ArticlePDF Available

Methodology for identification, characterization and removal of errors on yield maps

Authors:

Abstract and Figures

Precision farming is an approach that provides the capacity to manage the yield based on spacialized information. Yield map is an important information as it describes the answer of the crop to the inputs and soil conditions. Due to the automation of data collection some errors may occur and the elimination of those errors from the data represents information quality. This work proposes the development of a filtering routine of raw data to eliminate errors. The first step is the analysis and characterization of the errors present on data from six different commercial yield monitors. A methodology was developed to identify the running direction of the machine, allowing the estimation of the filling time error. Yield outlier limits were established and values over and under the limits were found in the data. Based on the characteristics of each error, a filtering routine was developed. The routine has seven steps, each one acting over different errors. The filtering process improved the semivariance analysis and the final quality of yield maps.
Content may be subject to copyright.
The authors are solely responsible for the content of this technical presentation. The technical presentation does not necessarily
reflect the official position of the American Society of Agricultural Engineers (ASAE), and its printing and distribution does not
constitute an endorsement of views which may be expressed. Technical presentations are not subject to the formal peer review
process by ASAE editorial committees; therefore, they are not to be presented as refereed publications. Citation of this work should
state that it is from an ASAE meeting paper. EXAMPLE: Author's Last Name, Initials. 2002. Title of Presentation. ASAE Meeting
Paper No. 02xxxx. St. Joseph, Mich.: ASAE. For information about securing permission to reprint or reproduce a technical
presentation, please contact ASAE at hq@asae.org or 616-429-0300 (2950 Niles Road, St. Joseph, MI 49085-9659 USA).
Paper Number: 021168
An ASAE Meeting Presentation
Methodology for identification, characterization and
removal of errors on yield maps
J. P. Molin, PhD.
PhD, Professor, Agricultural Machinery, Dept. of Rural Engineering, ESALQ, Un. Of São
Paulo. Av Pádua dias, 11, 13418-900 Piracicaba – Brazil. E-mail: jpmolin@esalq.usp.br
L. A. A. Menegatti, Agronomist MSc
Agronomist MSc, Dept. of Rural Engineering, ESALQ/USP, Piracicaba, SP Brazil, supported
by FAPESP scholarship laameneg@hotmail.com
Written for presentation at the
2002 ASAE Annual International Meeting / CIGR XVth World Congress
Sponsored by ASAE and CIGR
Hyatt Regency Chicago
Chicago, Illinois, USA
July 28-July 31, 2002
Research supported by FAPES, São Paulo, Brazil
Abstract. Precision farming is an approach that provides the capacity to manage the yield based on
spacialized information. Yield map is an important information as it describes the answer of the crop
to the inputs and soil conditions. Due to the automation of data collection some errors may occur and
the elimination of those errors from the data represents information quality. This work proposes the
development of a filtering routine of raw data to eliminate errors. The first step is the analysis and
characterization of the errors present on data from six different commercial yield monitors. A
methodology was developed to identify the running direction of the machine, allowing the estimation
of the filling time error. Yield outlier limits were established and values over and under the limits were
found in the data. Based on the characteristics of each error, a filtering routine was developed. The
routine has seven steps, each one acting over different errors. The filtering process improved the
semivariance analysis and the final quality of yield maps.
Keywords. Precision farming, Filter, Yield monitors
2
Introduction
The first yield maps were produced in the beginning of 90’s (Haneklaus et al., 2000). They were
generated using local positioning systems, with radios positioned in the borders of the field. In
that time, the GPS system became available for users, making possible its use in yield mapping
systems. Ever since, the yield mapping is considered for many researchers as essential part of
precision farming (Makepeace, 1996). Molin (2000) reinforces that the starting point considered
by users and researchers to evidence the spatial variability of yield is the yield map, which
materializes the answer of the crop to the inputs. The equipment responsible for gathering the
data for the generation of the yield maps is composed of several sensors. There are two types
of commercial yield monitors: those that measure mass directly, through load cells or
potentiometers, and, the measurement of the volume of grains that, through its density, is
converted into mass. Molin (2000) suggested the classification of the commercial yield monitors
in two groups, those formed by the equipments from the combine manufacturing companies,
usually, equipments for specific models of machines and not available for installation in other
machines. The other group includes the manufacturers of equipments available for installation in
any brand or combine model. The mass flow sensor is installed in the clean grain elevator or in
the sequence, before the grain enters into the grain tank of the combine.
A considerable number of systematic errors exist in a yield map. The programs supplied by the
manufacturers eliminate some of them. Moore, (1998), working with data from six years,
identified errors of suavisation, volume calibration, incorrect width and filling time. Blackmore &
Marshall (1996) still mention the positioning errors, due to the rethreshing, grain loss and the
accuracy of the flow sensor. According to Larscheid (1997), no automatic correction for grain
density was proposed to compensate the error. The accuracy of the flow sensor is another
source of errors on yield maps. The inclination of the combine influences the yield measure, as
reported by Reits et al. (1996). Several authors found differences between the actual yields and
that showed by the monitor (Molin el al., 2000; Auernhammer et al., 1993; Marcy, 1994; Birrel et
al., 1996). Sttot et al. (1993) found an average error of –3,78% in total yield measurement.
Blackmore and Moore (1999) classified the positioning errors in two types. In the first group,
they put those that affect a small amount of points (for example, loss of differential correction)
and in the second those that are real positioning displacements (points outside the field, for
example). Alleatory error from GPS is included in both groups.
