全 文 :植物保护学报 Journal of Plant Protection, 2016, 43(3): 353 - 361 DOI: 10 13802 / j. cnki. zwbhxb. 2016 03 001
基金项目:山西省科技攻关项目(20120311013⁃4),山西省留学基金(2013⁃重点 6)
∗通讯作者(Author for correspondence), E⁃mail: sxaulisc@ 126. com
收稿日期: 2015 - 08 - 01
Comparison of occurrence periods of wheat aphids based
on artificial neural network and wavelet neural
network prediction systems
Jin Ran Li Shengcai∗
(College of Agriculture, Shanxi Agricultural University, Taigu 030801, Shanxi Province, China)
Abstract: To build a more accurate and stable prediction model for insect pests, take precautions against
insect pests, reduce the crop loss and increase crop yields and quality, eight new independent variables
were developed from 42 basic meteorological factors by using the method of principal component analysis
and prediction models in this study. Key parameters were selected with cut⁃and⁃trial method. Wavelet
Neural Network (WNN) model was established with Morlet wavelet function as the transfer function and
compared with Back Propagation Neural Network (BPNN) model with Sigmoid function as the transfer
function. In WNN, the training data from six out of ten years showed a fitting precision of more than 90%
with an average fitting precision of 89% . The predicted mean absolute percentage error (MAPE) and
mean square error (MSE) value were 4 1939 and 5 9764, respectively. In BPNN, the fitting precision
was above 90% for four of ten years with an average precision rate of 81 07% . The predicted MAPE and
MSE values were 6 4694 and 8 2457, respectively. The comparison results of different models showed
that WNN more accurately described the developed pattern of wheat aphids in the field and displayed
better fitting ability than BPNN. Besides, WNN possessed stronger prediction accuracy and stability than
BPNN.
Key words: wheat aphid; Wavelet Neural Network; Back Propagation Neural Network; occurrence
period; prediction
基于小波神经网络和 BP神经网络的麦蚜发生期预测对比
靳 然 李生才∗
(山西农业大学农学院, 太谷 030801)
摘要: 为建立更准确、稳定的病虫害预测预报模型,减少农作物病虫害损失、提高农作物产量与质
量,运用主成分分析法从 42 个基础气象因子中整合形成 8 个新的自变量输入模型,采用试凑法对
网络关键参数进行筛选,用 2002—2011 年数据进行网络训练,建立了以 Morlet 小波函数为传递函
数的小波神经网络模型,并与以 Sigmoid函数为传递函数的 BP神经网络模型进行了比较。 在小波
神经网络训练过程中,有 6 年拟合精度在 90%以上,平均拟合精度为 89% ,预测结果 MAPE 值为
4 1939,MSE值为 5 9764;在 BP神经网络的训练过程中,有 4 年拟合精度超过 90% ,平均拟合精度
仅为 81 07% ,预测结果中 MAPE值为 6 4694,MSE值为 8 2457。 从训练结果看,小波神经网络更
能准确描述麦蚜发生期的变化规律,其拟合能力较 BP神经网络好;从预测精度和模型的稳定性来
看,小波神经网络好于 BP神经网络。
关键词: 麦蚜; 小波神经网络; BP神经网络; 发生期; 预测
During the past decades, worldwide entomologists
have committed to develop various methods for insect
pest prediction, including the prediction systems based
on personal experience, experiment data or statistic
analysis results (Zhang et al. ,1985). Due to the fact
of the complex features of insect pest outbreaks, such
as unevenness, otherness, diversity, abruptness, and
periodicity, traditional linear prediction methods are
not able to deliver desirable prediction results. There⁃
fore, modern non⁃linear theories were employed into
pest prediction. Coupled with traditional Kinetic Theo⁃
ry, mathematical statistics and modern computer tech⁃
nology, novel pest prediction methods have been devel⁃
oped, including Artificial Neural Network ( ANN),
phase⁃space reconstruction, Wavelet Analysis (WA)
and Support Vector Machine ( SVM) ( Ma et al. ,
2002;Yuan et al. ,2008).
