Surfaces and materials#

The ground of a scene is split into non-overlapping semantic surfaces, each mapped to a documented Sionna material. Where a class is placed and which RF material it receives are separate: Overture features decide the former, OWRT’s material profile the latter.

Surface classes#

Class

ID

Sionna material

Interpretation

Ground

0

itu_medium_dry_ground

Generic exposed ground

Vegetation

1

itu_wood

Wood as a reproducible proxy; foliage volume loss is not modelled

Paved

2

itu_concrete

Concrete as a proxy for paved surfaces

Water

3

p527_fresh_water, p527_sea_water

ITU-R P.527 water at 20 °C: fresh inland, sea water (35 g/kg) for bodies that touch an ocean, sea, bay or strait feature

Ground, vegetation and paved surfaces use ITU-R P.2040 surrogates; water uses the ITU-R P.527 water model, whose permittivity and conductivity follow the scene frequency. Sea water has its own mesh, mesh/terrain_sea_water.ply. Each mapping is recorded with a surrogate confidence label. material_profile="itu" applies this table, on flat ground as well as on terrain, so a flat scene with the ITU profile still distinguishes water and paved surfaces. On a terrain grid, material_profile="uniform" assigns the ground class everywhere; without terrain it compiles no class grid. The uniform profile uses concrete for every surface, building roofs included.

Sionna models every surface as a slab of its material with air on both sides. Terrain and water surfaces bound solid ground and water, so the scene declares their materials as 10 m layers (terrain-itu_medium_dry_ground, terrain-itu_wood, terrain-itu_concrete, terrain-p527_fresh_water, terrain-p527_sea_water). This is far thicker than the penetration depth of any of these materials between 1 and 10 GHz: the surfaces reflect as half-spaces and pass negligible power into the space under the terrain. Building walls (itu_concrete) and roofs (itu_metal) keep Sionna’s default 0.1 m slab. Scenes record these materials as RF material profile version 3.

Sionna RT has no water material, so OWRT registers owrt-water-material when it loads Sionna (owrt.load_scene and the radio-map engines). To open a compiled scene.xml with plain sionna.rt.load_scene, call openworld_radio_twin.simulation.water_material.register_water_material() after importing sionna.rt.

How classes are assigned#

Overture land_cover, land_use, water and transportation segment features are normalised into one aligned grid at the terrain mesh spacing using cell-centre classification. Evidence precedence is:

  1. water geometry,

  2. explicit road surface,

  3. inferred road surface,

  4. explicit land-use surface,

  5. physical land cover,

  6. semantic land use,

with ties resolved by the stable class ID. Overture class and subtype are evaluated independently against the published schema. Road width_rules are interpreted as edge-to-edge widths in metres and, together with scoped road_surface rules, split at their linear-reference boundaries. Missing road widths and surfaces use recorded deterministic road-class assumptions, so footways do not inherit motor-road dimensions and gravel or dirt roads remain ground.

What is exported#

Every scene with an environment grid writes:

  • terrain/surface_classes.npy, the cell labels consumed by the solver;

  • terrain/scene_info.json, with the class-to-material mapping, profile version, the semantic-source release, the class/subtype conversion table, per-basis feature counts, the road-width and surface assumptions, and the source licence summary.

Elevation arrays are grid nodes, while surface classes label the cells between them: an (N, M) elevation grid has (N - 1, M - 1) class cells. Both run south to north by row and west to east by column. The exported raster is exactly the geometry Sionna sees; no semantic polygon is exported separately.

Selective masking at solve time#

The Sionna RT solve can exclude individual surface classes without recompiling the scene. include_ground, include_vegetation, include_paved and include_water in engine_config remove the corresponding meshes from the active scene while the source meshes and provenance stay in the dataset. Roof and wall switches apply the same operation to buildings. See Sionna RT.