Juerschik and Giebel (1999) reported extreme values of yield present in the data collected in
one field. They considered the elimination of those data but the criteria of the elimination has to
be based on good criteria because very high or low yield may really occur on nature. In all yield
monitors the area measurement is necessary to convert grain flow (kg.s-1) into yield (kg.ha-1),
and this is a source of errors. Moore (1998) studying the effect of partial swath width on yield
maps and found a reduction of 2.4% in the total yield compared to the actual yield.
Thylén (1997) identified the filling time error on yield maps. It is present in the beginning of rows
and is caused by the difference of time that the grain takes to reach the flow sensor when the
threshing mechanism is full compared to that when the mechanism is empty. The yield
measurements done during the filling time are lower then the actual yield because the grain is
used to fill up the mechanism of the machine.
Blackmore and Marshall (1996) proposed the Potential Mapping, a technique that was
supposed to deal and remove the errors from area measurement. This technique was
abandoned due to the introduction of errors on the edges of the fields, but marked the first
filtering on yield maps. Thylén and Algerbo (2000) developed a filter that compares each yield
3
point with its neighbors, removing those beyond a stated coefficient of variation. This filter
removed up to 50% of the points and the authors concluded that the application of a filter is
essential to the interpretation of yield data.
Taylor et al. (2000) proposed a technique called Mult Purposing Grid Mapping, that create cells
and inside of which, the yield points are evaluated. Based on a standard coefficient of variation,
the cells were considered good or bad to create the final yield map. Molin and Gimenez (2000)
developed an algorithm to reduce errors on yield maps. The authors used as parameters the
average yield and the difference of yield between points at a defined distance apart. The
algorithm reduced the variance and standard deviation of the data and concluded that the
application of the algorithm must be done associated with the knowledge of the errors present
on the yield data.
The objective of this paper is to analyze yield maps produced by six different commercial yield
monitors, identify and characterize the error present and propose a filtering routine that can be
applicable to all produced maps.
Material and Methods
The yield data used are from six different commercial yield monitors available in the market:
RDS Ceres 2, RDS Pro Series 8000, AFS, FieldStar, GreenStar and New Holland, in
fields with area up to 42.2 ha. Each yield monitor has its own characteristics, including the flow
sensor and way to record the data, coded or in text format (Table 1). Table 2 summarizes the
data obtained with each yield monitor and the crop mapped.
Table 1. Technical characteristics of the yield monitors used in this study.
Flow Sensor
Monitor Manufacturer
Combine
Yield
measurement Sensor File Type
RDS Ceres 2 RDS
Technology NH TC 57 Volumetric Optic Text
RDS Pro
Series 8000 RDS
Technology NH TC 57 Volumetric Optic Text
GreenStar John Deere JD 9510 Mass
detection Impact plate Code
FieldStar AGCO MF 34 Mass
detection Radiometric Open code
AFS AgLeader Case 2166 Mass
detection Impact plate Open code
New Holland AgLeader NH TC57 Mass
detection Impact plate Open code
4
Table 2. Summary of data obtained with each yield monitor and crop mapped.
Monitor Crop
Area
(ha)
Number of
collected
points
Density of
points
(points.ha-1)
RDS Ceres 2 Maize 22.0 12022 546
RDS Pro Series 8000 Soybeans 22.0 9955 452
GreenStar Maize 17.7 29631 1674
FieldStar Soybeans 78.0 19309 247
AFS Soybeans 42.2 9047 214
New Holland Wheat 12.6 8356 663
The developed methodology considers errors that can be identified in the yield map after its
production. The sequence of steps involved on the whole filtering process is presented on
Figure 1. For each map, the first step was a visual analysis of the yield map in a Geographic
Information System (GIS), looking for points located out of field (coarse positioning errors). After
that, the raw data were analyzed using spread sheet to identify null yield, wrong readings of the
moisture sensor, represented by null values of moisture or the minimum accept by the system,
that represents absence of signal from the sensor. All points with partial cutting width were
considered errors due to the possible erroneous interpretation of the actual cutting width by the
operator of the combine. Such error possibly occurs when the partial cutting width informed to
the monitor is different from the actual cutting width. For its identification, it was looked for points
in the file whose platform width was different from that informed as theoretical.
Figure 1. Flow chart of the steps of filtering process.
Text File Code file
Specific
program
Raw data
(text)
Step 1 – removal of
coarse positioning
errors
Step 2 – removal of
points with null or no
yield value
Step 3 – removal of
points with cutting
width different from
the total cutting width
Step 4 – removal of
points with null or no
grain moisture value
Step 5 – removal of
points with null
distance
Step 6 – removal of
points recorded
during the filling time
interval
Step 7 – removal of
discrepant data
Step 8 – Clean data
Step 0 – prepare
the data for the
filtering process
Harvest
5
Geographical coordinates of points were converted to UTM and the distance between points
was calculated. Points with null distance were considered as erroneous due to the destructive
nature of the yield mapping process. The points considered as erroneous were selected and
removed.
UTM coordinates were used to calculate the displacement index (DI), which describes the
direction of the combine’s displacement. Equations 1 and 2 show the calculation of the index in
two directions and Table 3 describes the interpretation of the DI.
)1()()( += iXiXNDI (1)
)1()()( += jYjYLDI (2)
where:
DI(N) is the displacement index in the direction North-south;
DI(E) it is the displacement index in the direction East-west;
X is the North coordinate of the combine;
i is the point in the direction N-S and considered way;
Y is the coordinate East of the combine;
j is the point in the direction E-W and considered way.
