Why bother with artificial intelligence in agriculture? —because it can improve fertilisation management

Key messages

• Artificial intelligence algorithms can improve nitrogen decisions over current methods
• The current limitation is the availability of data
• An effort is needed to integrate data sources and farm data to aid farm decision-making
• Building such a database is one of the main challenges for the community

Aims

This work aims to clarify the possible uses of Artificial intelligence (AI) in farm management, particularly the use of machine learning to improve N fertilisation management in wheat.

Introduction

Deciding how much nitrogen (N) to apply to a crop is complicated. Farmers weigh up the cost of applying fertiliser N against the value of the extra yield produced. However, none of the factors determining N decisions are constant, with the productive potential of the field, soil nitrogen status, seasonal forecast, input prices and commodity prices all changing. Combined with AI methods, sensors can deliver a N recommendation and solve this complex problem (Colaço and Bramley 2018). AI algorithms are specifically designed to help make complex decisions, and studies with experimental and simulated data have shown that a N decision can be improved with an AI algorithm (Lawes et al 2019; Colaço et al 2021). These studies had access to vast quantities of information about the crop and could exploit the power of AI to make improved N recommendations. The challenge is to take the insights from these theoretical studies and adapt them to real-world situations, with data gathered on-farm where the N decision needs to be made. Growers and consultants may not have access to the vast array of data used by the researchers, and AI algorithms may perform poorly with the field data farmers have.

AI requires vast amounts of data to generate a N recommendation. In this study we translated research results into the realities of on-farm decision making. We explored how we could use AI to help farm decision-making by determining the variables needed for the farm to use AI to make an N decision. Finally, we assessed whether AI could generate a better N recommendation and generate a better economic outcome than the farmer’s current management system. We explorde each of these questions across paddocks in Western Australia and South Australia to determine if AI could improve upon farmers decisions.

Method

We examined the key variables needed for N decisions. Previously, with simulation data, we identified that the most important variables that enhance a prediction about the optimal N rate were the mean yield, soil moisture, and plant leaf N (see full publication by Lawes et al 2019). We extrapolated the simulation study to farmers’ paddocks with two paddocks, one in WA and one in SA grown to wheat, intensively sampled for two seasons. Using on-farm N strips across the paddocks ensured N should be non-limiting (Figure 1). N values varied from paddock to paddock and season to season. A total of 46 variables were measured at each sampling point, from rainfall to organic carbon. Some of these variables were easy to gather, such as the vegetation indices (NDVI and NDRE) that are freely available from a satellite. Other variables were more difficult to acquire and included factors like available soil N at two different depths. These variables were manually sampled at each point in the field. With all these data, we aimed to determine which variables were most useful in making a N decision.

We analysed different methods for predicting the mid-season N recommendation using data from nine large-scale trials across SA (2018–20), WA (2019–20), Vic (2018–20) and NSW (2020), generating over 1500 observations of crop response to N application. There are different ways to define an optimum. The optimum N recommendation was based on a quadratic function defined as ‘optimum partial profit’ based on the observed yield and protein from the adjacent strips. The grain price varied depending on the protein content, and grain price was $262/t if protein fell below 10.5%; $308/t  if the protein was between 10.5 and 11.5%; $323/t if the protein was between 11.5 and 13%; and $339/t if the protein was above 13%. The N price was 0.50 $/kg of urea. The optimal N rate was defined by the amount of N that delivered the highest profit from the quadratic function. Any method could then be compared by adding the N rate value into the equation that would provide a profit for that location for that season. A full and detailed comparison with other methods and the profit analysis from Future Farm will be available shortly (Colaço et al 2022).

Results

The AI from farmers’ fields evaluated data from 46 variables to make a N decision. The most important variable was the ratio between farmer and rich strips of NDVI. Other important variables included the ratio between farmer and rich strips of NDRE, the available mid-season N in the soil, and the ratio between the farmer and low strip with NDRE. Of these variables, soil N is the most difficult to gather because there is currently no direct soil N sensor. Although the other variables help improve accuracy, the gain in accuracy by measuring every possible variable might not be justified. Thus, farmers should concentrate on using the information from a N-Rich and N-Minus strip to formulate a N decision. Mid-season (around growth stage Z31) soil N is also useful, but other variables do not greatly improve the prediction.

