Advanced

Field-Scale
Experimentation
for Grain precision
A guide to using
agriculture

Growers
to improve trial
results

Author: Brett Whelan,

Precision Agriculture Laboratory,
University of Sydney
 Title: A
 dvanced Field-Scale Experimentation for Grain Growers:
A guide to using precision agriculture to improve trial results.

GRDC Project Code: US00044

Author: B
 rett Whelan, Precision Agriculture Laboratory,
University of Sydney for the GRDC.

© 2015 Grains Research and Development Corporation and

the University of Sydney.
All material published in this guide is copyright protected and may not
be reproduced in any form without written permission from the GRDC
and the University of Sydney.

Published December 2015

ISBN: 978-1-921779-55-8

In submitting this report to the GRDC, the Precision Agriculture Laboratory

and the University of Sydney have agreed to the GRDC publishing this
material in its edited form.

2
This booklet can be downloaded from the GRDC website at:
ADVANCED FIELD-SCALE EXPERIMENTATION FOR GRAIN GROWERS

www.grdc.com.au/AdvancedFieldScaleExperimentation

or for copies please contact:

Ground Cover Direct
Free phone: 1800 11 00 44
Email: ground-cover-direct@canprint.com.au

Or

Ms Maureen Cribb Disclaimer:

GRDC Publishing Manager Any recommendations, suggestions or opinions contained in this publication do not
PO Box 5367 necessarily represent the policy or views of the Grains Research and Development
Corporation (GRDC) or the University of Sydney. No person should act on the basis of the
KINGSTON ACT 2604 contents of this publication without first obtaining specific, independent professional advice.
Phone: 02 6166 4500 The Grains Research and Development Corporation and the University of Sydney may identify
products by proprietary or trade names to help readers identify particular types of products.
Email: maureen.cribb@grdc.com.au We do not endorse or recommend the products of any manufacturer referred to. Other
products may perform as well as or better than those specifically referred to. The GRDC and
the University of Sydney will not be liable for any loss, damage, cost or expense incurred or
arising by reason of any person using or relying on the information in this guide.
Design and production:
Caution: Research on Unregistered Pesticide Use
Coretext, www.coretext.com.au
Any research with unregistered pesticides or of unregistered products report in this guide
does not constitute a recommendation for that particular use by the authors or the author’s
organisations. All pesticide applications must accord with the currently registered label for
that particular pesticide, crop, pest and region.
ABOUT THIS GUIDE 3

Applying the precision agriculture (PA) philosophy to crop management requires access to site-specific information. The
amount and pattern of variability in soil, landscape attributes and production output, the reasons for the variability, the

ADVANCED FIELD-SCALE EXPERIMENTATION FOR GRAIN GROWERS

agronomic implications and the management opportunities all ultimately need to be identified locally on each farm and in
each field.

When each farm and field is explored in this detail it is desirable to obtain information on any variability in crop yield in
response to different application rates of inputs across an entire site in order to identify if worthwhile production gains and
increased profit can be made by implementing spatially variable application rates.

This guide is designed to complement the earlier GRDC publication Designing your own on-farm experiments: how
PA can help. With the now widespread use of global navigation satellite systems (GNSS) for autosteer vehicle navigation
and variable-rate application of inputs, it is hoped that further discussion on design criteria and analysis provided here will
stimulate more on-farm research and help grain growers get the most out of PA.
 CONTENTS
On-farm experimentation...........................................................................................................................................................................................................................4
Field-scale experimentation....................................................................................................................................................................................................................4
Planning a rate-response experiment........................................................................................................................................................................................5
Potential management class (PMC)..............................................................................................................................................................................................5
Designing rate-response experiments.......................................................................................................................................................................................7
Treatment rates........................................................................................................................................................................................................................................................7
Field plan.........................................................................................................................................................................................................................................................................7
Application equipment............................................................................................................................................................................................................................7
Measuring the production response......................................................................................................................................................................................8
Treatment plot type.....................................................................................................................................................................................................................................8
Treatment replication and randomisation........................................................................................................................................................................9
Implementation.................................................................................................................................................................................................................................................9
Summary..................................................................................................................................................................................................................................................................... 10
Analysis of rate-response experiments................................................................................................................................................................................ 10
Question 1: Is there a difference in production response?......................................................................................................................... 10
Question 2: What is the optimum application rate?............................................................................................................................................ 11
A nitrogen fertiliser case study........................................................................................................................................................................................................ 12
Using the results................................................................................................................................................................................................................................................ 15

ON-FARM EXPERIMENTATION
Grains industry R&D continues to contribute significantly the general principles and results from grains industry
4 to improvements in crop production across the country. R&D and test them in their own production systems.
Results and recommendations are often targeted for On-farm experimentation allows the impact of changes in
ADVANCED FIELD-SCALE EXPERIMENTATION FOR GRAIN GROWERS

use at the regional or district level. With an increasing the production process to be measured right down to the
understanding of the variability in production at the local individual field scale so growers can more accurately judge
level, many growers and farm advisers are keen to take any benefits for themselves.