Table 3. Displacement index interpretation (DI).
Value Direction Way
Negative N-S South->North DI(N) Positive N-S North->South
Positive L-O East->West DI(E) Negative L-O West->East
The displacement index was used as indicative of beginnings and ends of rows, what allowed
the characterization of the filling end empting time errors. For the characterization of the filling
time error, five rows were alleatory chosen on each field. The average yield for each distance
was plotted as dependent on the distance and the filling time interval was considered the
distance needed for the average yield of the five rows to reache 90% of the maximum yield.
Considering all the points of the file, the average distance between points was calculated and
the number of points to be eliminated at the beginning of each row was calculated using
equation 3. Values were always rounded upward without decimal places, in case of fractions.
AD
FTI
N= (3)
where:
N is the number of points to be eliminated after the beginning of the row;
FTI is the filling time interval, in meters;
6
AD is the average distance among points.
The number of points collected during the FTI (N) was inserted in an algorithm to identify and
eliminate such error (Figure 2). The algorithm can be divided into two structural parts to facilitate
the understanding. The first part of the algorithm tests the data to identify the beginning of rows.
The criterion of stability is the occurrence of five consecutive points with the same DI signal.
Alteration of the sign of any point inside of this interval results in false value, interpreted by the
algorithm as beginning of row. If true, the algorithm repeats the yield value and it proceeds for
the next point.
If the algorithm identified row head, the first providence is to test if the analyzed point is inside of
the FTI. The DI signal of N previous points is checked, where N is the number of points to be
eliminated after the beginning of the row. If in N previous points there is change of sign of DI,
then it results false and yield value alters to null. If there is no signal alteration inside the verified
interval, the point is out of FTI and the algorithm results true, maintaining the yield value. The
algorithm is exemplified on equation 4, with criterion of stability of 5 points and elimination of 4
points after the beginning of the row.
IF
(N(i)<0 AND N(i-1)<0 AND N(i-2)<0 AND N(i-3)<0 AND N(i-4)<0)
OR
(N(i)>0 AND N(i-1)>0 AND N(i-2)>0 AND N(i-3)>0 AND N(i-4)>0)
THEN
M(i)
ELSE
IF (N(i)>0 AND N(i-1)>0 AND N(i-2)>0 AND N(i-3)>0;N(i-4)>0)
OR
(N(i)<0 AND N(i-1)<0 AND N(i-2)<0 AND N(i-3)<0 AND N(i-4)<0)
THEN
M(i)
ELSE
0
END
END
where:
N(i) are values of DI and
M(i) are yield values.
(4)
7
Figure 2. Flow chart of the acting of the algorithm to eliminate the filling time interval from the
yield data.
After the removal of points recorded during the FTI, an exploratory analysis of the data,
according to the methodology proposed by Tukey (1977). Limits were established for the
discrepancy of the data. The upper limit was defined according to equation 5 and the lower limit
according to the equation 6. Points with yield values out of the established limits are eliminated.
The data set is reordered by the identity of each point, finishing the filtering process.
IRUQUL .5,1+= (5)
IRLQLL .5,1= (6)
where:
UL is the upper limit;
LL is the lower limit;
UQ is the upper quartile;
LQ is the lower quartile and
IR it is the quartile range.
The filtering methodology was applied to the available data. Histograms of frequency distribution
and semivariance analysis of raw and clean data were built. Because of computational
limitation, only about 1000 points of each file, located in a diagonal strip to the field were
selected, so that all the representative areas were sampled. The adjusted models of the raw
and clean data were compared by cross validation, using linear correlation as qualitative
measure of the adjustment.
Is there a DI signal
change in the last five
points?
NO Not a row
head
Repeats the
yield value Next point
YES
It is a row
head
Is the point
considered
inside the
FTI?
NO
YES
Turns null the
yield value of
the point
8
Results and Discussion
Table 4 summarizes the characteristics of positioning error, null values of yield, cutting width
different from the total, null distance between points and null grain moisture.
Table 4. Frequency of occurrence of errors found on each yield file according to the proposed
methodology.
Yield Monitor
Errors RDS
Ceres 2 RDS Pro
Series
8000
Field
Star Green
Star New
Holland AFS
Frequency (%)
Positioning 0.00 0.07 0.00 0.00 6.30 0.00
Null Yield 0.33 4.60 0.00 2.34 1.83 0.34
Cutting Width 4.10 10.80 NA* 5.60 4.30 1.29
Null Distance between
points 0.00 0.00 0.08 11.48 0.07 0.00
Null grain moisture 9.50 9.30 NA* 3.80 1.40 1.06
* NA: not available in the file.
The identification of moisture sensor errors can be defective, because it only identifies cases
were there was not moisture reading and, however, there was reading of the flow sensor,
resulting in a point with null grain moisture in the file. Occurrence of 9.50% of moisture null
values on the RDS Ceres 2 yield data is associated to the occurrence of values of yield of
9999.9 t.ha-1. In the data obtained with the monitor RDS Pro Series 8000, the occurrence of
null yield is associated with the occurrence of null moisture, but points with null cutting width
also have yield with the same value. Other associations of errors happened with the other
monitor files, evidencing that the occurrence of some errors are not alleatory, but result from
equipment inconsistency.
