Geo2SigMap

Overview

Geo2SigMap is an efficient framework for high-fidelity RF signal mapping leveraging geographic databases, ray tracing, and a novel cascaded U-Net model. Geo2SigMap features a scalable, automated pipeline that efficiently generates 3D building and path gain (PG) maps via the integration of a suite of open-source tools including OpenStreetMap (OSM) and Nvidia's Sionna Library. Geo2SigMap also features a cascaded U-Net model, which is pre-trained on pure synthetic datasets leveraging the building map and sparse signal strength (SS) map as input to predict the full SS map for the target (unseen) area. The performance of Geo2SigMap has been evaluated using large-scale field measurements collected using three types of user equipment (UE) across six LTE cells operating in the Citizens Broadband Radio Service (CBRS) band deployed on the Duke University West Campus. Our results show that Geo2SigMap achieves significantly improved root-mean-square error (RMSE) in terms of the SS map prediction accuracy compared to existing baseline methods based on channel models and ML.

Our repository has been divided into two primary components:

• Scene Generation: A pure Python-based pipeline for generating 3D scenes for arbitrary areas of interest. This new Python-based pipeline replaces the scene generation pipeline used in our DySPAN'24 paper, and is more scalable, efficient and user-friendly.
• ML-based Propagation Model: ML-based signal coverage prediction using our pre-trained model based on the cascaded U-Net architecture, also described in our DySPAN'24 paper.

Our newest release v2.0.0 has enhanced the original scene generation pipeline by incoporating LiDAR terrain data and downstream Digital Elevation Models (DEM). Data was sourced from the USGS 3D Elevation Program (https://www.usgs.gov/3d-elevation-program). The tool does not depend on this data for its operation. In areas with insufficient data, a flat terrain can still be assumed.