RESEARCH
Artificial Neural Network Models as a Decision
Support Tool for Selection in Sugarcane:
A Case Study Using Seedling Populations
Marvellous M. Zhou, Collins A. Kimbeng,* Thomas L. Tew, Kenneth A. Gravois, and Michael J. Pontif
ABSTRACT
Artificial neural network (ANN) models are
mathematical models based on biological neu-
ral networks; they are a supervised learning
method and use pattern learning from a train-
ing dataset that is a subsample of the whole
dataset to produce predictions of response
variables. We demonstrate the potential of an
ANN model as a tool for selection in sugarcane.
Cane yield components, namely stalk number,
stalk height, and stalk diameter, were measured
on individual seedlings and used as predic-
tor variables to produce a selection decision
(reject or select a seedling) based on an ANN
model. Compared with the currently used visual
method of selection, the difference in cane yield
between the mean of the selected and rejected
seedlings was greater for seedlings selected by
the ANN model. The difference increased when
similar selection intensity was applied in both
selection methods. The ANN model selected
fewer seedlings with cane yield lower than the
population mean and rejected fewer seedlings
with higher cane yield compared with the visual
method. The ANN model compels the breeder
to consider all traits simultaneously when decid-
ing whether to select or reject a clone, which is
likely to be more efficient than judging the merit
by considering each trait independently. The
ANN model can be a valuable tool to determine
selection rates to be applied in selecting sugar-
cane families during seedling selection.
CROP SCIENCE, VOL. 51, JANUARY FEBRUARY 2011
M.M. Zhou and C.A. Kimbeng, School of Plant, Environmental and
Soil Sciences, Louisiana State Univ. Agricultural Center, 104 M.B.
Sturgis Hall, Baton Rouge, LA 70803; M.M. Zhou, current address:
South African Sugarcane Research Institute, 170 Flanders Dr., Private
Bag X02, Mount Edgecombe, KwaZulu-Natal, ZA 4300. T.L. Tew,
USDA-ARS, Sugarcane Research Unit, 5883 USDA Rd., Houma, LA
70360. K.A. Gravois and M. Pontif, Sugar Research Station, Louisi-
ana State Univ. Agricultural Center, 5755 LSU Ag Rd., St Gabriel,
LA 70776. Received 4 Feb. 2010. *Corresponding author (ckimbeng@
agcenter.lsu.edu).
Abbreviations: AIC, Akaike information criterion; ANN, artificial
neural network; ASE, average squared error; FPE, fi nal prediction
error; LSU AgCenter, Louisiana State University Agricultural Center;
SBC, Schwarz Bayesian criterion.
A n artificial neural network model, often called an ANN
model, is a mathematical or computational model based on
biological neural networks (Nelson and Illingworth, 1991). The
ANN model is a supervised learning method and uses pattern
learning from training data to produce models that generate pre-
dictions of response variables (Nelson and Illingworth, 1991;
Masters, 1993). It consists of a layered, free-forward, and com-
pletely connected network restricted to a single direction of flow
(Nelson and Illingworth, 1991). It has an input layer, a hidden
layer, and an output layer. The ANN models complex relation-
ships between input variables and outputs (Fausett, 1994; Gurney,
1997). The model must be trained by processing data with input
and output patterns similar to the data to be predicted. The model
detects similarities in new input data and uses these similarities to
generate output predictions (Smith, 1993). The logistic function
calculates probabilities used to make predictions (Allison, 2003;
Published in Crop Sci. 51:2131 (2011).
doi: 10.2135/cropsci2010.02.0057
Published online 25 Oct. 2010.
Crop Science Society of America | 5585 Guilford Rd., Madison, WI 53711 USA
All rights reserved. No part of this periodical may be reproduced or transmitted in any
form or by any means, electronic or mechanical, including photocopying, recording,
or any information storage and retrieval system, without permission in writing from
the publisher. Permission for printing and for reprinting the material contained herein
has been obtained by the publisher.
21
Agresti, 2007). Multiple linear regression equations form
the linear predictors (Hertz et al., 1990; Agresti, 2007).
The ANN models have been used in financial risk
management (Huang et al., 2004; Sethuraman, 2006),
process control in manufacturing (Lee and Paik, 2006),
predicting credit scores and interest rates (Perkins and
Brabazon, 2006), and predicting fish abundance (Iglesias
et al., 2006). In predicting credit card scores and interest
rates using ANN models, for example, the applicants pay-
ment history and other variables are used to calculate the
probability that the applicant will make loan repayments
at a particular interest rate using data from other individu-
als as the training data set. The probabilities determine the
relative risk of the applicant and are used to determine the
interest rates on credit cards or loans.
Agronomic applications of ANN models include
yield prediction (Kaul et al., 2005; Ji et al., 2007), disease
prediction (Batchelor et al., 1997), soil-water retention
estimations (Schaap and Bouten, 1996), and crop devel-
opment modeling (Elizondo et al., 1994). Starrett et al.
(1997) reported that an ANN model performed better
(r 2 = 0.984) than a regression model (r 2 = 0.780) when
predicting applied N leaking below the root zone of turf
grass. According to Batchelor et al. (1997), ANN models
produced better results than traditional statistical methods
when predicting soybean rust. Recently, the ANN model
was used to classify green tea [Camellia sinensis (L.) Kuntze]
accessions into taxonomic groups using leaf morphologi-
cal measurements as input variables (Pandolfi et al., 2009).
The same group of researchers used the ANN model to
classify Japanese Camellia (Camellia japonica L.) using phyl-
lometric and fractal parameters (Mugnai et al., 2008).
In this paper, we demonstrate how the ANN model
concept can be applied as a decision support tool for selec-
tion in sugarcane. We use two seedling populations from
two different sugarcane breeding programs in Louisiana.
During seedling selection, stalk diameter, stalk length,
and stalk number are commonly used to evaluate cane
yield (Chang and Milligan, 1992). These yield compo-
nents (Kang et al., 1983, 1991; Milligan et al., 1990) can
be used as input variables in the ANN models to predict
the probability of either selecting or rejecting a seedling.
During selection, the decision to select or reject a seedling
depends on the combination and magnitude of these yield
components. The outcome or response variable would
be to either select (1) or reject (0) a seedling, which is
binary in nature. In this case, a training dataset consisting
of previously defined response variables (select or reject)
and the input or independent variables (cane yield com-
ponents) are used by the ANN model to determine the
logistic regression function. Then a new dataset consisting
of input variables is fed into the logistic regression func-
tion, which produces probabilities of either selecting or
rejecting a seedling as the output.
22 WWW.CROPS.ORG
The objective of this study was to evaluate the poten-
tial of using the SAS enterprise miner (SAS Institute, 2007)
ANN model as a decision support tool for selection in sug-
arcane. We built and used an ANN model to identify seed-
lings with high cane yield potential at the seedling stage of
two independent sugarcane breeding programs. The yields
of seedlings selected using the ANN models were compared
with those selected using the currently used visual method.
MATERIALS AND METHODS
Populations and Measurements
Data were collected from seedlings raised from true seed at the
USDA Agricultural Research Service (ARS) Ardoyne Research
Farm at Schreiver, LA., and Louisiana State University Agricul-
tural Center (LSU AgCenter) Sugarcane Research Station at St.
Gabriel, LA. Seedlings in sugarcane refer to the initial unselected
plants grown from true seed obtained after crossing. The seed-
lings from 17 crosses grown at the USDA and 5 crosses grown
at the LSU AgCenter (Table 1) were first raised in the green-
house and then transplanted into the field as single plants in the
spring (April) of 2002 along with seedlings in the regular breed-
ing program. In the USDA breeding program, 60,000 seedlings
are grown whereas in the LSU AgCenter program about 85,000
seedlings from about 130 crosses each year are grown. The seed-
lings from each cross are usually planted in unreplicated two-
row plots with varying number of seedlings per cross, depending
on the importance of the cross, with the elite crosses typically
having more than 500 seedlings. However, in this study, for the
USDA population, the seedlings of each cross were divided and
transplanted as two replications. In each plot, two rows were
planted, each row containing 16 seedlings. In 2003, eight seed-
lings (four from each row per plot) were randomly chosen from
each plot and used for data collection.
At the LSU AgCenter, five crosses (Table 1), each with
more than 500 seedlings, were selected from the seedling pro-
gram. Thirty seedlings per cross were randomly chosen in 2003.
The chosen seedlings from the two populations were evaluated
visually to determine if they would have been selected (1) or
rejected (0). The decision to select (1) or reject (0) a seedling in
this study was based on a consensus between two experienced
sugarcane breeders. Scrutiny by two breeders is not typically
practiced during routine selection and it usually takes less time,
than was the case in this study, to appraise seedlings. As is cus-
tomary, dry leaves of the seedlings at the LSU AgCenter were
stripped off before selection whereas those of the USDA seed-
lings were not. For the chosen seedlings in each of the popu-
lations (USDA and LSU AgCenter), the number of stalks per
stool was recorded, stalk height was measured from the base of
the stool to the topmost visible dewlap, and stalk diameter was
measured at the center of the stalk on three randomly chosen
stalks using a caliper and without reference to the bud.
The seedling cane yield was calculated based on an empiri-
cally derived formula (De Sousa-Vieira and Milligan, 1999)
(Eq. [1]). Their calculation assumed the sugarcane stalk was
a perfect cylinder with specific gravity of one as empirically
determined from previous studies (Miller and James, 1974; Gra-
vois et al., 1991; Chang and Milligan, 1992).
CROP SCIENCE, VOL. 51, JANUARY FEBRUARY 2011
Table 1. Cross showing female and male parents of sugarcane seedlings planted at the USDA and Louisiana State University
Agricultural Center (LSU AgCenter) sugarcane research farms.
Cross Female parent Male parent Cross Female parent Male parent
Crosses evaluated at the USDA research farm
306 Ho94-856 HoCP96-540 3255 HoCP00-945 Ho94-856
3055 HoCP00-945 HoCP99-866 3256 HoCP00-950 Ho94-856
3074 HoCP00-950 HoCP96-540 3257 L98-207 Ho94-856
3093 HoCP00-945 HoCP96-540 3276 TUCCP77-42 HoCP99-866
3101 HoCP99-866 HoCP96-540 3322 TUCCP77-42 L98-207
3107 HoCP00-950 LCP85-384 3328 HoCP91-555 LCP85-384
3111 HoCP99-866 LCP85-384 3345 HoCP91-555 L98-207
3174 HoCP00-945 LCP85-384 3417 HoCP91-555 TUCCP77-42
3249 N27 LCP85-384
Crosses evaluated at the LSU AgCenter research farm
XL01-001 HoCP92-624 HoCP91-552
XL01-050 LCP86-454 LCP85-384
XL01-059 HoCP95-951 HoCP96-540
XL01-215 TucCP77-42 LCP85-384
XL01-460 Ho95-988 L99-238

