03 · Verification C1 · 2023-11-30 — free forecast vs radar truth, control vs radar-cycled WRF

Setup. After DA cycling (12–15 Z, 4 hourly 3D-Var cycles), the pipeline releases a free forecast starting at 15 Z (valid to 12 Z+1, 21 h) on d01→d02→d03 via ndown. Both branches — control and radar-cycled — share the same GFS 12Z lineage (fair baseline: no cold start, no spin-up mixing).

The event was a weak afternoon convective episode, 16–21 Z. The d03 box (centred −21.99°, −41.31°) landed in a region with only 10 rain-gauge stations — station scores at d03 are empty; the signal lives on d02 and in the MODE/object-based metrics.

21-h accumulation scatter d02
21-h accumulated rainfall scatter, d02. MAE 20.5→14.6 mm (−29%).
Radar blind spots. The truth field here is a single radar, and ~19% of d03 is blocked to it (W/NW behind the Serra do Mar, plus man-made dead sectors). See the beam-blockage study — found two ways, the radar's own persistently dead sectors and a terrain beam-blockage simulation — for where rain can fall unseen and why those gates are masked out of the assimilation and the verification truth field.

Interactive CREF viewer

Obs × control × radar DA — composite reflectivity, per domain, per hour

Open viewer

Tables (from MET verification CSVs)

Stations — d02 (window / threshold)

windowthrPOD c→rCSI c→rBIAS c→rhits c→r

Figures

Scatter accum d02
Accumulation

Scatter — d02

21-h MAE 20.5 → 14.6 mm (−29%).

Scatter accum d03
Accumulation

Scatter — d03

6.6 → 5.3 mm. Sample limited (gauge-poor box).

HiRA BSS d02
HiRA

BSS — d02

HiRA BSS d03
HiRA

BSS — d03

ΔBSS d02
HiRA

ΔBSS — d02

ΔBSS d03
HiRA

ΔBSS — d03

ROC AUC d02
HiRA

ROC AUC — d02

ROC AUC d03
HiRA

ROC AUC — d03