Refine MRMS skill note and add printable comparison

Clarify evaluation and metric descriptions, emphasize StormScope Flash rows, and provide a printable PDF. Score values, registry, checkpoint weights, baseline assets, and the main branch remain unchanged.

mrms_forecast_skill.md

@@ -1,57 +1,34 @@
 # StormScope Flash: MRMS forecasting skill
 
-This note compares the 3 km, 10-minute StormScope baseline with its distilled
-StormScope Flash variant for composite radar reflectivity. Higher scores are better;
-1 is perfect. StormScope values are the published baseline scores; Flash values
-come from the archived full-domain evaluation.
+We evaluate composite radar-reflectivity forecasts from StormScope and its distilled StormScope Flash variant at 3 km spatial and 10-minute temporal resolution. Higher scores indicate greater forecast skill; a score of 1 denotes a perfect forecast.
 
 ## Evaluation
 
-The evaluation uses 929 WoFS-catalog initializations from March 4 through August 9,
-2025. Verification is observed MRMS composite reflectivity on the 3 km model grid
-at matching valid times. Flash runs coupled GOES and MRMS/GLM forecasts over the
-full 1024 x 1792 domain, then selects each catalog region for scoring. Forecasts
-contain 18 ten-minute leads and 16 ensemble members; GOES and MRMS use five and
-seven network calls per lead, respectively.
+The evaluation comprises 929 cases initialized between 4 March and 9 August 2025. Both models generate coupled GOES and MRMS/GLM forecasts over the full 1024 × 1792 CONUS grid. Forecasts are extracted over the corresponding evaluation regions and verified against observed MRMS composite reflectivity on the 3 km grid at matching valid times. Each forecast comprises 16 ensemble members and 18 ten-minute lead times, extending to 180 minutes. StormScope Flash requires five GOES and seven MRMS/GLM network evaluations per forecast step.
 
-Object CSI and neighborhood CSI use member 0, not the ensemble mean. pFSS uses all
-16 members. Contingency counts and pFSS numerator/denominator terms are summed
-over cases at each lead before computing the score; these are not averages of
-individual case scores.
+Deterministic metrics use the first ensemble member (member 0); probabilistic metrics use all 16 members.
 
 ## Metrics
 
-- **Object CSI** measures storm detection using connected regions at or above
-  40 dBZ, each containing a 45 dBZ core and at least 12 grid cells (108 km²).
-  One-to-one matching uses centroid and boundary distances with 40 km cutoffs,
-  area-ratio weighting, five matching passes, and total interest greater than 0.2.
-  CSI = hits / (hits + false alarms + misses). Redundant unmatched objects
-  classified as extras are excluded. The table pools all storm ages.
-- **CSI at 40/55 dBZ** is neighborhood-grid CSI, not object CSI or exact-pixel CSI.
-  Forecast and observation exceedances are each dilated with an 11 x 11 cell
-  (33 x 33 km) window before forming the same contingency ratio.
-- **pFSS at 40/55 dBZ** compares neighborhood exceedance fractions. Forecast
-  fractions are averaged across the 16 members and compared with the observed
-  fraction: pFSS = 1 - sum((forecast fraction - observed fraction)²) /
-  sum(forecast fraction² + observed fraction²). The window is also 11 x 11 cells.
-
-## Results
+**Object CSI.** Storm objects are identified as connected regions with reflectivity ≥40 dBZ, an embedded ≥45 dBZ core, and an area of at least 12 grid cells (108 km²). Forecast and observed objects are matched one-to-one based on centroid and boundary distances, using 40 km cutoffs, area-ratio weighting, and a total interest score >0.2. Object-based CSI is then calculated as H / (H + F + M), where H denotes hits, F false alarms, and M misses.
+
+**Neighborhood CSI at 40 and 55 dBZ.** Forecast and observed threshold-exceedance fields are each dilated using an 11 × 11-cell (33 × 33 km) window before calculating CSI from the resulting contingency counts.
+
+**Probabilistic Fractions Skill Score (pFSS).** Threshold-exceedance fractions are calculated over the same 11 × 11-cell window. Forecast fractions are averaged across the 16 members and compared with the observed fractions: pFSS = 1 − Σ(f − o)² / Σ(f² + o²), where f and o denote forecast and observed neighborhood fractions.
+
+## Forecast skill
 
 | Metric | Model | +30 min | +60 min | +120 min | +180 min |
-|---|---|---:|---:|---:|---:|
+|:--|:--|--:|--:|--:|--:|
 | Object CSI | StormScope | 0.685 | 0.533 | 0.335 | 0.220 |
-| Object CSI | StormScope Flash | 0.688 | 0.532 | 0.315 | 0.204 |
+| Object CSI | **StormScope Flash ⚡** | 0.688 | 0.532 | 0.315 | 0.204 |
 | CSI 40 dBZ | StormScope | 0.597 | 0.479 | 0.348 | 0.271 |
-| CSI 40 dBZ | StormScope Flash | 0.593 | 0.464 | 0.326 | 0.246 |
+| CSI 40 dBZ | **StormScope Flash ⚡** | 0.593 | 0.464 | 0.326 | 0.246 |
 | CSI 55 dBZ | StormScope | 0.274 | 0.170 | 0.068 | 0.028 |
-| CSI 55 dBZ | StormScope Flash | 0.286 | 0.167 | 0.060 | 0.024 |
+| CSI 55 dBZ | **StormScope Flash ⚡** | 0.286 | 0.167 | 0.060 | 0.024 |
 | pFSS 40 dBZ | StormScope | 0.907 | 0.788 | 0.588 | 0.449 |
-| pFSS 40 dBZ | StormScope Flash | 0.913 | 0.800 | 0.593 | 0.433 |
+| pFSS 40 dBZ | **StormScope Flash ⚡** | 0.913 | 0.800 | 0.593 | 0.433 |
 | pFSS 55 dBZ | StormScope | 0.719 | 0.502 | 0.201 | 0.064 |
-| pFSS 55 dBZ | StormScope Flash | 0.735 | 0.490 | 0.137 | 0.036 |
-
-These results describe the archived evaluation, not a new evaluation of the
-repackaged Earth2Studio checkpoints. Baseline and Flash scoring conventions were
-not fully harmonized, and statistical equivalence was not assessed.
+| pFSS 55 dBZ | **StormScope Flash ⚡** | 0.735 | 0.490 | 0.137 | 0.036 |
 
-[Download table values](mrms_forecast_skill.csv).
+[Download PDF](mrms_forecast_skill.pdf) · [Download table values](mrms_forecast_skill.csv).