Seedling cane yield (kg) = ndr 2L/1000 [1]
where n = seedling stalk number, d = density at 1.0 g cm3, r =
stalk radius (in cm), and L = stalk height (in cm).
Data Analysis Using Articial
Neural Networks
The training data consisted of 20% (30 seedlings grown at the
LSU AgCenter) and 10% (28 seedlings grown at the USDA) of
the original data. The input variables were stalk number, stalk
height, and stalk diameter and the response was either to select
(1) or reject (0) a seedling as determined by two experienced
sugarcane breeders at each location. The training data were
run in SAS enterprise miner (SAS Institute, 2007) to produce
the coefficients of the multiple linear regressions. The data col-
lected from 150 LSU AgCenter and 272 USDA seedlings con-
stituted the prediction data. In the prediction data, the response
values to either select (1) or reject (0) a seedling were coded
as missing values and needed to be estimated by the model.
The model selection criterion used was average error and the
network architecture was the generalized linear model. The
training technique used was the Levenberg-Marquadt set at
50 preliminary runs.
Coefcients of the Prediction Model,
Probability Values, and Fit Statistics
The ANN models use the training dataset to calculate coef-
ficients of the prediction model, which represents the relative
weighting of each input variable. The probability of either
selecting or rejecting a seedling is calculated by multiplying
the values of stalk number, stalk height, and stalk diameter by
their respective coefficient as shown in Eq. [2] (for the data from
the LSU AgCenter seedlings) and Eq. [3] (for the data from the
USDA seedlings).
p( Y = 1) = {exp[50.2 + (1.38 stalk no.)
+ (6.16 stalk height) + (11.2 stalk diam.)]}
/{1 + exp[50.2 + (1.38 stalk no.)
+ (6.16 stalk height) + (11.2 stalk diam.)]},
[2]
p( Y = 1) = {exp[18.1 + (0.04 stalk no.)
+ (2.73 stalk height) + (5.71 stalk diam.)]}
/{1 + exp[18.1 + (0.04 stalk no.)
+ (2.73 stalk height) + (5.71 stalk diam.)]},
[3]
Only one probability (p) can be modeled: in this case, the prob-
ability to select. The probability to reject was, therefore, 1 p.
To predict a response, a threshold probability must be specified.
If the probability to select is modeled, the response would be to
select when the probability value was equal to or greater than
the threshold value and to reject when the probability value
was less than the threshold. In SAS ANN models, the default
threshold is 0.5. Larger threshold values produce more stringent
selection criteria and vice versa.
The ANN model analysis produces six fit statistics that
represent parameters that can be used to compare the model
or equation for their ability to account for the variability in
the data. The average profit (prediction power) was estimated
as the correlation between the response variable (1 or 0) and
probability (Agresti, 2007). A higher profit would mean the
probability value was highly predictive of the response variable.
The misclassification rate was estimated as the proportion of
total observations that were classified by the model into differ-
ent response categories from what was observed. Lower values
would indicate correct model classification and accurate train-
ing dataset. The average squared error (ASE) was calculated as:
ASE = SSE/N = (OR Pp)2/N, [4]
in which SSE is the sum of the squared error, OR is the observed
response, Pp is the prediction probability, and N is the number
of observations in the training dataset. Smaller values would
indicate better model fit. The fi nal prediction error (FPE) was
estimated as:
FPE = [SSE(N + P)]/[N(N P)], [5]
in which P is the number of parameters including the intercept.
The FPE is an adjustment to ASE using (N + P)/(N P). The
adjustment penalizes for overparameterization (model complex-
ity) or the inclusion of too many input variables. Overparam-
eterization inflates FPE and increases prediction errors. It is
generally desirable to achieve the best model fit by specifying the

