PropSplatMap-Free RF Field Reconstruction via
3D Gaussian Propagation Splatting

William Bjorndahl, Maninder Pal Singh, Farhad Nouri, Joseph Camp

Southern Methodist UniversityUniversity of Houston

Abstract

Building a site-specific propagation model typically requires either ray-tracing over detailed 3D maps or dense measurement campaigns. Both approaches are expensive and often infeasible for rapid deployments where geographic data is unavailable or outdated. We present PropSplat, a map-free propagation modeling method that reconstructs radio frequency (RF) fields using 3D anisotropic Gaussian primitives. Each Gaussian encodes a scalar path loss offset relative to an explicit baseline path loss model with a learnable path loss exponent. Gaussians are initialized along observed transmitter–receiver paths and optimized end-to-end to learn the propagation environment without external information like floor plans, terrain databases, or clutter data. We evaluate PropSplat against wireless radiance field methods NeRF2, GSRF, and WRF-GS+ on two real-world datasets. On large-scale outdoor drive-tests spanning multiple topographical regions at six sub-6 GHz frequencies, PropSplat achieves 5.38 dB RMSE when training measurements are spaced 300 m apart and outperforms WRF-GS+ (5.87 dB), GSRF (7.46 dB), and NeRF2 (14.76 dB). On indoor Bluetooth Low Energy measurements, PropSplat achieves 0.19 m mean localization error, an order of magnitude better than NeRF2 (1.84 m), while achieving near-identical received signal strength prediction accuracy. These results show that accurate site-specific propagation reconstruction is achievable from sparse RF-native measurements. The need for geographic data as a prerequisite for scalable RF environment modeling is reduced.

PropSplat concept: learned Gaussian ellipsoids add or subtract path loss along a transmitter-receiver path; a waterfall plot combines these contributions with the baseline. Buildings are illustrative context and are not model inputs.
Fig. 1. Learnable Gaussian primitives model local propagation effects. Buildings are shown for context; PropSplat requires no map, building, or terrain data.

Outdoor RF field reconstruction

Ofcom drive-test measurements cover seven UK locations and six frequencies from 449 to 5850 MHz, spanning urban, suburban, rural, and mountainous environments. Training samples are spaced approximately 300 m apart, using less than 1% of the measurements.

London drive-test reconstruction at 5850 MHz. Panels show 375 training measurements spaced roughly 300 metres apart, predicted versus measured path loss, and held-out prediction errors along routes extending up to 8 kilometres from the transmitter.
Fig. 2. London at 5850 MHz: 375 training measurements are selected from more than 130,000 drive-test points. The panels show the sparse measurements, reconstructed field, and held-out test errors. View vector figure (PDF)

Results. PropSplat achieves 5.38 dB path loss RMSE, averaged across all city–frequency pairs. On the same 300 m spatial training split, WRF-GS+ achieves 5.87 dB, GSRF 7.46 dB, and NeRF² 14.76 dB.

Impact. Accurate coverage reconstruction from sparse measurements could reduce drive-test effort and support crowdsourced network optimization, including locations where building maps and terrain data are unavailable.

Indoor BLE signal modeling and localization

The public NeRF² Bluetooth Low Energy dataset contains measurements from 6,000 transmitter positions and 21 fixed gateways in a 15,000 sq. ft. nursing home. PropSplat predicts received signal strength (RSSI) and generates fingerprints for indoor positioning.

Two example indoor transmitter-to-gateway paths through learned Gaussians. Red Gaussians add attenuation and blue Gaussians reduce loss. Two plots trace cumulative predicted RSSI along each path, showing how local corrections bring the baseline prediction toward the measured signal strength.
Fig. 3. Each Gaussian's contribution can be traced along a propagation path. These two examples show opposite corrections to the baseline: 17.7 dB of added loss for TX₁ and a −22.3 dB correction for TX₂. The plots illustrate individual predictions; aggregate test results are reported below. View vector figure (PDF)

Results. Mean localization error is 0.19 m with a 70/30 train–test split and 0.44 m with only 758 training positions (12.6% of the dataset). NeRF² reaches 1.84 m and 7.33 m on the same respective splits. PropSplat's RSSI prediction RMSE is 4.74 dB with dense training and 6.51 dB with sparse training.

Impact. Similar signal-strength accuracy can lead to very different positioning accuracy: dense RSSI RMSE is 4.74 dB for PropSplat and 4.68 dB for NeRF². PropSplat preserves useful differences between nearby locations, supporting precise indoor positioning even with sparse measurements and no floor plan.

Citation

@INPROCEEDINGS{11571104,
  author={Bjorndahl, William and Singh, Maninder Pal and Nouri, Farhad and Camp, Joseph},
  booktitle={2026 IEEE International Symposium on Spectrum Innovation (DySPAN)},
  title={PropSplat: Map-Free RF Field Reconstruction via 3D Gaussian Propagation Splatting},
  year={2026},
  volume={},
  number={},
  pages={217-224},
  keywords={Modeling;Measurement;Equations;Training;Neural radiance field;Propagation;Radio frequency;Location awareness;Testing;Received signal strength indicator;RF field reconstruction;3D Gaussian splatting;sparse measurements;path loss prediction;machine learning},
  doi={10.1109/DySPAN69846.2026.11571104}}