FIELD-SCALE EXPERIMENTATION
With increased use being made of precision agriculture (PA) more alternative management practices or products (for
technologies for vehicle navigation and managing inputs, example, cultivation versus minimum or no-till, alternative
grain growers also have a greater opportunity to use PA chemicals, crop varieties, fertiliser products)?
to run their own field-scale trials. Being able to run field- n Rate-response experiment
scale trials on any/every field is one of the major benefits Is there a change in production response to different
of growers adopting PA technologies, and work continues rates of an input (for example, nutrient application rates,
on practical trial design and analysis to ensure meaningful irrigation water, seeding rates)?
results can be achieved by grain farmers. The first question is more easily dealt with in terms of
There are essentially two different field-scale experimental design and analysis. The second question
experimental types/questions that can be asked is more complex and the design and analysis potentially
by grain farmers. becomes more involved. This guide targets design and
n Approach-response experiment analysis for rate-response experiments but the techniques can
Is there a difference in production response to two or also be applied to designing approach-response experiments.
PLANNING A RATE-RESPONSE EXPERIMENT 5

For a field-scale rate-response experiment it is best to test Agriculture Manual, Applying PA Education and Training
rate changes in one single input or practice at a time, to avoid Modules for information on the stepwise implementation of PA

ADVANCED FIELD-SCALE EXPERIMENTATION FOR GRAIN GROWERS

potential confusion about the cause of any detected change in into grain crop management to explain this process in detail).
yield or profit. It keeps the design, analysis and interpretation Variation in production potential that remains after this step
of results simpler and allows for a greater number of treatment can be mainly attributed to natural variation in production
rates or replications of treatment rates within a given area. In prospects (for example, due to changes in soil depth or texture
order to get the best results from a rate-response experiment within the field) in combination with weather conditions. Weather
for a single input or practice, it is important that the field has conditions are generally considered to be uniform over a field
been investigated for variability in factors that may affect but in any case are difficult to predict before an experiment.
the production response to the input. Site-specific sampling What can be assessed and documented is any pattern in the
should have been carried out as part of a whole PA plan variation of the natural yield potential within a field. It is important
to diagnose causes of historic production variability and that this step is carried out before designing the field experiment
ameliorate any obvious soil issues (for example, areas of as it can be crucial to achieving an agronomically useful result.
hostile soil pH) or rectify management-induced issues such At present the most effective way to achieve this is to apply the
as areas of significant weed burden (see the GRDC Precision potential management class (PMC) approach.

Potential management class (PMC) FIGURE 1 The number of management classes

Fields are divided into areas (classes) that have shown affects the way variability is categorised.
differences in production potential which may require different low high
management treatments. A management class can be
assigned to one or more management zones within a field. 1 2
Figure 1 shows how this process breaks the production
variability into a number of relative categories, with more low average high
classes increasing the ability to describe the range of
potential production variability. 1 2 3
The boundaries of the classes/zones can be drawn
using a wide range of methods and initial information. The low medium low medium high high
methods range from growers drawing them by hand using
their knowledge of a field, such as in Figure 2, to statistical 1 2 3 4
processes that use one or more whole-field maps of
variability in yield/soil/terrain/reflectance (examples of which
are in Figure 3) to partition the production potential. Figure Low Average High
4 shows the result of combining a number of the layers in
Range of soil or crop measurements in a field
Figure 3 in a statistical process to form PMCs.
Using previously gathered maps of variability and
statistical processes takes the guesswork out of setting FIGURE 2 Hand-drawn class/zone boundaries
boundaries, but it is important to ensure that any maps of potential production variability within a field.
that are used correctly reflect the variability in the field and Medium
don’t have any problems caused by poor management
Low
or collection errors. As can be seen from the legends in
Figures 1 and 4, when the number of PMCs is increased the
differences between the production potential in each PMC
decreases. As these differences decrease, the potential
benefits of managing the classes differently are also reduced.
So any decision on how many management classes to use
should be made with consideration of both maintaining
an agronomically and financially significant difference in
production potential and ensuring the pattern of variation is
not overly broken up into many small zones.
In practice, one to four management classes for a
particular operation are used across Australia. Most
commonly one to three management classes are identified:
n one – uniform production potential;
n two – a division between high and low; or
n three – high, medium and low production potential identified. High
 FIGURE 3 Individual maps that may be used to help define boundaries of PMCs: crop yield maps
(a and b), elevation (c) and soil apparent electrical conductivity (d).
(a) (b)

Sorghum yield Chickpea yield

(t/ha) (t/ha)
3.05-3.47 0.06-0.28
3.48-3.88 0.29-0.50
3.89-4.30 0.51-0.72
4.31-4.71 0.73-0.93
4.72-5.13 0.94-1.15
5.14-5.55 1.16-1.37
N
5.56-5.96 1.38-1.59
5.97-6.38 1.60-1.81
6.39-6.80 1.82-2.03
6.81-7.21 2.03-2.25

0 150 300 450 600

(c) metres (d)