CROP SCIENCE, VOL. 51, JANUARY FEBRUARY 2011 WWW.CROPS.ORG 23

Table 2. Model t statistics from articial neural network
The probability to either select or reject a seedling as
analysis of sugarcane seedling data from the USDA and
Louisiana State University Agricultural Center (LSU AgCen- calculated using Eq. [2] (for the LSU AgCenter data) and Eq.
ter)populations. [3] (for the USDA data) are shown in Tables 3 and 4. The p(Y
Model t statistics LSU AgCenter USDA
= 1) is the probability to select and p(Y = 0) is the probability
Average prot 0.61 0.36 to reject a seedling. The threshold probability value used was
Misclassication rate 0.07 0.11 0.5. When p(Y = 1) was equal to or greater than 0.5, the
Average squared error 0.06 0.13 seedling was selected and categorized as select (1); otherwise
Final prediction error 0.08 0.17
Akaike information criterion (AIC) 19.79 31.72
it was categorized as reject (0). The column labeled ANN in
Schwarz Bayesian criterion (SBC) 25.52 37.05 Tables 3 and 4 represents the predicted categories. The input
variables are included in the output and can aid the breeder in
simplest or most parsimonious model. As with ASE, lower values deciding what to do with borderline seedlings.
would indicate better model fit. The Akaike information crite- Generally, seedlings were selected using higher probabil-
rion (AIC) (Akaike, 1974) and Schwarz Bayesian criterion (SBC) ity values from the LSU AgCenter (Table 3) than the USDA
(Schwarz, 1978) were used to compare the relative model fit for (Table 4) population probably because the USDA population
two or more models. Lower values indicated better model fit.
included a random sample of crosses from that program whereas
the LSU AgCenter population involved elite crosses. Eighteen
RESULTS of 30 LSU AgCenter seedlings were selected with probability
Model Fitness and Probability Values values ranging from 0.58 to 1.00 (mean = 0.88). Nine of 30
The fit statistics produced higher prediction power for the USDA seedlings were selected with probability values rang-
LSU AgCenter than the USDA dataset (Table 2). Misclassifi- ing from 0.53 to 0.91 (mean = 0.72). This indicated a greater
cation, ASE, FPE, AIC, and SBC values were greater for the precision of selection from the LSU AgCenter compared with
USDA population, indicating poorer model fit of the data the USDA population, reflecting the effect of the relative vari-
compared with that from the LSU AgCenter population. ability for cane yield in these populations, as depicted in Fig.
The distribution patterns of the two populations were 1 and 2. In Fig. 1, when the threshold for selecting a seedling
evaluated graphically by plotting the estimated seedling was set at a probability value of >0.5, seedlings meeting this
cane yield (on the x axis) against their corresponding criterion had an estimated seedling cane yield greater than 7.5
probability value (on the y axis). The LSU AgCenter data kg, whereas in Fig. 2, using the 0.5 threshold, seedlings with
followed closely the theoretical, logistic cumulative dis- estimated cane yield as low as 3.5 kg were included.
tribution function (Casella and Berger, 2003) compared
with the USDA data (Fig. 1). The distribution patterns Discriminating Ability of Articial Neural
depicted trends that were similar to the fit statistics (Table Network Models Versus Visual Selection
2), confi rming the larger variability found within the The means for seedling stalk number, stalk height, stalk
USDA than the LSU AgCenter data. diameter, and estimated cane yield were calculated for each
group of selected and rejected seedlings. The difference

Figure 1. The logistic cumulative distribution functions for estimated seedling cane yield (in kg) (x axis) plotted against posterior probabilities
(y axis) for the Louisiana State University Agricultural Center (LSU AgCenter) (a) and USDA (b) populations.

