Nitrogen management in the world of AI - lessons from the Future Farm project

Nitrogen management in the world of AI - lessons from the Future Farm project

Take home messages

  • AI and data-driven methods have the potential to increase grower confidence and profitability in fertiliser decision-making, but they are heavily dependent on extensive on-farm digital data.
  • On-farm experimentation is the core enabler of AI solutions in agriculture and should be adopted as a core element of farm business operations.
  • N management solutions must recognise the flat response around maxima.

Background

Over the past decades, many digital nitrogen (N) management methods have been developed, utilising different input data and recommendation frameworks. During the Future Farm project, thirteen methods for mid-season N recommendations in cereal production systems were evaluated and tested on farms. They ranged from simple mass balance techniques to non-mechanistic approaches based on artificial intelligence (AI) algorithms. To accomplish this, an extensive field research program was conducted, involving 21 N strip trials in wheat and barley fields across Australia over four cropping seasons between 2018 and 2021. This paper summarises the project's key findings, with further reflection on recent disruptive AI expectations.

The most common principle in agronomy to determine nutrient application rates is the nutrient budget or mass balance. This approach relies on predicting nutrient demand based on estimates of potential yield, assumptions regarding protein content, information on soil nutrient supply, and fertiliser efficiency. For the last thirty years, crop sensing technologies have facilitated the semi-automated implementation of frameworks for site-specific fertiliser application. Thus, the question is, can AI do better? We showed and proposed a new framework where AI can do better N recommendations than traditional sensor-based methods in a research environment (Colaço et al. 2021). But what about on a farm?

The project sought to understand and assess the accuracy of the N recommendation and its impact on farm profitability. All the details and comprehensive explanations of all the different methods, results and discussions can be found at Colaço et al. (2024). Here, we summarise those findings and reflect on the new wave of AI and its potential impacts on N decision-making processes.

Method

There were three key components in our methodology: experiments in real farms (Figure 1), optimal N rate analysis (Figure 2), and assessing the different strategies for making N fertiliser decisions (Figure 3).

Key elements of the work were on-farm experiments (OFE), which included ‘rich’, ‘zero’, and ‘field’ N rate treatments harvested with headers with yield and protein monitors. The ‘field’ strip was fertilised at the rate that the grower thought was best, the ‘rich’ was double that, and ‘zero’ was zero or as near as zero as the grower was comfortable with at sowing. All experiments were laid down by the growers using their equipment to ensure the concepts could be deployed at commercial scale. Twenty-one such field trials were sown between 2018 and 2021 in paddocks of wheat and barley from WA, SA, VIC, and NSW. The trials aimed to assess crop response to N application and determine optimal mid-season N rates across varying conditions. Field data were collected at various stages during the season, starting before sowing when soil samples were collected from the top 30cm depth layer at targeted locations along the length of each strip (example, Figure 1) for general soil characterisation, such as the analysis of soil texture, organic matter, pH, and analysis of mineral N. Depending on the field size, there were between 15 and 21 sample locations per trial (5–7 per strip). At Zadoks growth stage 31 (GS-31), the trial was scanned with a sensing system to measure various vegetation indices. Plant samples were collected at the same locations for dry above-ground biomass and leaf N concentration. Later, we manually harvested for yield and protein content at the same locations. To minimise farm management disruption and increase grower engagement in this project, in line with the participatory nature of this research between growers and researchers, a strip design was chosen (as described above) as opposed to a more randomised and replicated trial which could potentially allow more N rates to be used and better characterisation of crop N responsiveness.

Figure 1

Figure 1. Example of an N strip experiment and point sample location layout in a 357ha wheat field near Kalannie WA, 2019 (Colaço et al. 2024).

Using the harvested data, the site-specific crop response analysis to the different N strips was performed using a moving window regression analysis along the length of each strip trial. A 50m radius moved along the strip’s length in increments of 10m. A local regression with yield and protein determined the partial profit economical optimal N for that pixel; see the example in Figure 2. This generated the economical optimal N rate (EONR) based on average prices across all seasons.

Figure 2

Figure 2. Experimental strip design implemented in a 100ha barley field near Booleroo Centre SA, 2021, and a moving window analysis of crop response to N application. The graphs illustrate the local response curves and optimal N rates (ONR is depicted as red crosses) generated from data within one of the moving windows along the length of the strip trial (Colaço et al. 2024).