Elevation Soil ECa

(metres) (mS/m)
357-360 42-50
361-362 51-59
363-365 60-67
366-368 68-76
369-370 77-84
371-373 85-93
374-376 94-101
6 377-378 102-110
ADVANCED FIELD-SCALE EXPERIMENTATION FOR GRAIN GROWERS

379-381 111-118
382-384 119-127

FIGURE 4 Two (a) and three (b) potential management classes (PMCs) built using the soil ECa,
elevation and crop yield information for the field shown in Figure 3.
(a) (b)

Potential Potential
management class management class

Average Average Average Average

sorghum chickpea sorghum chickpea
yield yield yield yield
season 1 season 2 season 1 season 2
(t/ha) (t/ha) (t/ha) (t/ha)
1 4.8 1.1 1 4.7 1.1
2 5.8 1.5 2 5.7 1.3
3 5.9 1.5

0 150 300 450 600

metres
DESIGNING RATE-RESPONSE EXPERIMENTS 7

This diagnostic process undertaken to describe/explain variable costs of production. This might involve testing the
the extent of production variability within a field should profitability of increasing nutrients in some parts of the field,

ADVANCED FIELD-SCALE EXPERIMENTATION FOR GRAIN GROWERS

help define which input is the most useful candidate for reducing (or even eliminating) the amount added in other
including in a rate-response experiment. As an example, parts, or examining the financial benefit of redistributing the
a marked build-up or depletion of a soil nutrient identified same total amount of nutrient across the different PMCs; in
between classes could be used as a criterion, along with the each case, the aim might be to better match input rate to
magnitude of contribution the required input makes to the the water-limited crop potential.

Treatment rates
Once the input of interest has been chosen, it is time to the controlling system sends an appropriate signal to
decide on the number of treatment rates. The traditional or the actuator/motor/pump on the application implement
current best practice application rate that will be applied to to direct the application of the input at the required rate.
the majority of the field becomes one treatment. More than An as-applied map can be constructed from information
one alternative treatment rate is preferable, otherwise the gathered in a feedback loop through the variable-rate
best that can be hoped for is a decision that one treatment controller and the actuators/motors/pumps involved in the
is preferable to another. At least two alternative treatments application process. It allows managers to check what rates
can help provide information to decide the ‘best’ treatment were actually applied. This is very useful for identifying any
rate. mistakes and allows a record to be kept of the location and
So while the majority of the field is to be treated with treatment rate actually applied in the experiment.
a rate that is the current best practice, the alternative When using VRT, the smallest area in a field that can be
treatments could be calculated as multiples of that treated differently is related to:
application rate (for example, 0.25x, 0.5x, 1.5x, 2x). The n the minimum independent section width of the application
information from the experiment will be greatest when there equipment;
is a good size difference between the treatment rates. Where n the speed at which a rate change can be relayed to the
possible a zero rate (or very low) treatment should also be equipment;
included to help show the full scale of response. n the time it takes for the equipment to change up and
Using a low treatment rate obviously depends on the down rates; and
input being tested and the actual range of rates over n the speed of travel.
which the input would be applied in practice. However, it is This means that each equipment set-up and its operation
important to have low treatment rates where the experiment will provide a unique minimum area of application. So by
is testing whether a reduction in input rate is viable. choosing/changing equipment and speeds, growers have a
variety of options available to them for controlling the scale
of treatment application.
Field plan
The location and size of the applied treatments within
the field should consider the capability and width of input
application equipment, the method of measuring the
production response and the incorporation of any PMC
pattern. Depending on the range of treatment rates, it may
also be desirable to minimise the area/financial impact of the
experiment.

Application equipment
Treatments can be applied with traditional application
equipment using manual switching or multi-pass applications
to achieve different input rates. Alternatively, variable-rate
technology (VRT) is available for many rate-based input
operations. Using VRT makes the job less stressful and
allows more sophisticated positioning and rate adjustments
of the treatments.
When using VRT for experimental layout, a map of
desired application rates or actions (prescription map)
is produced for the field and loaded into the controlling
system prior to the actual operation. Figure 5 (page 8)
shows a general schematic of a mobile VRT system where
 FIGURE 5 Schematic description of the components required for a generic map-based
variable-rate technology (VRT) system.
Prescription As-applied
map map

Tank/bin
(product storage)

Return Pump
pipe

GNSS

Flow controller

VRT controller

Flow meter

Pipe work Electrical cable Digital media

Delivery of product

Measuring the production response Treatment plot type

Usually the production response being measured is This type of experiment can be run using field-length
crop yield, but it may be a quality parameter or biomass treatment strips, where each treatment rate is applied as
production. Whatever the measured response, it is important a strip over the length of the field (Figure 6a). Many grain
to understand the resolution of the chosen measurement growers have experience with this design as it is often used
system so that treatment plots are large enough to be for variety trials. Field-length strip designs are not reliant on
successfully monitored. VRT as the treatment rates can be set manually at the start
For grain crop yield, a harvester-mounted yield monitor of a strip and adjusted again at the beginning of the next
is ideal as it allows the whole field to be monitored run. However, where VRT is available the application of small
automatically. The impact of the resolution of the yield strip treatment plots is very feasible (Figure 6b).
monitor on design and analysis are discussed later. A weigh The minimum length of the ‘small strips’ will be
wagon or truck scales can be used where a yield monitor is constrained by the rate-command response of the
8 not available, but the harvesting of treatments needs to be application equipment and the resolution of the production-
performed carefully. response measurement equipment. For example, when
ADVANCED FIELD-SCALE EXPERIMENTATION FOR GRAIN GROWERS

