close

Вход

Забыли?

вход по аккаунту

?

j.compag.2018.08.012

код для вставкиСкачать
Computers and Electronics in Agriculture 153 (2018) 188–195
Contents lists available at ScienceDirect
Computers and Electronics in Agriculture
journal homepage: www.elsevier.com/locate/compag
Original papers
Identification of maize haploid kernels based on hyperspectral imaging
technology
T
⁎
Yaqian Wanga, Yingjun Lvc, Huan Liua,b, Yaoguang Weia, Junwen Zhangd, Dong Ana, ,
⁎
Jianwei Wue,f,
a
College of Information and Electrical Engineering, China Agricultural University, Beijing 100083, China
College of Electrical Engineering and Automation, Shandong University of Science and Technology, Qingdao 266590, China
c
Department of Electrical Engineering and Information Technology, Shandong University of Science and Technology, Jinan 250031, China
d
College of Agronomy and Biotechnology, China Agricultural University, Beijing 100193, China
e
Beijing PAIDE Science and Technology Development Co. Ltd., Beijing 100097, China
f
Beijing Research Center for Information Technology in Agriculture, Beijing 100097, China
b
A R T I C LE I N FO
A B S T R A C T
Keywords:
Maize
Haploid kernel identification
Hyperspectral imaging technology
Qualitative analysis
Joint modeling
Haploid breeding is a significant technology of maize breeding. Rapid and accurate haploid kernel identification
method has great significance to accelerating the efficiency of haploid breeding. At present, detecting genetic
markers on the embryo of kernels by machine vision and determining oil content of maize kernels by nuclear
magnetic resonance (NMR) are widely used to automatically identify haploid maize kernel. However, the machine vision method can only identify the haploid through embryo side of kernels, and the NMR method cannot
distinguish them when haploid and diploid have the overlap oil content. The study was aimed exploring a rapid
and accurate method to identify haploid maize kernel using near-infrared hyperspectral imaging technology to
overcome the limitations of current automated haploid identification and to achieve more accurate screening of
haploid. In terms of two representative varieties of maize (Zhengdan 958 and Nongda 616), the study adopted
spectral features of hyperspectral imaging to discuss the influence of embryonic orientation (embryo faces to or
against light source) on haploid identification model. Meanwhile, the separability of embryo and non-embryo
and identification accuracy of joint modeling of embryo and non-embryo were analyzed. The study showed that
the greater difference between embryo and non-embryo of haploid and diploid, but hyperspectral imaging
method could effectively distinguish haploid and diploid through embryo or non-embryo. At the same time, with
the qualitative analysis method, two maize varieties could accurately distinguished haploid and diploid with
overlapping oil content based on joint modeling. In this case, the test set of haploid and diploid achieved yielded
higher correct acceptance rate (CAR) of 99% and the false acceptance rate (FAR) were both below 1%, with a
high accuracy rate. The study showed that it is feasible to recognize maize haploid using hyperspectral imaging
technology, which can provide a reference for the later haploid sorting systems.
1. Introduction
Haploid breeding in maize breeding has very broad applications
(Ma et al., 2011). Haploid breeding approach combining biotechnology
with conventional methods, effectively accelerate the process of maize
breeding, thereby promoting greater efficiency of maize breeding while
saving the cost of maize breeding, and also have important significance
for germplasm improvement (Dwivedi et al., 2015; Sun et al., 2009).
Since the probability that maize produces haploid in natural conditions
is low, generally no more than 0.1%, even if the haploid produced by
artificially induced is also less than 10% (Li et al., 2016a; Qin et al.,
2016). That is to say, in practice, the number of diploid is more than
that of haploid, the ratio is about 9:1. In order to achieve haploid maize
breeding, it is necessary to select the haploid kernels from a large
number of mixed kernels of haploid and diploid, so it is very crucial to
rapidly and accurately recognize haploid maize kernels from a large
Abbreviations: NMR, nuclear magnetic resonance; CAR, correct acceptance rate; FAR, false acceptance rate; ROIs, region of interests; BULDP, biomimetic uncorrelated locality discriminant projection; BPR, biomimetic pattern recognition; PCA, principal component analysis
⁎
Corresponding authors at: College of Information and Electrical Engineering, China Agricultural University, 17 Tsinghua East Road, Beijing 100083, China (D.
An). Beijing PAIDE Science and Technology Development Co. Ltd., Beijing 100097, China (J. Wu).
