CAL: Calibrate reflectivity coverage thresholds in codebook

[rcjackson]

Summary

Recalibrates the three quantitative reflectivity-coverage reclassification thresholds in CODEBOOK.md (Section 3.1) against 1,212 hand-labelled CSAPR2 scenes (Bankhead National Forest). These thresholds drive the automatic label overrides applied by criteria_from_codebook / apply_criteria_to_labels.

Rule (source → reclassify target)	Field	Old	New	Basis
No Precipitation → Isolated Convection	pct_gates_50dbz	0.002%	0.005%	Old value overrode ~3% of genuine No-Precip scenes (observed max 0.0045%). pct_gates_50dbz does not separate these two classes, so this rule only guards against gross mislabels; raised just above the full observed No-Precip envelope.
Stratiform → Mesoscale Convective System	pct_gates_50dbz	0.02%	0.035%	Youden-optimal cut vs MCS (J=0.72). False overrides of genuine stratiform drop 13% → 6%, still catching ~79% of MCS.
Isolated Convection → Mesoscale Convective System	pct_gates_30dbz	1.0%	1.3%	Youden-optimal cut vs MCS (J=0.70). False overrides 16% → 13% at equivalent catch.

Method

For the two rules where the source and target classes genuinely separate on the field (Stratiform→MCS, Isolated→MCS), thresholds are set at the Youden-J optimal cut (maximizing true catch − false override). For No-Precip→Isolated Convection the field does not discriminate the classes (Isolated Convection's median 50 dBZ coverage ≈ 0), so the threshold is set just above the observed No-Precip envelope to avoid overriding genuine No-Precip scenes.

Verification

• criteria_from_codebook('CODEBOOK.md') parses all three new thresholds with correct reclassify targets.
• Applying old vs. new criteria to the ground-truth labels via apply_criteria_to_labels: spurious hard-overrides of correctly-labelled scenes drop from 82 (6.8%) → 43 (3.5%).
• Added a v1.1 changelog entry.

🤖 Generated with Claude Code