FIGURE 6 Two general options for rate-response experiments in a field with two PMCs.
Whole-field strips (a) and small strips (b).
(a) (b)
Application rate Application rate
(kg N/ha) (kg N/ha)
0 0
29 29
60 (rest of field) Class 2 60 (rest of field) Class 2
82 82

Class 1 Class 1

0 125 250 375 500

metres
using a harvester-mounted yield monitor a significant amount Implementation 9
of internal grain mixing occurs inside the harvester as it With either plot type, a minimum of two alternative treatment
travels along its operational path. This means that yield data rates (that is, three treatments in total including the current
gathered at the beginning and end of each strip should best practice applied to the rest of the field) and two
be regarded as contaminated by surrounding treatments replications per PMC are strongly recommended, regardless

ADVANCED FIELD-SCALE EXPERIMENTATION FOR GRAIN GROWERS

(usually standard field treatment). To compensate for this of the other constraints imposed on the production system.
and to cover the majority of VRT response specifications, Using the design parameters described, the number
the small strips should be a minimum of 100 metres long. of plots required for harvester-monitored rate-response
Twenty metres of yield data from each end of the harvested experiments with varying treatment and replicate numbers
treatment strips should be discarded from the analysis to can be calculated from Table 1. The minimum experimental
remove potential treatment contamination. This leaves a area required for the alternative treatment plots can be
minimum analysable ‘small strip’ length of 60m. calculated as:
Whether field-length or small strip plots are chosen,
the plots must be laid out in the direction of sowing and Field-length strip area (m2) =
harvesting if a harvester-mounted yield monitor is being plot number x 3 x harvester width (m) x
used. The width for both plot types should be at least field length (m)
three harvester widths to ensure that one full cutting width
can be obtained from each strip without the possibility of
Small strip area (m2) = plot number x 3
contamination from adjoining treatments. Therefore, the x harvester width (m) x 100 (m)
minimum width for either strip type will be determined by the
minimum multiple of the application machinery width that will TABLE 1 Minimum number of plots required for a
meet this target. given combination of treatments and replicates for
If the crop response is to be determined by canopy a field with either 2 or 3 management classes. Left-
sensing instead of harvester-mounted yield monitors then hand value indicates the number of plots required for
the resolution of the canopy sensor (proximal, aerial or 2 management classes and the right-hand value the
satellite) needs to be considered in determining the ‘small
number of plots for a 3-management-class field.
strip’ length and the width of either type of plot. Replicates Number of alternative treatments (excluding rest-of-field
treatment)
Two Three Four Five
Treatment replication and randomisation
Two 4/6 8/12 12/18 16/24
Replication refers to the repetition of the treatment rates
in more than one plot within a field. Randomisation is the Three 6/9 12/18 18/27 24/36
process of allocating the different treatment levels to different Four 8/12 16/24 24/36 32/48
field plots without any conscious choice or favour. These
two operations help to ensure the observed response
from a particular treatment is not specifically influenced by
other sources of variation and that it is representative of
the experimental area and not biased by decisions to put
high or low rates in a particular spot(s). For example, in an
experiment where a field is divided in half and one treatment
is applied to one half and an alternative treatment to the
other, it is impossible to tell if any difference in response
between the treatments reflects real treatment effects and
not inherent differences between the two halves of the field.
Replicating and randomising treatments across the two
halves of the field helps solve this problem.
However, with field-length strips the design can start to
take over the field and, where low or very high treatment
rates are included in the design, some significant loss
of production may be encountered in the name of
experimentation. This is where the use of small strips
becomes advantageous. The small strip plots allow a larger
number of true replications within a given area (increasing
the potential statistical value of the trial), more flexibility in
location of plots and greater coverage of the conditions
within each PMC in the field, and often result in less area
being subjected to potential yield-reducing treatments.
 Summary
The important points to consider when planning and rate should be included if possible to better describe the
designing any rate-response experiment where the response.
production response is going to be measured by a n The design should consider the capability and width of
harvester-mounted yield-monitoring system are: application equipment, any management class pattern
n Test changes in one single input or practice at a time. It is and the method of measuring the production response,
preferable to increase the number of different rates in the and should aim to minimise the financial impact of the
experiment, rather than include additional variables. experiment. The experiments can be run using whole-field
n Use more than one alternative treatment rate, otherwise treatment strips, but with variable-rate technology setting
the best that can be hoped for is a decision that one up small strip experiments is easy. Small strips provide
treatment is preferable to another. At least two alternative the opportunity for greater independent replication of
treatments (plus the standard rate) can help to decide the treatments, more flexibility in location and greater coverage
‘best’ treatment rate. of the conditions in the field, and often result in less area
n There should be a good size difference between the being subject to potentially yield-reducing treatments.
treatment rates. Multiples of the traditional application rate n Any strips should be at least 100m long and at least three
(for example, 0.25x, 0.5x, 1.5x, 2x) are suitable. A low harvester widths wide.

