Future farm - towards an improved sensor-based approach to nitrogen management

Take home messages

  • Future Farm is a large multi-institutional project which seeks to improve nitrogen (N) fertiliser decision making through the automated use of crop and soil sensors and other on- and off-farm data sources.
  • Early results from the field-based component of the research confirms the idea that an N fertiliser decision is a multivariate decision in which many factors contribute to the decision. That is, reliance on crop canopy sensing alone is unlikely to be effective.
  • Further work is aimed at developing appropriate multivariate models to support improved, site-specific N fertiliser decision making.

Background

Optimising the efficient use of nitrogen (N) fertiliser is an important driver of both profitability and productivity on cereal enterprises across Australia, typically representing approximately 30-40% of total input costs for a given season, with farmer risk shown to be closely related to the magnitude and effectiveness of expenditure on N fertiliser (Monjardino et al., 2013, 2015). On an industry-wide basis, the Australian grains sector applies approximately 1Mt N fertiliser annually (Angus and Grace, 2017), yet crop recovery of fertiliser N in the year of application is only around 45% (Angus et al., 2019). One way to optimise N use is to use the tools of Precision Agriculture (PA) to deliver on the ‘4 Rs’ – putting the right amount of the right product in the right place at the right time. However, in the absence of well defined, site-specific norms to underpin the ‘4 Rs’, implementing such strategies can require a substantial investment in time spent acquiring, processing and analysing data and may involve several steps that are not well integrated. Given the results of a recent survey (Bramley and Ouzman, 2018) which showed that farmer adoption of soil and crop sensors is low, and that confidence in the various decision aids which support N management is equivocal, there is an opportunity to re-examine and improve the way in which soil and crop sensors are used to inform decisions on N management.

The Future Farm project is supported by a joint investment by GRDC, CSIRO, the Universities of Sydney and Southern Queensland, Queensland University of Technology and Agriculture Victoria and is supported by a growing list of collaborating farmers. It aims to re-examine and improve the way in which soil and crop sensors are used to inform decisions on input management and to provide a way of automating the process from data acquisition, through analysis, to the formulation and implementation of decision options. Whilst the initial focus is on improving the efficiency and profitability of applied N, Future Farm is not a nutrition project per se. Rather, the main research focus is on the adaptive generation of site-specific management models through increased and improved use of in-season field monitored data (soil, crop, climatic), historic on-farm data, external public and private data and automation of decision rules in software that may potentially be linked to real-time application equipment. Further to a number of preliminary review projects (Bramley and Ouzman, 2018; Chlingaryan et al., 2018; Colaço and Bramley, 2018; Lawes et al., 2019), the development of such a decision aid is being based on the following identified operational targets:

  • N fertiliser application decisions should be supported by measures of plant N status (which in turn requires estimation of biomass), soil N status and soil water status/availability (i.e. a multi-sensor approach is required), together with assessment of the relative importance of measures of these attributes.
  • The decision support aid will use sensor data as a key input and employ machine learning methods of data integration for development of location-specific decision options.
  • Both remote and proximal sensing of the crop canopy will make an important contribution to N fertiliser decision-making, but need to be supported by some form of on-farm experimentation, with a zero N treatment (plot or strip); a critical enabler for interpretation.
  • Publicly available on- and off-farm data (soil survey data, weather and climate data sourced from the Bureau of Meteorology) along with historic yield monitor data and remotely sensed imagery (both on-farm and from adjacent areas) may provide valuable input to the decision tool.
  • The decision tool will be deployable in a way that will be complementary to the inclusion of other inputs/assessments that farmers and their advisers may also apply in decision making.

Key to the project is the recognition that in contrast to the univariate, plot-based, mechanistic approach used in much of the sensor-based N research (Colaço and Bramley, 2019), an N fertiliser decision is a multivariate decision in which the farmer and/or his/her adviser combines information from multiple sources; knowledge of historical paddock performance, crop and soil sensors, historical spatial data, publicly available datasets (for example; satellite imagery and weather data) and crop models, to predict the optimum N decision or variables that can be used for an N decision. Recognising the site-specific nature of the decision, it may be further informed by on-farm trials which are paddock-scale and spatially distributed (Bramley et al., 2013) and implemented and monitored using precision/digital agriculture tools (for example; variable rate applicators and yield monitors). In the present paper, we focus solely on the South Australian-based components of the Future Farm research, involving the use of proximal crop and soil sensing coupled with on-farm experimentation. We note however, that Future Farm is a national project with similar complementary field research to that described here being undertaken in each of the GRDC grain growing regions. Further information on the other components of the project are available in the proceedings of GRDC Updates (2020) held in other locations and, in the case of the use of off-farm data for on-farm decision support, in Fajardo (2019) and Fajardo et al., (2019).

