Monitoring Wheat Protein Content On-Harvester: Australian Experiences

Abstract

This paper was first published in Precision Agriculture '05, Proceedings of the 5 th European Conference on Precision Agriculture, ed. J.V. Stafford, Wageningen Academic Publishers. It is reproduced here with the permission of the editor. Please acknowledge the original source when referencing.] Abstract An on-harvester protein sensor has been tested for two seasons on a commercial combine harvester in Australia. Operators report that sensor and software were relatively easy to use especially since the model used is still a prototype set-up. Some problems with operation were noted and have been addressed for future commercial development. Output from the Zeltex NIT protein sensor was coherent and often strongly correlated to yield response, giving a good indication that the observed protein patterns are real. Absolute protein values however appeared suppressed and a new calibration curve for Australia has been developed for the Zeltex AccuHarvest ® sensor.

Full text

Precision Agriculture ’05 5th European Conference on Precision Agriculture
Monitoring Wheat Protein Content On-Harvester:
Australian Experiences
James Taylor1, Brett Whelan1, Lars Thylén2, Mikael Gilbertsson2 and James Hassall3
1Australian Centre for Precision Agriculture
McMillan Building A05, University of Sydney NSW 2006
j.taylor@agec.usyd.edu.au
2 Swedish Institute of Agricultural and Environmental Engineering (JTI)
Box 7033, SE-750 07 Uppsala
3 “Kiewa”, Gilgandra, NSW, Australia
[This paper was first published in Precision Agriculture ’05, Proceedings of the 5th European
Conference on Precision Agriculture, ed. J.V. Stafford, Wageningen Academic Publishers. It is
reproduced here with the permission of the editor. Please acknowledge the original source when
referencing.]
Abstract
An on-harvester protein sensor has been tested for two seasons on a commercial
combine harvester in Australia. Operators report that sensor and software were
relatively easy to use especially since the model used is still a prototype set-up. Some
problems with operation were noted and have been addressed for future commercial
development. Output from the Zeltex NIT protein sensor was coherent and often
strongly correlated to yield response, giving a good indication that the observed protein
patterns are real. Absolute protein values however appeared suppressed and a new
calibration curve for Australia has been developed for the Zeltex AccuHarvest® sensor.
Keywords: Real-time protein monitoring, Australia, nitrogen budget
Introduction
In Australia, protein content is an important consideration in grain sale price,
particularly wheat varieties. A bonus/discount payment is made on a 0.1% sliding scale
beyond the base rate for each grade. Knowledge of the variability in protein content
prior to marketing could be used in many tactical ways to optimise farm gate returns.
However, while the opportunity to raise profits does exist, the actual return to the
grower is dependent on market forces (Long et al. 2002) and the cost/effort required for
differential harvest, which may deter contract harvesters from the practice. Information
on nitrogen use and removal, in conjunction with yield data, would also be useful in the
strategic management of nitrogen fertiliser and the development of more accurate site-
specific gross margin analysis. This may be of greater benefit to growers in the short
term due to tangible savings in fertiliser cost, which may be upwards of 30% of
production costs.
The potential benefits of protein maps, particularly for nitrogen management, have
prompted a great deal of interest in the development of both on-the-go and remote-
sensing based protein measurement in Australia, as well as North America (Long et al.
1998. For the 2001, 2003 and 2004 winter cropping seasons, the Australian Centre for

