Feature Analysis

Ablation studies, feature importance, and data characteristics

Ablation Study — Feature Sets

RMSE comparison across 7 feature configurations (lower is better)

R² by Feature Set

R² comparison across 7 feature configurations (higher is better)

Key Finding: Lagged Yield Dominance

The operational model relies heavily on lagged yield history rather than weather features alone. Removing lagged yield increases RMSE from 0.658 to 0.889 (R² drops from 0.730 to 0.507). This suggests the model memorizes past yields rather than learning weather-yield relationships — a critical limitation for true early-warning capability.

Soil Attributes (10 features × 2 depths)

Soil properties at topsoil (0–30cm) and subsoil (30–100cm) depths

AWC

Available Water Capacity

BULK-DENSITY

Bulk Density

CLAY

Clay Content

SAND

Sand Content

SILT

Silt Content

SOC

Soil Organic Carbon

TOTAL_N

Total Nitrogen

TOTAL_P

Total Phosphorus

PHC

pH (Calcium Chloride)

ECEC

Exchangeable Cation Sum

Data Coverage

RegionsNSW, QLD, SA, TAS, VIC, WA
CropsBarley, Canola, Lupins, Oats, Wheat
Years1989–2021 (33 years)
Observations966 unique × 5 windows = 4,830
Weather records82,008 daily records
Weather stationsAustralia Silo network

Regional Climate Summary

RegionRainfallAvg TempEvaporationSolar
NSW520mm21.3°C4.8mm22.1MJ/m²
QLD610mm24.7°C5.6mm24.3MJ/m²
SA310mm22.8°C6.2mm23.5MJ/m²
TAS980mm15.6°C3.4mm18.7MJ/m²
VIC580mm17.4°C4.1mm20.2MJ/m²
WA420mm23.1°C5.8mm23.8MJ/m²

Forecast Window Details

Each window aggregates daily weather from May through the end month, capturing conditions during critical growth stages.

May–Jun

Earliest forecast — 12 weeks before harvest

1/5

~12 weeks before harvest

May–Jul

Early season outlook

2/5

~16 weeks before harvest

May–Aug

Mid-season prediction

3/5

~20 weeks before harvest

May–Sep

Late season forecast

4/5

~24 weeks before harvest

May–Oct

Pre-harvest — most accurate

5/5

~28 weeks before harvest