E-mail addresses: andong@cau.edu.cn (D. An), cmwjw@163.com (J. Wu).
https://doi.org/10.1016/j.compag.2018.08.012
Received 24 February 2018; Received in revised form 30 June 2018; Accepted 6 August 2018
0168-1699/ © 2018 Elsevier B.V. All rights reserved.
Computers and Electronics in Agriculture 153 (2018) 188–195
Y. Wang et al.
corn seeds (Williams et al., 2012). Huang et al. used hyperspectral
imaging to analyze maize kernels of different years, and found by
means of model updating could achieved the classification accuracy of
94.4% for maize kernels of different years (Huang et al., 2016). Serranti
et al. studied the rapid classification method of three types of wheat
(yellow berry, vitreous and fusarium damaged) using hyperspectral
imaging techniques, and obtained good classification performance
(Serranti et al., 2013). To overcome the existing problems of automatic
identification of haploid, thus, in this study, near-infrared hyperspectral
imaging technology is used to explore the feasibility of haploid maize
kernel identification.
To rapidly and correctly recognize haploid maize kernels using
hyperspectral imaging technology, two aspects should be considered:
number of mixed kernels for haploid breeding.
In haploid maize kernel identification field, two conventional and
non-destructive haploid identification technologies have been developed, respectively genetic marker method based on R1-nj pigment gene
(Nanda And Chase, 1966) and oil content marker method based on oil
xenia effect (Chen and Song, 2003). Genetic marker method uses pigment gene to express pigment in induced hybrid kernels, and the embryo of haploid has no color marker, the embryo of diploid has a color
marker, and therefore, human eyes can directly and easily select haploid. However, the selection of haploid by human eyes is highly based
on subjective experience. Meanwhile human eyes are apt to be fatigue,
and kernels with unclearly expressed color of embryo and endosperm
aleurone layer are hard to be recognized (Li et al., 2016a; Qin et al.,
2016). Thus, such selection method consuming excessive time and efforts is unable to meet breeding demand. Therefore, the haploid maize
kernel identification technology where machine vision is used to recognize genetic marker of color was presented (De La Fuente et al.,
2017; Li et al., 2016b; Zhang et al., 2013). Such technology recognizes
embryonic color marker by acquiring embryonic images of maize kernels. However, such technology applies to the circumstance that the
embryo of maize kernel is upward only. And kernels with unclear color
marker expression are challenging for both machine vision and selection with human eyes. This limits generalization of haploid classification and selection technology to some degree. Oil content marker
method makes haploid and hybrid diploid kernels prominently different
in oil content through hybrid between high oil inducer and non-high oil
material. At present, nuclear magnetic resonance (NMR) is used to recognize haploid by the oil content of kernel (Liu et al., 2012;
Melchinger et al., 2014; Wang et al., 2016). However, the oil content of
haploid and diploid is overlapped, in order to exclude diploid as far as
possible, the choice of oil content threshold will sacrifice the number of
haploid to a certain extent and cause the waste of haploid kernels (Li
et al., 2016a). Therefore, the problems encountered by the current
methods limit the application of haploid automated identification
technology.
Hyperspectral imaging technology is a new non-destructive inspection method which combines imaging technology and spectral
analysis technology. Hyperspectral image consists of images at a series
of wavelengths. There is a specific two-dimensional image at each
wavelength. Additionally, the grey information of the same pixel point
at different wavelengths also provides spectral information (Zhao et al.,
2011). Compared with conventional machine vision and NMR, hyperspectral imaging technology can obtain color, morphology, texture and
other external features of samples, and information about chemical
elements in kernels to analyze kernels in numerous aspects (Tang et al.,
2015). Williams et al. combined hyperspectral imaging with multivariate data analysis to achieve early fungal contamination detection of
(1) To fulfill rapid haploid maize identification and analyze without
human involved, impacts of different parts of maize kernel (embryo
and non-embryo) on identification performance of the system was
studied, and explore the method to accurately identify the different
poses and placements of kernels, and identification model was established to test and evaluate performances of the system.
(2) The situation that haploid and diploid have overlap of oil content
will exist when oil markers were used. The quantitative analysis
method with oil as an indicator, and threshold value selection
method would cause low accuracy, so it’s necessary to explore the
qualitative analysis method to extract the effective information
sufficiently and realize the classification accuracy of haploid and
diploid with overlapping oil content. The identification model
should ensure that haploid can be identified as much as possible
and have enough ability to reject diploid in the mixed kernels.
