Ambrosia2024/yield-autoformer-cybench

Model Description

Autoformer model trained for long-term crop yield forecasting as part of the Yield Hub Benchmark project. The model predicts seasonal yield using daily aggregated weather and agronomic features. These models are part a large-scale evaluation framework studying geo-generalization of crop yield forecasting models across multiple countries and wheat and maize crops.

Training Configuration

• Aggregation: Daily
• Lag years: 0
• Max epochs: 50
• Learning rate: 1e-4
• Weight decay: 1e-5
• Batch size: 64
• Seed: 42
• Test horizon: Last 5 years
• CWB feature: Enabled
• Spatial features: Disabled
• Recursive lags: Disabled
• GDD / Heat stress / RUE / Farquhar: Disabled

Training Runtime

• GPU: NVIDIA A100 40GB
• CPU: 32 cores
• RAM: 128 GB
• Training time: 10-11 minutes per model

Results

Crop	Country	NRMSE	R²	SMAPE
Wheat	DE	0.1328	0.2527	0.1070
Maize	DE	0.2106	0.2619	0.1822
Wheat	US	0.2535	0.4769	0.2154
Maize	US	0.2360	0.0136	0.2162