Skip to content
menu icon

Computing the impact of soil trait variability on yields

The most important soil traits affecting yields can now be identified using soil mapping data and software newly developed by machine learning experts Dr Yuhang Liu (left), Dr Rakesh David (centre) and Dr Dong Gong of the Australian Institute of Machine Learning.
Photo: University of Adelaide

Potentially important applications for machine learning (ML) within agricultural innovation and production systems are under investigation across Australia.

Investment by GRDC is targeting key challenges in agronomy, breeding, remote monitoring, disease control and more through a series of interlinked two-year scoping studies.

As the head of an agronomy study, Dr Rhiannon Schilling recently demonstrated that ML models can take existing soil maps and identify the most important factors that are constraining yields, while taking seasonal conditions into account.

Dr Schilling is the program leader of agronomy at the South Australian Research and Development Institute (SARDI) and the deputy director of the South Australia Drought Resilience Adoption and Innovation Hub.

To develop the ML model, she partnered with the Australian Institute of Machine Learning (AIML) in this University of Adelaide-led project.

AIML is a globally recognised centre of ML excellence. Dr Dong Gong (AIML), Dr Yuhang Liu (AIML) and Dr Rakesh David (University of Adelaide) were involved in the development of the ML models, as were project partners from the Australian Plant Phenomics Facility (APPF), the University of Queensland, Federation University, Thomas Elder Institute, Consilium Technology, Precision Agriculture, Riverine Plains Inc, Australian Grain Technologies, SARDI and Agronomy Solutions.

The project got underway in February 2020.

Soil factors driving yield variability

To train the ML software in the agronomy application, Dr Schilling drew on pre-existing soil mapping data that was made available by project partners.

“A key component of the project was standardising the paddock mapping layer data into a uniform format,” Dr Schilling says. “The data we used covered about 150 soil traits as well as elevation and climate data that had been captured over multiple seasons for multiple paddocks.”

Maximising the on-farm value of this data was then a two-step process. It was achieved by first developing a ‘causality-based’ ML model and then attaching it to an ‘attention model’. This is ML jargon for using the soil mapping data to first predict yield across all the variable zones in a paddock and then imputing which soil factors – and which complex combination of interacting factors – are key to understanding yield variability.

The ultimate result is extraordinary. Software visually represents the mapped paddock but rather than displaying variation in soil pH or water content, the map displays the soil traits that are having the largest impact on yield.

“Basically, the ML model ranks soil traits at various depths in terms of their importance to yield and does this spatially across the whole paddock across years,” Dr Schilling says. “So, the ML model is assigning ‘importance scores’ to the soil data.”

The maps are sophisticated. For example, they can show changes in the rank of various soil factors that arise due to varying climate conditions, such as a wet versus dry season.

“These ML-generated maps translate soil mapping data into a form that allows for targeted management decisions to remediate poor-performing patches or decisions about whether such an investment is worthwhile,” Dr Schilling says. “This is a practical way to use ML to achieve more-uniform yield across a paddock using a targeted, cost-effective strategy.”

Adopting ML technology

How best to roll out the ML technology for use by the grains industry was also explored by Dr Schilling.

Working with Consilium Technology, a prototype of a web-based application has been developed. This app would allow growers or their consultants to upload their soil data and generate the maps showing the importance scores.

The key requirement for using such a platform is the format of the input soil data. It needs to be ML-ready, with the formatting requirements now well understood by Dr Schilling.

“Grain growers who are interested in the ML model’s capability are more than welcome to reach out to me and I will talk them through how to collect and store data so it is ML-ready and also walk them through the demo application,” Dr Schilling says.

In addition, there are requirements for consistent forms for the accompanying metadata (information about who collected the data, when and where). This is required to make packages of data findable, interpretable and usable.

Overall, the project has validated the use of ML for soil management applications. The project demonstrated the feasibility of pre-processing a large number of historical datasets, of developing new ML models to determine traits that cause crop variability and then using these ML models to develop a demonstration web-based application.

“Rather than simply identifying and mapping variable areas of our cropping paddocks into low and high-yielding zones, we can now determine and spatially and temporally map which trait (or combination of traits) is driving this variability in each zone,” Dr Schilling says. “With further tweaks, the web-based application developed in this project could be used to simulate impacts on yield from various soil treatment strategies to identify the most significant options. This could ultimately also come to include costs associated with investment in different management strategies.”

These are all uses that Dr Schilling is keen to see developed in the near future.

Like many people in agriculture, she confesses to not fully understanding the potential of ML technology at the project’s outset. She adds: “At the same time, the ML experts didn’t understand the challenges of our variable cropping paddocks and the many historical data layers we have. Bringing these different fields together has been challenging, but exciting and I think, ultimately, worthwhile.”

The GRDC investments into machine learning were intentionally set up precisely to enable cross-pollination of ideas and expertise between Australian grains researchers and machine learning experts.

Dr Jeff Cumpston, Manager, Data Analytics at the GRDC says that for each investment, researchers have reported having to overcome a steep initial learning curve, reaching a mutual understanding of the important problems facing Australian grains research and the role that machine learning analytics can provide in solving them.

“In this way, the machine learning investment have not only delivered tangible outcomes that have benefits for Australian growers, but have increased the capacity and capability for further work using machine learning based analytics in Australian grains research,” Dr Cumpston says.

More information: Rhiannon Schilling, rhiannon.schilling@sa.gov.au

back to top