ANN is developed by mimicking the structure of
human brain with a physical abstraction and simplifica⁃
tion, and works effectively in multiple fields including
information processing, model identification, biological
signal monitoring and analysis, disease diagnosis, mar⁃
ket price prediction, risk evaluation, automation con⁃
trol and transportation ( Kialashaki & Reisel, 2014;
Nyhan et al. ,2014).
WA shows good quality in temporal frequency and
variable focal length. The need⁃based adjustments can
be done with temporal frequency window which has been
successfully applied in signal and image compression,
engineering technology and signal analysis (Mostafapour
et al. ,2014;Moya⁃Martínez et al. ,2015). Combining
wavelet transformation theory and ANN, Zhang & Ben⁃
veniste (1992) proposed a new neural network, WNN
in which wavelet function was used as the transfer func⁃
tion (Zhang & Benveniste, 1992;Bakshi & Stephanopo⁃
ulos,1993). Thus a new area was developed for re⁃
search in the field of neural network and has been ap⁃
plied in signal processing, image compression, model i⁃
dentification and system identification ( Falamarzi et
al. ,2014;Taghavifar & Mardani,2014).
So far, ANN has been applied in insect pest pre⁃
diction, while forecasting pest occurrence with WNN
was just began in both China and other parts of the
world (Li & Peng,1999;Wang et al. ,2011). The ob⁃
jectives of this study are to develop and verify the prac⁃
tical WNN models by predicting occurrence period of
wheat aphid in the crop fields and to compare the ad⁃
vantages and disadvantages of WNN with BPNN models
in terms of pest prediction.
1 Materials and Methods
1 1 Data source
Since 1990s, wheat aphid has been ranked the
third important crop pest in terms of damage area and
yield loss in China (Cao et al. ,2006). And in south⁃
ern Shanxi Province, wheat aphid is also one of the
most danger insect pests in the fields. Therefore, the
wheat aphid was selected as target insect, and related
emerging data of the insect were obtained from Shanxi
Provincial Station of Plant Protection and Quarantine.
The main sampling site was located in Beiguan
Village, Guwei Town, Ruicheng County of Yuncheng
City, Shanxi Province (34°36′ - 48°30′ N, 110°36′ -
42° 30′ E ). The annual average temperature of
12 77℃ with 250 frost⁃free days and annual rainfall of
513 mm. As the major place of wheat production, the
town is plain and fertile with around 5 000 hm2 of ara⁃
ble land.
Data collection: data were collected from five
sampling sites along a diagonal line with 50 plants of
each site. When the number of aphids exceeds 500 for
each hundred plants, the number of plants for each site
was decreased to 20. The number of plants damaged by
aphids, the species and the number of aphids were in⁃
vestigated in designated locations regularly with the
same method.
Scope of data: data was obtained during February
2002 and June 2014, and was collected every four days
between late February and early June of each year.
Meteorological data: the meteorological data used
453 植 物 保 护 学 报 43 卷
in this study was obtained from Shanxi Meteorological
Bureau. As wheats grow and the temperature rises in
April, wheat aphids are less likely to die due to climat⁃
ic conditions. Thus the density of wheat aphids per
plant increases with growing occurrence of wheat
aphids. Entering May, with wheats heading, flower⁃
ing, postulating, the climatic conditions are more suit⁃
able for the occurrence of wheat aphids, leading to the
dramatic growth of its intensity. Therefore, with the
purpose of building a short⁃term wheat aphid prediction
model, the following singular meteorological indicators
were collected based on the meteorological factors of
every April between 2002 and 2014, including five⁃day
average temperature, highest temperature and lowest
temperature, average five⁃day relative humidity, pre⁃
cipitation, sunlight hours and wind speed (Table 1).
Table 1 Comparison of meteorological factor in this study
Date
Average
temperature
(℃)
Average
highest
temperature
(℃)
Average
lowest
temperature
(℃)
Average
relative
humidity
(% )
Average
precipi⁃
tation
(mm)
Average
sunshine
duration
(h)
Average
wind
speed
(m / s)
4 01—4 05 x1 x2 x3 x4 x5 x6 x7
4 06—4 10 x8 x9 x10 x11 x12 x13 x14
4 11—4 15 x15 x16 x17 x18 x19 x20 x21
4 16—4 20 x22 x23 x24 x25 x26 x27 x28
4 21—4 25 x29 x30 x31 x32 x33 x34 x35
4 26—4 30 x36 x37 x38 x39 x40 x41 x42
1 2 Modeling of neural network
1 2 1 Neural network and its prediction procedure
Modelling after biological neural networks, ANN
is a physical abstraction, simplification and simulation
of human brains and has strong information processing
capability in non⁃linear ways. And artificial neurons
are formed into networks through different means of
connection. Neural network has four basic properties of
non⁃linearity, generalizing from limited information,
adaptability to environment changes and convexity.
