Pilot scope
Parcel boundaries · click a row to zoom the map
Bright outlines = inner-buffered zonal-stats polygon (used for EVI extraction). Faint dashed = full LPIS block (sum of all dissolved components). Pin = farmer-supplied centroid.
Per-parcel summary · click a row to zoom
| Parcel | Dept | Producer / commune | Contracted (ha) | Block (ha) | Buffered (ha) | Variety | 2026 sown | Status |
|---|
Per-parcel EVI history 2017–2024 (zonal mean over polygon)
Mean EVI per Sentinel-2 scene over the buffered polygon (SCL ∈ {4 vegetation, 5 bare soil, 6 water, 7 unclassified}). 2017 sparse (pre-S2B). Yearly maize peaks visible around DOY 210–230 in better years.
Within-year tier matrix vs dept-33 random sample
Each parcel-year is ranked against ~30 random dept-33 maize parcels for that year and mapped to a tier band on the Agreste 2010–2024 yield quintiles (Q20/40/60/80). T0 = severe (≤ Q20) … T4 = none (> Q80). Color = dept-33 percentile (red = worst, green = best in the dept-33 sample for that year). Tier rank uses centroid-pixel signal vs the dept-33 calibration sample (also centroid-based) so the comparison is apples-to-apples; the EVI curves above use polygon zonal-mean (cleaner per-parcel signal).
Per-département Agreste seed-maize yield reference (2010–2024)
Dept 33 (Gironde): mean 37.4 q/ha, 2022 worst at 31 q/ha. Dept 82 (Tarn-et-Garonne): mean 34.9 q/ha, 2022 worst at 28 q/ha. Both dept's 2022 was the SW France drought.
Ground-truth yields (Geoffroy 2026-05-11)
Parcel-level actual yields and `atteinte objectif` (target ratio) from the broker for dept-33. Loss bands: T0 ≥30%, T1 20-30%, T2 10-20%, T3 0-10%, T4 over-target. Rotation crops (potato, grain maize) shown for context but don't feed the maize calibration.
Predicted target ratio vs actual (DigiFarm FB · peak_evi linear fit)
Linear fit of zonal-mean peak_evi → target ratio on the 5 dept-33 maize-semence parcel-years (2024+2025). r and slope come from `outputs/reports/marifloor2026_signal_to_ratio.json`. n=5 is small; "défaut de fertilité sur les mâles" (the 2024 male-sterility issue on D3+D2) is not EVI-visible — flagged as the main known calibration miss.
Retro-prediction for the 5 dept-33 contracted parcels (Ask 1 — Geoffroy 2026-05-12)
Predicted T0–T4 per parcel × year 2017–2024, treating each year as "as-if maize-semence" and applying the peak_evi → target_ratio linear fit calibrated on the 2024+2025 GT (n=5, r=+0.55 on DigiFarm FB). 2024 and 2025 also show the actual tier from Geoffroy's GT for comparison. Pre-2017 = Landsat (to follow).
As-if portfolio — 30 random RPG maize parcels per maize-semence dept (Ask 2)
For each département in the Marifloor contract (33, 46, 47, 64, 81, 82): 30 random RPG-2024 maize parcels, EVI extracted 2017–2024, peak_evi → predicted target_ratio applied. Cells show % of the 30 parcels predicted to over-perform (T3 mild + T4 none) in that year — higher = better year for that dept's maize. Caveat: "seed maïze" not separable in RPG, so we use all maize as a proxy.
Calibration honesty note
Year-pooled signal→yield linear fit on dept-33 has LOYO r ≈ −0.89 across all three EVI signals on the 6 available years (2019–2024). The 2022 drought year had average EVI but bad yield (31 q/ha) — classic late-season grain-fill stress that EVI misses. Therefore the per-parcel tiers above use the within-year percentile approach, not a yield prediction. This matches the hybrid-policy structure where EVI is a guardrail on declared loss, not a primary parametric trigger. v1.0 calibration will use parcel-level historical yield records (still outstanding from Limagrain) to break the tie.
Data & reproducibility
- Polygons:
s3://digifarm-limagrain-yield/polygons/marifloor2026_pilot_polygons.geojson - Zonal EVI:
s3://digifarm-limagrain-yield/evi/marifloor2026_zonal_evi.parquet(1,401 rows, Saga 17978637) - Calibration sample:
s3://digifarm-limagrain-yield/evi/marifloor2026_dept33_calib_parcel_evi.parquet(~210 parcel-yrs, Saga 17974907) - Per-parcel zonal features:
s3://digifarm-limagrain-yield/evi/marifloor2026_pilot_zonal_features.parquet - Tier table:
s3://digifarm-limagrain-yield/outputs/reports/marifloor2026_pilot_per_parcel_tiers.csv - Methodology: Marifloor 2026 first-pass GDoc