Aurum — AI Corn Yield
Forecasting at County Scale
Multi-source geospatial AI pipeline fusing NASA HLS satellite imagery, PRISM climate grids, and USDA SSURGO soil surveys to issue 4-stage growing season yield forecasts with calibrated uncertainty — delivered 4–8 weeks before official USDA NASS estimates.
0.789
Model R²
12.5 bu/ac
MAE
20,536
Training Records
4–8 wks
Forecast Lead Time
Powered by
NASA
HLS Satellite · Prithvi
USDA
NASS Yield · NRCS gSSURGO
AWS
SageMaker Training
Google Earth
Earth Engine · NDVI Export
Esri
GIS Infrastructure
Databricks
Data Platform
NASA
HLS Satellite · Prithvi
USDA
NASS Yield · NRCS gSSURGO
AWS
SageMaker Training
Google Earth
Earth Engine · NDVI Export
Esri
GIS Infrastructure
Databricks
Data Platform
Model R²
0.789
Train 2005–22 · Val 2023
MAE
12.5 bu/ac
Mean absolute error
Counties
263
5 states
Years
2005–2024
20-year window
Features
35
satellite + climate + soil
Training Rows
20,536
county-year records
2025 State Yield Forecasts
XGBoost county-level aggregated · vs USDA 2024 actual
184bu/ac
USDA 2024: 211
XGB 184 · KNN 183
Top county: Ida, IA
197 bu/ac predicted
73 counties · 3 rising
175bu/ac
USDA 2024: 184
XGB 178 · KNN 172
Top county: Phelps, NE
210 bu/ac predicted
60 counties · 19 rising
171bu/ac
USDA 2024: 162
XGB 164 · KNN 178
Top county: Grant, WI
197 bu/ac predicted
59 counties · 33 rising
166bu/ac
USDA 2024: 175
XGB 158 · KNN 173
Top county: Atchison, MO
183 bu/ac predicted
53 counties · 8 rising
154bu/ac
USDA 2024: 131
XGB 146 · KNN 163
Top county: Delta, CO
206 bu/ac predicted
18 counties · 13 rising
Colorado range reflects irrigated (Delta 206) vs dryland (eastern plains 66–88) county variation — both are valid predictions
2025 Yield Forecast — Cone of UncertaintyUSDA overlay
p10/p50/p90 from K-NN analog-year retrieval (K=10) · dashed line = USDA NASS published forecasts
Aug 1
183
bu/acre
±19 uncertainty
Early Grain Fill
Sep 1
185
bu/acre
±15 uncertainty
Late Grain Fill / Dough
Oct 1
185
bu/acre
±14 uncertainty
Maturity / Early Harvest
Final
183
bu/acre
±15 uncertainty
End of Season
Iowa — End of Season 2025
183 bu/ac model
210 bu/ac USDA NASS finalΔ -27
p10–p90 analog range
173 – 202
USDA outside the cone — model miscalibrated for this state
Key Advantage
4–5 months ahead of USDA
First forecast issued August 1— USDA NASS final estimates release in December. That's a 4–5 month lead window for grain traders, insurers, and food-security planners.
Aug 1
First forecast
4–5 mo
Lead time
4×
Updates per season
Forecast Timeline — Four Growth Stages
Each forecast issued 4–8 weeks before the next, using cumulative satellite + climate data
Aug 1
Sep 1
Oct 1
Final
Early Grain Fill
Kernels accumulating starch. VPD and heat stress are critical signals.
Late Grain Fill / Dough
NDVI declining. Silking-stage satellite features heavily weighted.
Maturity / Early Harvest
GDD accumulation complete. Black layer formation. Highest model confidence.
End of Season
Post-harvest reconciliation against USDA NASS reported values.
Uncertainty range shrinks from ±26 bu/ac in August to ±5 bu/ac at season end as new satellite passes are ingested
Historical Yields + 2025 Forecast
USDA NASS actuals · XGBoost county-mean forecast for 2025 Actual Forecast
Colorado 2024 drop (188→131) reflects drought conditions · Iowa 2024 (211) was a record high
Model Iterations
Higher R² and lower MAE = better · LOCO-CV validation
Final model: XGBoost · 35 features · LOCO-CV R² 0.789 · MAE 12.5 bu/ac
Feature Importance
XGBoost feature importance (gain) · model8 · 35 features
State Performance Profile
Normalized agricultural factors by state · 2025 XGBoost county-mean forecasts
| State | 2025 Fcst | Soil | Irrigation | Model R² |
|---|---|---|---|---|
Iowa | 184 | A+ | 12% | 0.79 |
Nebraska | 178 | A | 64% | 0.72 |
Wisconsin | 164 | B+ | 8% | 0.75 |
Missouri | 158 | B | 7% | 0.71 |
Colorado | 146 | B+ | 78% | 0.68 |
2025 Yield Predictions — County Level
263 countiesPredicted corn yield with 90% CI · XGBoost model8 · sorted by forecast
scripts/nass_2025_county_finals.pyafter publication and the table below will gain a “USDA 2025” column with model deltas.| # | County | State | Predicted | 90% CI | Trend |
|---|---|---|---|---|---|
| 1 | Phelps, NE | Nebraska | 210 bu/ac | 198–222 | ↑ Rising |
| 2 | Hamilton, NE | Nebraska | 207 bu/ac | 194–220 | — Stable |
| 3 | Delta, CO | Colorado | 206 bu/ac | 193–218 | ↑ Rising |
| 4 | Brown, NE | Nebraska | 206 bu/ac | 194–218 | — Stable |
| 5 | Adams, NE | Nebraska | 205 bu/ac | 192–217 | — Stable |
| 6 | Kearney, NE | Nebraska | 205 bu/ac | 192–217 | ↑ Rising |
| 7 | York, NE | Nebraska | 204 bu/ac | 192–216 | — Stable |
| 8 | Dawson, NE | Nebraska | 200 bu/ac | 188–213 | — Stable |
| 9 | Hall, NE | Nebraska | 199 bu/ac | 186–211 | — Stable |
| 10 | Yuma, CO | Colorado | 197 bu/ac | 185–210 | ↑ Rising |
Geospatial Intelligence Layer
County-level predictions mapped across the US corn belt — 847 counties, 5 states
2025 State Forecasts
90% confidence intervals · Random Forest
County Prediction Map
Circle = state coverage zone · Numbered pins = predicted yield (bu/ac) · Click for details
Loading map...