With good self⁃adaptability, self⁃organization and self⁃
learning ability, it uses a parallel distributive system,
overcoming the flaws of traditional logic symbols⁃based
artificial intelligence in dealing with intuitive and non⁃
structural information. There are various types of neu⁃
ral networks. The most commonly used is BPNN.
WA was first put forward by French engineer Jean
Morlet in 1974. It is able to extract information effec⁃
tively in time and spatial transformations, making de⁃
tailed multi⁃dimensional analysis of signals in the scale
and shift algorithms. In a permanent window where the
time and frequency windows can change, the low⁃fre⁃
quency part has narrow frequency window but wide
time window while the opposite is true with high⁃fre⁃
quency part. There are two structures for neural net⁃
works based on wavelet analysis. One is loose⁃type
WNN which means to preprocess signals with wavelet
analysis before the output of signals through ordinary
neural networks. The other is compact⁃type WNN,
which is a widely⁃applied structure that replace hidden
functions in normal neural networks with wavelet func⁃
tions, the weights between inputs and hidden layer
with the scales of wavelet functions, and threshold val⁃
ues of hidden layer with the translator parameters of
wavelet functions (Fu et al. ,2010).
This study took the wheat aphid occurrence peri⁃
ods of 2002—2011 as the training set and those of
2012—2014 as the test set. The prediction procedures
were as shown in Fig. 1. In writing neural network pro⁃
grams with MATLAB Starter Application, it took the
following modeling procedures, data normalization,
network training, network prediction, error analysis
and result construction. WNN and BPNN prediction
models were built respectively with three layers of
structures including an input layer, a hidden layer and
an output layer.
1 2 2 Selection of meteorological factors with princi⁃
ple component analysis (PCA)
Selected meteorological factors determine the ac⁃
curacy of pests and disease prediction. The independ⁃
5533 期 Jin Ran, et al. : Comparison of occurrence periods of wheat aphids based on different prediction systems
Fig. 1 The morphological characterization of pathogen
causing kiwifruit rot disease
ent variables must be highly related to the occurrence
period of wheat aphid. With PCA, a correlation matrix
was gained via data matrix formed with original multidi⁃
mensional input variables. Accumulated variance con⁃
tribution rate was obtained based on the proper value of
the correlation matrix, then principal components were
identified based on the proper vector of the correlation
matrix.
There were as many as 42 annual meteorological
factors as the input variables of neural network in this
study. Given the correlation among factors, including
them all as the input variables may lead to the overuse
of information, increasing the complexity of neural net⁃
work model, prolonging training process, reducing
learning efficiency and weakening generalization per⁃
formance. Therefore, in this paper, PCA was used to
screen meteorological factors of collinear nature on
SPSS platform and new component factors were selected
for feeding into the model (Liu et al. ,1997).
1 2 3 Processing data with normalization method
As different component factors have different di⁃
mensions, neural network is most sensitive to data
within [0, l] (Ye & Wei,2015). Thus, component
factors were processed into the range of [0, l] before
modeling. The formula is as follows: x = (xi - xmax) /
(xmax - xmin). In the equation, xi is the value of origi⁃
nal data, x is the normalized data, xmax and xmin repre⁃
sents the maximum and minimal value of each compo⁃
nent factor respectively.
1 2 4 Screening of parameters with cut⁃and⁃trial method
There are various parameters for ANN, such as
input layer nodes, hidden layer nodes, output layer
nodes, transfer functions, training functions and learn⁃
ing rates etc. (Liu et al. ,2006). When the network
structure, weights and threshold value are the same,
the proper number of hidden layer nodes, transfer
functions, training functions directly affects the learn⁃
ing and generalization ability of the network.
