The challenge

Soil constraints such as acidity, sodicity and compaction limit the productivity of around 77% of Australia’s agricultural soils, costing the grains industry billions of dollars in lost production potential each year. Constraints often occur together and can vary in severity both across a field and down the soil profile. This makes accurate diagnoses difficult without intensive soil sampling and laboratory testing, which is expensive, slow, and labour intensive.
As a result, many growers do not collect the information needed to accurately identify soil constraints. Without a clear diagnosis, it is difficult to target amelioration effectively and potential productivity gains may be missed.

Our research

This project developed and tested a framework for diagnosing multiple soil constraints using publicly available data and farmers’ own data.

To diagnose physical constraints, the project focused on plant available water capacity (PAWC), which measures how much water a soil can store for crop use and is a useful indicator of physical limitations to root growth. PAWC was estimated using the Agricultural Production Systems sIMulator (APSIM) with an inverse modelling approach. Rather than predicting crop performance from known soil properties, the model worked backwards from crop performance to estimate the PAWC most consistent with observed yields. The model combined information from the Soil and Landscape Grid of Australia, Scientific Information for Land Owners (SILO) climate data, and multi-year farm management and yield records. The framework was tested across Australia’s grain growing regions.

For chemical constraints, the project developed machine learning models to predict the soil properties associated with common constraints, including pH (acidity and alkalinity), electrical conductivity (salinity), exchangeable sodium percentage (ESP, sodicity), cation exchange capacity (nutrient holding capacity), and exchangeable aluminium and boron (toxicity).

The models were trained using existing soil datasets together with soil, environmental and crop information. Their accuracy was assessed using cross validation and independently collected soil data. Several machine learning methods, including Cubist and Random Forest, were compared to identify the best approach for each soil property. Where local soil data were limited, national datasets were used to improve model performance.

Research findings

Read our research findings fact sheet to learn more.