ACML 2026 Research Project

Stage-Aware Early Warning forAustralian Winter Crop Yield

Forecast yield shortfalls using 38 years of daily weather data (1989–present) and soil properties across 6 Australian regions. Run models directly in your browser.

6 Australian Regions

NSW, QLD, SA, TAS, VIC, WA

5 Major Crops

Wheat, Barley, Canola, Oats, Lupins

ML-Powered Models

Ridge, ElasticNet, HistGradientBoosting

Uncertainty Quantified

Conformal prediction intervals

82,008

Data Points

Daily weather records

966

Observations

Region × Crop × Year

5

Forecast Windows

May-Jun through May-Oct

4830

Total Rows

Expanded panel dataset

About the Project

A stage-aware early-warning system built for Australian winter crop management.

What We Predict

The system forecasts Australian winter crop yield shortfalls — when actual yields fall below the expected linear trend. We predict both the continuous yield shortfall (t/ha) and classify the risk of low-yield events for 5 major crops across 6 states.

Why It Matters

Australian agriculture faces increasing climate variability. Early warnings of yield shortfalls give farmers, policymakers, and supply chain operators critical lead time (12–28 weeks before harvest) to make informed decisions about planting, irrigation, and market strategy.

Data Sources

Daily weather data from the Australia Silo network (1989–present, 82,008 records) combined with 10 soil attributes across 2 depth layers (topsoil 0–30cm, subsoil 30–100cm). Yield data from AustCropRrt covering 966 unique region-crop-year observations.

Methodology

Stage-aware feature engineering captures weather conditions during critical crop growth stages. Two model families: history-free (weather + soil only) and operational (with lagged yield history). The operational model achieves R² = 0.730 for the May-Oct window.

Data Timeline & Splits

89
90
91
92
93
94
95
96
97
98
99
00
01
02
03
04
05
06
07
08
09
10
11
12
13
14
15
16
17
18
19
20
21
Train (1989–2012) Validation (2013–2016) Test (2017–2021)

Forecast Windows

Five lead-time windows from May, each providing progressively more information as the growing season advances.

Order 1

May–Jun

Earliest forecast — 12 weeks before harvest

~12 weeks lead
Order 2

May–Jul

Early season outlook

~16 weeks lead
Order 3

May–Aug

Mid-season prediction

~20 weeks lead
Order 4

May–Sep

Late season forecast

~24 weeks lead
Best
Order 5

May–Oct

Pre-harvest — most accurate

~28 weeks lead

Model Families

Two task types, multiple algorithms — all evaluated across 5 forecast windows with rolling-origin cross-validation.

Regression — Yield Prediction

Predict expected yield (t/ha) and yield shortfall (t/ha)

Ridge

Linear model with L2 regularization

Best R²: 0.5 at May-Oct

Elastic Net

Combined L1 + L2 regularization

Best R²: 0.5 at May-Oct

HistGradientBoosting

Gradient-boosted decision trees

Best R²: 0.53 at May-Oct

Classification — Low-Yield Risk

Classify whether yield will fall below crop-specific thresholds

Logistic Regression

Linear classifier for low-yield risk

Best AUC: 0.772

HistGradientBoosting

GBDT classifier for risk prediction

Best AUC: 0.734