✈ Contrail Formation Predictor

python predictor.py --latlon 52.52 13.40 "Berlin, DE"

Click a column header to sort; type above to filter.

Each pixel = 1 hour. Hover for time + status.

T = temperature at pressure level • RH = relative humidity w.r.t. ice • [N/3] = persistent level count

6/7 cities show higher high-cloud agreement for PERSISTENT than EPHEMERAL. Aggregate: PERSISTENT = 73% vs EPHEMERAL = 49% (Δ = +24pp, ex-London). Miami is a special case: tropical convective cirrus saturates both classes at ~85%. This layer cannot separate contrail cirrus from natural cirrus — see the GOES-19 satellite layer below, which can.

When we predict PERSISTENT, GOES-19 finds a linear contrail overhead 67% of the time, vs 49% for EPHEMERAL (Δ = +18pp). 6/7 cities show the correct ordering — the same score the independent METAR layer reaches, from a different instrument.

Null test: the classes are very unequal (54 PERSISTENT vs 352 EPHEMERAL), so a gap this size could in principle be small-sample noise. Shuffling the predicted labels against the observations 20,000 times reproduces a gap this large in p = 0.010 of shuffles — so the signal is unlikely to be chance.

Honest caveat: the city-level rates above (does this city-hour show a linear contrail overhead, yes/no) are a coarser measure than pixel-level precision/recall, so they are not restated as calibrated numbers here. The same detector has since been calibrated pixel-by-pixel against OpenContrails human labels — see the calibration panel below (6.3% precision, 1.9% recall, 32× lift over chance at the shipped threshold) — so the detector overall is no longer uncalibrated, even though this panel's own city-level rates are not the calibrated figures. Only the relative ordering between classes at this city-level granularity is claimed from this panel alone. Los Angeles is the outlier in both layers — June marine-layer subsidence leaves it with almost no upper-level ice cloud for either class to work with.

The panel above establishes that PERSISTENT scenes score higher than EPHEMERAL ones. It never established that what the detector flags is a contrail — so the standing caveat was “uncalibrated; only the class ordering is meaningful.” This panel replaces that caveat with a measured number by scoring the detector directly against all 151 human-labelled scenes (40 containing contrails) in the OpenContrails validation shards on disk — OpenContrails (Ng et al. 2023, CC BY 4.0). Labels are pixel-level; we score the aggregated ground truth (a pixel counts as contrail when >50% of annotators marked it).

What this says, plainly. At the shipped operating point the detector reaches 6.3% precision and 1.9% recall (F1 0.030). In absolute terms that is a weak segmenter — it finds a small minority of contrail pixels and most of what it flags is not one. But contrail pixels are only 0.20% of all pixels, so that precision is 32× better than chance. Both halves of that sentence are the result: the detector is emphatically not a random number generator, which is what licenses the class-ordering claim above — and it is also nowhere near good enough to call a contrail segmenter.

How sure are we? Bootstrapping over scenes (10,000 resamples with replacement, seed 0) puts a 95% confidence interval of 2.0–11.3% on precision and 0.7–3.2% on recall. The interval is wide — contrail pixels are rare, so the numbers ride on the 40 labelled scenes that contain one — but even the lower precision bound of 2.0% is about 10× the 0.20% base rate, so “far above chance” survives the uncertainty; “a good segmenter” does not.

It behaves like a real detector. Sweeping the anisotropy threshold traces a clean precision/recall trade-off — precision climbs monotonically from 1.1% to 100% as the threshold tightens, while recall falls. A noise source has no such curve. That monotonic ordering, not any single point, is the strongest single piece of evidence that the line filter is keying on real linear structure.

Where the signal comes from. The brightness-temperature candidate gate alone scores 0.14% precision — at or below the 0.20% base rate, i.e. no better than guessing. The oriented line filter supplies essentially all of the discriminative power. That is a load-bearing detail: the physics gate finds cold thin ice, which over CONUS is mostly ordinary cirrus; only the anisotropy test asks the question that separates a contrail from the cirrus around it.

