The challenge
Evaluating flowering windows for optimal crop sowing
Canola is Australia's third most valuable grain crop and a critical and profitable element in crop rotation for sustainable farming. Australia's canola growers are keen to adopt the best technology available for oilseed yield and quality. A wide selection of canola varieties with diverse attributes are continuously being made available from breeding and seed companies.
When selecting a variety growers need to be confident of when to sow so that the optimal flowering window is met for their farm environment. This will ensure that costly yield losses from excess water, frost and heat stresses are minimised.
Currently, determining the optimal flowering window evaluates the flowering time for each variety across a range of environments and sowing times. We also considered are key factors for flowering such as cumulative high and low temperature exposure. These values then feed into crop growth models to predict flowering time over an even broader range of farming environments and locations. This allows breeders, seed companies and growers to better target the optimal flowering window for a variety in a particular environment.
While this method works well, it is expensive and slow due to field-based evaluations. This delays adoption by a few years. And this is an ongoing process because new varieties are continuously released.
Our response
Combining data to optimise canola flowering models
With support from the Grains Research and Development Corporation (GRDC), we combined our expertise in crop process modelling, agronomy, genomics, phenomics and computational analytics, to develop a new method to calculate optimal the flowering window for canola crops.
We've merged genomics with crop development estimates and crop modelling to make the flowering optimisation process faster. The overall goal was to be able to estimate the flowering parameters for canola varieties in specific environments using readily accessed genomic information in combination with machine learning computer models trained on data from extensive large-population field trials, avoiding continuous experimental investigations for each new variety.
To generate the necessary crop development data we needed, the team collected a large and diverse panel of 350 canola varieties from Australian and global sources.
In an ambitious four-year field campaign, plant development and flowering was measured in key environments across the canola-growing regions of Australia for different sowing times. More than 400,000 observations of plant development were made.
These data were integrated into a hybrid machine learning - Agricultural Productions Systems Simulator (APSIM) framework to estimate flowering time across sites and the model's performance was evaluated against our observed data.
Combining this data with a dense and diverse genomic data set for the population into the new hybrid genomic model enabled predictions of flowering for different environments.
The results
A new hybrid genomic model to predict flowering
The new hybrid model was evaluated against the slower benchmark approaches and was found to predict flowering for known varieties grown in known or new sites. This delivered accuracy of 95 percent and 93 percent respectively
Using the hybrid genomic model, it is now possible to estimate when new canola varieties that have not been so well studied will flower in a particular environment. This prediction is based only on their genomic profile as a data input, and it works for environments that have not been tested before. The accuracy of the hybrid genomic model to predict flowering time is 87% for known environments and 86% for new ones. These scenarios have slightly lower accuracy, but their average performance window, between 5-8 days, is good enough to be very useful for growers and breeders. They can now quickly adapt sowing dates of new varieties to their farms, which they could not do before, this will make the future canola industry more agile in its ability to adopt new varietal technologies.
The new hybrid genomic model performed better for Australian genetics than for international germplasm because the former were more highly represented in model training. In this overall effort, we have collaborated directly with breeding companies and the National Varietal Trial project to add their current varieties to the model and ensure it is based on 'up-to-date' genetics, including hybrid material. Getting access to genomic data on the canola breeder's proprietary genetic material has been a significant achievement. Trusting CSIRO and taking this step with us shows the value the industry sees in what this new crop data science tool will deliver.
A new Grain Research and Development Corporation (GRDC) partnership project is now developing the tool as a web-based app for growers and breeders that will run at high resolution across the canola growing region. This app is called ‘Canola Flowering Calculator’ and is currently in its development and testing phase for release later in 2024 or early 2025.