The biggest difference between WNN and BPNN
lies in the difference in transfer functions. WNN takes
Morlet wavelet function as its transfer function. It is a
sine function of the single frequency ratio under Gauss
function (Chen & Feng,1999), C is the normalized
constant, which is defined as follows: ψ( t) = Ce - ( t^2) / 2
cos(5x). Morlet wavelet function is not one of the
transfer functions of MATLAB. Custom transfer func⁃
tions must be created. The preparation of the two sub⁃
routines are: (1) function y = mymorlet( t), y = exp
( - ( t. ^2) / 2)∗cos(1 75∗t); (2) function y = d_
mymorlet( t), y = - 1 75∗sin(1 75∗t). ∗exp( -
( t. ^2) / 2) - t∗cos(1 75∗t). ∗exp( - ( t. ^2) / 2);
Normally, Sigmoid is chosen as the transfer function of
BPNN, which includes three major types of functions,
logsig, tansig, and purelin. Sigmoid function is smooth
and differentiable, more accurate than linear functions
with better fault tolerance. It’s defined as follows: in⁃
creasing the number of nodes in the hidden layer can
effectively improve the training accuracy of neural net⁃
work, but too many hidden layer nodes can also reduce
the training speed. Consult the following formula for
the number of hidden layer neuron nodes: (1) l =
| mn | ( Li et al. ,2006); (2) l < | m + n | + a
(Chen,2009); (3) l = 2n + 1(Wang et al. ,2012);
(4) l = log2n; (5) l < n - 1(Wang et al. ,2013).
The learning rate determines the variation of weights
generated in each circuit training. An overly high
learning rate leads to reduced stability of the model
while an overly low learning rate leads to prolonged
653 植 物 保 护 学 报 43 卷
training time and reduced rate of convergence. There⁃
fore, in selecting learning rate, the range was set be⁃
tween 0 01 and 0 8 (Feng,2007).
In this study, optimal parameters are selected for
hidden layer nodes, transfer functions, and learning
rates with cut⁃and⁃trial method. The best original value
combination was gained after screening while other pa⁃
rameters were achieved based on experience.
1 2 5 Model verification method
In this study, model performances are assessed
with MAPE and MSE indicators (Xiang & Zhou,2010;
Li et al. ,2013).
MAPE = 1n
n
i = 1
| (yi - y^i) / yi | ; MSE =
1
n
n
i = 1
(yi - y^i) 2, in the formula, yi is the actual value in oc⁃
currence period, y^i is the predicted value by the mod⁃
el, and n is the training sample number. Lower MAPE
value means higher precision of prediction by the model
and lower MSE value means more stable prediction.
2 Results and Analysis
2 1 Results of PCA
The accumulative percentages were gained by ana⁃
lyzing 42 independent variables in Table 1 with PCA
method. The accumulative contribution rate of the first
eight components was 91 851% . These eight new
components included basically the information of all
factors that were not correlated. The accumulative con⁃
tribution rate of the rest 34 components was only
8 149% and may considered noise items and not be
included into the modeling process.
Take April 10th as the number “1” and the daily
increment as “1”, by that analogy, May 1st would be
referred to as the “22”. The new sample set compri⁃
sing the new component factors z1 - z8 based on compo⁃
nent matrix and occurrence period of wheat aphid were
shown in the Table 2.