24 WWW.CROPS.ORG CROP SCIENCE, VOL. 51, JANUARY FEBRUARY 2011

Table 3. Probability of selecting [p(Y = 1)] or rejecting [p(Y = 0)] a seedling, the predicted selection decision by the articial
neural network (ANN) model, the selection decision by the visual method (Visual), stalk number (Stalks), stalk height (Height),
stalk diameter (Diameter), and seedling cane yield (Cane) for the rst 30 seedlings derived from the Louisiana State University
Agricultural Center (LSU AgCenter) population.
Seedling p(Y = 1) p(Y = 0) ANN Visual Stalks Height (cm) Diameter (cm) Cane (kg)
1 1.00 0.00 1 1 23 237 1.70 12.38
2 0.09 0.91 0 0 11 225 1.68 5.49
3 0.31 0.69 0 0 9 250 1.93 6.59
4 1.00 0.00 1 1 25 240 2.12 21.19
5 1.00 0.00 1 1 20 250 1.70 11.35
6 0.01 0.99 0 0 12 190 1.51 4.08
7 0.00 1.00 0 0 12 170 1.34 2.88
8 1.00 0.00 1 1 16 240 1.84 10.21
9 0.00 1.00 0 0 4 240 2.11 3.36
10 0.87 0.13 1 1 11 235 2.00 8.12
11 0.03 0.97 0 0 10 240 1.63 5.01
12 1.00 0.00 1 1 17 260 2.33 18.85
13 0.91 0.09 1 1 10 240 2.13 8.56
14 0.13 0.87 0 0 6 240 2.25 5.73
15 1.00 0.00 1 1 14 245 2.08 11.66
16 0.90 0.10 1 0 14 220 1.75 7.41
17 1.00 0.00 1 1 23 230 1.90 15.00
18 0.34 0.66 0 0 10 220 1.98 6.78
19 0.01 0.99 0 0 10 205 1.72 4.77
20 0.00 1.00 0 0 8 240 1.68 4.26
21 0.91 0.09 1 1 16 230 1.45 6.08
22 1.00 0.00 1 1 19 230 1.72 10.16
23 0.33 0.67 0 1 11 215 1.88 6.57
24 1.00 0.00 1 1 13 255 2.23 12.95
25 0.58 0.42 1 1 10 235 1.98 7.24
26 0.00 1.00 0 0 8 230 1.48 3.17
27 1.00 0.00 1 1 19 240 1.63 9.52
28 0.74 0.26 1 1 14 230 1.58 6.32
29 1.00 0.00 1 1 18 220 1.66 8.57
30 0.84 0.16 1 1 13 235 1.74 7.27

between the means of the selected and rejected seedlings
was calculated and expressed as a percentage of the rejected
seedlings (Table 5). This metric was used to describe and
evaluate the discriminating ability of the ANN models
and the visual method. A large percentage of the difference
between the means of the selected and rejected seedlings
was used as an indicator of greater discriminating ability.
The ANN models produced greater discrimination
between the selected and rejected seedlings than the visual
method (Table 5, Fig. 2). The ANN models were twice (for
the LSU AgCenter population) and 1.5 times (for the USDA
population) more discriminating between the selected and
rejected seedlings than the visual method. The seedlings
selected by the ANN models produced more stalks than
those selected by the visual method. These selected seed-
lings also produced thicker stalks for both populations and
longer stalks for the USDA population.
Further evaluation of the discriminating ability was
done for each of the five families from the LSU AgCenter
population (Table 6). The ANN model produced greater
discrimination between the selected and rejected seedlings
than the visual method for all the families. The seedlings
selected by the ANN model also produced more stalks than
those selected by the visual method. The stalks were thicker
than those selected by the visual method. The magnitude of
the discrimination of the ANN model was greater than that
of the visual method, where the ANN model selected more
seedlings than the visual method, for example, families
XL01-001, XL01-050, XL01-059, and XL01-460 (Table
6). Where the number of seedlings selected was equal, for
example, family XL01-215, the discriminating ability of the
ANN model was very similar to that of the visual method.
Selection Efciency of Articial Neural
Network Models Versus Visual Selection
Improving selection efficiency is a challenge shared by
sugarcane breeders. Selection efficiency is the ability to
discard a seedling that would eventually produce low
cane yield and/or select a seedling that would produce
high cane yield. The number of seedlings selected by
one method and rejected by the other and the number
that performed better or worse than the population mean

