Blue outline: selected tree's model crown envelope
Leaf-on vs leaf-off / 着葉・落葉を同じ条件で比較
Pause and capture this epoch. Both cases use the same receiver, satellite endpoints, trees and wood. Crown loss is a model component, not measured signal strength or positioning accuracy.
Ready to capture a paired comparison.
Live conditions changed. This result remains the captured comparison; run again to update.
Selected ground path · model relative power
Only the selected satellite's shortest valid ground path is compared. Missing paths remain unavailable. Fixed bounce loss is held constant.
Signal
Leaf-on dB
Leaf-off dB
Δ off − on dB
Geometry
Crown loss only. WOOD/TERRAIN still block the direct path in both cases; their total power is unavailable. No measurement-noise or receiver-tracking model is added.
Artificial single-path ranges: straight, robust straight, and latent direct/ground-path models. Known terrain + 1.6 m receiver height. Uniform random arrival selection is not receiver tracking. The matched-map case is an ideal self-consistency test, not field accuracy.
Uses displayed GPS satellites; at least three feasible observations required. Running pauses playback. Spatial search is local, ±12 m around a deliberately perturbed prior.
Model
Horizontal error
Clock error
Status
Inference receives ranges, source endpoints, prior and map only; true position/path labels are used separately for generation and scoring. All methods retain the same observation set. CPU binary64 in a cancellable worker; WebGL renders the scene. Route compares both density profiles at the selected time step, with all-GNSS attenuation/path counts and fixed-σ GPS positioning. Density is not used as a noise or tracking model, so density alone cannot establish positioning gains. Export includes assumptions, search limits, generation truth, failures and all results.
Model & limits
¹ In linear comparison mode, illustrative canopy loss = specific loss × integrated crown density × foliage multiplier. The synthetic radial/height profile is rho0 × (1 − 0.65 |u|²) × (1 + 0.45 u_z) / 0.61 inside each ellipsoid. It has the same total mass as uniform density and is integrated analytically along each chord. Crown path is the union length; densities add where crowns overlap. These coefficients are uncalibrated and frequency independent. This is not an implementation fitted to ITU-R P.833, C/N0, code delay, a receiver model or a positioning-accuracy result. The extinction reference uses the fitted sigma_t values in ITU-R P.833-10 Table 8: silver maple at 1.3 GHz (0.241 m⁻¹ in leaf, 0.139 m⁻¹ out of leaf), Korean pine at 1.5 GHz (0.215 m⁻¹) or Himalayan cedar at 1.5 GHz (0.271 m⁻¹). Power loss is 10 / ln(10) × sigma_t × relative-density path integral. Bare broadleaf crown volumes retain fine-branch extinction while their leaf meshes remain absent; evergreen references retain foliage. Relative spatial density and the growth-form-to-species mapping are synthetic. The reference has no fitted water-content or continuous seasonal law. Published frequency anchors are held fixed without interpolation or per-satellite carrier assignment. This is only the ballistic extinction component, not the full RET or slant-tree solution, diffuse/scattered gain, diffraction or calibrated RHCP GNSS power. Its use on both reflected legs is an unvalidated transfer. Foliage scaling and manual specific loss are disabled for the reference; density profiles remain synthetic sensitivity controls. Wood intersections and terrain obstruction make total direct power unavailable. Broadleaf crowns are rounded ellipsoid density volumes; evergreen conifers use five vertically disjoint tapering ellipsoid tiers and whorled branches. The same component volumes and wood cylinders drive rendering, direct rays and both reflected legs. Leaf-off removes broadleaf leaf meshes and, in linear mode, leaf-density loss; the extinction reference retains the out-of-leaf crown coefficient; evergreen foliage and all wood remain. These growth forms and binary leaf-retention assumptions are synthetic, not species/season calibration. Crown densities are not mass-matched across tree types. Trunks and branches are finite cylinders; drawing tessellation is used only for display. Leaf branch cards are distributed around each density envelope; their cutout silhouettes are display approximations, not RF boundaries. Both attenuation modes share exact crown/wood geometry and use CPU binary64 integrals; the reference does not modify positioning noise, visibility or path selection. Botanical leaf and needle cutouts, tapered fine branches, root flares, fallen wood, stones, moss, litter, understory, atmospheric horizon and filtered sunlight shadows are visual approximations; no new RF attenuation or occlusion is assigned to them. The trail view follows the same completed route epoch from 6 m behind and 1 m above the receiver so incoming rays remain visible. The physical antenna remains 1.6 m above terrain. Direction labels can lie at the viewport edge; geometric signal paths are drawn over the scene, with blocked alternatives dashed. The near clipping plane is 0.08 m. The selected satellite also has finite-triangle, single-bounce ground paths. Both complete segments are checked against the loaded terrain and wood; canopy density is integrated along both. Occluded candidates are retained separately in exports. Only the shortest valid path is drawn; this does not imply the receiver tracks it. Bounce loss is a constant uncalibrated power loss. The satellite endpoint is held at the direct-flight-time broadcast state; reflected flight times are not iterated. No coherent summation, polarization, bark reflection, diffraction, diffuse scattering or receiver tracking is implemented.
The forest, terrain and 1.6 m-high walking route are synthetic and generated from the exported seed/parameters. Only the GNSS broadcast orbits and GPST span come from the UrbanNav 2018-12-19 dataset used by the city lab. The ENU patch is placed at that dataset's geodetic anchor; it is not an observed forest at that location. Terrain is piecewise planar within the finite patch, and geometry outside it is not modeled. Geometry uses CPU binary64; WebGL uses Float32 only for rendering.