Method

The four-year field program initiated in 2018 comprises two types of on-farm experiments; referred to here as ‘core’ and ‘satellite’ sites. Both employ spatially distributed, strip-based designs and ‘target’ calibration points where plant and soil samples are collected. Experiments at ‘core’ sites were designed with three specific objectives:

  • To provide on-farm estimates of the optimum N rate (ONR) against which a multivariate sensor-based model can be calibrated;
  • to enable investigation of the value of ‘N-rich’ and ‘N-minus’ strips as a ‘business as usual’ approach to in-season prediction of ONR and thus, fertiliser decision making; and
  • to provide a range of crop and soil conditions from where sensor calibration data can be taken.

On-farm experimentation

‘Core’ sites

Figure 1 shows the experiment implemented in 2018 at our the ‘core site’ near Tarlee, South Australia (SA). In a 64ha paddock, N-rich and N-minus strips were established using a liquid fertiliser sprayer of 39m width, split into three sections such that the N-rich strip was 26m wide, whilst the N-minus strip was 13m wide. After crop emergence (early June), the N-rich strip received 85kg N/ha. Additional N was applied to the paddock on three further occasions until flag leaf emergence; in early (28kg N/ha) and mid-July (38kg N/ha), and late August (34.5kg N/ha), the first two of these excluding the N-minus area and the last one across the entire paddock area. Given that mono-ammonium phosphate (MAP) applied at sowing resulted in an initial application to the whole paddock of 10kg N/ha, the final N rates applied were 195, 110 and 45kg N/ha for the ‘rich’, ‘paddock’ and ‘minus’ area, respectively. The strips were located such that they crossed the different management zones in the paddock, which were previously defined based on a cluster analysis of historical yield and soil electrical conductivity maps. For crop sensor calibration, 21 target locations spread across zones and strips were defined for soil and plant sampling. Three soil moisture probes (sensing to 1m depth) were installed, one for each management zone of the paddock; these were in addition to one already installed in the paddock. More details on the data collected is available in Table 1.

Figure 1. Trial layout at the 64ha South Australian ‘core’ site near Tarlee, in 2018.

Table 1. The range of data collected at the South Australian ‘core’ site.

Variable

Source

Sampling intensity

Crop stage

Vegetation indices

Proximal on-the-go (CropCircle™) and remote sensing (Sentinel imagery)

Whole paddock

GS22-23, GS31 and GS33-34

Visible and near-infrared crop reflectance

Proximal hyperspectral sensing (hand-held ~300-1800nm in ~1.5nm increments)

Multiple target points

GS31

Grain yield and protein

Yield and protein monitors at harvest on-the-go

Whole paddock

Harvest

Crop height and biomass

Light detection and ranging (LiDAR) on-the-go

Whole paddock

GS31

Biomass, plant N concentration, and other plant nutrition status

Plant sampling

Multiple target points

GS31 and at harvest

Soil N, other fertility status and texture

Soil sampling

Multiple target points

Pre-sowing, GS31 and harvest

Soil moisture

Soil moisture probe (insert type)

One target point in each zone

Daily across season

Soil drained lower and upper limits, bulk density

Soil profile characterisation

One target point in each zone

-

Soil electrical conductivity, soil gamma radiation, historical yield, etc.

Historical data base from farm records and previous research projects

Whole paddock

-

‘Satellite’ sites

The ‘core’ experimental sites were supplemented from 2019 by numerous ‘satellite’ sites based on a simpler experimental design and less intense monitoring; these are farmer-initiated N strip trials used to guide their mid-season N decision. The purpose of these trials to Future Farm is to broaden the range of biophysical conditions (soils and climate, especially rainfall) over which we collect data for calibrating the crop sensors for prediction of variables that can be used in N decision models; for example, mid-season crop biomass, plant N uptake and yield potential. In the southern region one ‘core’ site and seven ‘satellite’ sites were implemented in 2019.

