Abstract

During the past few decades, a range of digital strategies for Nitrogen (N) management using various types of input data and recommendation frameworks have been developed. Despite much research, the benefits accrued from such technology have been equivocal. In this work, thirteen methods for mid-season N recommendations in cereal production systems were evaluated simultaneously, ranging from simple mass balance through to non-mechanistic approaches based on machine learning. To achieve this, an extensive field research program was implemented, comprising twenty-one N strip trials implemented in wheat and barley fields across Australia over four cropping seasons. A moving window regression approach was used to generate crop response functions to applied N and calculate economically optimal N rates along the length of the strips. The N recommendations made using various methods were assessed based on the error against the optimal rate and expected profitability. The root mean squared error of the recommendations ranged from 15 to 57 kg/ha. The best performing method was a data-driven empirical strategy in which a multivariate input to characterise field and season conditions was abundantly available and used to predict optimal N rates using machine learning. This was the only approach with potential to substantially outperform the existing farmer management, reducing the recommendation error from 42 to 15 kg/ha and improving profitability by up to A$47/ha. Despite being reliant on extensive historical databases, such a framework shows a promising pathway to drive production systems closer towards season- and site-specific economically optimum recommendations. Automated on-farm experimentation is a key enabler for building the necessary crop response databases to run empirical data-driven decision tools.