Machine Learning for Crop Yield Forecasting
Crop yield forecasting asks a simple question with expensive consequences: how much grain, vegetables or fruit will a field produce, and when will it be ready? Using machine learning to predict crop yields means combining satellite imagery, weather, soil measurements and management records into estimates of production. For Australian growers, that can support irrigation timing, harvest labour, contract negotiations and stock orders. The useful lesson from IASSIST 2017 is less about one algorithm than the shared language needed to connect data with farm decisions.
Hosted at the University of Kansas in Lawrence from 23 to 26 May 2017, “Data in the Middle: The Common Language of Research” brought together discussions on big data, deep learning, digital agriculture and global food security. Its relevance to Australian research lies in that mix of technical expertise and practical interpretation. Yield models become more credible when agronomists, data specialists and industry participants agree on what is being predicted, which records count as truth and where uncertainty matters most.
From Field Signals to Harvest Forecasts
A forecast begins with signals captured at different scales. Multispectral images reveal canopy colour and vigour, often through indices such as the normalised difference vegetation index. Weather records add rainfall, temperature, evapotranspiration and frost risk. Soil sensors describe moisture, while field logs record crop variety, sowing date, fertiliser use and pest pressure. Satellite and drone observations are especially valuable when they line up with dates of flowering, senescence and harvest.
Those inputs become training examples only after someone reconciles units, locations and time periods. A rainfall total from one week cannot simply be paired with an image collected three weeks later, and irrigation applied through several pivots may not represent the whole block. Historical yields harvested by a combine monitor are useful labels, but they may mix saleable moisture with field weight. Cleaner supervision comes from separating production tonnes, moisture and harvested area.
Australian conditions make these alignment problems visible. Growers commonly check Bureau of Meteorology rain radar and outlook apps before spraying or committing to irrigation, yet seasonal rainfall remains less predictable than in many benchmarking datasets. Frost, heat and dry spells can affect the same district within a single season. Around Brisbane, market gardens and peri-urban orchards face a different pressure: high land values and tight delivery windows make scheduling as important as tonnes per hectare.
Models, Evidence and Error
Linear or generalised additive regression provides a useful benchmark because its coefficients expose the assumed relationship between rainfall, soil and yield. Random forests and gradient boosting then capture nonlinearity and mixed data types, often with less manual feature construction. Deep models add spatial or sequential structure, but their appetite for labelled harvests creates a familiar constraint in agriculture: many farms, seasons and commodities are represented, yet reliable field-level observations remain scarce.
Validation must match the intended use. Randomly splitting records can leak nearby paddocks or future weather into training, producing flattering accuracy. Holding out farms, regions and seasons gives a harsher and more relevant test. Mean absolute error and root mean square error describe average performance, while prediction intervals show whether a forecast is useful for a decision. Researchers looking for related applied-informatics events can consult the ICAIE conference site to follow discussions beyond the archived programme.
The conference theme also changes how disagreements are handled. A model may predict tonnes accurately yet fail to reproduce a known management response, such as a yield drop after frost. Agronomists can identify that mismatch; software specialists can test whether the model missed a lag, threshold or interaction. This common language turns model evaluation into a shared diagnostic exercise rather than a contest between black boxes and spreadsheets.
A Practical Comparison of Prediction Methods
No single method is best for every crop, region and decision horizon. The comparison below separates common modelling choices by their strengths and constraints. It does not assign performance scores from the IASSIST archive, because conference contributions used different datasets, targets and evaluation protocols.
| Method | Typical inputs | Main strength | Main limitation | Suitable role |
|---|---|---|---|---|
| Linear or additive regression | Historical yield, rainfall and temperature | Transparent assumptions and quick updates | Limited representation of complex interactions | Baseline and simple seasonal outlooks |
| Random forest or gradient boosting | Imagery, weather, soil and management records | Handles nonlinear patterns and mixed variables | Depends on representative training coverage | Regional forecasts across many fields |
| Convolutional neural network | Spatial imagery from satellites or drones | Detects canopy patterns and spatial structure | Data hungry, sensitive to domain shift and difficult to interpret | Yield maps and targeted scouting |
| Long short-term memory network | Ordered weather, soil moisture and vegetation series | Captures seasonal lags and changing conditions | Requires tuning and extensive longitudinal data | In-season monitoring and early warnings |
| Hybrid or probabilistic ensemble | Process-model outputs plus observed and remote-sensing data | Combines agronomic knowledge with learned relationships | More complex to validate and maintain | Operational planning with explicit uncertainty |
Each row answers a different question. A benchmark regression model may be sufficient for a national tonnes estimate, while a recurrent network can reveal the weeks when a crop is entering a rapid dry-down. Image-based models excel at showing where risk is concentrated, allowing managers to inspect only the blocks that warrant attention. Hybrid approaches are attractive when a simulation supplies physical context and machine learning corrects errors the simulation does not capture.
Targets must also be stated precisely. Total production, marketable yield, protein content and grain moisture are related, yet they are not interchangeable. A model trained on one sensor or region may perform differently after a new cultivar, field boundary or weather pattern. For Australian supply chains, forecasts are most valuable when they distinguish an early signal from a harvest-ready estimate and retain confidence ranges for storage, freight and wholesale planning.