2. Materials and methods
2.1. Experimental samples
The experiment utilized two representative samples, namely,
Zhengdan 958 and Nongda 616, respectively. Each variety includes
haploid and diploid kernels. The experimental samples were produced
from high oil hybrid induction with R1-nj genetic marker was induced
and supplied by National Maize Improvement Center of China
Agricultural University. One hundred kernels of haploid and one hundred kernels of diploid were selected from Zhengdan 958 as research
object, one hundred kernels of haploid and one hundred kernels of
diploid were selected from Nongda 616 as verification object. Each
kernel was numbered in the experiment. The difference of genetic
marker between haploid and diploid of Zhengdan 958 and Nongda 616
is shown in Fig. 1. It can be seen that there are no color genetic markers
on the haploid embryo of the two maize varieties, and color genetic
Fig. 1. The genetic markers of haploid and diploid: (a). Zhengdan 958; (b). Nongda 616.
189
Computers and Electronics in Agriculture 153 (2018) 188–195
Y. Wang et al.
Fig. 2. The distribution of oil content ratio of haploid and diploid: (a). Zhengdan 958; (b). Nongda 616.
In this study, when images were ensured to be free of distortion, the
movement speed of platform was set at 0.27 cm/s and exposure time at
35 ms. Each acquired image was three-dimensional image cube. In
order to explore impacts of placement of kernels on the model, images
of embryo and non-embryo were acquired respectively for each kernel.
To reduce impacts of parameter drift during the measurement, the way
of alternative acquisition was adopted, that is, each image covers 10
haploid and 10 diploid kernels. Alternative acquisition was carried out
successively in accordance with the embryo of haploid and diploid
faced the light source in the first image, the non-embryo of haploid and
diploid faced the light source in the second image. To reduce disturbance from the external environment, images were acquired in the
close black box.
markers are present in the diploid embryo. The difference of oil content
between haploid and diploid (the oil content of single kernel was
measured by NMI20-015V-I nuclear magnetic resonance imaging analyzer by Shanghai Newman Electronic Technology Co., Ltd.) can be
seen in Fig. 2, which shows that some haploid and diploid of the two
varieties have the overlap of oil contents.
2.2. Hyperspectral image data acquisition
Since maize kernels are small, push broom hyperspectral imaging
system used in the study, shown in Fig. 3. Core components of the
system are composed of even light source, spectral camera, mobile
control platform, computer and control software. The spectral camera
adopts Image-λ-N17E “spectrum and image” near-infrared improved
hyperspectral camera (Zolix Instruments Co., Ltd., Beijing, China),
which integrates the Imspector imaging spectrometer series and InGaAs
CCD camera. Its spectral region ranges from 862.9 to 1704.2 nm (it
contains 256 bands), the spectral resolution is 5 nm, the camera pixel is
320 * 256, and the slit width is 30 μm. The mobile control platform is
controlled by step motor and images are acquired by the image acquiring software SpecView (Zolix Instruments Co., Ltd., Beijing, China).
2.3. Hyperspectral image data processing
2.3.1. Hyperspectral image calibration
In the experiment, to narrow measurement errors, such as light
source fluctuation and dark current in camera, white and black calibration must be carried out for hyperspectral image. Through the Eq.
(1) realized the black and white calibration:
Camera controller
CCD camera
Image spectrograph opƟcs
Len
Computer
Lamp
plaƞorm control
Maize
Fig. 3. Hyperspectral imaging system.
190
Computers and Electronics in Agriculture 153 (2018) 188–195
Y. Wang et al.
R ci =
sampleci−darkci
whiteci−darkci
2.4.3. Discrimination model
The biomimetic pattern recognition (BPR) method proposed by
Wang Shoujue (Wang, 2002; Wang and Wang, 2002) was used in this
research to establish haploid and diploid identification models of two
varieties. Firstly, BPR is different from traditional pattern recognition
based on optimal division of multi-class, BPR according to the “ cognition ” of each class of samples, it obtains the best covering of each
class of samples in the feature space which is more consistent with the
process of human cognition on objects; Secondly, BPR uses a closed
geometry to cover each category of samples, which makes it more
compact to cover the same class of samples and effectively avoid the
heterogeneous samples being mistakenly attributed to the same class;
Finally, BPR can realize the effective rejection of other categories of
samples, so that the other samples will not be mistakenly classified as
haploid and diploid. Therefore, BPR can effectively meet the requirements of this research.