County Field Locator
Click any card to expand satellite field view
Boone County, IA
211 bu/ac forecast
Click to expand
Platte County, NE
208 bu/ac forecast
Click to expand
Kewaunee Co., WI
196 bu/ac forecast
Click to expand
Yield Signal Heatmap
5-State Corn Belt Coverage — 847 counties, 10 years of training data.
State-level predictions with calibrated 90% confidence intervals.
90% CI: 203–213
90% CI: 185–195
90% CI: 177–187
90% CI: 161–171
90% CI: 187–197
90% CI: Historical
Modeled growing season trajectory. Yield signal builds from vegetative to harvest.
30m Resolution Harmonized Landsat Sentinel-2 at 4 phenological stages per season.
20 Features GDD, VPD at silking, PRISM precipitation, and soil productivity index.
Model Validation vs USDA
2024 BacktestLOCO-CV: trained on 2015–2023, predicted 2024 without seeing that year's data
9.2
bu/ac MAE
4–8 wks
before USDA
-14 bu/ac
6.6% err
-4 bu/ac
2.2% err
+4 bu/ac
2.5% err
-17 bu/ac
9.7% err
+7 bu/ac
5.3% err
USDA NASS Crop Production Annual Summary (Jan 2025) · Our forecast issued Aug 1, 2024 — 4 months before USDA final
Data Pipeline
Five data streams fused into a dual-model forecast — issued 4–8 weeks before USDA official estimates
MODIS + HLS
NASA satellite
NDVI/EVI · 4 stages
gridMET + USDM
4km daily weather
GDD · VPD · Drought
gSSURGO Soils
USDA NRCS
NCCPI · AWC · SOC
Feature Fusion
merge_all.py
847 counties · 20 yrs
XGBoost + KNN
dual-model
Point est. + p10/p90 cone
Project Roadmap
CSU Hackathon 2026 · April 25 presentation — go/no-go gates at end of B, C, and D.1
Methodology
Multi-source data fusion pipeline — satellite, climate, drought, and soil features combined for county-level yield forecasting
Satellite Imagery (NDVI)
MODIS NDVI (2005–2024, all 21 years) CDL-masked to corn pixels via Google Earth Engine, plus Harmonized Landsat-Sentinel HLS v2.0 (2013–2024) for NDVI/EVI at four phenological windows: vegetative, silking, grain fill, and peak season.
Climate Data (gridMET)
Daily 4km gridded weather from gridMET (Univ. of Idaho): Tmax/Tmin/Precip/Solar Radiation/VPD, 2005–2024. Derived features: Growing Degree Days (50–86°F base), Extreme Degree Days above 86°F, Vapor Pressure Deficit at silking, cumulative precip windows. Drought severity from US Drought Monitor (D0–D4 weekly).
Soil Quality (gSSURGO)
USDA gridded SSURGO Valu1 table aggregated to county level: National Commodity Crop Productivity Index for corn (nccpi3corn), available water storage (aws0_100), soil organic carbon (soc0_30), and droughty soil indicator. Static across years — joined to every (county, year) row.
Dual-Model Architecture
Point estimate: XGBoost trained on 2005–2022 county-year records (val: 2023, holdout: 2024). Cone of uncertainty: K-NN analog-year retrieval (K=10) — standardized features find the 10 most similar historical seasons; their actual yields produce the p10/p50/p90 cone. The cone naturally narrows as more of the season is observed.
Why It Matters
Faster forecasts. Smarter decisions.
From field to global markets.
The US corn belt produces enough grain to feed 10 billion people for a year. Our model delivers county-level yield estimates 4–8 weeks before USDA official reports — enabling earlier, better-informed decisions across the entire agricultural value chain.
40%
of global corn supply
comes from the US
$60B+
US corn market value
annually
4–8 wks
earlier than USDA
official estimates
91M
acres planted
across the US in 2024
Who Benefits
Commodity Traders
Early yield signals reduce market volatility. A 10 bu/ac forecast error can move corn futures by $0.30/bushel — that's $2.7B in market exposure.
Farm Managers
County-level granularity lets individual producers benchmark their operation against regional forecasts and adjust grain storage or forward-contract decisions.
Food Security Policy
USDA, FEWS NET, and WFP use early-season yield signals to trigger food aid procurement before harvest shortfalls materialize into price spikes.
Crop Insurance
Sub-county yield estimates enable more accurate actuarial pricing for MPCI and revenue protection products, reducing systemic basis risk for insurers.