The occurrence of null distance between points in the data obtained with the GreenStar yield
monitor is associated to null or simply absent yield, and to null grain moisture. Part of the points
that presented null distance to each other also presented unlikely and very high yields. The
errors presented by the data obtained with the monitor New Holland were shown independent
to each other, and error with larger occurrence was the positioning error, represented by a
group of points located out of the field. The same absence of relationship among the errors is
shown by the data obtained with the monitor AFS. The errors due to inaccuracy of GPS are not
identifiable by this methodology.
Specifically related to the monitor GreenStar, the high frequency of occurrence of points with
null distance to each other is probably due to the high collection rate of points, associated to
small positioning errors and also to the low harvesting speed. Many points are collected in a
short time interval, insufficient for the GPS accuracy to detect position change. As the
harvesting speed is low, the distance traveled by the combine during the collection interval of
data is small and it is inside of the circle of probable error of most of GPS receivers used in
precision farming, making possible the new position being erroneously equivalent the previous.
The result of the application of the DI in one of the yield files can be seen in the Figure 3. Such
index allows the exact identification of the course of the combine, the beginning and ending of
9
each row. However, it can only be applied when there is only one predominant harvesting
direction, North-south, East-west or intermediate. When the rows have more than one way, as
for instance, in accentuated curves, the methodology is shown limited in the identification of the
displacement way.
Figure 3. Application of the Displacement Index (DI) to the data collected with RDS Ceres 2
yield monitor.
The results of FTI characterization may be seen in the Figure 4, where are represented the five
alleatory chosen harvest lines and the average line for two files of the yield data analyzed, as
example. The characterization of the filling time interval provided the results shown in Table 5.
The data obtained with the monitors GreenStar and AFS did not present filling time interval,
while for the other monitors, the FTI ranged from 14 to 25 m after the beginning of the row. The
elimination of points collected during the combine filling time interval is shown on Figure 5,
where the action of the algorithm is exemplified with the yield map produced with data from the
FieldStar monitor. The algorithm acted strongly in the zones of headboard of the fields,
eliminating points collected during the FTI. The first 5 points of each beginning of them were
eliminated.
Dis
p
lacement Index
South – north
North – south
10
RDS Ceres 2
0
2
4
6
8
10
01020304050
Distânc ia (m)
Produtividade (t/ha)
Linha 1 Linha 2 Linha 3
Linha 4 Linha 5 Média
FieldStar
0,0
1,0
2,0
3,0
4,0
0 1020304050
Distânc ia (m)
Produtividade (t/ha)
Linha 1 Linha 2 Linha 3
Linha 4 Linha 5 Média
Figure 4. Example of determination of the filling time interval (FTI) for data from the yield
monitors FieldStar (left) and RDS Ceres 2 (right).
Table 5. Characterization of the filling time interval (FTI) for the files analyzed.
Yield monitor
GreenStar AFS New
Holland RDS
Ceres 2 RDS Pro
Series 8000 FieldStar
FTI (m) 0 0 14 20 22 25
Average distance
between points (m) 0 0 4 5 5.67 5.45
N1 (eq 3) 0 0 4 4 4 5
1 number of points to be eliminated after the beginning of each harvest row.
Figure 5 - Results of the application of the sixth step of the filtering process applied to the data
gathered with the FieldStar yield monitor (right) that corresponds to the elimination of the
points recorded during the filling time interval (FTI), in comparison with the map produced
starting from raw data (left).
RD
Ce
r
es
2 FieldStar
Yield (t/ha)
Yield (t/ha)
Line 1 Line 1
Line 2 Line 2 Line 3
Line 4 Line 5
Line 3
Line 4
Average Line 5 Average
Distance
(
m
)
Distance
(
m
)
11
In all the example files analyzed, points were eliminated in the tails of the distribution curve
starting from the inflection demonstrated by the distribution of the raw data. The definition of the
upper and lower limits (Table 6) for discrepant data allowed the elimination of unlikely points.
The calculated limits coincided with the inflections of the frequency distribution curves of the raw
data.
Table 6. Statistical limits for data discrepancy calculated for the data analyzed for each yield
monitor.
Yield monitor
Factor GreenStar AFS New
HollandRDS
Ceres 2
RDS pro
Series
8000 FieldStar
Yield (t.ha-1)
Upper quartile 8.68 3.73 4.53 5.04 5.34 3.36
Lower quartile 6.46 2.71 3.37 4.35 3.54 2.74
Quartile range 2.22 1.02 1.15 0.69 1.80 0.62
Upper limit 12.02 5.26 6.26 6.08 0.84 1.82
Lower limit 3.13 1.18 1.64 3.32 8.03 4.28
The first step of the process was regarding the elimination of points with coarse positioning
errors. The second step removed points with null yield. It is important to point out that mistakes
on measuring the yield or inaccuracy of the sensor are not detected by that methodology, just
points in that there were not yield readings. The same logic can be applied to the elimination of
points with null moisture; the process does not identify mistakes of reading of the sensor, just
absence of value or null values. On the other hand, the elimination of errors due to the
erroneous interpretation of the cutting width through the elimination of points with partial cutting
width eliminates just the points marked by the operator with partial width. The process is not
capable to identify points that were generated with partial width of platform and that were not
informed to the system. Table 7 presents the summary of the filtering process, indicating the
retreat of points in function of each applied stage to the data.
12
Table 7. Synthesis of the filtering process indicating the amount and the percentage of points
removed from the original raw data file.