Where the recall goes. The candidate gate keeps only 25% of human-labelled contrail pixels before the line filter ever runs, so it — not the line filter — is the ceiling on recall. The gate demands cold (BT13 < 273 K) and thin-ice (BTD > 0.5 K); a young or semi-transparent contrail over warm ground satisfies neither cleanly. Raising recall means loosening the physics gate, not sharpening the anisotropy test.

Not results-shopping. The +18pp separation above was measured at the shipped threshold (1.0) and is not restated at any other value. Best F1 over the sweep is 0.035 at threshold 0.75; that is reported as an observation for future work, not retro-fitted to the published headline.

Honest caveat — GOES-16 labels, GOES-19 detector. OpenContrails labels GOES-16 scenes; the dataset predates GOES-16's retirement as GOES-East on 2025-04-07. Our detector runs on GOES-19, which took the same 75.2°W slot with the same ABI instrument and the same bands — so the calibration should transfer, but it was measured on GOES-16 and applied to GOES-19. These are not numbers measured on our own scenes.

Session 28 — a second satellite (GOES-West / GOES-18). Western and Pacific-facing cities (LA, Seattle, Phoenix, San Francisco, Portland, Salt Lake City, San Diego, Las Vegas) sit deep in the GOES-East limb, where a fixed-degree box foreshortens badly. They are now scored by GOES-18 (137.2°W), which gives them near-nadir geometry. Each city is assigned to its best-geometry satellite by the same derived limb test that keeps Tokyo out — never a hand-kept list. G18 carries the identical ABI instrument, product and C13/C15 bands, so the detector is band-for-band applicable, and as of Session 29 the transfer is measured, not assumed: the same city-hours in the FOV overlap (8 western cities, 2026-07-18 to 2026-07-24) were scored through both satellites with the shared calibration — Cohen's kappa 0.64 (95% CI [0.57, 0.71]), concordance 83.0%, n=448 paired city-hours, detection rate G19 43.5% vs G18 31.5%. That is at least moderate agreement at the interval floor — the transfer holds on real scenes, with the interval still to tighten. West cities enter as new-roster strata and accrue forward-only, so they accelerate future evidence rather than tightening today's CI.

The satellite panels above are a one-time check over a fixed June window. This one turns that into an ongoing, append-only track record: every day a scheduled job takes the forecast that was already archived before that day (a genuine forecast with real lead time, not a hindcast) and scores it against what the calibrated GOES-19 detector actually observed over each CONUS city's footprint. It grows one day at a time.

Read this as small-sample, weak-detector evidence — because it is. The detector is a weak segmenter (~6% pixel precision; see the calibration panel), so a single day's per-city detection bool is noisy. Daily samples are a handful of cities × a few hours, so early numbers carry almost no statistical power — only the accumulated contrast means anything. At cruise altitude in summer nearly every hour forecasts EPHEMERAL and NONE is essentially never forecast, so the literal PERSISTENT-vs-NONE table is near-empty; the useful signal is whether the detection rate rises from EPHEMERAL to PERSISTENT. And the detector is calibrated on GOES-16 labels but run on GOES-19.

Cumulative over 48 day(s) (2026-06-28 → 2026-08-29) — how often GOES saw a linear contrail, by forecast class:

The detection rate rises from 37% (EPHEMERAL, n = 4467) to 66% (PERSISTENT, n = 951), a +29pp gap (bootstrap 95% CI [+26, +33]pp) that is now statistically distinguishable from zero — two-sided Fisher's exact p < 0.001, and the bootstrap gap CI excludes zero. That pooled number is confounded, though — see the city-stratified check below, which is the one to believe.

New-roster watch (since 2026-07-22): 29 additional GOES-East cities joined the roster but hold no honest history — a forecast must predate its observation, so they can only validate forward via the daily timer, never by backfill. So far they have accrued 38 obs day(s) and 3576 city-hour(s), and on the current sample they are corroborating it — folding them in lifts the stratified gap (a shift of +1.6pp). This line updates itself from scorecard.json as the new cities accumulate.

Per-day record (detected / forecast, newest first):

The full 2×2 (PERSISTENT vs not-PERSISTENT), per-day samples, and lead times are in scorecard.json. Refreshed daily by the contrail-scorecard systemd timer; no manual step.