CROP SCIENCE, VOL. 51, JANUARY FEBRUARY 2011 WWW.CROPS.ORG 25

Table 4. Probability of selecting [p(Y = 1)] or rejecting [p(Y = 0)] a seedling, the predicted selection decision by the articial
neural network (ANN) model, the selection decision by the visual method (visual), stalk number (Stalks), stalk height (Height),
stalk diameter (Diameter), and seedling cane yield (Cane) for the rst 30 seedlings derived from the USDA population.
Seedling p(Y = 1) p(Y = 0) ANN Visual Stalks Height (cm) Diameter (cm) Cane (kg)
1 0.53 0.47 1 0 19 231 1.95 13.11
2 0.05 0.95 0 0 10 188 1.67 4.12
3 0.08 0.92 0 0 22 168 1.78 9.20
4 0.53 0.47 1 1 20 206 2.07 13.87
5 0.33 0.67 0 0 19 201 1.95 11.41
6 0.05 0.95 0 0 6 188 1.72 2.62
7 0.12 0.88 0 0 15 206 1.73 7.27
8 0.11 0.89 0 0 5 208 1.77 2.56
9 0.75 0.25 1 1 15 224 2.18 12.55
10 0.07 0.93 0 0 4 178 1.83 1.87
11 0.48 0.52 0 0 14 193 2.13 9.63
12 0.91 0.09 1 1 6 239 2.40 6.49
13 0.03 0.97 0 0 11 163 1.70 4.07
14 0.85 0.15 1 1 24 208 2.30 20.75
15 0.29 0.71 0 0 7 229 1.87 4.40
16 0.02 0.98 0 0 11 163 1.67 3.93
17 0.35 0.65 0 1 10 213 1.97 6.49
18 0.02 0.98 0 0 15 191 1.45 4.73
19 0.28 0.72 0 0 20 226 1.78 11.25
20 0.23 0.77 0 0 9 216 1.86 5.28
21 0.13 0.87 0 0 11 203 1.78 5.56
22 0.87 0.13 1 0 5 216 2.43 5.01
23 0.75 0.25 1 0 5 249 2.13 4.44
24 0.14 0.86 0 0 10 224 1.72 5.21
25 0.12 0.88 0 0 2 185 1.93 1.08
26 0.29 0.71 0 0 4 224 1.92 2.60
27 0.11 0.89 0 0 24 211 1.62 10.44
28 0.48 0.52 0 0 2 224 2.08 1.52
29 0.70 0.3 1 1 8 211 2.25 6.71
30 0.61 0.39 1 0 10 229 2.08 7.78

Figure 2. Comparison of mean cane yield (in kg) for the seedlings selected and rejected using visual and articial neural network models
for the Louisiana State University Agricultural Center (LSU AgCenter) (a) and USDA (b) populations.

are shown in Table 7, whereas the mean performance of included more lower-yielding seedlings than the ANN
these seedlings is shown in Table 8. Generally, the visual model, indicating lower selection efficiency of the visual
method rejected more higher-yielding seedlings and method (Table 7, Fig. 3). Seedlings selected by the ANN

26 WWW.CROPS.ORG CROP SCIENCE, VOL. 51, JANUARY FEBRUARY 2011

Table 5. Means for stalk number (Stalks), stalk height (Height), stalk diameter (Diameter), and cane yield (Cane) for seedlings
selected (S) and rejected (R) by the visual and articial neural network selection methods and the means expressed as a
percent of rejected [(S R)/R%] for the Louisiana State University Agricultural Center (LSU AgCenter) and USDA populations.
Selection method
Visual Articial neural network
Population Trait Rejected Selected (S R)/R% Rejected Selected (S R)/R%
LSU AgCenter Stalks 9.74 15.58 60 7.83 14.28 82
Height (cm) 211 228 8 216 219 1
Diameter (cm) 2.17 2.17 0 1.99 2.24 13
Cane (kg) 7.62 12.62 66 5.08 12.01 136
USDA Stalks 12.17 11.89 2 11.89 13.08 10
Height (cm) 207 222 8 205 226 10
Diameter (cm) 1.73 2.13 24 1.70 2.17 27
Cane (kg) 6.02 9.37 56 5.65 10.44 85

Table 6. Difference between the means of the selected and rejected seedlings expressed as a percent of the rejected seed-
lings for the seedlings selected using the visual method (Visual) and the articial neural network (ANN) model for stalk num-
ber (Stalks), stalk height (Height), stalk diameter (Diameter) and cane yield (Cane) and the number of seedlings selected (No.
Selected) for the individual crosses derived from the Louisiana State University Agricultural Center (LSU AgCenter) population.
XL01-001 XL01-050 XL01-059 XL01-215 XL01-460
Trait Visual ANN Visual ANN Visual ANN Visual ANN Visual ANN
Stalks 89 104 72 76 50 88 71 77 45 60
Height (cm) 1 1 9 5 4 9 7 7 0 2
Diameter (cm) 3 6 7 16 2 6 7 6 14 14
Cane (kg) 104 126 115 166 59 144 119 126 73 100
No. Selected 16 21 6 16 10 14 18 18 7 27

Table 7. Number of seedlings that were rejected using arti-

models but rejected by the visual method (Table 8) pro- cial neural network (ANN) (after being selected during visual)
duced 75% (in the LSU AgCenter population) and 51% and selected using ANN (after being rejected during visual),
(in the USDA population) more cane yield than seedlings and numbers rejected (after producing higher cane yield than
population mean) or selected (after producing lower cane
that were rejected by the ANN models but selected by the than population mean) for the Louisiana State University
visual method (Table 8). The ANN model was, however, Agricultural Center (LSU AgCenter) and USDA populations.
less effective when selecting from the USDA population.
Rejected and Selected and
Recognizing that the USDA data provided a poorer fit to Seedling produced higher produced lower
the model (Table 2, Fig. 1) than did the LSU AgCenter numbers cane yield cane yield
data, with experience, one could learn to adjust the thresh- LSU AgCenter
old probability for the ANN model to further improve Rejected 3/57 Visual = 21 Visual = 17
Selected 42/93 ANN = 0 ANN = 17
selection efficiency when dealing with this type of data.
USDA
Rejected 13/46 Visual = 88 Visual = 14
Seedling Cane Yield Increased Selected 20/226 ANN = 79 ANN = 12
with Increasing Probability Value
We investigated the relationship between the estimated increased with probability rankings. The trend for stalk
probability and cane yield value assigned to each seedling. number showed no clear pattern across probability rank-
The ANN output data were ranked in ascending order ings. Stalk number was less important in determining
of probability. The 150 LSU AgCenter seedlings and the cane yield in this population, probably because dry leaves
272 USDA seedlings were divided into 10 groups each. were not stripped off the stalks as was done with the LSU
Group 1 had the lowest probability of selecting a seedling AgCenter population. This made this trait more difficult
and group 10 the highest. The means of each group for and less efficient to appraise.
each trait were calculated. The means (on the y axis) were
plotted against group probability rankings (on the x axis). Articial Neural Network Models Versus
The trends for cane yield and stalk number from the LSU Visual Method at Identical Selection Rates
AgCenter population were very similar and increased Comparison of the ANN model and the visual method at
with probability rankings (Fig. 4). The trends for stalk different selection rates obscured their impact on selection.
height and diameter were less similar to that for cane A more balanced comparison should use identical selec-
yield and marginally increased with probability rankings. tion rates. Therefore, to produce a balanced comparison of
From the USDA population, the trends for diameter and the ANN models and the visual method during seedling
stalk height were very similar to that for cane yield and selection, identical selection rates were used within each