Precision Agriculture ’05 5th European Conference on Precision Agriculture
Precision Agriculture (ACPA), in conjunction with growers in Conservation Farmers
Incorporated (CFI), has been collaborating with Zeltex and the Swedish Institute of
Agricultural and Environmental Engineering (JTI) to test a prototype Near-Infrared
Transmission (NIT) on-harvester grain protein and moisture sensor. This paper reports
on some of our experiences with the sensor and the development of a calibration for
Australian conditions.
Sensor Mounting and Operation
In Australia, grain tends to be harvested at a lower moisture content (generally <13%) to
avoid post-harvest drying. Climatic conditions at harvest are generally dry and dusty
with temperatures often above 40ºC. In 2001, dust and light contamination created
problems with sensor operation. As a result very little useful data was recorded. The
sampling system was subsequently redesigned to minimise dust and light
contamination.
The sampling system has been designed to mount on the side of the clean grain elevator
housing. Grain is sampled from the up elevator shaft and deposited into the down
elevator shaft. The inlet and outlet are controlled by two trapdoors driven by
windscreen wiper motors. The trapdoors close tightly to seal the chamber and avoid
contamination in the sensor. Figure 1 illustrates the mounting and sampling
mechanism. The trapdoors are activated by LED fill sensors at the top and bottom of
the NIT sensor measurement chamber. When the top fill sensor is triggered the top door
closes and the NIT sensing protocol is initiated. Once the NIT sensing is complete the
bottom trapdoor opens to purge the NIT sensor of grain. The bottom trapdoor closes
once the lower fill sensor is activated and the top trapdoor opens completing the cycle.
If the cycle gets stuck at a particular stage for >30s, the sensor has an override function
to open the bottom trapdoor and restart the cycle. In initial testing this option was not
available and the sensor had to be manually restarted if it stopped functioning. The
sensor takes a reading at approximately 12s intervals, which equates to 65-70
points/hectare at normal harvesting speeds. The actual NIT sensor is a 14-band Near
Infrared Transmission whole grain analyser (Zeltex AccuHarvest®) operating 14
wavelengths between 893 and 1045nm.
The data from the sensor is currently being logged onto a laptop computer installed in
the cabin of the harvester. The software is written in LabView. The software is easy to
use and provides a graphical and numerical indication of how grain protein and
moisture is varying over a 3-4 minute window. The growers had no problem interacting
with the software and the only drawback was the current use of a laptop to log the data.
The cabin environment in combine harvesters is not particular suited to laptops and a
more robust, simpler data logger will be required for any commercial release.
At the time of writing the second full harvest of protein data has been collected with the
Zeltex sensor. The same system has now been used for 2 Australian and 2 European
harvests. For the 2004 harvest in Australia, the sensor was run at two locations over a
period of 6 weeks. The sampling system that was used is an early prototype that has not
been updated. As a result some hardware fatigue issues were identified from overuse
that will be addressed in future models. In general, however, the system worked well

Precision Agriculture ’05 5th European Conference on Precision Agriculture
and for the second year running a large amount of data has been collected. Both
growers that operated the sensor had no problems apart from the hardware fatigue.
Without having analysed, visualised and discussed the data from 2004, the information
displayed in real-time on the software interface generally concurred with the growers’
knowledge of the field.
Figure 1: The sensor and sampling system mounted on the clean grain elevator (left)
and a close-up of the sensor and sampling system (right)
Sensor Calibration
The Zeltex NIT sensor is currently calibrated using data from the Northern Hemisphere
(North America and Europe). One of the aims of this study was to determine if the
calibration is applicable to Australian conditions. The null hypothesis was that the
calibration is the same for both Australian and Northern Hemisphere wheat production.
This hypothesis was initially tested in the field and later in the laboratory.
Field Testing
For the field situation, two transects in a wheat field near Gilgandra, NSW, were used.
As the combine harvester harvested the two transects, the Zeltex sensor was used to
measure grain protein on-harvester. At the same time 15 samples were manually
collected from the bubble-up auger near the top of the clean grain elevator. These
samples were taken at approximately the same recording interval (12s) as the on-
harvester protein sensor. The manually sampled grain was analysed using desktop NIT
spectrophotometers at the Gilgandra silo and the Australian Bread Research Institute
(BRI) using the FOSS Infratec 1229 (Global calibration No. WH000003). The mean
protein values of the two transects from the different measurements are given in Table
1.

Precision Agriculture ’05 5th European Conference on Precision Agriculture
The comparison of the desktop results with the on-harvester sensor indicated that the
on-harvester sensor was underestimating both grain protein and grain moisture content.
The mean protein difference between the desktop sensors (at the Gilgandra Silo and
BRI) and the Zeltex was 0.94%. The mean moisture difference was 0.62%. There may
be some error in this approach as it was not possible to manually collect the same grain
that was sampled by the Zeltex sensor. It is hypothesized that this underestimation is a
bias from the use of the Northern Hemisphere calibration curve. To test this a
laboratory experiment was conducted.
Table 1: Comparison of protein results from samples taken along two transects and
analysed with three different NIT sensors. (NB. The Gilgandra silo and BRI sensors
analysed the same grain samples. *The adjusted Zeltex response is discussed in the
following section)
Sensor Transect Protein % (σ) Moisture % (σ)
Silo 1 14.64 (0.58) 10.32 (0.13)
2 14.50 (0.96) 9.58 (0.12)
BRI 1 14.70 (1.13) 10.30 (0.15)
2 14.46 (0.62) 9.63 (0.16)
Zeltex 1 13.70 (1.11) 9.33 (0.26)
2 13.57 (1.19) 9.35 (0.13)
Zeltex (adj)* 1 14.54(1.07) 10.39 (0.35)
2 14.88 (0.72)
Laboratory Experiment
The results from the field trial indicated that a new calibration would be required for
Australian conditions. This was not unexpected given the different climate conditions,
wheat varieties and moisture content at harvest in Australia. The new Australian
calibration for the Zeltex sensor was derived using 99 Australian grain samples sourced
from different regions of the Australian grain belt, including North-West NSW, the
Riverine district on the NSW/Victoria border and the Yorke Penisula in South Australia.
The grain samples were analysed using the FOSS Infratec 1229 protein content under
standard conditions (25ºC) at the Australian Bread Research Institute (BRI) at North
Ryde, NSW. The FOSS Infratec 1229 analyses the grain at 2nm intervals between 850
and 1048nm. The spectra were extracted and a protein and moisture value determined
using the standard calibration (WH000003) for Australian wheat developed for the
FOSS Infratec 1229
The same samples were then run through the Zeltex sensor, mounted in a laboratory at
the University of Sydney, at two temperatures, 25°C and 40°C. The two temperatures
were selected as temperature has a known effect on NIT and it was desirable for the
calibration to encompass the probable temperature ranges at harvest. The raw output
from the sensor (14 wavelengths) was extracted. From this analysis 5 readings gave
strange values and were discarded. This left 193 readings (from the 99 samples at two
temperatures). A Multiple Linear Regression (MLR) was performed to predict the BRI
protein % using the wavelength response from the Zeltex sensor. Similarly the moisture

