HarvestNet
Crop yield prediction model trained on agricultural data from Brazil,
Argentina, and Mexico. Reference model for the HarvestYield-LAC benchmark.
Model Description
Models trained on U.S. Midwest data systematically underperform on
Southern Hemisphere growing conditions, soil types, and crop varieties.
HarvestNet is trained on regional data from CONAB, SIAP, INTA, and
INMET/IDEAM climate records.
- Architecture: Gradient boosted trees with neural feature extraction
- Inputs: climate variables, soil type, crop variety, management
practices, satellite-derived vegetation indices
- Output: yield prediction in tonnes/hectare with confidence interval
Crops Covered
Soy (Brazil, Argentina), Corn (Brazil, Mexico, Argentina),
Sugarcane (Brazil), Wheat (Argentina), Coffee (Brazil, Colombia)
Performance
| Crop | Region | RMSE | MAPE | R2 |
|---|
| Soy | Brazil | 0.31 | 4.2% | 0.87 |
| Soy | Argentina | 0.28 | 3.9% | 0.89 |
| Corn | Brazil | 0.44 | 5.1% | 0.83 |
| Corn | Mexico | 0.51 | 6.3% | 0.79 |
| Wheat | Argentina | 0.38 | 4.7% | 0.85 |
Usage
from genia.agri import HarvestNet
model = HarvestNet.from_pretrained("genia-americas/harvest-net-v1")
prediction = model.predict(
crop="soy",
region="mato_grosso_br",
climate=climate_data,
soil=soil_data
)
print(prediction.yield_estimate, prediction.confidence_interval)
Citation
@techreport{genia2025harvestnet,
title={HarvestNet: Crop Yield Prediction for Latin America},
author={GENIA Americas Corporation},
institution={GENIA Americas / RaceFor.AI},
year={2025},
url={
https://github.com/GENIA-Americas/multimodal-ai-americas}
}
Links