CROP SCIENCE, VOL. 51, JANUARY FEBRUARY 2011 WWW.CROPS.ORG 27

Figure 3. The mean cane yield (in kg) for the seedlings rejected by the articial neural network (ANN) model and selected by the visual
method (Rejected) and seedlings selected by the ANN model and rejected by the visual method (Selected) for the LSU AgCenter and
USDA populations.

Table 8. Means of the rejected and selected, and the difference of the means of selected (S) and rejected (R) expressed as a
percent of rejected [(S R)/R%] for stalk number, stalk height, stalk diameter and seedling cane yield for the Louisiana State
University Agricultural Center (LSU AgCenter) and USDA populations.
LSU AgCenter USDA
Trait Rejected Selected (S R)/R% Rejected Selected (S R)/R%
Stalk number 10.00 12.21 22 11.31 14.65 30
Stalks height (cm) 215 205 4 212 226 7
Stalk diameter (cm) 1.91 2.39 25 1.97 2.12 8
Cane yield (kg) 6.17 10.78 75 7.09 10.74 51

Rejected refers to seedlings selected by the visual method and rejected by the ANN model.

Selected refers to seedlings rejected by the visual method and selected by the ANN model.

population. From the LSU AgCenter population, 57 out
of 150 seedlings (38%) were selected by the visual method
whereas from the USDA population, 46 out of 272 (17%)
were selected. From the LSU AgCenter population, the
ANN model selected 96 out of the 150 seedlings (64%) and
from the USDA population, 53 out of the 272 seedlings
(19%). To produce identical comparisons, the visual selec-
tion rates within each population were used as standard
for the ANN models. The number of the ANN model-
selected seedlings was adjusted to equal that of the visual
method after ranking the probability values. The means of
the highest 38% for the LSU AgCenter population and 17%
for the USDA population were used for the comparison
(Table 9). The seedlings selected by the ANN model pro-
duced 16% (in the LSU AgCenter population) and 8% (in
the USDA population) more cane yield than those selected
by the visual method. The seedlings selected by the ANN
model produced 8% more stalks that were thicker than
those selected by the visual method. The selection rates
achieved in this study (38% for the LSU AgCenter and 17%
for the USDA populations) are atypical of the 5% (at the
LSU AgCenter) and 10% (at the USDA) rates practiced in
the regular breeding programs. In this study seedlings were
selected for cane yield components alone and not for cane
yield components and Brix as is usually the case during rou-
tine seedling selection.
DISCUSSION
The ANN model was superior to visual selection in identi-
fying seedlings with high cane yield potential, as evidenced
by several comparisons between the two selection meth-
ods. For example, the proportion of high yielding seedlings
selected by the ANN model was greater than that selected
by the visual method. This proportion increased when sim-
ilar selection rates were used for both methods. Generally,
seedlings selected by the ANN model produced more stalks
that were thicker and longer than those selected by the visual
method. The visual method rejected a greater proportion of
seedlings that produced estimated cane yields higher than the
population mean compared with the ANN model. A good
number of these seedlings rejected by the visual method
were selected by the ANN model. Conversely, the ANN
model rejected low-yielding seedlings that were selected by
the visual method. Because only a limited numbers of seed-
lings can be advanced to the next stage, the low efficiency of
the visual method would greatly reduce the overall efficiency
of a selection program. The ANN uses fast and automated
computations and was superior to the visual method even for

28 WWW.CROPS.ORG CROP SCIENCE, VOL. 51, JANUARY FEBRUARY 2011

Figure 4. Trends for means of seedling stalk number, stalk height, stalk diameter, and cane yield (in kg) (y axis) plotted against the group
probability rankings (x axis) for the Louisiana State University Agricultural Center (LSU AgCenter) (a) and USDA (b) populations.

Table 9. Means for stalk number, stalk height, stalk diameter and estimated seedling cane yield of seedlings selected by the
articial neural network (ANN) models and the visual method (Visual), and of seedlings selected by the ANN method expressed
as a percent of seedlings selected by the visual method (ANN % Visual) for the Louisiana State University Agricultural Center
(LSU AgCenter) (38% selection rate) and USDA (17% selection rate) populations.
LSU AgCenter USDA
Trait Visual ANN ANN % Visual Visual ANN ANN % Visual
Stalk number 15.58 16.77 108 11.89 12.87 108
Height (cm) 228 225 98 222 228 102
Diameter (cm) 2.12 2.24 106 2.13 2.19 103
Cane yield (kg) 12.62 14.65 116 9.37 10.45 108

a variable dataset with poor model fit, as was the case with
the USDA population. A good aspect of the ANN model is
that as the breeder gains experience, he/she will be in a better
position to recognize data with a poor model fit and adjust
the probability threshold accordingly.
The ANN model selected seedlings based on those traits
that exhibited the largest variability within the population.
Conversely, traits with low variability would be less associ-
ated with the estimated seedling-cane yield and have little
influence in determining the probability value assigned to
each seedling. Therefore, the ability of the ANN model
to use the most genetically variable traits during seed-
ling selection leads to higher selection efficiency than was
the case with the visual method. In this study, the ANN

