# Observation reconciliation — gridded vs station Tx

Compared 189 station Tx records against the nearest 0.25 deg gridded cell over
2002-2023 (common days only; 29 Feb
dropped; EHD via diagnostic min(Tx90p,35) climatology). Full numbers in
`results/tables/obs_reconciliation_*.csv`.

## Per-region summary (station minus gridded)

 region     region_name  n_stn  elev_mean  bias   mae  rmse  corr  ehd_freq_stn  ehd_freq_grid  ehd_freq_diff
      1      R1 Tay Bac     19    629.816 2.242 2.949 3.258 0.937         0.421          0.109          0.312
      2     R2 Dong Bac     48    304.408 1.633 2.160 2.402 0.977         0.362          0.099          0.263
      3    R3 DB Bac Bo     26     30.962 1.521 1.677 1.883 0.984         0.311          0.103          0.208
      4 R4 Bac Trung Bo     30     78.000 1.541 1.842 2.102 0.976         0.330          0.122          0.208
      5 R5 Nam Trung Bo     16     17.625 1.989 2.081 2.287 0.905         0.503          0.109          0.394
      6   R6 Tay Nguyen     21    567.333 1.928 2.010 2.196 0.934         0.477          0.103          0.374
      7       R7 Nam Bo     29     23.641 1.670 1.795 2.011 0.832         0.587          0.119          0.468

## Decision

- **Calibration target = gridded Tx** (`Tx_EC_0.25_2002-2023.nc`). The forecast product is
  gridded over all 541 land cells and every lead block; only the gridded obs provides the
  spatially complete, temporally aligned target the calibration and verification need.
- **Station Tx = independent validation.** Stations are used to audit the gridded product at
  point scale, especially where the two disagree.
- **Where they disagree, station is weighted for point realism, gridded for spatial
  coverage.** The disagreement is dominated by a **general warm offset**: stations read on
  average ~1.5-2.2 degC warmer than their 0.25 deg cell in every region (figure panel a),
  with elevation-dependent scatter (a few high peaks, ~1500 m, run colder). Day-to-day
  correlation stays high (0.83-0.98), so this is a systematic representativeness/offset gap,
  not a timing error. Because the shared threshold is derived from the cooler gridded field,
  the warmer stations exceed it far more often, so **station EHD frequency (0.31-0.59) is 2-5x
  the gridded (~0.10)** (panel c). Consequence for the product: skill and thresholds are fit
  and verified in the gridded world; the station comparison bounds how far point-scale EHD
  risk may exceed the gridded product, and is reported as a limitation.

## Documented data-quality handling

- Missing values (`-99.0`) are dropped, never imputed. Per-station missing fractions are in
  `results/tables/station_metadata.csv`; stations with <365 matched days are excluded here.
- Station periods and lengths vary; matching is done on the intersecting dates only.
- Obvious inhomogeneities (station relocation, instrument change) are not corrected; flagged
  as a limitation for point validation.