Table 2 Sample set of the occurrence period of wheat aphid after PCA
Year z1 z2 z3 z4 z5 z6 z7 z8 y
2002 1 1365 - 0 5356 - 2 1458 0 0961 0 2019 1 6742 - 0 9562 - 0 8976 31
2003 - 0 9930 - 1 5111 0 0948 - 0 0365 2 2906 - 0 6515 - 0 7314 0 2136 51
2004 1 4852 1 1312 - 0 1054 0 8580 0 6623 - 0 9882 - 0 7980 1 8989 21
2005 0 6536 1 0559 0 5340 - 2 0150 0 4239 0 6693 0 4660 0 3745 41
2006 - 0 2646 0 7190 - 0 2178 - 1 2140 0 8783 - 1 0655 0 6522 - 1 6553 31
2007 0 5356 0 2172 0 2892 1 0089 - 0 6614 - 0 0830 0 1917 - 1 1282 26
2008 - 1 1223 0 3928 1 4415 - 0 5028 - 0 5857 1 5542 - 1 6766 0 3558 36
2009 - 0 3323 - 0 1469 0 5937 1 4932 0 1493 - 0 0061 - 0 1047 - 0 3949 26
2010 - 1 0291 0 2831 - 0 7556 - 0 2376 - 1 6931 - 1 6492 - 1 1684 - 0 3463 40
2011 1 0826 - 0 8308 1 7453 0 5666 - 0 2479 0 0899 0 7843 - 0 7889 50
2012 0 1991 0 7949 - 0 3559 0 1045 - 0 3191 - 0 2028 0 7419 0 2742 40
2013 0 3681 - 2 1378 - 0 2757 - 1 0155 - 1 2152 - 0 3049 0 7975 1 2672 36
2014 - 1 7195 0 5680 - 0 8422 0 8941 0 1161 0 9635 1 8018 0 8269 46
2 2 Results of model parameters screening
2 2 1 Results of WNN parameters screening
On the basis of hidden layer nodes formula calcu⁃
lation, chose n = 8 (input layer node number), m = 1
(output layer node number), a was between [1, 10],
the number of hidden layer nodes of WNN was adjusted
based on experience and tested between 1 and 20. The
test suggested that when the number of hidden layer
nodes was 15, the value of MSE and MAPE hit the
lowest level (Table 3).
The learning rate was set between 0 01 and 0 8,
and Lr1 and Lr2 were paired. When Lr1 = 0 1, Lr2
was set within the range of 0 01 - 0 09 and 0 1 - 0 8
respectively for testing. when Lr2 was 0 1, the values
of MSE and MAPE were the lowest, which meant the
best value for Lr2 was 0 1 when Lr2 = 0 1, Lr1 was
tested between 0 01 - 0 09 and 0 1 - 0 8 respective⁃
ly. When Lr1 was 0 3, the lowest MSE and MAPE
values were obtained. Thus the best Lr1 value was 0 3
(Table 4).
7533 期 Jin Ran, et al. : Comparison of occurrence periods of wheat aphids based on different prediction systems
Table 3 Forecast errors corresponding to different
hidden layer nodes in WNN
Node number of
hidden layer neurons MSE MAPE (% )
1 10 5113 12 1055
2 30 1390 10 5113
3 348 7269 43 0379
4 53 2544 15 6134
5 168 8766 22 1060
6 20 4457 10 8626
7 85 2218 18 8572
8 125 3540 24 8216
9 169 1621 30 8143
10 414 0848 45 0219
11 452 6489 43 3986
12 496 5482 53 9865
13 211 9083 26 0557
14 50 7474 16 2135
15 16 5651 9 5909
16 86 1652 19 5693
17 107 4177 23 9009
18 68 0656 19 4685
19 476 3302 48 8502
20 512 1153 52 0258
After selection, three⁃layer WNN structure 8⁃15⁃1
was built up with the learning rate Lr1 = 0 3, Lr2 =
0 1, iterations of 150 and error accumulation of 0.
2 2 2 Results of BPNN parameters screening
The three major types of Sigmoid functions were
paired consecutively to obtain the corresponding pre⁃
diction errors of different transfer functions. Shown the
lowest values for both MSE and MAPE in the test, the
logsig⁃logsig combination was selected as the transfer
function for both the hidden and input layers.
Testing learning rate between 0 01 - 0 09 and
0 1 - 0 8 respectively found that the MSE and MAPE
values of BPNN were the lowest when the learning rate
was 0 02. It was found that in selection of the hidden
layer node numbers when the node number was 20, the
MSE and MAPE values of BPNN hit the lowest level.