(1)
where Rci are the calibrated images, the calibrated images are used for
the later spectral extraction; sampleci are the original images of the
samples; darkci is the dark reference image, the equipment is sealed
with lens cap when dark reference image is acquired; whiteci is the white
reference image, whiteboard is placed at the same height where the test
object is placed.
2.3.2. Image processing and feature extraction
Hyperspectral imaging technology has the feature of combining the
image with spectrum. Due to the complexity of image information extraction, it cannot meet the requirements of real-time processing, and
also cannot extract the essential information of haploid and diploid (Jia
et al., 2013). Therefore this research only used spectral information of
samples. The information of kernel must be extracted from the image
after white and black calibration. The extraction procedure of average
reflectance spectrum: (a). Region of interests (ROIs) were extracted at
1064.8 nm with the highest contrast ratio between kernel and background. (b). The maximum value of background was selected as the
threshold for image binaryzation, and the boundary coordinates of each
sample were acquired, and generate binary masks; (c). By binary masks
were generated to acquire ROIs at 256 bands; (d). The mean grey values
of the maize kernel in ROIs were extracted at 256 bands, and the mean
reflectance spectrum of each kernel was obtained. The software employed for spectral feature extraction was MATLAB 2016a (USA,
MathWorks Company).
2.4.4. Elimination of abnormal samples
BPR uses the best coverage unit. The existence of abnormal samples
will affect classification performance of BPR. In order to avoid from
making abnormal samples affect the modeling, the study applies the
quadratic discriminate method to eliminate abnormal samples after
spectra preprocessing. Quadratic discriminate method, an elimination
abnormal samples method in near-infrared spectroscopy analysis, utilizes the correlation and proximity among samples to judge whether the
sample is abnormal. The quadratic discriminate method firstly uses the
correlation of the same class of samples to select suspected abnormal
samples. The correlation of the samples is determined by the correlation
coefficient between the samples. The smaller the correlation coefficient
means the smaller the correlation between sample and the greater
possibility of an abnormal sample. The suspected abnormal samples
were determined by the correlation between the samples. Then convert
the original data space to a low-dimensional data space via principal
component analysis (PCA) to reduce the distance among normal samples and to increase the distance between abnormal samples and normal
samples. Then determining the suspected sample is abnormal or not by
means of the proximity of the sample in low-dimensional space. The
proximity of the sample is determined by calculating the relationship
between the suspect sample and its proximity sample. When the distance between two nearest normal samples of suspected sample and the
center of gravity are greater than the distance between the two nearest
normal samples of suspected sample and the suspected sample, the
sample is considered to be normal sample, otherwise the sample is
considered to be abnormal sample and remove it. This method can effectively remove abnormal samples instead of excessive removing the
samples.
2.4. Spectral processing and discrimination model
2.4.1. Spectra preprocessing
Spectrum preprocessing plays a crucial role in establishing accurate
and stable model (Yan et al., 2005). In this study, a moving average
window smoothing treatment (width = 9), first derivative method
(width = 9), vector normalization to preprocess spectral data. The
preprocessing methods and its parameters were obtained by repeated
experiments. After spectra were preprocessed, the background interference and random noises were eliminated, amplify the difference of
spectral signals and improve the spectral resolution.
2.4.2. Feature extraction of spectral data
It is a key link of qualitative analysis to correctly acquire spectral
features (Yan et al., 2005). After original spectra were preprocessed, the
similarity was still larger than the difference among spectra of haploid
and diploid. There was a strong correlation among different wavelength
points of spectra. Thus, on the basis of preprocessing, further feature
extraction and dimensionality reduction was required.
Ning et al. (2018) put forward a method: Biomimetic Uncorrelated
Locality Discriminant Projection (BULDP), which is a kind of supervised
feature extraction method based on Unsupervised Discriminant Projection. Different from the conventional feature extraction algorithms
where reasoning plays a leading role, BULDP incorporates human
image cognitive into feature extraction. When constructing the similarity coefficient, it not only considers the neighborhood of the samples,
but also applies the bionic thinking about homologous continuity and
heterogeneous similarity. Different from the other feature extraction
methods that use the neighborhood of the sample to construct a similarity coefficient, the similarity coefficient of BULDP structure can reflect the continuity between the same class and the similarity between
the different classes, which is more in line with human cognitive. At the
same time, in order to reduce the redundancy of discriminant vectors,
BULDP introduces an uncorrelated space to ensure that there is no
correlation between discriminant vectors.