Yield monitor
Step GreenStar AFS New HollandRDS
Ceres 2 RDS Pro
Series 8000 FieldStar
Remaining
points (%) Remaining
points (%) Remaining
points (%) Remaining
points (%) Remaining
points (%) Remaining
points (%)
Raw data 29631 100.0 9047 100.0 8356 100.0 12022 100.0 9955 100.0 19309 100.0
Coarse
positioning
errors 29631 100.0 9028 99.8 8033 96.1 11918 99.1 9182 92.2 19299 99.9
Null yield 28958 97.7 8998 99.5 7895 94.5 11883 98.8 9182 92.2 19299 99.9
Partial cutting
width 27394 92.5 8891 98.3 7606 91.0 10412 86.6 8858 89.0 19299 99.9
Null moisture 26201 88.4 8797 97.2 7577 90.7 10377 86.3 8500 85.4 19299 99.9
Null distance 23408 79.0 8797 97.2 7573 90.6 10374 86.3 8500 85.4 19299 99.9
FTI 23408 79.0 8797 97.2 7096 84.9 9991 83.1 8099 81.4 18548 96.1
Discrepancy
data 21537 72.7 8635 95.4 7006 83.8 9706 80.7 7830 78.7 18127 93.9
Points
removed (%) 27.3 4.6 16.2 19.3 21.3 6.1
The filtering process acted mainly in the points located in the extremities of the yield frequency
distribution. As example, Figure 6 displays the histograms of yield frequency distribution for the
data obtained with the GreenStar yield monitor before and after the filtering process, during
which 27,3% of the points were removed. The filter application did not alter the distribution
tendency, maintaining the highest yield occurrence in the frequency classes from 6.5 to 10 t.ha-
1.
GreenStar (Milho)
0
2
4
6
8
10
12
0 2 4 6 8 10 12 14 16 18 20
Produtividade (t/ha)
Freqüência (%)
Green Star
0
2
4
6
8
10
12
14
0 2 4 6 8 10 12 14 16 18 20
Produtividade (t/ha)
Freqüência (%)
Figure 6. Yield frequency distribution for the example of the data obtained with the monitor
GreenStar before (left) and later (right) of the filtering process.
The histogram of yield frequency distribution of the data obtained with the monitor AFS can be
seen in the Figure 7 and it exemplifies the yield frequency distribution in a situation of soft action
of the filter. In this case, only 4,6% of the points were removed (Table 7).
GreenStar GreenStar
Yield (t/ha)
Fre
q
uenc
y
(
%
)
Yield (t/ha)
Fre
q
uenc
y
(
%
)
13
AFS (Soja)
0
5
10
15
20
25
30
35
012345678910
Produtividade (t/ha)
Frequência (%)
AFS
0
5
10
15
20
25
30
35
012345678910
Produtividade (t/ha)
Freqüência (%)
Figure 7. Yield frequency distribution for the data obtained with AFS monitor before (left) and
after the filtering process (right).
Figure 8 displays an example of the selection of data used in the semivariance analysis. The
adjusted coefficients to the semivariograms produced with raw and clean data can be seen on
Table 8. As example, the adjusted semivariogram to the raw and clean data of the monitor
AFS is presented on Figure 9.
Table 8. Adjusted semivariogram parameters for the raw and clean data.
Parameter
Monitor File Model
Nugget Sill Range
Cross
validation
ϒ* ϒ (m)
Raw Gaussian 0.32 1.32 550.00 0.56
AFS Clean Gaussian 0.14 1.00 500.00 0.68
Raw Gaussian 1.20 3.20 850.00 0.38
New Holland Clean Spherical 0.27 0.70 17.36 0.58
Raw Exponential 1604000.00 3209000.00 171.00 0.47
RDS Ceres 2 Clean Exponential 0.09 0.17 64.50 0.62
Raw Nugget - - - -
RDSPro Series
8000 Clean Nugget - - - -
Raw Exponential 0.02 0.17 117.43 0.85
FieldStar Clean Exponential 0.00 0.09 28.05 0.83
Raw Spherical 1.45 3.95 18.87 0.71
GreenStar Clean Exponential 0.58 2.02 42.90 0.75
ϒ* is the semivariance
AFS AFS
Yield (t/ha) Yield (t/ha)
Fre
q
uenc
y
(
%
)
Frequency (%)
14
Figure 8. Example of points selected for the semivariance analysis. The selected points are
displayed in gray in this RDS Ceres 2 file.
Figure 9. Adjusted semivariograms for the raw (left) and clean (right) data collected with the
AFS yield monitor.
In a general way, the filtering process contributed to the characterization of the spatial
dependence, reducing the variability not explained by the adjusted models to the data. Only the
data obtained with the monitor RDS Pro Series 8000 did not present spatial dependence. In
the Figure 9 the reduction of the nugget effect from 0,32 to 0.14 may be observed due to the
performance of the filtering process (Table 6). Especially with relationship to the data obtained
with the monitor RDS Ceres 2, the filtering process was quite beneficial, reducing the nugget
effect of 1604000 for 0.09. Such reduction is due to the presence of values 9999.9 t.ha-1 in the
raw yield file. The presence of these values, usually in the heads of the field, elevates the
semivariance considerably. The filtering process improved the cross validation of the adjusted
models, to the data obtained with the monitors RDS Ceres 2, AFS, New Holland and
GreenStar, and it reduced from 0.85 to 0.83 for the data obtained with the monitor FieldStar,
being this the smallest alteration due to the application of the filter.