After selection and empirical evaluation, the following
data were selected: 8⁃20⁃1 as the structure of BPNN;
logsig as the transfer function of both the hidden and
output layers, trainlm as the training function, learng⁃
dm as the weight learning function, 0 9 as factor of
momentum, 7 as training frequency, Lr = 0 02 as
Table 4 Forecast errors corresponding to different
training rates in WNN
Number Lr1 Lr2 MSE MAPE (% )
1 0 10 0 01 826 9894 66 3142
2 0 10 0 02 590 4091 49 5207
3 0 10 0 03 390 8248 42 7918
4 0 10 0 04 157 9846 31 0323
5 0 10 0 05 89 1868 23 2455
6 0 10 0 06 210 6638 32 0327
7 0 10 0 07 784 7746 58 6597
8 0 10 0 08 652 6718 62 3337
9 0 10 0 09 191 1578 32 3730
10 0 10 0 10 29 8538 10 4489
11 0 10 0 20 153 3896 30 1089
12 0 10 0 30 100 5360 21 7400
13 0 10 0 40 292 6553 42 4632
14 0 10 0 50 62 6902 18 7671
15 0 10 0 60 269 0985 30 3073
16 0 10 0 70 171 2278 29 0347
17 0 10 0 80 56 0535 18 3636
18 0 01 0 10 32 3502 12 8018
19 0 02 0 10 464 9620 50 5647
20 0 03 0 10 467 2970 53 8074
21 0 04 0 10 261 7130 37 7740
22 0 05 0 10 158 5949 25 3539
23 0 06 0 10 707 6330 53 4514
24 0 07 0 10 731 0330 61 9555
25 0 08 0 10 436 2818 50 1377
26 0 09 0 10 112 3579 23 5084
27 0 10 0 10 29 8538 10 4489
28 0 20 0 10 49 6233 17 3152
29 0 30 0 10 5 9764 4 1939
30 0 40 0 10 469 8798 51 9498
31 0 50 0 10 51 3823 16 7606
32 0 60 0 10 293 8904 33 5394
33 0 70 0 10 273 3855 40 5395
34 0 80 0 10 164 1060 28 0749
learning rate, 1 × 10 - 5 as the lowest expected target
error.
2 3 Fitting results of occurrence periods of wheat
aphid during 2002—2011
The occurrence periods of wheat aphid of 2001—
2011 were trained using WNN, the fitting precision
was above 90% in six of the years with an average pre⁃
cision rate of 89% (Fig. 2). In BPNN training, four
years showed a fitting precision of 90% with an average
precision of 81 07% . The test demonstrated that WNN
was better at describing the change pattern of occur⁃
853 植 物 保 护 学 报 43 卷
Fig. 2 The simulation effect of study of WNN and BPNN
rence period of wheat aphid with better fitting capabili⁃
ty than BPNN.
2 4 Occurrence period of wheat aphid prediction
between 2012 and 2014
The following can be learnt by comparing the pre⁃
diction accuracy and performance of WNN and BPNN:
WNN being 95 81% and BPNN being 93 53% . A
contrast of both model ’ s MSE suggested the MSE
value of WNN ( 5 9764 ) was smaller than BPNN
(8 2457 ). The result demonstrated that WNN was
better than BPNN in the accuracy and stability of pre⁃
diction (Table 5).
Table 5 Results of occurrence period of wheat aphid prediction based on PCA⁃WNN and PCA⁃BPNN
Year Actualvalue
PCA⁃WNN PCA⁃BPNN
Predicted
value
Error
absolute
value
Fitting
precision
(% )
Predicted
value
Error
absolute
value
Fitting
precision
(% )
2012 40 00 36 77 3 23 91 93 36 00 4 00 90 00
2013 36 00 36 13 0 13 99 65 38 89 2 89 91 98
2014 46 00 44 09 1 91 95 84 46 64 0 64 98 61
Average prediction accuracy 95 81 93 53
3 Discussion
The occurrence and development of crop pests are
complicated and non⁃linear in the field. By using the
principal component analysis, WNN and BPNN models
were built based on the monitoring data of the popula⁃
tion of wheat aphid occurred during different period of
time in Shanxi Province. The prediction accuracy and
stability of two models were compared. The result indi⁃
cated that the average prediction accuracy of both mod⁃
els was greater than 90% , and WNN showed better
prediction accuracy and stability than that of BPNN.
WNN also showed the property of temporal frequency
simultaneous analysis of wavelet analysis and the capa⁃
bility of self⁃learning, self⁃organization and non⁃linear
mapping of the neural network. WNN model has poten⁃
tial to be used as a novel method for insect pest predic⁃
tion in the field by gradually optimizing the influence of
insect pest factors on the model such as overwintering
population, natural enemy types, pest control strategy,
and so on.