ANALYSIS OF RATE-RESPONSE EXPERIMENTS

The interpretation of the output from these experimental designs focuses on two main questions.
1. Is there a difference in the production response to the different treatment rates?
2. What is the input application rate that optimises production or profit?

Question 1: Is there a difference in experiments described here requires some caution. The
production response? tests use the average response to each treatment for
Two commonly applied ‘standard’ statistical analyses can be comparisons, and include an assessment of the accuracy
used to answer question 1. They are: of the averages when making a decision on significance.
n t-test – used to compare two different treatments; and The accuracy is assessed by the amount of variation in all
n analysis of variance (ANOVA) – which can be applied to the measurements and the number of measurements used
test for differences between more than two treatments. to calculate the averages. In these tests, as the number of
When used to compare two treatments, it is the same as measurements increases, the accuracy of the estimates for
10 the t-test. the averages and the overall variation is deemed to increase,
Both these tests assess the overall variation in the and therefore it is easier to find significant differences.
ADVANCED FIELD-SCALE EXPERIMENTATION FOR GRAIN GROWERS

response data and calculate how much can be attributed However, each measurement is expected to be from an
to effects from the treatments and how much is occurring independent sample of the property being measured (for
between the individual measurements for each treatment. example, crop yield). When measuring the yield response
The result is then expressed according to whether or not using a harvester-mounted yield monitor, the internal
there is a ‘significant difference’ between the average mixing of grain during mechanical harvesting means that
response for each treatment. neighbouring samples taken by the monitor are correlated
The finding of a ‘significant difference’ is an assessment and so each measurement does not constitute a new
that the average results from the treatments are statistically piece of information in a statistical sense. Given that yield
different enough that the probability of such a difference monitors can collect quite a number of measurements
happening by chance is below a set threshold. The in most treatment plots, if this is not recognised and
thresholds are known as ‘levels of significance’. Common each yield measurement is considered independent in a
levels of significance are 5% (p=0.05), 1% (p=0.01) and statistical analysis, then the chance of finding statistically
0.1% (p=0.001). For example, if someone argues that ‘there significant treatment effects is artificially inflated simply by
is only one chance in a thousand this could have happened using a yield monitor to harvest the experiment.
by chance’, a 0.001 (high) level of statistical significance is One way to improve the reliability of the results
being implied. Choosing a level of significance is an arbitrary obtained from using these tests on field-length strip
task, but for many agricultural applications, a level of 5% experiments is to break the response data into sections
is common. However, grain growers may be content to and then use the average response values in each
operate at lower levels of significance (10% to 20%). The section for analysis. When the response measurements
higher the significance level, the stronger is the required are yield-monitor data, the sections can be created
evidence of a difference. by imposing 20-metre buffers along the strip at
Using these tests on the types of field-scale response 50-metre intervals (Figure 7). While this does not fully
FIGURE 7 Whole-field treatment strips will be dependent on the risk profile of the producer. But 11
segmented. operating the analysis in this way produces a usable
methodology for applying the ANOVA/t-test for commercial
field-scale experimentation.
Class 2

ADVANCED FIELD-SCALE EXPERIMENTATION FOR GRAIN GROWERS

So, the results of using yield-monitor data in
these standard statistical tests can provide
information for decisions on rate changes but
they should be considered with caution.