Precision Agriculture ’05 5th European Conference on Precision Agriculture
% was also predicted. Alternative regression models, Multiple Stepwise Linear
Regression (MSLR) and Partial Least Squares Regression (PLS) were also tried with no
additional benefit.
Results and Discussion
The protein and moisture calibration resulting from the MLR is shown in Figure 2. The
r2 and standard error of prediction (SEP) are given on the graphs. The SEP value for
protein is similar to those derived using the Northern Hemisphere data set (Thylén and
Algerbo, 2001), however the moisture SEP is higher. This may be due to the lower
moisture contents at harvest or the different grain varieties. The protein shows a strong
1:1 linear response over a large range of protein values (9-17%).
Figure 2: Plots of measured vs predicted protein (left) and moisture (right) using
derived calibration equations.
For the general calibration curve the transect data from the within field transects was
used. The transect data was excluded from the dataset and a new calibration curve
derived (n=134). The new calibration curve was then applied to the transect data to
predict protein and moisture for the transect data above. The mean results are shown in
Table 1 (Zeltex adj.). After transformation the absolute mean difference in percent
protein measurement between the standard laboratory measurements and the Zeltex
instrument decreased from 0.94% to 0.14%. For moisture the absolute mean difference
was slightly increased from 0.29% to 0.41%. The Australian calibration curve appears
to be giving a better protein prediction than the Northern Hemisphere calibration for the
data from these two transects.
For the 2004 harvest, field samples were again taken to help validate the accuracy of the
protein sensor. These data were not analyzed at the time of this writing and cannot be
presented here.
Nitrogen Budgeting.
One of the principal benefits of a protein monitor identified by Australian growers is the
ability to better identify nitrogen use within fields and variably replenish nitrogen. The

Precision Agriculture ’05 5th European Conference on Precision Agriculture
amount of nitrogen removed from a system is a function of the amount of grain (yield)
and the amount of nitrogen in the grain. For Australian conditions the relationship
between protein, yield and nitrogen removed in wheat has been quantified by Kelly et.
al. (2003) as;
1.75 (%)Protein Grain (Mg/ha) YieldGrain (Mg/ha) removal N
•
•
= (1)
The relationship in North America has also been quantified (see Long et al. 1998) but is
not used here due to different varieties and growing conditions. Since both yield and
protein data have been collected on-the-go the data can be interpolated (block kriged
using local variograms) onto a common grid and the nitrogen removed from the
cropping system calculated using Equation 1.
Figure 3: Interpolated maps of Grain Yield (Mg/ha) (left) and Grain Protein (%) right.
The interpolated grain yield and protein maps are shown in Figure 3. Both maps show
similar spatial patterns with a negative response between grain and protein (highlighted
by the boxed areas). This negative relationship is not unusual in Australia and generally
reflects either insufficient nitrogen or soil moisture to achieve yield potential. The
nitrogen removed map is shown in Figure 4.
As well as calculating the amount of nitrogen removed a site-specific nitrogen budget
can be determined using Equation 2.
removed N -input N present N budget N
+
= (2)