CROP SCIENCE, VOL. 51, JANUARY FEBRUARY 2011 WWW.CROPS.ORG 29

models selected seedlings that produced more stalks than
visual selection. Research done on early selection stages in
Zimbabwe showed that stalk number was positively associ-
ated with cane yield (Zhou, 2004b). In Louisiana, seed-
lings producing high stalk number are routinely selected to
enhance cane yield and ratooning ability.
Because land is always a limiting resource in most breed-
ing programs, the breeder has little choice but to design the
best allocation of this resource. The ANN model offers the
breeder greater flexibility for adjusting the number of seed-
lings to advance during seedling selection. The breeder can
increase or decrease the number of seedlings to advance by
decreasing or increasing the threshold probability. These
adjustments can be used to refine the selection process using
trait values that can be included in the output, for example,
Brix, disease and insect resistance scores that were not used
in this study to develop the ANN prediction model.
Efficiency is also increased when using the ANN
model. For example, to reduce the number of seedlings
to be advanced using the visual method, the breeder
would have to go back to the field and review all the
selected seedlings and decide on the seedlings to discard.
To increase the number of seedlings to be advanced, the
breeder would have an equally daunting task of physi-
cally reviewing all the rejected seedlings to identify those
seedlings that would have been included. With the ANN
model, this can be easily accomplished by adjusting the
probability threshold and associated trait values.
Applying the ANN model for seedling selection
would require the measurement of variables such as stalk
number, stalk height, stalk diameter, Brix, and disease and
pest resistance. This could be an added cost in breeding
programs, for example, in the United States and Austra-
lia, where these traits are not routinely measured at this
stage of the program. These traits could be measured
before selection on a sample of seedlings from each cross
and the ANN model used as a tool to provide guidelines
as to what selection rates to apply in each cross during
selection. However, in some programs, for example, the
Zimbabwe sugarcane breeding program where the cost
of labor is relatively cheap and mechanical harvesting
of seedlings is sometimes unavailable these variables are
routinely measured as a means to evaluate families at the
seedling stage (Zhou, 2004a). Measurement costs can be
reduced by excluding seedlings that are too inferior and
would probably never be selected. Visual scores for stalk
number, stalk height, and stalk diameter can be used as
input variables. Visual scores would be easier and quicker
to collect but may reduce precision. Scores may be more
useful as a validation tool and their precision may improve
with time as the staff gets more experienced.
The success of the ANN model is highly dependent
on the amount, suitability, and precision of measure-
ments of the training dataset used (Pandolfi et al., 2009).
30 WWW.CROPS.ORG
Pandolfi et al. (2009) noted that the ANN model train-
ing data should capture the variation in the population to
attain the best results. However, they reported that even
when the training data were not statistically representa-
tive of the target population, the ANN models appeared
capable of generalizations beyond the training data and
produced correct results even in different populations.
Pandolfi et al. (2009) applied the ANN model for classifi-
cation of tea accessions. In their study, it was important to
capture the variation in the population. This may not be
entirely necessary when the ANN model is used as a deci-
sion support tool for selection, as the intent is to shift the
variation of the population toward seedlings with higher
cane yield values.
In sugarcane seedling selection, the training data can
be collected from part of the seedling population or from
special populations created from some elite families. This
population would be a reservoir of the ideal trait combina-
tions. With selection, the objective is to shift the popula-
tion toward a desired direction of trait values such as high
cane yield. Therefore, the ideal training data need not
have variability similar to that of the target population.
Rather, the training data should be a population with the
desired combination of trait values that will be mimicked
by the selection process. In this case, the ANN models
provide the added advantage of allowing the breeder to
directionally shift the population toward high cane yield
more objectively than the visual method does.
In summary, the ANN model is a statistical tool that
can be used to increase selection efficiency in a sugar-
cane breeding program. Here, we have demonstrated the
potential of the ANN model as a decision support tool
for selection in sugarcane using the seedling population.
The ANN model was superior to the visual method in
discriminating between seedlings with high and low
cane yield. The magnitude of the difference between the
selected and rejected seedlings was greater for the ANN
model than for the visual method. The magnitude of the
difference increased when similar selection rates were
applied for the visual method and the ANN model and
would have been even larger had we weighed the seed-
lings to estimate cane yield for the training dataset. The
computations in the ANN model are automated by the
SAS software and therefore large numbers of seedlings can
be evaluated quickly. The output in neural network mod-
els provides a decision to select or reject a seedling based
on a threshold probability that is user defined. The ANN
model compels the breeder to consider all traits simultane-
ously when deciding whether to select or rejects a clone
as opposed to judging the merit of a clone by consider-
ing multiple traits independently. Kaul et al. (2005) found
that ANN models consistently produced more accurate
yield predictions than multiple regression models in soy-
beans [Glycine max (L.) Merr.] and corn (Zea mays L.). This
