Take home message

• Current methods for calculating in-season nitrogen requirements based on a single sensor and simplistic agronomic decision frameworks can match the performance of good farmer management decisions at the uniform field-rate level
• The way to improve the accuracy and profitability of nitrogen rate decisions when the uniform-rate decisions are near optimal is by increasing the spatial resolution of management
• Empirical, multivariate, data-driven methods for predicting the site-specific, economically optimum nitrogen application rate (EONR) have potential to successfully increase the spatial resolution of decisions and reduce the error and increase profitability of fertiliser management (~$50/ha in this study)
• Data availability is critical to enable data-driven prediction methods and increase profitability. On-farm experimentation (OFE) is a critical enabler of these data-driven decision tools as they allow the automated collection of large digital datasets of crop response to applied nutrients that are needed to train the algorithms. Such OFE should be adopted as a core element of farm business operation to support decision optimisation
• Farm businesses that collect and maintain relevant production response and resource data will be able to push closer towards season- and site-specific economically optimum operations.

Introduction

Future Farm is a research program combining skills from CSIRO, USYD, USQ, QUT and AgVIC. It aimed to re‐examine and improve the way in which on-farm soil and crop sensors, and digital data from elsewhere, could be used to improve decisions about input management and explore automation of the process from data acquisition, through analysis, to the formulation and implementation of decision options. The research focused on nitrogen application decisions, but the concept could be applied to any rate-based inputs (e.g., lime, gypsum and other nutrients).

Methods

A program of on-farm experiments (OFE) was established at sites across the three GRDC regions that were designed to document the local yield and protein response to applied nitrogen. The trials included three application rates: a zero (or reduced) nitrogen rate, a farm decision nitrogen rate, and a high nitrogen rate that ensures nitrogen should be non- limiting. The nitrogen rate treatments were placed adjacent to each other in strips or plots and the treatments were applied to run through zones of predetermined potential management classes in each field site (Figure 1). Data was gathered from the OFE in-season and the harvest data from the OFE was used to evaluate and compare the benefits of a range of different methods that the team designed that use digital data from a range of sources (Appendix 1) to ultimately predict the N requirement in different agro-ecological zones across Australia.

Figure 1. An example of a field trial layout (left) and grain quality data (right) for a site in Northern NSW.

Assessment of nitrogen recommendation models

The assessment process was designed to provide a comparison between the average ability of the different methods designed by the team to predict nitrogen requirement across the sites, in terms of both average accuracy and average profitability. The assessment process also allows the developed methods to be compared against the current farm decision approach (a whole-paddock uniform rate decision) and also against a number of currently used ‘benchmark’ methods. A number of simple methods for predicting nitrogen requirement using Normalized Difference Vegetation Index (NDVI) and Normalized Difference Red Edge (NDRE) indices calculated from ground-based Crop Circle sensors (CC) and satellite systems (Sentinel) have also been included. These provide an evaluation of a digital, single sensor approach to N estimation. In all, 15 different models were produced and tested, however Table 1 shows the models employed in the final comparison process.

Yield and protein data were obtained from the OFE using commercially-available, harvester-mounted yield and protein sensors. A regression analysis was performed between applied N rates and the harvest data (grain yield and protein) along the length of the trial strips, fitting a quadradic model to the response data using a moving window approach (Figure 2). This process provided site-specific functions at a 10 m scale for each site (site-specific management analysis) which could then be aggregated up to an average response function for each potential management class where possible (potential management class analysis) or an average single response function for the whole paddock (whole-paddock uniform management analysis).

From these functions the economically optimal N rate (EONR) was calculated as the rate that maximised partial profit (harvest income minus expenditure on N fertiliser). The harvest income was calculated based on the average grain price between 2018 and 2020 adjusted for protein premiums based on Table 2. The EONR is regarded here as the ultimate nitrogen application rate decision against which all the recommendation and benchmark methods were compared. The comparison was made in terms of the prediction accuracy of each method (root mean squared error, RMSE) and also the resultant partial profit. The partial profits for the different prediction methods were obtained by inserting the nitrogen recommendations from each into the respective partial profit response functions for each site.

The results of the comparative assessment are tabulated in Table 3 where the methods are ranked on normalised partial profit (NPP) achieved. The NPP is the partial profit achieved for each prediction method normalised against the EONR partial profit values for each site at the different management scales. In Figure 3, each method is plotted based on the average achieved across all sites for RMSE and NPP. The results are only shown here for the site-specific and whole-paddock management levels to enable a clearer view of the comparison.

Table 1. Methods included for in-season prediction of nitrogen requirement.

Figure 3 shows that as accuracy in prediction of nitrogen required increases (decreasing RMSE), partial profit increases, but the rate of increase diminishes as the methods become more accurate. This result reflects the often ‘flat’ profit response to applied N around the optimum rate which can limit the improvement in profitability through greater accuracy. However, while the rate of increase in NPP slows, the increase in accuracy means an application rate closer to target is more often achieved, bringing a commensurate decrease in the risk associated with N management. Thus, reducing the error improves the chance of getting the management decision correct (or less chance of making an incorrect decision) and increases farmer confidence in the decision. Better targeting of nitrogen application rates to optimum also has implications for minimising potential environmental impact.