Considering that some errors are function of the dynamics of the relationship among machine,
monitor and crop, it is expected that the characteristics of each error keep constant for the same
machine, with the same monitor working in the same crop, facilitating data filtering work.
Evidently, the filtering process does not eliminate the need of correct calibration of the flow
sensor and the pre-harvest routines that are necessary for good quality of the data.
Separation distance (m) Separation distance (m)
Semivariance
Semivariance
15
The proposed routine was based on the removal of problematic points whose problems were
previously characterized. Other solutions could be applied to increase the final quality of the
yield maps. When working with points with problems in the reading of the grain moisture,
instead of classifying as problematic and eliminating them, the neighboring points could be used
to elaborate an average and to attribute it to the points whose reading was problematic,
avoiding, the exclusion of points with some valuable information. The same solution could be
applied to points with null yield, taking advantage of the other information contained in the string
that composes the file.
The incorrect calibration of the lag time generates a displacement of the points collected in the
direction of the combine displacement, and the way depends on the value of the lag time. If
overestimated, the displacement happens in the way of the movement and if underestimated,
the displacement happens in opposite way. In a more elaborated process, the displacement of
the points could be calculated starting from measurements in the head of the field and, with this
information associated to the displacement index and trigonometric equations, all the points of
the maps could be reallocated in a new position that should be closer to the place where the
points were really collected. Points with small positioning error could also have it altered to their
more probable position. Deviations of harvest rout and rows superposition could be altered
based on an imaginary line that would simulate the real course of the combine.
The developed filtering process is of easy application to the data, and the time of application is
not influenced by the file size or field area on ordinary computers. As the routine is fully applied
in computers, the creation of a program based on the guidelines of the process would reduce
the time spent during the application of the filter in each file that now is of approximately 40
minutes. The automation of the error characterization and its removal based on the
characteristics of each error can be considered as the natural evolution of the process.
Conclusions
Information quality in precision farming is a primordial factor for the decision making process.
Six different types of errors and discrepant data were found and characterized in the raw files
analyzed. The errors varied according to the monitor and the interaction between it and the
combine.
Based on the individual characteristics of each error present in each raw data file, the proposed
filtering routine is efficient on removing errors and improving the quality of information. The
routine acts differently for each data file and the intensity of points removal depends on the
errors present in the raw data.
Rebuilding the routine as an automated program can be considered as next step regarding error
removal and improvement of quality of the information, reducing its application time.
Acknowledgements
The authors would like to thanks São Paulo State Research Foundation – FAPESP for financial
support and to ABC Foundation, AGCO, CNH and John Deere for helping with the data
gathering.
16
References
Auernhammer, H.; Demmel, M.; Muhr, K.; Rottmeier, J.; Wild, K. Yield measurement on
combine harvesters. In: ASAE WINTER MEETING, Chicago, 1993. Chicago:
ASA;CSSA;SSSA, 1993. p.15. (ASAE paper, 01-1503).
Birrel, S.J.; Sudduth, K.A.; Borgelt, S.C. Comparison of sensors and techniques for crop yield
mapping. Computers and Electronics in Agriculture, v.14, n.2, p.215-223, 1996.
Blackmore, B.S.; Marshall, C.J. Yield mapping; errors and algorithms. In: INTERNATIONAL
CONFERENCE ON PRECISION AGRICULTURE, 3., Minneapolis,1996. Proceedings.
Madison: ASA;CSSA;SSSA, 1996, p.403-415.
Blackmore, B.S.; Marshall, C.J. Yield mapping; errors and algorithms. In: INTERNATIONAL
CONFERENCE ON PRECISION AGRICULTURE, 3., Minneapolis,1996. Proceedings.
Madison: ASA;CSSA;SSSA, 1996, p.403-415.
Blackmore, S.; Moore, M. Remedial correction of yield map data. Precision Agriculture, v.1,
n.1,p. 51-66, 1999.
Haneklaus, S.; Lilienthal, H.; Schnug, E.; Panten,K.; Haverech, E. Routines for efficient yield
mapping (Compact disc). In: INTERNATIONAL CONFERENCE ON PRECISION
AGRICULTURE, 5., Madison, 2000. Proceedings. Madison: ASA, 2000.
Juerschik, P.; Giebel, A. Processing of point data from combine harvesters for precision
farming. In: EUROPEAN CONFERENCE ON PRECISION AGRICULTURE, 2.,
Denmark, 1997. Precision agriculture’99. Denmark: BIOS Scientific Publishers, 1999.
v.1, p. 297-307.
Larscheid, G.; Blackmore, B.S.; Moore, M. Management decisions based on yield maps, In:
EUROPEAN CONFERENCE ON PRECISION AGRICULTURE, 1., Warwick, 1997.
Precision agriculture’97. Warwick: BIOS Scientific Publishers,1997. p. 895-902.
Makepeace, R.J. Benefits and limitations of precision farming, In: CONFERENCE ON PESTS
AND DISEASES, 1., Brighton, 1996. Proceedings. Brington: Brington University
Express, 1996. p. 1235-1242, 1996.
Marcy, T.S.; Theckery, D.L.; Macy, N.C. Yield monitoring experiences. In: ASAE WINTER
MEETING,1., Atlanta, 1994. abstracts. Atlanta: ASAE, 1994. p.48.
