Phase 2-D.1 · Hybrid Production Model
2024 Holdout — Out-of-Sample Predictions
Per-state routing across two regressors. Iowa & Colorado pull from a Ridge linear probe over the 1024-D Prithvi-EO-2.0-300M embedding (foundation-model linear probe + XGBoost). Missouri, Nebraska, Wisconsin use the engineered baseline. The hybrid was trained on 2013–2022, validated on 2023, and the 2024 yields shown below were never seen during training.
263
Counties forecast
5
States covered
170 bu/ac
5-state avg
18.65
EOS RMSE (val 2023)
+6.50%
Lift vs canonical Row A
2024
Holdout year
2024 Holdout — Hybrid Model State Forecasts
Per-state routing · Prithvi-aware (Row B) for CO, IA · engineered (Row A) elsewhere
201bu/ac
Actual 2024: 211-10
Row A 202 · Row B 201
Top county: Woodbury, IA
220 bu/ac predicted
73 counties · 18 above actual
EOS RMSE — hybrid 18.15 · Row A 16.49
180bu/ac
Actual 2024: 184-4
Row A 180 · Row B 181
Top county: Phelps, NE
215 bu/ac predicted
60 counties · 24 above actual
EOS RMSE — hybrid 14.78 · Row A 14.78
159bu/ac
Actual 2024: 162-3
Row A 159 · Row B 158
Top county: Lafayette, WI
197 bu/ac predicted
59 counties · 26 above actual
EOS RMSE — hybrid 16.12 · Row A 16.12
161bu/ac
Actual 2024: 175-14
Row A 161 · Row B 158
Top county: Atchison, MO
195 bu/ac predicted
53 counties · 11 above actual
EOS RMSE — hybrid 24.64 · Row A 24.64
148bu/ac
Actual 2024: 131+17
Row A 143 · Row B 148
Top county: Delta, CO
217 bu/ac predicted
18 counties · 13 above actual
EOS RMSE — hybrid 37.57 · Row A 32.6
Hybrid uses Row B (Ridge probe on 1024-D Prithvi + engineered) for CO, IA; Row A (engineered only) for the others. 2024 yields were withheld from training.
Model Comparison — Val 2023 EOS
RMSE in bu/acre (lower = better) · Hybrid is the production model
canonical reference
engineered only
Prithvi-aware
production
Per-Forecast-Date Validation
Val 2023 RMSE per stage · n_train, n_val are county-level rows
| Stage | Hybrid | Row A | Row B | n_train | n_val |
|---|---|---|---|---|---|
| Aug 1 | 17.81 | 17.78 | 18.11 | 3,069 | 308 |
| Sep 1 | 20.31 | 20.92 | 20.93 | 3,069 | 308 |
| Oct 1 | 20.79 | 20.82 | 20.50 | 3,069 | 308 |
| Final | 18.65 | 19.01 | 19.42 | 3,069 | 308 |
Hybrid Predictions vs Actual — 2024
State-mean predicted vs USDA NASS final estimate · bu/acre
2024 Holdout Predictions — County Level
263 countiesHybrid model · 80% CI ≈ ± hybrid EOS RMSE · sorted by predicted yield · actuals shown when reported
| # | County | State | Hybrid | Actual | Δ | 80% CI | Row A · Row B |
|---|---|---|---|---|---|---|---|
| 1 | Woodbury, IA | Iowa | 220 | 206 | +14 | 202–239 | 216 · 220 |
| 2 | Marshall, IA | Iowa | 219 | 222 | -3 | 200–237 | 217 · 219 |
| 3 | Fremont, IA | Iowa | 218 | 231 | -13 | 199–237 | 211 · 218 |
| 4 | Delta, CO | Colorado | 217 | 188 | +29 | 199–236 | 211 · 217 |
| 5 | Mesa, CO | Colorado | 217 | 202 | +15 | 198–236 | 213 · 217 |
| 6 | Cherokee, IA | Iowa | 217 | 207 | +10 | 198–236 | 206 · 217 |
| 7 | Jones, IA | Iowa | 217 | 233 | -16 | 198–235 | 217 · 217 |
| 8 | Plymouth, IA | Iowa | 217 | 196 | +21 | 198–235 | 210 · 217 |
| 9 | Sac, IA | Iowa | 217 | 229 | -12 | 198–236 | 216 · 217 |
| 10 | Dubuque, IA | Iowa | 216 | 222 | -6 | 197–234 | 224 · 216 |
| 11 | Clinton, IA | Iowa | 215 | 236 | -21 | 196–234 | 216 · 215 |
| 12 | Phelps, NE | Nebraska | 215 | 199 | +16 | 196–233 | 215 · 217 |
How the Hybrid Model Works
A frozen-feature-extractor pipeline with per-state routing
- CDL corn masks — annual binary corn rasters (5 states × 12 years, 2013–2024) sourced from USDA CropScape and reprojected to EPSG:5070 / 30 m. Annual masks avoid corn-soy rotation contamination.
- HLS chips on /mnt/d/ — 43,724 corn-richest 224×224 chips extracted from cap=48 cleanest cloud-cover granules per (state, year, phase). Stored on a 1.9 TB external SSD; ~1.4 TB cumulative download.
- Prithvi-EO-2.0-300M-TL frozen embedding — 1024-D mean-pooled feature vector per (county, year, forecast_date), 14,560 rows total.
- Ridge linear probe — α=10.0, fit on chip-bearing 2013–2022 rows. Compresses the 1024-D embedding into a single yield-prediction scalar that XGBoost consumes alongside engineered features. Every Prithvi dim contributes via the linear weights.
- Per-state routing — at predict time, route by
state_alpha: CO & IA → Row B (engineered + Ridge probe); MO, NE, WI → Row A (engineered only). Selected from the cap=48 ablation breakdown — Prithvi adds noise where corn fraction is uniform across years. - Gate result — hybrid val 2023 EOS RMSE 18.65, vs canonical Row A 19.95 = +6.50% lift. Threshold +5%; PASS.