CROP SCIENCE, VOL. 51, JANUARY FEBRUARY 2011
was apparently because in addition to linear relationships,
nonlinear relationships overlooked by other methods can
be determined by ANN models with little a priori knowl-
edge of the functional relationship (Elizondo et al., 1994).
Where there are labor constraints in the breeding pro-
gram, the training data can be collected from part of the
seedling population. The ANN model can then be used
to predict selection rates to be applied in each cross dur-
ing seedling selection as well as to provide trait weighting
insight into the subjective selection process.
Acknowledgments
This research was supported by a grant from The American
Sugar Cane League of the USA, Inc. The sugarcane breed-
ing staff at the USDA and LSU AgCenter provided much
needed technical assistance during the field work. Two anony-
mous reviewers provided critical reviews as well as tremendous
insights and input that helped to strengthen the manuscript. All
are gratefully acknowledged.
References
Agresti, A. 2007. An introduction to categorical data analysis. 2nd
ed. John Wiley & Sons, Hoboken, NJ.
Akaike, H. 1974. A new look at the statistical model identification.
IEEE Trans. Automatic Control AIC 19:716723.
Allison, P.D. 2003. Logistic regression using the SAS system: The-
ory and applications. SAS Inst., Cary, NC.
Batchelor, W.D., X.B. Yang, and A.T. Tschanz. 1997. Develop-
ment of a neural network for soybean rust development.
Trans. ASAE 40:247252.
Casella, G., and R.L. Berger. 2003. Statistical inference. 2nd ed.
Thomson Publications, Belmont, CA.
Chang, Y.S., and S.B. Milligan. 1992. Estimating the potential of
sugarcane families to produce elite genotypes using univariate
cross prediction methods. Theor. Appl. Genet. 84:662671.
De Sousa-Vieira, O., and S.B. Milligan. 1999. Intra-row spacing
and family x environment effects on sugarcane family evalua-
tion. Crop Sci. 39:358364.
Elizondo, D.A., R.W. McClendon, and G. Hoogenboom. 1994.
Neural network models for predicting flowering and physi-
ological maturity of soybean. Trans. ASAE 37:981988.
Fausett, L. 1994. Fundamentals of neural networks. Prentice Hall,
Englewood Cliffs, NJ.
Gravois, K.A., S.B. Milligan, and F.A. Martin. 1991. Indirect
selection for increased sucrose yield in early sugarcane testing
stages. Field Crops Res. 26:6773.
Gurney, K. 1997. An introduction to neural networks. UCL Press,
London, UK.
Hertz, J., R.G. Palmer, and A.S. Krogh. 1990. Introduction to the
theory of neural computation. Perseus Books, New York, NY.
Huang, S., K.K. Tan, and K.Z. Tang. 2004. Neural network: The-
ory and applications. Research Studies Press Ltd. Baldock,
Hertfordshire, UK.
Iglesias, A., B. Arcay, and J.M. Cotos. 2006. Connectionist sys-
tems for fi shing prediction. p. 265296. In J.R. Rabunal and
J. Dorado (ed.) Artificial neural networks in real-life applica-
tions. Idea Group Publishing, London, UK.
CROP SCIENCE, VOL. 51, JANUARY FEBRUARY 2011 WWW.CROPS.ORG
Ji, B., Y. Sun, S. Yang, and J. Wan. 2007. Artificial neural net-
works for rice yield prediction in mountainous regions. J.
Agric. Sci. 145:249261.
Kang, M.S., J.D. Miller, and P.Y.P. Tai. 1983. Genetic and phe-
notypic path analysis and heritability in sugarcane. Crop Sci.
23:643647.
Kang, M.S., P.Y.P. Tai, and J.D. Miller. 1991. Genetic and pheno-
typic path analyses in sugarcane: Artificially created relation-
ships. Crop Sci. 31:16841686.
Kaul, M., R.L. Hill, and C. Walthall. 2005. Artificial neural net-
works for corn and soybean yield prediction. Agric. Syst.
85:118.
Lee, K.C., and T.Y. Paik. 2006. A neural approach to cost mini-
mization in a production scheduling setting. p. 297313. In
J.R. Rabunal and J. Dorado (ed.) Artificial neural networks
in real-life applications. Idea Group Publishing, London, UK.
Masters, T. 1993. Practical neural network recipes in C++. Aca-
demic Press, San Diego, CA.
Miller, J.D., and N.I. James. 1974. The influence of stalk density
on cane yield. Proc. Int. Soc. Sugarcane Technol. 15:177184.
Milligan, S.B., K.A. Gravois, K.P. Bischoff, and F.A. Martin. 1990.
Crop effects on broad sense heritabilities and genetic vari-
ances of sugarcane yield components. Crop Sci. 30:344349.
Mugnai, S., C. Pandolfi, E. Azzarello, E. Masi, and S. Mancuso.
2008. Camellia japonica L. genotypes identified by an artificial
neural network based on phyllometric and fractal parameters.
Plant Syst. Evol. 270:95108.
Nelson, M.N., and W.T. Illingworth. 1991. A practical guide to neu-
ral nets. Addison-Wesley Publishing Company, Reading, MA.
Pandolfi, C., S. Mugnai, S. Bergamasco, E. Masi, and S. Mancuso.
2009. Artificial neural networks as a tool for plant identifica-
tion: A case study on Vietnamese tea accessions. Euphytica
166:411421.
Perkins, R., and A. Brabazon. 2006. Predicting credit ratings
with a GA-MLP hybrid. p. 220237. In J.R. Rabunal and J.
Dorado (ed.) Artificial neural networks in real-life applica-
tions. Idea Group Publishing, London, UK.
SAS Institute. 2007. The SAS system for Windows version 9.1.3.
SAS Inst., Cary, NC.
Schaap, M., and W. Bouten. 1996. Modeling water retention
curves of sandy soils using neural networks. Water Resour.
Res. 32:30333040.
Schwarz, G. 1978. Estimating the dimension of a model. Ann.
Stat. 6:461464.
Sethuraman, J. 2006. Soft computing approach for bond rating
prediction. p. 202219. In J.R. Rabunal and J. Dorado (ed.)
Artificial neural networks in real-life applications. Idea Group
Publishing, London, UK.
Smith, M. 1993. Neural networks for statistical modeling. Van
Nostrand Reinhold, New York, NY.
Starrett, S.K., S.K. Starrett, and G.L. Adams. 1997. Using artifi-
cial neural networks and regression to predict percentage of
applied nitrogen leached under turfgrass. Commun. Soil Sci.
Plant Anal. 28:497507.
Zhou, M.M. 2004a. Strategies for variety selection in the breeding
program at the Zimbabwe Sugar Association Experiment Sta-
tion. Proc. South Afric. Sugar Technol. Assoc. 78: 125131.
Zhou, M.M. 2004b. Stalk population control of yield, quality and
agronomic traits of sugarcane population in early selection
stages. Sugar Cane Int. 22:1420.
31