Figure 2 shows the experimental design implemented in a ‘satellite’ trial located near Loxton, SA. At seeding, the farmer applied a 12m width strip with 46kg N/ha alongside a zero N strip, to guide his mid-season N decision. The paddock was divided into productivity zones based on historical yield maps. Twelve target points across the strips/zones were selected for crop scanning and plant sampling.

Data analysis

Interpolation of yield maps, establishment of management zones and other basic analysis of PA data followed accepted methods predominantly based on those outlined in Taylor et al., 2007 and implemented using PAT (Ratcliff et al., 2019). Here, the focus is on the analysis of the experimental data.

Two main approaches were followed for analysis of experimental data:

  • Prediction of ONR (i.e. building an N decision model); and
  • prediction of relevant variables for N management (mid-season crop biomass, plant N uptake, yield potential and grain N uptake) which can be used for a given N decision model.

Both approaches draw on our multivariate dataset.

The dataset from the 2018 core site in SA and the calibration dataset from SA in 2019 (‘core’ and ‘satellite’ sites) are used here to illustrate some of the analysis being explored in this project.

With the 2018 ‘core’ dataset, the difference in grain N removal between the N-minus and N-rich strips was calculated along the length of the strips in increments of 10m using the point data from the yield and protein monitors. These paired comparisons were also analysed using the moving window t-test of Lawes and Bramley (2012). The difference between the N-minus and N-rich strips was also used to calculate an N recommendation (in this case, the N rate which will maximise grain N removal) along the length of the paddock based on a given fertilisation efficiency factor (N rate = N removal difference between strips / fertiliser efficiency factor). Other options for N rate calculations will include an N budget approach based on soil data collected around each of the target points and crop modelling. In time, grain and fertiliser prices may also be added for the estimation of economically optimum N rates.

Figure 2. Trial layout at a South Australian ‘satellite’ site near Loxton, in 2019.

The normalized difference vegetation index (NDVI) and the normalized difference red-edge index (NDRE), measured at GS31 and obtained from both the CropCircle™ proximal canopy sensor and Sentinel satellite imagery, were used to calculate response indices (ratios between N-rich and N-minus strips; Raun et al., 2005) and examined as predictors of the final crop response to N and N requirement.

For demonstration of the second analytical approach, sensor calibrations for mid-season crop biomass, plant N uptake, grain yield, grain protein concentration and grain N uptake were generated based on the 2019 dataset using the target sampling locations.

For this paper, the results of simple linear regressions for both approaches are presented. It is the intention that these analyses will be enhanced using machine learning techniques to implement multivariate prediction algorithms. For the development of an N decision algorithm, such analyses will allow the assessment of which combination of variables can best predict ONR and where they should be measured (within reference areas (for example; N-rich, N-minus), under normal field conditions or both). The focus here on simple linear regression is to emphasise the need for a more multivariate approach.

Results and discussion

Figure 3 illustrates the results obtained from the strip analysis for the ‘core’ 2018 trial. Based on the difference observed between the N strips, it is seen that the crop responded to N mostly in terms of grain protein concentration and less so in terms of grain yield. As expected, grain N removal was greater in the N-rich compared to the N-minus strip, although to a varying extent along the strip. Consequently, the recommended N rate was also variable along the examined area. The relationships between vegetation indices and crop parameters were generally weak, especially for the Crop Circle data (Table 2), although the general trends of significant yield differences between strips (mainly between the 900 and 1250m marks, Figure 3) were identified by the sensors mid-season. Simple ratios between vegetation indices measured in the N-rich and N-minus strips were also poor predictors of crop response to N and of N requirement (Table 2), which might be partially due to the difficulty of predicting grain protein concentration by the sensors. However, and as expected, given the previous work of Colaço and Bramley (2019), mid-season predictions of harvest parameters (grain yield, grain protein and grain N uptake) and N demand based solely on vegetation indices were not successful. That is, a univariate approach based on sensor data alone, is not a sound basis for N fertiliser decision making. Further analysis will investigate the benefit of combining more prediction variables for multivariate models.

