Pilot scope — from 2026-04-15 call

Which EVI feature best predicts yield?

We evaluate candidate signals against real Agreste yields. Higher |r| means a stronger relationship; sign matters for interpretation.

Takeaway

Peak EVI is the wrong signal — negative correlation in Coop 63 (r=−0.30) means high peak EVI can coincide with low seed maize yield (non-seed maize types may be overrepresented in RPG random sampling). EVI Integral (area under the seasonal curve, r=+0.62 within Coop 63) and Half-Max DOY (r=−0.56, earlier = better) are much stronger candidates, consistent with Konstantin's insight on the call.

Heat, rain, GDD vs. yield

Growing-season aggregates (Apr-Sep) from ERA5-Land via the ARCO point-query endpoint, at the centroid of Puy-de-Dôme. Climate is now the single strongest signal we have — heat-day count captures drought years (2019, 2022) without depending on satellite cloud cover.

Peak EVI & EVI Integral vs. Agreste seed maize yield

2022 dropped — no vegetation-class Sentinel-2 observations (Massif Central cloud cover).

Vegetation-class EVI by day-of-year (mean across sampled fields)

Only SCL=4 (vegetation-classified) pixels shown; sparse points in some years reflect persistent cloud cover over the Massif Central.

EVI Integral vs. yield index across 82 French départements

Each point is one département-year (2017–2024). Colors are by dept; the red cluster highlights Coop 63 (dept 63).

Ridge regression performance by held-out year

Using total-maize yield anomaly as target. R² negative in most years reflects the fundamental challenge: EVI features alone cannot predict yield anomaly at dept level without irrigation and climate covariates. This motivates adding ERA5 climate features (growing-degree-days, drought index) and parcel-level targeting via Coop 63 contract IDs.

Data sources

• Agreste SAA Séries Longues 2010–2024 — département-level yields for Maïs semence (seed) and Maïs grain+semence (total). 91 départements × 15 years.
• RPG 2023 (IGN) — 707,724 maize parcels nationally; 8,084 in Coop 63 (28,899 ha). Parcel IDs preserved.
• Sentinel-2 L2A — Earth Search STAC, cloud-filtered (SCL=4 vegetation only for peak EVI), 199,551 observations extracted.
• ERA5-Land (in progress) — daily 2010–2024, aggregated to monthly + growing-season GDD / precip / drought-index / heat-days.

Deliverables in S3 bucket

s3://digifarm-limagrain-yield/
├── agreste/maize_yield_dept.csv          Yield data (91 depts × 15 years)
├── rpg/2023/maize_parcels.parquet         707K maize parcels + IDs
├── samples/centroids.parquet              240K sampled field centroids
├── evi/evi_timeseries.parquet             199K Sentinel-2 observations
├── outputs/
│   ├── features.parquet                   Aggregated dept × year × signal
│   ├── model/yield_model.pkl              Trained model
│   └── reports/coop63_calibration.md      This report
└── client_docs/                           Vivadour contracts + 2025 polygons

Code repository

github.com/digifarmio/limagrain-yield (private)