Drought-impact prediction benchmark (terraflow.drought)¶

Predict insured drought loss (USDA RMA Cause of Loss) from within-season climate/vegetation anomalies — a decision-relevant impact target, distinct from drought severity (USDM D2+) and crop yield benchmarks. This notebook walks the fetch → build → evaluate workflow.

See the guide for task details.

1. Configuration¶

Point feature_table at the flashdry feature_table.parquet and rma_dir where the RMA Cause of Loss ZIPs live (fetch them with terraflow drought fetch).

In [ ]:

from terraflow.drought.config import DroughtConfig

cfg = DroughtConfig(
    states=['17', '18', '19', '27', '29', '31'],  # 6-state Corn Belt
    crop='CORN', year_min=2000, year_max=2023,
    cutoff_doy=212,                # early-warning: signal available by ~Jul 31
    test_years=[2012, 2017, 2022, 2023], train_max_year=2015,
    rma_dir='data/drought/rma',
    feature_table='PATH/TO/flashdry/data/processed/feature_table.parquet',
    output_dir='outputs/drought_v0',
)
cfg.years[:5], cfg.years[-3:]

from terraflow.drought.config import DroughtConfig cfg = DroughtConfig( states=['17', '18', '19', '27', '29', '31'], # 6-state Corn Belt crop='CORN', year_min=2000, year_max=2023, cutoff_doy=212, # early-warning: signal available by ~Jul 31 test_years=[2012, 2017, 2022, 2023], train_max_year=2015, rma_dir='data/drought/rma', feature_table='PATH/TO/flashdry/data/processed/feature_table.parquet', output_dir='outputs/drought_v0', ) cfg.years[:5], cfg.years[-3:]

2. Build the benchmark table¶

Parses RMA Cause of Loss → drought-loss labels, aggregates predictors up to the cutoff, and joins them into benchmark.parquet with a deterministic build fingerprint.

In [ ]:

from terraflow.drought.dataset import assemble_benchmark

benchmark = assemble_benchmark(cfg, write=True)
print(len(benchmark), 'county-years;', benchmark.GEOID.nunique(), 'counties')
print('positive rate:', round(benchmark.significant_drought_loss.mean(), 3))
benchmark[['GEOID', 'year', 'drought_loss_ratio', 'significant_drought_loss']].head()

from terraflow.drought.dataset import assemble_benchmark benchmark = assemble_benchmark(cfg, write=True) print(len(benchmark), 'county-years;', benchmark.GEOID.nunique(), 'counties') print('positive rate:', round(benchmark.significant_drought_loss.mean(), 3)) benchmark[['GEOID', 'year', 'drought_loss_ratio', 'significant_drought_loss']].head()

3. Evaluate the leaderboard¶

Naive, severity-only, and Ridge/RF/GBM climate models across the temporal and spatial regimes.

In [ ]:

from terraflow.drought.evaluate import run_leaderboard

report = run_leaderboard(benchmark, cfg, write_dir=cfg.output_dir)
for model, m in report['temporal']['classification'].items():
    print(f"{model:26s} ROC-AUC={m['roc_auc']:.3f}  PR-AUC={m['pr_auc']:.3f}")

from terraflow.drought.evaluate import run_leaderboard report = run_leaderboard(benchmark, cfg, write_dir=cfg.output_dir) for model, m in report['temporal']['classification'].items(): print(f"{model:26s} ROC-AUC={m['roc_auc']:.3f} PR-AUC={m['pr_auc']:.3f}")

Findings (v0)¶

• Within-season signal predicts insured drought loss well (ROC-AUC ~0.91-0.93).
• USDM severity is a strong baseline; climate anomalies match/slightly beat it and arrive earlier.
• Temporal extrapolation to extreme years (2012) is the hard regime: tree models collapse while linear models hold — the benchmark's most interesting open challenge.