Question 2: What is the optimum

application rate?
Application rate (kg N/ha)
0 In an experiment where there are at least two alternative
Class 1 29 treatment rates plus the traditional treatment, the average
60 (rest of field) production response for each treatment rate can be used to
82 N
build production response functions. These can be used to
provide an agronomic assessment of the input application
0 100 200 300 400 rate that optimises production or profit. In this case it is
metres not the maximum yield that is sought, but the yield that
optimises the cost of production and financial returns.
Ordinary least squares (OLS) regression, which is
accessible to the general scientific and agronomic
communities, is suggested as the method for fitting a
function to the production response data. Once a production
address the statistical issues, by using averages of response function has been defined then the point where
separated sections as the measurements, the number production is optimised can be located using marginal rate
of ‘observations’ is reduced, which helps make the analysis.
significance assessment more realistic. For this analysis the cost of the input and the price
To deal with this issue in the small-strip design, the received for the crop yield are required to calculate the
average of the yield-monitor data from within each marginal revenue (MR) and marginal cost (MC) along the
strip plot (replicate) should be used as an estimate production function. MR and MC are defined as either the
of the production response to the treatment rate. For change in revenue or cost as each additional unit of input is
this reason the greater the number of plots that can applied. For the majority of agricultural inputs, MC of each
be established for each treatment level, the better the additional unit of input is constant (for example, the price of
chance of discovering a statistical difference. one kilogram of nitrogen) and MR is the amount of money
It should be noted that the discovery of a ‘statistical’ received for the output (yield) gained from using that unit of
significant difference does not necessarily equate to input.
an agronomically or financially significant difference. When the return (MR) for applying an additional unit of
Agronomic significance is defined here as a production input is greater than the cost (MC), then applying the unit of
difference, relative to input rate change, that is input is profitable. If MR is less than MC, the application of
large enough to compel crop producers to modify the additional unit of input is not profitable. Since total profit
management actions in order to gain the observed increases when marginal profit is positive and total profit
benefits. And the reverse is also true. The fact that a decreases when marginal profit becomes negative, it must
statistically significant difference is not found does not reach a maximum where marginal profit is zero – or where
necessarily mean that an agronomically or financially MC=MR. This is because the producer has collected positive
significant difference is not present. profit up until the point where MC=MR.
So when using the methods described here to analyse Finally, an ANOVA/t-test analysis or the response function
both the field-length-strip and small-strip designs for analysis would be aimed at exploring the differences in the
differences in the production response to different production responses between any defined PMCs. However,
treatment rates, a lowering of the level of significance the analysis can be undertaken at the whole-field scale by
to 10% to 20% or taking a pragmatic ‘agronomic’ aggregating the management class responses. Comparing
significant difference approach to the results may the whole-field response to the responses in each PMC
therefore be required. will allow the assessment of the effectiveness of the PMC
The latter will introduce some subjectivity into the process in partitioning production variability.
analysis and should include the opinion of the producer These processes are best understood through a case
and/or local agronomist. The choice of value for the former study.
 A NITROGEN FERTILISER CASE STUDY
Figure 8 shows the actual fertiliser application map and When the average yield values for the treatment strips
the subsequent yield map for the nitrogen response trial are used in the same tests, the results are less decisive
designed in Figure 6b. The traditional application rate for the (Figure 10). In the whole-field and PMC analyses it is only the
field was 60 kilograms of nitrogen per hectare. The average response to the 0kg/ha treatment that can be considered
yield from each treatment plot was calculated, along with the statistically different from the others. The change in the
average for each treatment rate in each class. The values statistical difference assessment is due to the decrease in
are shown in Table 2. the number of measurements used and the resultant effect
The results show that the two classes are obviously on the calculated accuracy of the averages in the test
responding differently to the applied nitrogen fertiliser and procedures.
the traditional rate may not be best for either class. The These results highlight the potential issues with using
season was above average in rainfall, with growing-season these ‘standard’ statistical tests on this type of practical
rain (June to November) of 350 millimetres, or in the 70th experimental data. In Figure 9 (page 13), where each yield-
centile. data point is treated as an individual measurement, statistical
Figure 9 shows the results of an ANOVA using the differences between yield values that would probably not be
individual yield values obtained from the harvester in each considered agronomically different (4.486t/ha and 4.684t/
treatment strip. In Figure 9a all the treatments are considered ha in Figure 9a) are found. When the data is reduced to
statistically different when the trial is analysed as a whole. averages for each treatment and the perceived number of
When the field is analysed in PMCs, all the treatments are replications is reduced, the reverse is the case. Differences
considered statistically different in Class 1. Class 2 shows in yield response that would be considered agronomically
that 0kg/ha and 60kg/ha are statistically different from each significant are not identified as statistically different.
other and from the 29kg/ha and 82kg/ha treatments. The These issues aside, it is possible to identify which of
response to the 29kg/ha and 82kg/ha treatments are not the four rates produces the most yield using the ANOVA
statistically different from each other. tests, but not the rate that optimises production. To identify
the optimum application rate from the trial, yield response
TABLE 2 Average yield values for the treatment strips functions for the two classes were built using the average
in Figure 8. yield data from the strips (Figure 11a, page 15). Applying
Rate Class 1 Class 2 marginal rate analysis to the two functions using prices for
(kg N/ha) Rep 1 Rep 2 Average Rep 1 Rep 2 Average the season (Figure 11b) locates the points where MR=MC
0 3.32 3.76 3.54 3.41 3.25 3.33 and shows that the optimum rates of applied nitrogen
29 4.68 4.15 4.41 4.09 4.07 4.08
fertiliser are 109kg N/ha and 39kg N/ha for Class 1 and 2
respectively.
60 – – 4.84 – – 3.97
82 4.97 5.50 5.24 4.11 4.16 4.13
12
ADVANCED FIELD-SCALE EXPERIMENTATION FOR GRAIN GROWERS

FIGURE 8 (a) the as-applied map for the nitrogen response experiment shown in Figure 6b –
traditional field treatment was 60kg N/ha; (b) the wheat yield map.
(a) (b)
Nitrogen as-applied Wheat yield 29
(kg N/ha) (t/ha)
0 0.27-0.87 82
29 0.88-1.47 0
60 1.48-2.08
82 2.09-2.68 29 Class 2
2.69-3.29 82
3.30-3.89
3.90-4.49 0
4.50-5.10
5.11-5.70 82
5.71-6.30 82
N
0 29

