [cs.IT] doi:10.48550/arXiv.2303.11103 [9] Mirko Mälicke. 2022. SciKit-GStat 1.0: A SciPy-flavored Geostatistical Variogram Estimation Toolbox Written in Python. Geoscientific Model Development 15 (2022), 2505–2532. doi:10.5194/gmd-15-2505-2022 [10] OpenStreetMap contributors. n.d.. OpenStreetMap. Map data. https: //www.openstreetmap.org/ [11] OpenStreetMap contributors. n.d.. Overpass API. Used to query OpenStreetMap data. https://overpass-api.de/ [12] Province of Nova Scotia. n.d.. Nova Scotia Geographic Data Directory. Nova Scotia Topographic Database buildings. https://nsgi.novascotia. ca/gdd/ [13] Hamid Rezatofighi, Nathan Tsoi, JunYoung Gwak, Amir Sadeghian, Ian Reid, and Silvio Savarese. 2019. Generalized Intersection over Union: A Metric and a Loss for Bounding Box Regression. In 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition. 658– 666. doi:10.1109/CVPR.2019.00075 [14] Nigel M. Roberts and Humphrey W. Lean. 2008. Scale-Selective Verification of Rainfall Accumulations from High-Resolution Forecasts of Convective Events. Monthly Weather Review 136, 1 (2008), 78–97. doi:10.1175/2007MWR2123.1 [15] Anton Schösser, Friedrich Burmeister, Philipp Schulz, Mohd Danish Khursheed, Sinuo Ma, and Gerhard Fettweis. 2025. Advancing Spectrum Anomaly Detection through Digital Twins. IEEE Communications Magazine 63, 4 (2025), 40–46. doi:10.1109/MCOM.001.2400221 [16] Susruth Sudhakaran, Javier Perez-Ramirez, Dave Cavalcanti, Cosmin Cazan, Nicholas Olson, Rafael Rosales, and Valerio Frascolla. 2024. Wireless Network Digital Twin Calibrated by Real Time Telemetry and XR Feedback Interface. In 2024 IEEE 20th International Conference on Factory Communication Systems (WFCS). 1–8. doi:10.1109/WFCS60972. 2024.10540718 [17] Paolo Testolina, Michele Polese, Pedram Johari, and Tommaso Melodia. 2024. Boston Twin: The Boston Digital Twin for Ray-Tracing in 6G
DTwin ’26, October 26–30, 2026, Austin, TX, USA Networks. In Proceedings of the 15th ACM Multimedia Systems Conference (MMSys 2024). Association for Computing Machinery, New York, NY, USA, 441–447. doi:10.1145/3625468.3652190