0003. VLM gate: per-station vertical rates vs GNSS up¶
- Status: accepted (owner, 2026-10-13): median and NMAD gate, RMSE reported
- Changes:
requirements.VLM_GATE(rate_abs_median_max_mm_yr = 1.0,rate_nmad_max_mm_yr = 2.0,min_stations = 20);disp2vlm-validate vlm.
Context¶
VLM (opera_vlm) writes one product per target pair (T0, Tk). The owner's
target (2026-10-11): the VLM vertical rate within 1-2 mm/yr of GNSS.
First real runs (Venti plan T56.1, cal_disp_e2e/t56, t57):
| run | stations (both looks) | median | NMAD | RMSE (mm/yr) |
|---|---|---|---|---|
| LA F16940 + F18904, 2016-2023 | 172 | −0.42 | 1.44 | 2.67 (1.67 without 6 local outliers) |
| Houston F08882 + F38238 (edge case) | 52 | +0.43 | 2.48 | 3.51 |
| NYC F08622 + F30175 (projection only, data gap) | 34 | −0.60 | 1.29 | 4.86 |
Decision¶
- The gate is on station rates (slope of the VLM vertical series minus
the slope of GNSS up at the same epochs), not on per-pair displacements:
the per-pair error carries the shared
T0and per-epoch atmosphere, which σ covers (opera_vlmsigma_rate) and rates average. - Gated stations: both looks (mode 1), outside deforming areas and event footprints (as decision 0002). Projected (mode 2) stations and the excluded classes are reported.
- Limits: |median| ≤ 1 mm/yr and NMAD ≤ 2 mm/yr, on at least 20 stations. RMSE is reported, not gated: a handful of local sites (landslides, wells, coast) set it (LA: 2.67 with, 1.67 without six of them).
- σ realism (coverage of |z| < 1.96 and robust z spread per span) is reported for T57.
Consequences¶
LA passes; Houston fails on NMAD (edge case, final refinement pass); NYC is not assessed in the both-looks class (its descending stack has a gap). RMSE stays reported, not gated (owner, 2026-10-13). A handful of local sites set it, and an outlier rule would itself need a decision.