Molin, J.P. Geração interpretação de mapas de produtividade para agricultura de precisão, In:
BORÉM,A.; GIÚDICE, M.P. del; QUIROZ, D.M. de; MANTOVANI, E.C.; FERREIRA,
L.R.; VALLE, F.X.R. do; GOMIDE, R.L. Agricultura de precisão. Viçosa:Universidade
Federal de Viçosa, 2000. p. 237-258.
Molin, J.P.; Gimenez, L.M. Desenvolvimento e um algoritmo para redução de erros em mapas
de rendimento obtidos em agricultura de precisão (compact disc). In: CONGRESSO E
MOSTRA DE AGROINFORMÁTICA, 1., Ponta Grossa, 2000. InfoAgro 2000. Ponta
Grossa, 2000.
Moore, M. An investigation into the accuracy of yield maps and their subsequent use in crop
management. Silsoe, 1998. p. 379. Thesis (Ph.D.) – Silsoe College.
Reitz, P., Kutzback, K.H. Investigations on a particular yield mapping system for combine
harvesters. Computer and Electronics in Agriculture, v. 14, n. 2/3, p. 137-150, 1996.
Sttot, B.L., Borgelt, S.C., Sudduth, K.A. Yield determination using an instrumented Class
combine. In: ASAE WINTER MEETING, Chicago, 1993. Chicago: ASA;CSSA;SSSA,
1993. p. 1507-1521. (ASAE Paper 93-1507).
Taylor, R.K.; Kastens, D.L. ; Kastens, T.L. Creating yield maps from yield monitor data using
multi-purpose grid mapping (mpgm) (compact disc). In: INTERNATIONAL
17
CONFERENCE ON PRECISION AGRICULTURE, 5., Madison, 2000. Proceedings.
Madison: ASA, 2000.
Thylén, L.; Jurschik, P.; Murphy, D.L.P. Improving the quality of yield data, In: EUROPEAN
CONFERENCE ON PRECISION AGRICULTURE, 1., Warwick, 1997. Precision
agriculture’97, Warwick: BIOS Scientific Publishers, 1997. v.1, p. 743-750.
... Furthermore, recent indicators have been proposed in the literature to improve the analyses of within-field spatial variability. However, those latter Basso et al. (2007), Junior et al. (2006), Molin and Menegatti (2002), Reza et al. (2010), and Simbahan et al. (2004) CV and Cambardella and/or MCD index Baluja et al. (2012), Cambardella and Karlen (1999) are poorly reported in recent scientific papers. Are those approaches too complicated for practitioners to understand and/or implement? ...
Article
Full-text available
In agricultural fields, observations near in space share more similarities than observations far apart. This phenomenon of closely related points, the so-called spatial autocorrelation or in that case, the within-field spatial variability, is well-recognized and needs to be characterized to consider site-specific management. Quantifying spatial dependency is fundamental for understanding the underlying factors affecting field productivity. An examination of multiple scientific papers was carried out to assess why and how practitioners were evaluating the spatial variability across their fields. An analysis of the existing descriptors of within-field variability was performed to identify the most relevant indicators to use based on (i) the case studies that practitioners employed, (ii) the different nature of data to which users were confronted, and (iii) the complexity and ease of access to information in support of these available approaches. Finally, this paper provides users with a comprehensive decision tree that should help them select an appropriate index of spatial variability for their work. Results also highlight the needs for further investigation, especially regarding the implementation of more general and standardized approaches that will enable cross-study comparison.
... At the same time the first attempts occurred to measure grain flow in combine harvesters to generate yield maps. Coffee, as e permanent crop, followed a specific trend and its mechanization started too late compared to the majority of the global crops. Mechanized harvesting started in 1979, in Brazil, and only recently precision agriculture has been added to the crop practices (Molin et al., 2002). In 2000, an automatic system to measure the volume flow of mechanically harvested coffee was developed and incorporated into the Jacto ® K3 ...
Article
Coffee was introduced in Brazil in the 18th century and it was, for a long time, the most important economical activity influencing the Brazilian colonization and development. It still involves about ten million people in its production chain, even though it was partially supplanted by several other economical activities. The search for higher yield and quality demands the adoption of new technologies such as precision agriculture. In order to demonstrate usefulness of precision agriculture, this study compared the effects of nitrogen, phosphorus, and potassium fertilization, at variable rates of fertilizer compared to constant rate in the coffee yield in two experimental fields. The experiment was conducted constant and variable rate in each sequential row in the entire field, using a customized fertilizer application machine. In two fields, the variable rate of coffee fertilizer management offered a significant increase in yield when compared to the constant rate normally used by farmers, indicating that precision agriculture techniques may be a solution for yield improvements.
... Revista Brasileira de Milho e Sorgo, v.6, n.3, p. 309-324, 2007 A produtividade de grãos da cultura (Prod, kg ha -1 ), corrigida para teor de água igual a 13%, foi medida com monitor de produtividade NH PLMS, produzido pelas empresas AgLeader ® e GPS Trimble ® , modelo AG-132, com correção por algoritmo interno, instalado em uma colhedora NH 59 (New Holland ® ). Das produtividades de grãos amostradas, foram eliminados os pontos com valores nulos, os pontos localizados fora do perímetro da gleba e os pontos localizados na borda da gleba, correspondentes à fase de equalização do fluxo de grãos internos à colhedora, conforme descrito por Molin & Menegatti (2002). ...
