geospatial ai pipeline forecasting corn yields months ahead of the usda across 263 counties. 0.789 r² on an honest holdout. i owned the aws s3 data lake, model training on sagemaker, and the operator dashboard.
- moved corn-yield forecasts from 4 months after harvest to 4 months before it, across 263 counties in 5 states, by fusing 5 satellite and climate sources.
- raised accuracy from a 0.271 r² baseline to 0.789 (12.5 bu/ac mean absolute error) by replacing a single regressor with a dual-model design and blocking data unavailable at forecast time.
- on a 2024 backtest trained only on 2015 to 2023, state forecasts landed within 9.2 bu/ac of the usda final estimate on average.
- python
- aws
- databricks
- next.js
github.com/Yumstezy/CSUHackathon2026