2.4.5. Evaluation indexes of model
Haploid identification model was used to identify haploid and diploid, and the system performance was tested. In this study, correct
acceptance rate (CAR), false acceptance rate (FAR) were used to evaluate the model and data. The higher the CAR is, the lower the FAR is,
and the better the system performance is. This study measured the separability of haploid and diploid by the ratio of inter-class and intraclass. The intra-class distance refers to the mean distance from each
sample to the center of gravity of the class, while the inter-class distance means the distance between the centers of gravity of haploid and
diploid. The larger the ratio, the stronger the separability. The detailed
definition of the parameters is as follows:
191
CAR =
The number of samples received correctly
The number of samples that should be accepted
(2)
FAR =
The number of samples received falsely
The number of samples that should be rejected
(3)
Computers and Electronics in Agriculture 153 (2018) 188–195
Y. Wang et al.
Fig. 4. The system framework of identification model of haploid maize kernel.
R=
Inter-class distance
Intra-class distance
second harmonic generation in the vicinity of 1037 nm is derived from
oil. On the whole, the spectral curves have a similar trend, but also
exists the different degrees of difference. There were obvious differences in the spectra of embryo (the spectra extracted from the images
taken by the embryo faces to light source are called the spectra of
embryo) of haploid and diploid, and the difference of the spectra of
non-embryo (the spectra extracted from the images taken by non-embryo faces to light source are called the spectra of non-embryo) of
haploid and diploid was significantly smaller than that of the spectra of
embryo. At the same time, it also can be observed that the reflectivity of
the embryo was higher than non-embryo, and the spectra of embryo
and non-embryo have obvious differences. The differences between
these spectra provide the basis for the identification of haploid using
hyperspectral imaging.
(4)
The work flow chart of the haploid identification system is shown in
Fig. 4. The software employed for the data processing and discrimination model was MATLAB 2016a (USA, MathWorks Company).
3. Results and discussion
3.1. Analysis of spectral characteristic
In view of the similarity of the two maize varieties in the spectral
curves, the paper takes the original reflectance spectral data of the
haploid and diploid of Zhengdan 958 as representative, involving the
spectral data of the embryo and non-embryo for each haploid and diploid (Fig. 5a). Due to the large noise at both ends of the original
spectral curve, 880.2–1669.6 nm (a total of 240 wavebands) was selected for the study. As indicated in the raw spectral data of haploid and
diploid in Fig. 5a, the spectral data of haploid and diploid overlap,
regardless embryo or non-embryo. Therefore, it is difficult to classify
haploid and diploid by raw data.
In order to analyze the differences between haploid and diploid
spectra more clearly, the average spectra of embryo and non-embryo of
haploid and diploid were calculated, as shown in Fig. 5b. It can be seen
from the plot that the troughs in the O-H third harmonic generation
around 990 nm due to carbohydrates absorption, C-H third harmonic
generation around 1200 nm due to starch-based carbohydrates, O-H
second harmonic generation around 1450 nm due to the absorption of
moisture respectively (Tang et al., 2015). In addition, owing to difference in oil content between haploid and diploid, the internal chemical
compositions of haploid and diploid are different. Among them, the C-H
3.2. Influence of embryonic orientation on identification model
In the process of measurement, the sample placement position exists
in two situations, the embryo facing the light source and opposite light
source, so the two kinds of measurement situations need to be analyzed
in order to realize the later automatic selection.
The PCA was conducted on the data of Zhengdan 958 to study the
influence of embryonic orientation of haploid and diploid on the
identification model, and selected the first two dimensions (PC1 and
PC2) to draw the sample scoring vector diagram, as shown in Fig. 6,
wherein the horizontal and vertical coordinates represent the first
principal component (PC1) and the second principal component (PC2)
respectively. The variance contribution rate of the first two principal
components was 90.67%, that of PC1 being 80.69% and that of PC2
reaching 9.98%. Most of the raw spectral data was contained in the first
192
Computers and Electronics in Agriculture 153 (2018) 188–195
Y. Wang et al.
Fig. 5. The reflectance spectra plots of embryo and non-embryo of haploid and diploid of Zhengdan 958: (a). raw reflectance spectra; (b). average raw spectra.