Lessons Preserved in the IASSIST Archive
The archived IASSIST programme remains a record of how research teams framed common problems across disciplinary boundaries. Its digital agriculture and deep learning strands illustrate the value of combining domain knowledge with flexible analytical tools. Plenary themes on global food security add a necessary scale check: a regional production forecast can inform supply estimates, but it does not by itself reveal household access, affordability or the distribution of risk.
Communication was practical as well as conceptual. The archive includes check-in, transport and local recommendations for visitors to Lawrence, showing the kind of operational detail that often determines whether knowledge travels between a laboratory and its users. The archived hotel information is one example of that wider conference record, where session content and attendee logistics sit together.
Reading the archive also requires restraint. A presentation can establish a promising workflow without proving that a model will transfer to another soil type, cultivar or management system. Reproducible comparisons need documented inputs, clear baselines, permission to use data and checks against leakage. Those conditions are especially important when an application concerns growers whose commercial decisions depend on a forecast being both intelligible and dependable.
Priorities for Australian Growers
The conference’s examples translate into an Australian research agenda measured in records, decisions and local constraints. A practical programme starts with data that can be audited and forecasts tied to a specific operational choice.
Data worth collecting
- Keep a block-level harvest ledger that records gross tonnes, moisture, tare, combine pass and final saleable grade, rather than retaining only an annual average.
- Attach dates and locations to imagery, rainfall, soil moisture and irrigation records so the model can reconstruct growth stages instead of treating them as interchangeable observations.
- Record cultivar, sowing date, fertiliser, pest control and water applications, since management history often explains differences between fields with similar satellite signals.
- Label early canopy indicators separately from final yield and flag hail, frost, flood and equipment failures that can otherwise distort the training set.
- Maintain a transparent baseline based on historical averages and local weather, allowing complex models to demonstrate measurable value before operational use.
Decisions and safeguards worth testing
- Define the decision before training the model: a harvest date, irrigation allocation, labour roster, contract quantity or wholesale order each requires its own target and horizon.
- Test performance on unseen farms and seasons, then report uncertainty alongside the headline error so managers know when a forecast falls outside familiar conditions.
- Check contractual ownership of imagery, machinery and farm records, and apply the Privacy Act 1988 whenever personal information about operators or households enters a service.
- Account for Murray–Darling water entitlements and state extraction limits before converting a yield forecast into an irrigation recommendation.
- Compare production grades with buyer specifications in Sydney and Melbourne wholesale markets and retailer requirements from chains such as Coles and Woolworths, while meeting the Australia New Zealand Food Standards Code and relevant state food-safety requirements.
Public datasets can anchor these workflows. ABARES outlooks provide Australian agricultural context, while Australian Bureau of Statistics releases help separate production trends from sampling and revision effects. Operational datasets still need local checks: Sydney Markets and Melbourne Wholesale Markets reveal how quality bands, order timing and transport constraints shape value. A prediction of tonnes has little commercial meaning until it connects with those downstream conditions.
The IASSIST 2017 legacy is practical rather than futuristic. Its strongest contribution is a disciplined method for combining observation, seasonal forecasting and local knowledge, giving Australian growers clearer expectations about production, timing and risk.
At the Conference
What attendees experienced in Lawrence
Plenary Sessions
Keynotes from Daniel Reed on data, technology, and culture, and Jennifer Clarke on digital agriculture and the Midwest Big Data Hub.
Workshops
Full-day technical workshops on Tuesday, May 23. Morning sessions ran 9:00–12:00 and afternoon sessions 13:00–16:00. Laptops were required.
Social Events
An opening reception at The Oread, a banquet, and an optional post-conference tour of Kansas City including Crescent Moon Winery.
Venue & Accommodations
Where the conference took place
Kansas Union
University of Kansas campus, Lawrence. Main conference venue with check-in on the 4th and 5th floor lobbies.
The Oread
1200 Oread Avenue, Lawrence. Hosted the opening reception and offered a room block for attendees.
The Eldridge
701 Massachusetts Street, Lawrence. A partner hotel with a reserved room block for conference guests.
Springhill & TownePlace Suites
Marriott properties in Lawrence with room blocks reserved under the "KU IASSIST Conference" name.
Program Highlights
Sessions and activities
Getting Here
Lawrence, Kansas
Kansas Union · University of Kansas
1301 Jayhawk Blvd, Lawrence, KS 66045
Kansas City International Airport (MCI) is approximately 50 minutes by car. Lawrence Transit Routes 10 and 11 served the area ($1 exact change).
Plan Your Stay
Accommodation options that were available
The Eldridge
701 Massachusetts Street, Lawrence, KS 66044. Room block now closed.
The Oread
1200 Oread Avenue, Lawrence, KS 66044. Room block now closed.
Springhill Suites
Marriott property. Room block reserved under "KU IASSIST Conference."
TownePlace Suites
Marriott property. Room block reserved under "KU IASSIST Conference."