Key grains events in Western Australia in February will hear how machine learning and artificial intelligence (AI) could help create the next generation of nitrogen (N) fertiliser decision aids for grain growers.
CSIRO researcher Roger Lawes will discuss with participants at the Grains Research and Development Corporation (GRDC) Grains Research Updates in Perth and Kendenup the potential for digital technologies to assist with N management.
Expenditure on N fertiliser is one of the largest annual costs for Australian grain growers and is challenging to get right in terms of achieving maximum profitability.
Dr Lawes says conventional crop N fertiliser decision aids require information about the supply of N from the soil and the N requirements of the crop.
"These tools are often complex to use, as growers need to measure the starting amount of N present in the soil, accounting for mineralisation and then estimate the requirements for N by the crop," he said.
"These factors vary with the season and the soil type, complicating decision making."
Dr Lawes said that, as part of a GRDC investment, CSIRO had developed a framework that could help modernise and simplify growers’ N decisions using machine learning, remote sensing, on-farm trialling and crop modelling.
"We found that modern analytical approaches, combined with on-farm experimentation and the sensing of multiple crops, may enhance N fertiliser recommendations for wheat crops," he said.
Dr Lawes said previous research by CSIRO had found applying N according to the long-term mean yield of a paddock could provide an economically sensible approach to N fertiliser management.
"Since that study, artificial intelligence and machine learning techniques have evolved and it is now possible to process large volumes of information about crops, using satellite imagery or information generated from crop models like the Agricultural Production Systems sIMulator (APSIM)," he said.
"In essence, on-farm strip trials, combined with crop modelling and satellite imagery analysis, could help growers understand whether a crop in a particular paddock will respond to an application of N, regardless of crop type, soil or season."
Dr Lawes said the study aimed to identify the variables that needed to be monitored to help growers make a profitable N fertiliser decision.
"The machine learning technique ‘Random Forest’ was used to determine which variables were the most important and useful for predicting the optimal N fertiliser rate for a paddock," he said.
"As expected, this analysis found the most important variable was the long term historical mean site yield, as estimated from APSIM, the crop simulation model used in this study.
"Extractable soil water to a depth of 150cm and leaf N content determined by the ‘N minus strip’ were the two next most important variables."
Dr Lawes said future studies would aim to integrate information from models, satellites and on-farm trials to develop a grower-ready package for making N decisions.