Figure 2. An example of the moving window regression analysis at one point along a strip trial at a site in SA.

Table 2. Average grain grade and nitrogen prices applied in partial profit analysis.

The best future farm method was the data-driven model (DD data abundance), which represents a situation where a database is available for fields that provides information on past OFE and production response and associated environmental data that encompasses the variation in production that can be achieved in that field. This method succeeds in providing better estimates than the average farm- based decision (farmer) at all management scales. At the whole-field uniform and site-specific management scales, the DD (data abundance) can improve NPP by 2% ($14/ha) and 5% ($47/ha) respectively, over the average Farmer method (Table 3; Figure 3).

Table 3. Results of the comparison of methods for predicting N requirement along with benchmark methods for comparison. Methods ranked on normalised partial profit achieved.

The farmer recommendation is on average 2% ($14/ha) lower in NPP than the average uniform EONR (Table 3). All the farmers involved in the project were skilled in using Precision Agriculture technologies in farm management and the result here confirms that they were very good at calculating nitrogen requirement for the seasons under study. Aside from the top performing (DD data abundance) method, a number of the other digital mechanistic sensor-based methods (e.g., ‘Yield prediction (RF)’, ‘Response function (NDVI Sent)’ and ‘N sufficiency MV’) approached within 1% ($13 - $16/ha) of the Farmer decision level of average profit and accuracy at the uniform application scale (Table 3; Figure 3). The response function approach was sensitive to the type of input data, that is, some combinations of vegetation indices and their data sources may match a profit response function better than others.

For the fields in this study, the sensor-based response function that most resembled the final profit function was the one derived from Sentinel 2 NDVI. Since Sentinel data can be accessed for free, the fact that it out-performed the methods which used proximal crop sensors represents an important result for farmer adopters of such PA technologies. The ‘N sufficiency’ method, a less common sensor-based approach, had a similar performance to the ‘yield prediction’ methods.

Figure 3. Profitability (y axis) verses accuracy (x axis, root mean squared error) of methods for N recommendation averaged across trial sites at uniform and site-specific management scales. All relevant sites utilised.

However, apart from the DD (data abundance) approach, all the other methods performed worse on average when implemented at the site-specific scale as compared to the unform application scale (Figure 3) because of their larger recommendation error. In some cases, using the methods to calculate a uniform application rate reduced the error substantially, suggesting that methods in which the recommendation error is expected to be large are better implemented as the average for the field instead of site-specifically.

The variables that proved of most importance for the DD (data abundance) method were the Sentinel NDVI from each N strip, historic yield monitor data, accumulated rainfall to GS31 and the historic NDRE (95^th percentile). For the mechanistic models that predicted yield in order to calculate a nitrogen recommendation rate, the most important variables were historic yield monitor data, accumulated rainfall to GS31, Sentinel NDRE at GS31 from all strips and historic NDRE (5^th percentile). From this it is clear that information derived from the OFE are crucial for these prediction methods.

The average results shown in Figure 3 also suggest that. Moving towards a more site-specific nitrogen management approach may be more achievable in all locations using an approach that includes more information that is able to describe the range of site and season variables that impact the response to nitrogen.

Conclusions

From assessing these varied model types over multiple sites in five States, it appears the farmers in this study are operating near the optimum management level at the uniform paddock scale, for the seasons included, and they would need to move towards site-specific decisions using a data-driven approach with more data in order to improve the accuracy and profitability of nitrogen requirement decisions. The methods built here require further development and testing. However, the data-driven approach with increasing levels of OFE data for directly predicting ONR appears to be a promising target methodology for improving site-specific decisions. Using the methodology at the uniform paddock scale would also be a viable approach to improve uniform management decisions on farms where uncertainty in decisions at that level remain high.

The data-driven approach relies on data availability to ensure the method performs at its optimum. Its success at all management scales in this assessment provides a significant pointer towards a future where farm businesses that collect and maintain relevant production response and resource data will be able to push closer towards season- and site-specific economically optimum operations.

A system in which farmers share OFE data across larger communities may also play a crucial role in building the necessary database for empirical DD approaches. Improvements can be gained as soon as more OFE data is collected and made available from farms. However, as formalised OFE is an exception rather than a rule across farming operations, this means it is not an approach that can currently be employed by every farmer.

Acknowledgements

The research undertaken as part of this project is made possible by the significant contributions of growers through trial cooperation and the support of the GRDC, the authors thank them for their continued support. This research was cofounded by the Commonwealth Scientific and Industrial Research Organisation (CSIRO), the University of Sydney, Agriculture Victoria, the University of Southern Queensland, the Queensland University of Technology and the GRDC. The authors are also most grateful to the various collaborating farmers who have provided us with commercial cereal 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 Shephard and Peter Bell.

Contact details

Brett Whelan
Precision Agriculture Laboratory
University of Sydney, 1 Central Avenue, South Eveleigh, NSW 2015
Ph: 02 8627 1132
Email: brett.whelan@sydney.edu.au

Date published: March 2023

Appendix 1. List of variables available for use in digital N recommendations.

GRDC Project Code: CSP1803-020RMX,