Article
Full-text available
It is important to consider agricultural production with emphasis on sustainability and it is fundamental to consider spatial and temporal variability of yield factors and conduct crop managements under this variability. The objective of the present study was to define management zones in an experimental field cultivated with corn considering the spatial variability of macronutrients and saturation of bases on the soil and their interactions with grain yield and population of normal plants. Soil data was derived from grid sampling, plant population was collected on the same grid, yield was obtained from yield monitor on the whole area and geostatistics tools were used for analysis. Results allowed estimating the levels of phosphorus, potassium, calcium, magnesium and saturation of bases in the soil and establishing management zones for soil fertility based on the population of normal plants and grain yield, considering that
Article
Full-text available
The utilization of Normalized Difference Vegetation Index (NDVI) data obtained through satellite images can technically improve the process of delimiting management zones (MZ) for annual crops, resulting in socioeconomic and environmental benefits. The aim of this study was to compare delimited MZ, using crop productivity data, with delimited MZ using the NDVI obtained from satellite images in areas under a no-tillage system. The study was carried out in three areas located in the state of Rio Grande do Sul, Brazil. Three crop productivity maps, from 2009 to 2015, were used for each area, whereby the NDVI was calculated for each crop productivity map using images from the Landsat series of satellites. Descriptive and geostatistical analysis were conducted to determine the productivity and NDVI data. The MZ were then delimited using the fuzzy c-means algorithm. Spearman's correlation matrix was used to compare the methodologies used for delimiting the MZ. The MZ based on NDVI calculated from the satellite images correlated with the MZ based on crop productivity data (0.48 < r < 0.61), suggesting that the NDVI can replace or be complementary to productivity data in delimiting MZ for annual cropping systems.
Article
Full-text available
Yield maps are a powerful tool with regard to managing upcoming crop productions but can contain a large amount of defective data that might result in misleading decisions. The objective of this work is to help improve and compare yield data filtering algorithms by generating simulated datasets as if they had been acquired directly in the field. Two stages were implemented during the simulation process (i) the creation of spatially correlated datasets and (ii) the addition of known yield sources of errors to these datasets. A previously published yield filtering algorithm was applied on these simulated datasets to demonstrate the applicability of the methodology. These simulated datasets allow results of yield data filtering methods to be compared and improved.
Article
Full-text available
Up to date, the use of precision agriculture concepts on sugar cane is limited and one reason is due to difficulties in detecting the spatial yield variability and the effects of localized input applications. The main objective of this work was to explore the use of simple techniques to collect data in the field, during harvesting, to produce yield maps based on the harvesting systems used today, without interfering on it. The main hypothesis was that the weight load average on a small series of loads, represented by one truck or wagon is sufficiently representative for detecting yield variability based on the loads spacing in the row. Samples were taken in the fields from two loaders as treatments in 2002 and load coordinates from one field were collected in 2003 to calculate the area and yield of each load. Estimation of the total dispersion of load weights showed that it can be done by an interval of ± 2,5s from the weight average and treatments did not show the same trend on normality. Loaders averages differed, indicating that frequent calibrations would be necessary for using the average weight as an estimator of the load weights. It may be possible if using small population of loads. Data from 2003 allowed the estimation of yield points that were interpolated resulting in a surface map of sugar cane yield showing the variability within the field.
Article
Precision agriculture (PA) involves using electronic technology to collect a large amount of data in the field for use in site-specific crop management. Major issues in the implementation of PA include interpreting the huge amount of data collected, understanding the causes of variability, and being able to propose sound strategies for field variability management. To help address these issues, a software infrastructure is proposed. It adopts concepts from software engineering with the aim of providing a basis for the development of information systems for PA based on open platforms, and on data communication and software interoperability standards. The infrastructure has two main components and a reference architecture for the development of distributed service-oriented systems. This paper presents and discusses the infrastructure and its components and presents a prototype application named yield data filtering, which filters yield monitor data and was developed as a proof of concept.
Article
Yield monitor data offer the potential to improve management decisions through increased knowledge of production variability within a field. However, yield monitor data contain numerous sources of error that can lead to misrepresentative maps and inappropriate inferences. Many techniques have emerged to reduce these errors, including surface contouring and potential mapping. Each of these techniques has advantages, however its use is limited in many situations, as when terraces or grassed waterways are present, when point-row data exist, or when multiple combines harvest a field and every machine does not contain a yield monitoring system. This research evaluated a new technique called Multi-Purpose Grid Mapping (MPGM) for generating yield maps using geographical techniques and an understanding of combine operation. The technique is similar to potential mapping, but differs in that the rule based cell yield determination allows for missing data. The MPGM procedure was evaluated on 12 fields where known harvest width errors existed and yield monitor data were available for the entire field. Results were generalized to create rules for the MPGM procedure regarding data quality thresholds, grid cell size, and methods for determining cell yields.
Yield measurement on combine harvesters
  • H Auernhammer
  • M Demmel
  • K Muhr
  • J Rottmeier
  • K Wild
Auernhammer, H.; Demmel, M.; Muhr, K.; Rottmeier, J.; Wild, K. Yield measurement on combine harvesters. In: ASAE WINTER MEETING, Chicago, 1993. Chicago: ASA;CSSA;SSSA, 1993. p.15. (ASAE paper, 01-1503)
An investigation into the accuracy of yield maps and their subsequent use in crop management
  • M Moore
Moore, M. An investigation into the accuracy of yield maps and their subsequent use in crop management. Silsoe, 1998. p. 379. Thesis (Ph.D.) – Silsoe College.