0 29

Class 1

0 125 250 375 500 0 125 250 375 500

metres metres
FIGURE 9 ANOVA of wheat yield (t/ha) by applied nitrogen rate (kg N/ha): (a) individual yield data 13
from the trial plots used as input in the analysis of the trial response over the whole field;
(b) management Class 1; and (c) management Class 2.
(a)
Wheat yield (t/ha)

ADVANCED FIELD-SCALE EXPERIMENTATION FOR GRAIN GROWERS

6.0
a) Analysis of variance (ANOVA) – whole field
5.5 Source Sum of Mean
DF squares square F Ratio Prob > F
5.0 Applied N rate 3 119.13 39.71 181.72 <.0001*
(kg/ha)
4.5 Error 524 114.51 0.22
Total 527 233.64
4.0
Comparison of means (student’s t-test)
3.5
Rate Mean (t/ha)
3.0 82 A 4.684
60 B 4.486
2.5 29 C 4.246
0 29 60 82 Each pair
student’s t-test 0 D 3.435
Applied N rates (kg/ha)
0.05 Rates not connected by same letter are significantly different.

(b)
Wheat yield (t/ha)
6.0
b) Analysis of variance (ANOVA) – Class 1
5.5 Source DF Sum of Mean F Ratio Prob > F
squares square
Applied N rate 3 96.40 32.13 229.58 <.0001*
5.0 (kg/ha)
Error 261 36.53 0.14
4.5 Total 264 132.93

4.0 Comparison of means (student’s t-test)

Rate Mean (t/ha)
3.5 82 A 5.236
60 B 4.835
3.0 29 C 4.415
0 29 60 82 Each pair
student’s t-test 0 D 3.541
Applied N rates (kg/ha)
0.05 Rates not connected by same letter are significantly different.

(c)
Wheat yield (t/ha)

5.0 c) Analysis of variance (ANOVA) – Class 2

Source DF Sum of Mean F Ratio Prob > F
squares square
4.5 Applied N rate 3 41.56 13.85 113.24 <.0001*
(kg/ha)
Error 259 31.69 0.12
4.0 Total 262 73.25

Comparison of means (student’s t-test)

3.5
Rate Mean (t/ha)
82 A 4.132
3.0 60 A B 4.079
0 29 60 82 Each pair 29 B 3.975
student’s t-test 0 C 3.336
Applied N rates (kg/ha)
0.05 Rates not connected by same letter are significantly different.
 FIGURE 10 ANOVA of wheat yield (t/ha) by applied nitrogen rate (kg/ha): (a) average yield data
from the trial plots used as input in the analysis of the trial response over the whole field;
(b) management class 1; and (c)).
(a)
Wheat yield (t/ha)

5.0 a) Analysis of variance (ANOVA) – whole field

Source Sum of Mean
DF squares square F Ratio Prob > F
4.5 Applied N rate 3 3.46 1.15 5.52 0.0129*
(kg/ha)
Error 12 2.51 0.21
4.0 Total 15 5.97

Comparison of means (student’s t-test)

3.5 Rate Mean (t/ha)
82 A 4.685
60 A 4.405
3.0 29 A 4.247
0 29 60 82 Each pair
Applied N rates (kg/ha) student’s t 0 B 3.435
0.05 Rates not connected by same letter are significantly different.

(b)
Wheat yield (t/ha)

5.5 b) Analysis of variance (ANOVA) – Class1

Source DF Sum of Mean F Ratio Prob > F
squares square
5.0 Applied N rate 3 3.16 1.05 10.92 0.0214*
(kg/ha)
Error 4 0.39 0.096
4.5 Total 7 3.55

Comparison of means (student’s t-test)

4.0
Rate Mean
82 A 5.235
3.5 60 A 4.835
0 29 60 82 Each pair 29 A 4.415
14 Applied N rates (kg/ha) student’s t 0 B 3.540
0.05 Rates not connected by same letter are significantly different.
ADVANCED FIELD-SCALE EXPERIMENTATION FOR GRAIN GROWERS

(c)
Wheat yield (t/ha)
4.2
4.1
c) Analysis of Variance (ANOVA) – Class2
Sum of Mean F Ratio Prob > F
4.0 Source DF squares square
3.9 Applied N rate 3 0.83 0.28 71.665 0.0006*
(kg/ha)
3.8 Error 4 0.016 0.0039
3.7 Total 7 0.85
3.6
3.5
Comparison of means (student’s t-test)
Rate Mean
3.4
82 A 4.135
3.3
60 A 3.975
3.2 4.080
0 29 60 82 Each pair 29 A
Applied N rates (kg/ha) student’s t 0 B 3.330
0.05 Rates not connected by same letter are significantly different.

FIGURE 11 Yield response (a) to applied nitrogen fertiliser from the trial in Figure 6. Traditional
field application was 60kg N/ha. Marginal rate analysis (b) shows where MC=MR and that
Class 1 optimum = 109kg N/ha; Class 2 optimum = 39kg N/ha.
(a) (b)
Yeld (t/ha) Price ($)
5.0
Source Sum of Mean
DF squares square F Ratio Prob > F
4.5 Applied N rate 3 3.46 1.15 5.52 0.0129*
(kg/ha)
Error 12 2.51 0.21
4.0 Total 15 5.97

Comparison of means (student’s t-test)

3.5 Rate Mean (t/ha)

USING THE RESULTS

3.0
82
60
A
A
4.685
4.405
15

0 29 60 82 Each pair 29 A 4.247

Applied N rates (kg/ha) student’s t 0 B 3.435
Comparing gross margins 0.05
Continued experimentation
Rates not connected by same letter are significantly different.