Lastly, an EONR database of observations was built, and the assessment of the various N recommendations was possible against a benchmark of maximising profit (Figure 3). Then, we tested 24 different strategies for N recommendation split into two spatial scales, uniform and site-specific (precision agriculture). For clarity, the site-specific EONR was the reference and, therefore, its root mean square was zero, and the normalised partial profit was one (or 100%). The strategies included, but were not limited to, AI data-driven with abundant data and limited data based on response functions, simplified mass balance, N sufficiency, and yield prediction. Each field trial had an average of 112 windows for generating crop response functions and EONR observations, totalling 2 352 response observations across all paddocks and seasons. The full description of all the methods is available at Colaço et al. (2024).

Figure 3

Figure 3. The process of building a database with observations of economically optimal N rates (EONR) using local response curves along the length of an N strip trial (1), and the assessment of the error and expected profitability of different N recommendation methods using the local response function (2). The illustration of the strip experiment is out of scale, and values in the tables are hypothetical (Colaço et al. 2024).

Results and discussion

The AI with an abundant dataset (DD in Figure 4) was the best-performing strategy for site-specific and uniform management. However, the growers who collaborated with Future Farm achieved strong performance. Their uniform management was only 6% shy of the site-specific EONR. This means that for those high-performing growers to improve their N management, an enhancement of spatial resolution is required. The results show that as the performance of the different methods approached the optimal input N rate, the impact on profitability was reduced. This illustrates the plateau effect (flatness) often seen at the peak of response curves that depict the relationship between production and input use. A wide range of N rates near the optimal level similarly impacts crop performance and profitability, possibly limiting the benefit of more accurate N recommendations. This also explains the good performance of most recommendation methods, which were between 85 and 95% of the maximum partial profit. AI is heavily dependent on data for its training and evaluation processes (Richetti et al. 2023); and the ag sector is still far from having abundant data compared to other industries.

Figure 4

Figure 4. Error by profit biplot showing the average results of various N recommendation methods across 21 large-scale on-farm trials.

Conclusion

AI and data-driven methods hold great promise for reducing errors and boosting the profitability of fertiliser management. They, therefore, offer an important avenue for increasing grower confidence in N rate decision-making. However, their effectiveness is highly dependent on extensive on-farm digital data availability. On-farm experimentation is crucial for enabling AI solutions in agriculture and should be integrated into core farm business operations. Additionally, it's important for site-specific nitrogen management solutions to account for the flat response observed around maximum application rates.

Acknowledgements

The research undertaken as part of this project is made possible by the significant contributions of growers through both trial cooperation and the support of the GRDC, the authors would like to thank them for their continued support. The authors also thank other Future Farm partners – CSIRO, University of Sydney, University of Southern Queensland, Queensland University of Technology and Agriculture Victoria – for their continued support. The technical and other assistance provided by Damian Mowat and Christina Ratcliff (CSIRO), Logan Torrance, Dr Jacob Humpal and Lachlan Day (USQ), Russell Argall (AgVic) and Aarshi Bhargav, Fabio Manca, Daniele De Rosa and Max De Antoni (QUT) is also gratefully acknowledged.

References

Colaço AF, Richetti J, Bramley RGV, Lawes RA (2021) How will the next-generation of sensor-based decision systems look in the context of intelligent agriculture? A case-study. Field Crops Research 270, 108205. https://doi.org/10.1016/j.fcr.2021.108205

Colaço AF, Whelan BM, Bramley RGV, Richetti J, Fajardo M, McCarthy AC, Perry EM, Bender A, Leo S, Fitzgerald GJ, Lawes RA (2024) Digital strategies for nitrogen management in grain production systems: lessons from multi-method assessment using on-farm experimentation. Precision Agriculture 25, 983–1013. https://doi.org/10.1007/s11119-023-10102-z

Richetti J, Diakogiannis FI, Bender A, Colaço AF, Lawes RA (2023) A methods guideline for deep learning for tabular data in agriculture with a case study to forecast cereal yield. Computers and Electronics in Agriculture 205, 107642. https://doi.org/10.1016/j.compag.2023.107642

Future Farm – the potential value in data-driven N decisions (https://grdc.com.au/resources-and-publications/grdc-update-papers/tab-content/grdc-update-papers/2023/02/future-farm-the-potential-value-in-data-driven-n-decisions)

PCT Agcloud (https://pct.ag/)

Contact details

Jonathan Richetti
CSIRO
147 Underwood Ave, Floreat WA 6014
0411 150 425
jonathan.richetti@csiro.au

GRDC Project Code: CSP1803-020RMX,