Model Overview
OYPT-CA uses a Linear Mixed-Effects (LME) model to predict almond yield from satellite-derived vegetation and moisture indices. The model uses previous year's data (lag-1) to predict the current year's harvest — meaning predictions are available before the growing season begins.
The LME framework accounts for orchard-level variability through random intercepts, allowing the model to learn each orchard's baseline productivity while sharing information about how spectral features drive year-to-year changes across all orchards.
Prediction Features
Each variety uses a simple set of 2–3 spectral indices plus orchard age:
- NDVI (June) — Normalized Difference Vegetation Index. Measures overall canopy greenness and photosynthetic activity. Used for NP and MO.
- CIRE (June) — Chlorophyll Index Red Edge. Captures photosynthetic capacity via red-edge reflectance. Used for WC.
- NDMI (August) — Normalized Difference Moisture Index. Late-season canopy moisture status and water stress. Used for all varieties.
- Orchard Age — Trees ramp up to age 8-12, then gradually decline. Used for all varieties.
- Last Year Yield (optional) — If known, the previous year's yield can be included. This is most helpful for Wood Colony (coefficient 0.28), with smaller effects for NP and MO.
Why Simple Models?
We tested 13 feature combinations (from single-index to 4-index sets) across 7 model types (LME, Ridge, Lasso, RF, Gradient Boosting, SVR) for each variety. We also compared against complex models with 170+ features from 17 spectral indices, temporal trends, and climate data.
Key finding: Simple 2–3 feature models significantly outperformed complex models on the 2025 holdout test. LME with BLUP was the most consistent performer for known orchards.
Model Performance
Trained on Orchard Futures Initiative (OFI) data, 2020–2025, California Central Valley. Evaluated on 2025 holdout test:
| Variety | Features | Test R² | RMSEP | Train R² | Orchards |
|---|---|---|---|---|---|
| NP | NDVIJun + NDMIAug + age | 0.626 | 333 | 0.826 | 92 |
| WC | CIREJun + NDMIAug + age | 0.707 | 461 | 0.625 | 67 |
| MO | NDVIJun + NDMIAug + age | 0.364 | 480 | 0.775 | 68 |
Confidence Intervals
The prediction interval reflects: (1) residual model error and (2) unknown orchard-level effects. For new orchards (not in the training set), the interval is wider because orchard-specific adjustments cannot be applied.
Data Sources
- Sentinel-2 L2A (Copernicus) — 10–20m spectral indices
- OFI Dataset — UC Davis, ~170 orchards, 2019–2025
Limitations
- Predictions use fixed effects only (no orchard-specific BLUP adjustments for unknown orchards).
- Trained on Central Valley almonds only.
- Cloud cover may cause missing satellite data for June or August.
- MO variety has lower prediction accuracy (R²=0.36) — use with caution.
- Use the confidence interval alongside your own field knowledge.
Model version: March 2026. Coefficients updated annually.