Figures 4 and 5 show the relationships between Crop Circle indices and crop parameters at mid-season (N concentration, dry weight and N uptake) and at harvest (grain yield, grain protein and grain N uptake) for the 2019 SA trials. Overall, sensor calibrations for individual sites were poor. Whilst the sensor indices are sensible to variations in some crop parameters (particularly to crop biomass), relationships between sensor and crop variables can be site-specific. Nonetheless, global calibrations were produced reaching R2s of up to 0.65 (for NDRE vs mid-season N uptake). As expected, predictions of harvest parameters were more difficult than prediction of mid-season crop measures. Again, further analysis will explore the use of multivariate models to improve predictions of such parameters.

Figure 3.Grain yield, grain protein, grain N removal and mid-season crop normalized difference vegetation index (NDVI) and the normalized difference red-edge index (NDRE from the Sentinel satellite across the length of N-rich and N-minus strips (top five graphs)) (results of strip comparison based on moving window t-test shown in different background colours) and N rate recommendation (bottom graph) along the length of the strips.

Figure 4.   Crop Circle calibration data for each of the 2019 experimental sites in South Australia (top three rows refers to crop parameters around GS31-33).

Crop Circle sensor calibrations using pooled data for all South Australian sites in 2019

Figure 5.    Crop Circle sensor calibrations using pooled data for all South Australian sites in 2019.

Table 2.Correlation (r) between vegetation indices and harvest parameters. ‘Response’ was calculated as the ratio between N-rich and N-minus values.

 

Grain yield

Grain protein

Grain N uptake

Grain yield response

Grain protein response

Grain N uptake response

Recommended N rate

Crop Circle NDVI

0.34

0.18

0.46

-

-

-

-

Crop Circle NDRE

0.33

0.27

0.51

-

-

-

-

Sentinel NDVI

0.56

-0.10

0.50

-

-

-

-

Sentinel NDRE

0.44

0.23

0.59

-

-

-

-

Crop Circle NDVI response

-

-

-

0.31

-0.13

0.18

0.14

Crop Circle NDRE response

-

-

-

0.32

-0.10

0.18

0.15

Sentinel NDVI response

-

-

-

0.59

-0.25

0.40

0.37

Sentinel NDRE response

-

-

-

0.38

-0.04

0.25

0.20

The strip analysis shown for the 2018 ‘core’ site (Figure 3) demonstrates the approach’s ability to generate data (observations of crop parameters and crop response to N) that covers a range of biophysical conditions within a single paddock for the calibration of sensor-based decision models. The study was able to capture an even greater range of variability for the sensor calibrations through the farmer-led ‘satellite’ trials; both ‘core’ and ‘satellite trials highlighting the value of on-farm experimentation in developing a basis for site-specific decision making. Thus, just as the technologies of precision and digital agriculture (PA/DA) promote an ability for these new spatially distributed approaches to field experimentation, the successful adoption of PA/DA (for N management in this case) is likely reliant on such experimentation for the development of management norms appropriate to the farming system at any given location. The approach being used in Future Farm is reflective of this, in that our focus is on the development of an appropriate process for the acquisition and analysis of multivariate data to inform site-specific management. Thus, we are using the techniques of PA/DA to move away from the idea that norms for fertiliser management are ubiquitously applicable. It is hoped that the merits of this approach will be demonstrated in future updates to industry.

Conclusion

N fertiliser decisions are a multivariate issue which therefore require multivariate input. Future Farm is seeking to develop an automated, sensor-based approach to the delivery of site-specific decision support. Results to date confirm that a univariate approach based solely on either satellite imagery or proximal crop canopy sensor data alone is unlikely to deliver value to farmers. Moving forward, the focus will be on adding value to such sensor data through the development of multivariate N decision models.