Compared with the study of the predicated poten⁃
tial of BPNN with cotton aphid, Aphis gossypii Glover,
on sugarcane (Ou et al. ,2008) and the study of WNN
predication activity on egg laying peak of corn ear⁃
worm, Helicoverpa armigera ( Hübner) ( Zhu et al.
2010), the present study not only determined the pre⁃
diction accuracy of both BPNN and WNN, but also op⁃
timized the modeling procedure, and verified the ad⁃
vantage of the WNN.
9533 期 Jin Ran, et al. : Comparison of occurrence periods of wheat aphids based on different prediction systems
The reasons of WNN model possessing better
predicated potential of the occurrence period of wheat
aphid were, firstly, prior to build a neural network,
the input variables were preprocessed with PCA meth⁃
od, minimized the interplay effect of associated factors,
and reduced the dimensions from 42 to eight, which ef⁃
fectively avoided the dimension disaster caused by high
dimensions. Secondly, the node numbers of the hidden
layer were screened. The structures of 8⁃15⁃1 for WNN
and 8⁃20⁃1 for BPNN were set up, in which WNN had
15 nerve cells and BPNN had 20. The simpler struc⁃
ture of WNN provided the shortened training time
which ensured optimum points accessible and avoided
falling into the local minimum points and the over⁃fit⁃
ting phenomenon during network training, by which the
predication accuracy of WNN was effectively improved.
Thirdly, WNN chose the common wavelet function
Morlet as the transfer function, while BPNN used
logsig as the transfer function of the hidden and output
layer. The former had better capability in extracting
non⁃linear qualities, better Robustness, lesser error,
and better generalization ability to cope with the inter⁃
ferences. Finally, WNN showed the capability to effec⁃
tively extract and modify the key information of signals
by multi⁃dimensional analysis of signals with scale ex⁃
pansion factors and time shift factors.
References
Bakshi BR, Stephanopoulos G. 1993. Wave⁃net: a multiresolution,
hierarchical neural network with localized learning. AIChE
Journal, 39(1): 57 - 81
Cao YZ, Yin J, Li KB, Zhang KC, Li XQ. 2006. Discussion on
wheat aphids continuously rampant reasons and control counter
measures. Plant Protection, 32(5): 72 - 75 (in Chinese)
Chen Z, Feng TJ. 1999. Research development and prospects of
wavelet neural networks. Journal of Ocean University of Qingd⁃
ao, 29(4): 663 - 668 (in Chinese)
Chen ZG. 2009. Application of genetic algorithm and BPNN in GDP
forecasts. Computer and Digital Engineering, 37 (9): 172 -
175 (in Chinese)
Falamarzi Y, Palizdan N, Huang YF, Lee TS. 2014. Estimating
evapotranspiration from temperature and wind speed data using
artificial and wavelet neural networks (WNNs). Agricultural
Water Management, 140: 26 - 36
Feng ZY. 2007. The comparison and study based on Wavelet neural
network and BPNN. Master Thesis. Chengdu: Chengdu
University of Technology, pp. 26 (in Chinese)
Fu HX, Zhao H, Wang YC, Song HY. 2010. Application design of
neural network on MATLAB. Beijing: China Machine Press,
pp. 193 - 208 (in Chinese)
Kialashaki A, Reisel JR. 2014. Development and validation of arti⁃
ficial neural network models of the energy demand in the indus⁃
trial sector of the United States. Energy, 76: 749 - 760
Li QQ, Wang CQ, Zhang WJ, Yu Y, Li B, Yang J, Bai GC, Cai
Y. 2013. Prediction of soil nutrients spatial distribution based
on neural network model combined with goestatistics. Chinese
Journal of Applied Ecology, 24(2): 459 - 466 (in Chinese)
Li WC, Song DM, Chen B. 2006. Artificial neural network based on
genetic algorithm. Computer Engineering and Design, 27(2):
