ClearCast: experimental underwater visibility estimate (Finistère)
EXPERIMENTAL. Not validated against any visibility observation in Brittany. No open Finistère horizontal-visibility dataset exists (Ifremer Surval Secchi stops in 2006; diver logs are not open).
Pipeline
- Forecast (
zsd_models.joblib): LightGBM quantile models (P10/P50/P90) predicting log of the 4 km Copernicus Marine daily L3 Secchi depth (ZSD, global multi-sensor) 1, 2, 3 days ahead at 12 Finistère points, from Open-Meteo wave/wind/rain features (up to day t), season, position and the last observed ZSD. - Visibility link: visibility ≈ 0.76 × ZSD, a fixed physical scaling checked against 6175 diver surveys of the IMOS/NRMN Reef Life Survey (8007 surveys since 2020, matched to the same-day satellite pixel). Uncertainty from the log residual quantiles (P10 ×0.48, P90 ×1.82), combined in quadrature with the forecast spread.
Forecast of ZSD, test 2026-01 → 2026-08 (MAE, log units)
| horizon | persistence | climatology | LightGBM P50 | skill vs persistence | P10–P90 coverage raw | after conformal |
|---|---|---|---|---|---|---|
| h1d | 0.203 | 0.258 | 0.192 | 5.4% | 82% | 88% |
| h2d | 0.233 | 0.257 | 0.213 | 8.5% | 83% | 88% |
| h3d | 0.259 | 0.256 | 0.232 | 10.6% | 82% | 86% |
ZSD → diver visibility (Reef Life Survey, 2020–2026, same-day matchups, n=6175)
- Spearman ρ(visibility, ZSD) ≈ 0.65 overall, 0.83 over the temperate subset (Canada, New Zealand, France, Spain, Italy; n=613).
- Temperate subset MAE: 4.8 m with visibility = 0.76·ZSD vs 7.9 m for a constant median; within ×1.5: 55% vs 32%.
- Cell-grouped CV (all data): MAE physics 5.0 m, LightGBM 4.5 m, median 6.7 m. LightGBM on optical features beats the physical rule in grouped CV but not in leave-one-country-out transfer, so the physical rule is used.
- Leave-one-country-out P10–P90 coverage of the residual interval (target 0.80):
| country | n | coverage | median log bias |
|---|---|---|---|
| Australia | 4652 | 0.84 | +0.02 |
| Belize | 60 | 0.80 | -0.11 |
| Canada | 124 | 0.59 | +0.27 |
| France | 59 | 0.71 | +0.09 |
| Italy | 50 | 1.00 | +0.10 |
| New Zealand | 157 | 0.76 | +0.07 |
| Panama | 54 | 0.85 | +0.09 |
| Spain | 223 | 0.78 | -0.37 |
| Tonga | 90 | 0.76 | +0.30 |
| United States | 273 | 0.66 | -0.18 |
| Wallis and Futuna | 73 | 0.96 | -0.12 |
Limitations
- Diver visibility is an integer-rounded subjective estimate; surveys at a site share satellite pixels and days (not independent). Mostly Australian data.
- 4 km pixels near coasts are coarse; Finistère coastal pixels may be contaminated or masked. Valid ZSD on about 60% of days.
- Gain over persistence for the ZSD forecast is only 5–11%. The 100 m satellite turbidity matched in-situ REPHY turbidity weakly (r=0.38; 0.17 in clear water), see
experiments/run1. - Training weather is reanalysis; live use feeds real-time analysis.
Source code: github.com/cpoisson/clearcast (currently private, to be opened after review). Licences: Copernicus Marine (free, attribution), Open-Meteo (CC BY 4.0), RLS/IMOS NRMN (CC BY 3.0 AU; cite Edgar & Stuart-Smith 2014, Sci Data 1:140007).