Acknowledgements

This work is supported by a joint investment by CSIRO, the University of Sydney, University of Southern Queensland, Queensland University of Technology, Agriculture Victoria and GRDC, whose input is critically dependent on the significant contributions of growers. The work is also critically dependent on our collaborating growers (and their advisers) who have provided access to their farms, laid down trials and otherwise enabled the research to proceed. In this regard, we are most grateful to Mark Branson, Bob Nixon, Rob Cole, Jessica and Joe Koch, Ashley Wakefield, Ben Pratt and Sam Trengove, Ed Hunt, Mark Swaffer, Stuart Modra and Robin Schaefer. We are also indebted to Damian Mowat (CSIRO) for his excellent technical assistance.

Useful resources

A video describing much of the above work at our SA ‘core’ site is available at Future Farm - GRDC - YouTube

References

Angus JF, Grace PR. 2017. Nitrogen balance in Australia and nitrogen use efficiency on Australian farms. Soil Research 55, 435-450.

Angus J, Bell M, McBeath T, Scanlon C. 2019. Nutrient management challenges and opportunities in conservation agriculture. In: Pratley J and Kirkegaard J. (Eds). Australian agriculture in 202: From conservation to automation. pp. 221-236. Agronomy Australia and Charles Sturt University, Wagga Wagga.

Bramley RGV, Ouzman J. 2018. Farmer attitudes to the use of sensors and automation in fertilizer decision-making: Nitrogen fertilization in the Australian grains sector. Precision Agriculture 20 157-175.

Bramley RGV, Lawes RA, Cook SE. 2013. Spatially distributed experimentation: tools for the optimization of targeted management. In: Oliver MA, Bishop TFA, Marchant BM. (Eds). Precision Agriculture for Sustainability and Environmental Protection. Earthscan, Food and Agriculture Series. Routledge, Abingdon, UK. pp. 205-218.

Chlingaryan A, Sukkarieh S, Whelan BM. 2018. Machine learning approaches for crop yield prediction and nitrogen status estimation in precision agriculture: A review. Computers and Electronics in Agriculture 151, 61-69.

Colaço AF, Bramley RGV. 2018. Do crop sensors promote improved nitrogen management in grain crops?  Field Crops Research 218, 126-140.

Colaço AF, Bramley RGV. 2019. Site-year characteristics have a critical impact on crop sensor calibrations for nitrogen recommendations. Agronomy Journal 111 2047-2059.

Fajardo M. 2019. Progress in using off-site data for PA. Proceedings of the 22nd Symposium of Precision Agriculture in Australasia. 9-10 September, Launceston. SPAA / University of Sydney. pp. 35-39.

Fajardo M, Whelan BM, Filippi P, Bishop TFA. 2019. Wheat yield forecast using contextual information. In: J.V. Stafford (ed.) Precision Agriculture ’19, Wageningen Academic Publishers, pp 713-720.

Lawes RA, Oliver YM, Huth NI. 2019. Optimal nitrogen rate can be predicted using average yield and estimates of soil water and leaf nitrogen. Agronomy Journal 111 1155-1164.

Monjardino M, McBeath TM, Brennan L, Llewellyn RS. 2013.  Are farmers in low-rainfall cropping regions under-fertilising with nitrogen? A risk analysis. Agricultural Systems 116 37-51.

Monjardino M, McBeath T, Ouzman J, Llewellyn RS, Jones B. 2015. Farmer risk-aversion limits closure of yield and profit gaps: A study of nitrogen management in the southern Australian wheatbelt. Agricultural Systems 137, 108-118.

Ratcliff C, Gobbett D, Bramley R. 2019. PAT - Precision Agriculture Tools. V2. CSIRO. Software Collection. 
CSIRO Data Access Portal - Precision Agriculture Tools

PAT - Precision Agriculture Tools

Raun W, Solie JB, Stone, ML, Martin KL, Freeman KW, Mullen RW, Zhang H, Schepers JS, Johnson GV. 2005. Optical sensor-based algorithm for crop nitrogen fertilization. Communications in Soil Science and Plant Analysis 36 2759-2781.

Taylor JA, McBratney AB, Whelan BM. 2007. Establishing management classes for broadacre agricultural production. Agronomy Journal 99, 1366–76.

Contact details

Dr Rob Bramley
Senior Principal Research Scientist
CSIRO, Waite Campus
Locked Bag 2
Glen Osmond
SA 5064
08 8303 8594
rob.bramley@csiro.au

GRDC Project Code: CSP1803-020RMX,