ADVANCED FIELD-SCALE EXPERIMENTATION FOR GRAIN GROWERS

With this response information it is possible to compare the By maintaining the treatment plots in position for a number
gross margin (GM) under traditional uniform management of seasons and measuring the production response to total
to the GM achievable under variable-rate management input level rather than applied input rate, it is possible to better
(b)
scenarios. The GM for each management scenario can be quantify the optimal input level required in each PMC at the
Wheat yield (t/ha)
calculated by multiplying the yield achieved by the grain time when decisions are made. This is especially useful for
price
5.5 b)and deducting the cost for the quantity of fertiliser Analysisexperiments
nutrient-response of varianceand (ANOVA) is achieved– Class1by pre-season
applied: sampling of the treatment Sum of Mean
Source DF squares square F Ratio and
plots for resident nutrient Prob > F
adding the applied treatment quantity to calculate the actual
Gross margin (GM) ($/ha) =
5.0 Applied N rate
nutrient presented (kg/ha) to the 3crop in3.16 each strip.
1.05 10.92 0.0214*
(yield (t/ha) x grain price ($/t)) – It is important Error to wait at 4 least 0.39two years0.096 to let the zero/low
4.5
(fertiliser applied (t/ha) x fertiliser cost ($/ha)) treatment plots Total
reduce in7resident 3.55 nutrient before beginning
the in-plotComparison
nutrient monitoring. of means With(student’s
this approach t-test)the higher
4.0
From the case study, two simple management response treatment plots Ratecan also be monitored for any build-up
Meanin
scenarios can be considered. resident nutrient 82 levels due to
A the experimental treatments.
5.235
3.5 Scenario one would maintain the total amount of fertiliser Operating the 60 rate-response A experiments in this manner 4.835
applied to0the field, but 29 moves the
60over-application
82 on means
Each pair that yield29 response for
A crops in the management 4.415
Class 2 to Class 1. Applied This would keep
N rates (kg/ha) fertiliser costs constant rotation
student’s t can be 0 examined and information B on seasonal3.540
and result in yield gains in Class 1 that would improve GM 0.05conditions can be matched with the responses. An archive
Rates not connected by same letter are significantly different.

by $11.50/ha across the field when compared with the of this information over a number of seasons and crops
traditional application of 60kg N/ha across the whole field. would provide greater information to tailor future application-
(c) Scenario two would aim to apply the optimum amount to rate decisions to expected seasonal conditions.
each Class, requiring
Wheat yield (t/ha) an additional 1.4t of nitrogen in total Input response data from individual fields may then also
to
4.2 achieve the yield goals of 5.4t/ha in Class 1 and 4.0t/ha be used as a replacement for generic response models in
in
4.1
c) 2. However, the increased yield would mean that
Class crop growthAnalysis of Variance
simulation programs. (ANOVA) – Class2
Sum of Mean
the
4.0 GM for the field would be improved by $25/ha over the Source DF squares square F Ratio Prob > F
traditional
3.9 application of 60kg N/ha across the whole field. Applied N rate 3 0.83 0.28 71.665 0.0006*
(kg/ha)
3.8 Running rate-response experiments and undertaking Error 4 0.016 0.0039
a3.7GM analysis will show if there are worthwhile financial
Total 7 0.85
gains
3.6 to be made by exploring the optimisation of fertiliser
application
3.5 rates within a field. Any projected gains will Comparison of means (student’s t-test)
be site-specific, but the management-class response Rate Mean
3.4
information may be used to help inform decisions on 82 A 4.135
3.3
management of class-specific yield targets and fertiliser 60 A 3.975
3.2 4.080
strategies 0in subsequent 29 years. 60 82 Each pair 29 A
Applied N rates (kg/ha) student’s t 0 B 3.330
0.05 Rates not connected by same letter are significantly different.

FIGURE 11 Yield response (a) to applied nitrogen fertiliser from the trial in Figure 6. Traditional
field application was 60kg N/ha. Marginal rate analysis (b) shows where MC=MR and that
Class 1 optimum = 109kg N/ha; Class 2 optimum = 39kg N/ha.
(a) (b)
Yeld (t/ha) Price ($)
5.5 4.5
4.0
5.0 3.5
Class 1
3.0
4.5 2.5 MR class 1
4.0 2.0
Class 2 1.5 MR class 2
3.5 1.0 MC
0.5
3.0 0 90 100 110
0 10 20 30 40 50 60 70 80 90 100 0 10 20 30 40 50 60 70 80
Applied nitrogen (kg N/ha) Applied nitrogen (kg N/ha)
GRDC, PO Box 5367, Kingston ACT 2604 T 02 6166 4500 F 02 6166 4599