Also, when comparing the partial profit from the optimum N rate, the farmers were, on average, missing out on $12/ha while the AI model was missing out on just $3/ha from the optimal profit. This highlights that the collaborating farmers are already making excellent N decisions. On average, the AI model generated an additional $9/ha profit over and above the farmers’ management decisions. However, at some points, the difference between the farmer and the AI was much higher, with the AI generating an additional $86/ha. At another point, the farmer was $200/ha better than the AI model. The reasons for the poor performance of the AI on a particular paddock were explored to help determine when to use AI to make a N decision.

We evaluated the interquartile range in crop yield for individual paddocks. The interquartile range is the difference between the 75% highest values and the 25% lowest. It excludes outliers and is a more consistent way of describing the amount of variation. The paddock where AI performed best had an interquartile range for yield of 930kg/ha. The interquartile range for the paddock where the AI performed poorly was just 285kg/ha. This indicates that if there is no variation in the paddock, AI is not required. In fact, this is well known for precision agriculture with Lawes and Robertson (2011) showing on a WA wheat farm that the number of paddocks warranting variable rate nutrients was influenced by the amount of yield variation within the field and the starting levels of the nutrient concerned. Commodity and input prices influence the number of fields that will benefit from variable rate, but these factors are overshadowed by yield variability and soil fertility effects.

A caveat for this study is that the data used to train the AI model had comparatively few observations compared to the research studies. That is, the model was trained with only two seasons from one paddock in WA and one season from a SA paddock. AI’s main ability is to capture patterns in data. Future AI developments should use data from other sources and from other paddocks as this will more closely mimic how farmers and agronomists make decisions about a field. Such experience could be translated into some sort of pattern recognition, which is the main function of AI algorithms. Thus, expanding the AI algorithm to different seasons and places will improve the algorithm’s ability. In time, AI decision support systems will improve.

Conclusions

Vegetation indices are not enough to make a N decision as these are only useful when coupled with on-farm experimentation. Other variables, such as soil N, are also important to a certain extent. Secondly, and most importantly, AI methods with enough data (many observations of key variables) can improve mid-season N management by reducing error and bias if there is enough yield variability and, therefore, increasing the profitability of the decision. Finally, AI can be deployed on any farm and enhance the N decision-making prowess of the entire industry.

Acknowledgments

In addition to the authors, the Future Farm team comprises André Colaço, Damian Mowat and Rob Bramley (CSIRO), Mario Fajardo, Asher Bender, and Brett Whelan (University of Sydney), Alison McCarthy, Anand Pothula and Craig Baillie (University of Southern Queensland), Eileen Perry, Alex Clancy and Glenn Fitzgerald (Agriculture Victoria) and Stephen Leo, Daniele De Rosa and Peter Grace (Queensland University of Technology); their valued input is gratefully acknowledged. We are also most grateful to the various farmers who have provided us with fields in which to develop and test our methods – Ashley Wakefield, Ben Pratt, Bob Nixon, Ed Hunt, Jessica and Joe Koch, Mark and Sam Branson, Mark Swaffer, Rob Cole, Robin Schaeffer, Stuart Modra, Kieran Shepard and Peter Bell.

Bibliography

Colaço, A.F., B. Whealan, and R.G.V. Bramley. 2022. Better targeted, more precise fertiliser decisions as a counter to rising fertiliser prices – focussing on 3 of the 6 Rs. Precision Agriculture News 18 in press

Colaço, A.F., and R.G.V. Bramley. 2018. Do crop sensors promote improved nitrogen management in grain crops? F. Crop. Res. 218: 126–140. doi: 10.1016/j.fcr.2018.01.007.

Colaço, A.F., J. Richetti, R.G.V. Bramley, and R.A. Lawes. 2021. How will the next-generation of sensor-based decision systems look in the context of intelligent agriculture? A case-study. F. Crop. Res. 270(June): 108–205. doi: 10.1016/j.fcr.2021.108205.

Lawes, R.A., Y.M. Oliver, and NI. Huth. 2019. Optimal nitrogen rate can be predicted using average yield and estimates of soil water and leaf nitrogen with infield experimentation. Agron. J. 111(3): 1155–1164. doi: 10.2134/agronj2018.09.0607.

Lawes, R.A., and M.J. Robertson. 2011. Whole farm implications on the application of variable rate technology to every cropped field. F. Crop. Res. 124(2): 142–148. doi: 10.1016/j.fcr.2011.01.002.

Contact details

Jonathan Richetti
CSIRO
147 Underwood Av. Floreat 6014 - WA
Ph: 0411 150 425
Email: Jonathan.Richetti@csiro.au