Precision Agriculture ’05 5th European Conference on Precision Agriculture
This is similar to the approach of Long et al. (1998) except that the inherent soil
nitrogen pre-sowing is considered. In Australia this is important as the variable rainfall
means that crop failure is quite possible and in drought situations there may be a large
amount of residual nitrogen stored in the soil. For Field 3, the pre-sowing soil nitrogen
levels were only available as a field average of 60kgN/ha. A pre-sowing application of
80kg/ha Monoammonium phosphate (MAP) (12%N) was applied. The majority of the
field, except for the panhandle in the southeast corner, was also top-dressed with
50kg/ha urea (46% N). This means that the majority of the field received 32.6kg N/ha
(assuming an even distribution) and the panhandle received 9.6kg N/ha. The panhandle
was not top dressed due to the spreader running out of fertiliser. The crop response to
this missed application is clearly evident with grain protein suppressed in the panhandle
although grain yield was not (Figure 3).
Figure 4: Map of Nitrogen removed (left) and the Nitrogen Budget (right) for Field 3
derived from the interpolated yield and protein maps (Figure 5)
Figure 4 shows a continuous nitrogen budget map for Field 3. In reality, field
management, including fertiliser, is being managed at a zone level. To facilitate
adoption the ACPA intends to develop a protocol to incorporate this information into
management class response functions as proposed by Whelan et al. (2005). The
nitrogen budget is currently limited by the use of a mean response for initial soil N.
Without real-time soil N-sensors, it is an expensive soil sampling exercise to obtain an
accurate map of soil available N prior to sowing. Management zones represent an
approach that allows the grower to determine the mean zone ‘N present’ in a cost and
time effective manner at a similar resolution to which management is being applied
(Whelan et al. 2002).

Precision Agriculture ’05 5th European Conference on Precision Agriculture
While Equation 2 allows for the calculation of a nitrogen budget, the actual level of
nitrogen applied will be dependent on the availability of soil moisture. Determining
rainfall and soil moisture availability is difficult, however, in Southern Australia where
top dressing (within season fertilising) is common, earlier season rainfall can be a strong
indicator of total seasonal rainfall (Peter Stone, CSIRO Land and Water, pers. comm.)
Conclusions
The output from the protein sensor shows strong spatial patterns that are consistent with
what the grower expects, observed yield variations and management decisions.
Growers are enthused about the quality of data being generated by the sensor while
combine operators are happy with the ease of performance of the sensor and data
collection software on-harvester. There seems little reason from an engineering
perspective not to commercially release a limited number of units for the next harvest.
Preliminary investigations into deriving an Australian based protein and moisture
calibration curve appear to give better results, in Australian conditions, than the current
global calibration curves. However further validation is needed over a wider
distribution of grain samples. Accurate site-specific determination of protein content
will provide growers with confidence in calculating site-specific or zonal nitrogen
budgets as well as trying to determine the reasons for spatial patterns in their crop
production.
Acknowledgements
The authors would like to acknowledge the assistance and in-kind support given to this
project by Todd Rosenthal and the staff at Zeltex Inc. Maryland, USA.
References
Kelly, R., Strong, W., Jensen, T., Butler, D., Town, B. and Adams, M. 2003 Recurrence
of yield and protein variation in the northern grains region. Proceedings of the
11th Australian Agronomy Conference, Australian Society of Agronomy,
Geelong, February, 2003 (http://www.regional.org.au/au/asa/2003/c/15/kelly.htm)
Long, D.S., Carlson, G.R. & Engel, R.E. 1998 Grain protein mapping for precision
manangement of dryland wheat. In P.C. Robert, R.H. Rust & W.E. Larson (eds)
Precision Agriculture, Proceedings of the 4th International Conference on
Precision Agriculture ASA/CSSA/SSSA, Madison, WI, USA pp787-796
Long, D.S., Carlson, G.R. & Engel, R.E. 2002 Gross value of spring wheat under
precision nitrogen management as influenced by grain protein. In P.C. Robert,
R.H. Rust & W.E. Larson (eds) Precision Agriculture, Proceedings of the 6th
International Conference on Precision Agriculture, ASA/CSSA/SSSA, Madison,
WI, USA, CD-ROM
Thylén, L. and Algerbo, P.A. 2001 Development of a protein sensor for combine
harvesters. In G. Grenier & S. Blackmore (eds) ECPA 2001, Proceedings of the
3rd European Conference on Precision Agriculture, Montpellier, June, 2001,
pp869-873

Precision Agriculture ’05 5th European Conference on Precision Agriculture
Whelan, B.M., Cupitt, J. & McBratney, A.B. 2002. Practical definition and
interpretation of potential management zones in Australian dryland cropping. In
P.C. Robert, R.H. Rust & W.E. Larson (eds) Precision Agriculture, Proceedings
of the 6th International Conference on Precision Agriculture, ASA/CSSA/SSSA,
Madison, Wisconsin, pp325-329
Whelan, B.M. and Taylor, J.A. 2005. Local response to inputs: advancing SSCM in
Australia. In Proceedings of the 5th European Conference on Precision
Agriculture. Ed J.V. Stafford, Wageningen Academic Publishers, The
Netetherlands