HalifaxDT: A Wireless Digital Twin from Open Geospatial Data

1 Introduction

• HalifaxDT complements simulation platforms, calibrated propagation models, and city-scale scenes by combining open-data scene reconstruction with sparse LoRaWAN field-measurement evaluation.The paper identifies Sionna RT as its ray-tracing engine and BostonTwin as an example of detailed geometry when available.Source: §1 · pp. 1–2
• Many cities lack detailed 3D models that can be directly converted into wireless simulation scenes, while building-height sources vary in coverage, age, alignment, and height definition.The sources include elevation rasters, explicit height attributes, floor-count estimates, and external topographic records.Source: §1 · p. 2
• HalifaxDT reconstructs a ray-tracing-ready Halifax Peninsula scene from publicly available geospatial and spectrum data rather than an existing 3D city model.It integrates footprints, LiDAR-derived elevation, building attributes, and spectrum-licence information, and exports a Sionna RT-compatible scene.Source: §1 · p. 2
• The workflow combines heterogeneous building-height evidence while preserving provenance, then evaluates the HalifaxDT instance through coverage mapping and interference detection.The stated contributions include a reusable construction workflow and jammer detection without labeled jammer data.Source: §1 · p. 2

2 HalifaxDT Construction Methodology

• HalifaxDT follows a reusable four-stage workflow: identify geospatial inputs, align sources, convert them into a ray-tracing-ready scene, and integrate that scene with a wireless simulator.The paper states that the procedure can be adapted to other cities where similar open or institutional sources are available.Source: §2 · p. 2

2.1 Geospatial Inputs and Source Discovery

• The input layers include building footprints, elevation products, and building metadata, which are spatially aligned and transformed into a common metric coordinate system.Table 1 summarizes each layer's role, typical sources, and corresponding HalifaxDT data.Source: §2.1 · p. 2, Table 1 · p. 2
• Co-registered DSM and DEM products provide building height above local terrain, while HalifaxDT supplements these candidates with explicit heights, floor counts, and external records.The DSM–DEM approach depends on elevation-layer availability and resolution, temporal consistency, and spatial alignment.Source: §2.1 · p. 2, §2.2 · p. 3

2.2 Preprocessing and Height Inference

• For each footprint, the DSM-derived candidate uses the 95th percentile of normalized DSM–DEM samples to approximate the upper roof surface with less sensitivity to isolated outliers.Other source-derived candidates are scored; agreeing plausible estimates are combined by weighted median, otherwise the highest-scored plausible candidate is selected, with metadata-based fallback when direct evidence is unreliable.Source: §2.2 · pp. 3–4
• The height-selection output records its source or consensus, decision strategy, confidence, candidates, and rationale, supporting inspection and later correction.Most buildings retain DSM-derived heights, while 14.5% have a final selected height more than 2 m from DSM-P95.Source: §2.2 · pp. 3–4

2.3 Scene Generation for Wireless Simulation

• Scene generation uses a 5 m DEM for terrain, extrudes height-enriched footprints with bases following local elevation, and assigns approximate RF material classes.The 5 m terrain choice balances mesh complexity against limited benefit from a 1 m surface; coverage evaluation uses one building material to avoid uncertainty from unvalidated material differences. The protocol can also be applied per spatial tile in larger cities.Source: §2.3 · p. 4

2.4 HalifaxDT Instance and Sionna Integration

• The exported scene links terrain and building meshes with material definitions in a Mitsuba-compatible XML file that Sionna RT loads for radio maps or propagation paths.The same environment scene can support multiple analyses with experiment-specific radio configurations kept separate.Source: §2.4 · p. 4

3.1.1 Coverage-Map Construction.

• The reference map treats non-receptions as censored at −140 dBm and uses regression kriging to combine a distance-and-coordinate trend with locally kriged residuals.Residual weights use up to 18 neighboring measurement cells within 300 m; the reference is an interpolation product, not independent ground truth.Source: §3.1.1 · p. 5
• Sionna RT generates HalifaxDT predictions on a 25 m outdoor grid, while Okumura–Hata provides an analytical baseline evaluated with matched frequency, gateway, antenna, and RSSI-floor settings.The ray-tracing setup uses geometry and terrain with simplified materials; HalifaxDT maps alone receive spatial smoothing.Source: §3.1.1 · p. 5