316 - 318 (in Chinese)
Li ZY, Peng LH. 1999. Prediction model of agricultural plant disea⁃
ses and insect pests based on artificial neural network and its
verification. Acta Ecologica Sinica, 19(5): 759 - 762 (in Chi⁃
nese)
Liu G, Yang XH, Ge YB, Miao YX. 2006. An artificial neural net⁃
work⁃based expert system for fruit tree disease and insect pest
diagnosis. Proceedings of the 2006 IEEE International Confer⁃
ence on Networking, Sensing and Control. Florida: IEEE Pub⁃
lisher, pp. 1076 - 1079
Liu LF, Feng DY, Zhang HC. 1997. Prediction of amount of wheat
aphid occurrence by application of principal component analy⁃
sis. Entomological Knowledge, 34 (5): 260 - 263 ( in Chi⁃
nese)
Ma F, Xu XF, Zhang XL, Cheng XN. 2002. A prediction model in⁃
tegrating phase space reconstruction and artificial neural net⁃
work and its application in pest forecasting. Acta Ecologica Sin⁃
ica, 22(8): 1297 - 1301 (in Chinese)
Mostafapour A, Davoodi S, Ghareaghaji M. 2014. Acoustic emission
source location in plates using wavelet analysis and cross time
frequency spectrum. Ultrasonics, 54: 2055 - 2062
Moya⁃Martínez P, Ferrer⁃Lapeña R, Escribano⁃Sotos F. 2015. In⁃
terest rate changes and stock returns in Spain: a wavelet analy⁃
sis. Business Research Quarterly, 18: 95 - 110
Nyhan M, McNabola A, Misstear B. 2014. Evaluating artificial neu⁃
ral networks for predicting minute ventilation and lung deposited
dose in commuting cyclists. Journal of Transport & Health, 1:
305 - 315
Ou ZR, Tan ZK, Su YX. 2008. Study on the meteorological grade
for the development of sugarcane aphid in Guangxi with the BP
063 植 物 保 护 学 报 43 卷
neural network. Journal of Anhui Agricultural Sciences, 36
(21): 9141 - 9143, 9152 (in Chinese)
Taghavifar H, Mardani A. 2014. Wavelet neural network applied for
prognostication of contact pressure between soil and driving
wheel. Information Processing in Agriculture, 1: 51 - 56
Wang DM, Wang L, Zhang GM. 2012. Short⁃term wind speed fore⁃
cast model for wind farms based on genetic BPNN. Journal of
Zhejiang University ( Engineering Science), 46 (5): 837 -
841, 904 (in Chinese)
Wang GC, Wang HL, Lü WY, Qing XF, Yu H. 2011. Pest insects
foecasting based on artificial neural networks. / / The 2nd Asia⁃
Pacific Conference on Information Network and Digital Content
Security, pp. 110 - 112 (in Chinese)
Wang XC, Shi F, Yu L, Li Y. 2013 43 cases analysis on MATALB
neural network. Beijing: Beihang University Press, pp. 9 ( in
Chinese)
Xiang CS, Zhou ZY. 2010. Application of ARIMA and SVM hybrid
model in pest forecast. Acta Entomologica Sinica, 53 ( 9 ):
1055 - 1060 (in Chinese)
Ye QL, Wei LX. 2015. The prediction of stock price based on im⁃
proved wavelet neural network. Open Journal of Applied Sci⁃
ences, 5: 115 - 120
Yuan ZM, Zhang YS, Xiong JY. 2008. Multidimensional time series
analysis based on support vector machine regression and its ap⁃
plication in agriculture. Scientia Agricultura Sinica, 41 (8):
2485 - 2492 (in Chinese)
Zhang QH, Benveniste A. 1992. Wavelet networks. IEEE Transac⁃
tions on Neural Networks, 3(6): 889 - 898
Zhang XX, Zhai BP, Mou JY, Zhang GA, Liu XD. 1985. Insect
ecology and forecast. Beijing: China Agriculture Press, pp.
205 - 207 (in Chinese)
Zhu JS, Zhai BP, Dong BX. 2010. Forecasting model for the ovipo⁃
sition peak day in the second generation of Helicoverpa armigera
(Lepidoptera: Noctuidae) based on radial basis wavelet net⁃
work. Acta Entomologica Sinica, 53 (12): 1429 - 1435 ( in
Chinese)
(责任编辑:高 峰)
1633 期 Jin Ran, et al. : Comparison of occurrence periods of wheat aphids based on different prediction systems