3.1.2 Limited-Data Calibration.

• Fifty random draws of 30 campaign records calibrate both maps with a ridge-regularized Tobit model that accounts for censored non-receptions at −140 dBm.Penalties shrink the global offset toward zero and RSSI scale toward one, limiting adjustment unless supported by data.Source: §3.1.2 · pp. 5–6

3.1.3 Spatial Interpretation.

• Okumura–Hata’s smooth radial trend misses site-specific shadowing and irregular boundaries, whereas HalifaxDT variations align with the measurement-derived map.Missing vegetation in the southern park leaves the modeled area relatively open, contributing to RSSI overestimation and indicating where added scene detail could help.Source: §3.1.3 · p. 6
• HalifaxDT MAE is 47.2% lower than Okumura–Hata raw and remains 15.6% better after 50 repeated 30-point Tobit calibrations; city-scale average errors are similar.The evaluation mask limits metrics to cells within 300 m of a raw field-measurement record.Source: §3.1.4 · p. 6

3.1.4 Quantitative Comparison.

• HalifaxDT obtains MAE/RMSE of 7.96/9.84 dB, while calibrated Okumura–Hata obtains 7.57/9.38 dB; its advantage reflects measurement-supported coverage-map structure, not uniformly better packet-level RSSI prediction.Non-receptions are excluded from this MAE/RMSE calculation and handled through censored Tobit calibration.Source: §3.1.4 · p. 7
• After calibration, HalifaxDT has 71.5% lower V-RMSE, 9.2% higher FSS, 13.6% higher IoU, and 55.0% lower boundary distance than Okumura–Hata.Together, these metrics indicate better preservation of the measured map’s site-specific spatial structure.Source: §3.1.4 · p. 7

3.2.1 Detection Method.

• The detector compares the received signal with a digital-twin prediction of the legitimate aggregate signal and declares a jammer when residual power exceeds a threshold.Unlike energy detection, which uses received power alone, the residual test uses the predicted legitimate signal to distinguish legitimate and interference components.Source: §3.2.1 · p. 7

3.2.2 Experimental Setup.

• The idealized evaluation assumes exact transmitter and sensing-unit inputs and uses the same ray-tracing scene for observations and the digital-twin reference, excluding model mismatch.The residual also reflects independently drawn noise in the observed and predicted signals.Source: §3.2.1 · p. 7, §3.2.2 · p. 8
• The simulation uses three fixed street-level legitimate transmitters, one rooftop sensing unit, and five independently sampled jammer locations in a 500 × 500 m HalifaxDT area.Thresholds target a 1% false positive rate using 400 legitimate samples; each location is tested with 500 legitimate and 4,000 jammer-contaminated samples.Source: §3.2.2 · pp. 7–8, Figure 4 · p. 8

3.2.3 Results.

• At the lowest JNR (J4), the DT-based detector reached 48.9% TPR versus 4.5% for energy detection, and it outperformed energy detection at all five positions.At moderate JNR (J0, J1, J3), DT-based TPR remained above 99%, while energy detection ranged from 70.8% to 95.2%; both reached 100% at the highest JNR (J2).Source: §3.2.3 · p. 8, Table 4 · p. 8

4 Limitations and Future Work

• The coverage evaluation uses one concrete building material because city-scale material labels are sparse and coarse, and 2018 LiDAR may miss newer or modified buildings.These input constraints limit how fully the scene represents the current built environment.Source: §4 · p. 8
• Omitted vegetation can affect propagation accuracy, as observed in park areas.Source: §4 · p. 8
• The digital twin lacks independent RF measurement validation, and the interference study assumes its ray-tracing scene represents the real environment.Validation under digital-twin model mismatch remains future work.Source: §4 · p. 8

• The reproducible workflow converts heterogeneous 2D and elevation sources into a ray-tracing-ready scene with inferred building heights, terrain, optional coarse material metadata, and gateway metadata.Source: §5 · p. 8
• Coverage mapping and interference detection show that the environment-aware twin provides spatial and signal-level information not captured by coarse analytical models or received-power-only detection.Source: §5 · p. 8
• Missing vegetation, simplified materials, and limited RF calibration can create systematic errors, motivating measurement-driven calibration, richer environmental modeling, and synchronization with live observations.Source: §5 · p. 8