3.3. Separability analysis on the embryo and the non-embryo
First, the spectral data of embryo of haploid and diploid were used
to investigate whether the embryo of haploid and diploid had separability. Two varieties (Zhengdan 958 and Nongda 616) were selected to
establish the haploid identification models and each variety had 200
spectra, of which 100 were diploid embryonic spectra and 100 were
haploid embryonic spectra. The number of samples of the training set
and test set was distributed according to 2:1. The raw data was preprocessed and then the abnormal samples of the training set were rejected, after which BULDP was used to extract the spectral features to
change the spectral data from high-dimensional space to two-dimensional space. Haploid identification model was established by BPR.
The embryonic spectral features of Zhengdan 958 extracted by
BULDP were projected into a two-dimensional space (Fig. 7a). It can be
seen from the figure that the spectral data of embryo of haploid and
diploid had obvious separability and large differences in the feature
space. The separability of haploid and diploid in the original space and
feature space was calculated respectively, and R = 1.0340 in the original space but R = 6.5453 in the feature space, which indicated that
there were significant differences between haploid and diploid by feature extraction. In order to reduce the impact of sample selection, and
the identification results more representative, test samples were selected by random sampling method, and the average value of results for
10 runs was calculated as final results, and 3 abnormal samples were
removed on average through 10 runs. The embryonic data of two
varieties were tested respectively, and the results of the test were shown
in Table 1.
Table 1 shows the final results of the classification of haploid and
diploid using the embryonic data of maize kernels. As shown in Table 1,
for Zhengdan 958, the test set of haploid and diploid achieved the CAR
of 99.68% and 99.34%, the FAR of 0.61% and 0%, respectively. For
Nongda 616, the test set of haploid and diploid achieved the CAR of
100%, the FAR of 0%. It is indicated that the haploid and diploid can be
classified accurately by the embryonic data of maize kernels.
As a result of the oil difference of maize kernel is almost contained
in embryo (about 85%), the main component of the non-embryo is
starch. Thus, it is necessary to discuss the separability of the non-embryo on the basis of the separability of the embryo. By using the same
batch of maize kernels, the spectral number and processing method the
same as the embryo, and haploid identification models were established
by using spectra data of non-embryo. The non-embryonic spectral features of Zhengdan 958 extracted by BULDP were projected into a twodimensional space (Fig. 7b). It can be seen from the figure that the nonembryonic spectral data of haploid and diploid had obvious separability
and large differences in the feature space. The separability of haploid
and diploid in the original space and feature space was calculated
Fig. 6. Two-dimension map of the sample distribution of Zhengdan 958.
two principal components.
It can be seen from the distribution of sample points in the PCA
feature space that the maize kernels of haploid and diploid can be divided into A and B intervals through the division of the first principal
component PC1, in which the embryo of haploid and diploid were
distributed in the Zone A, and the non-embryo of haploid and diploid
were distributed in the Zone B. It is indicated that the distance between
the spectra of embryo and non-embryo was greater than that between
the haploid and the diploid, and the difference between the spectra of
embryo and non-embryo become the main difference. The A and B
areas were analyzed respectively. First, it can be seen from Zone A that
the distance between haploid and diploid embryo was far, which
showed obvious separability. In Zone B, the distance between the nonembryo of haploid and diploid was significantly smaller than that of the
embryo, and it is necessary to prove further whether the non-embryo is
separable.
The above results showed that the difference between haploid and
diploid was smaller than the difference between embryo and non-embryo, it cannot establish identification model by one side of the surface
to predict the other surface. At the same time, in the process of production, it cannot guarantee that all the embryo of kernels can be upward or downward. The method for only modeling one side limits the
application of haploid recognition. Therefore, in order to realize the
automation of haploid sorting, the separability of embryo and nonembryo of haploid and diploid must be analyzed.
193
Computers and Electronics in Agriculture 153 (2018) 188–195
Y. Wang et al.
Fig. 7. Samples distribution by BULDP: (a). embryonic data of Zhengdan 958; (b). non-embryonic data of Zhengdan 958.
effectively distinguish haploid and diploid by means of embryo and
non-embryo. Therefore the haploid identification models were built
through the method of jointly modeling the embryo and non-embryo.
As can be seen from the distribution of oil content ratio of samples in
the Fig. 2, there was an overlap of oil between haploid and diploid of
two varieties. The overlap of oil will affect the threshold of NMR or the
identification of quantitative method using oil as an indicator. Qualitative analysis not only uses the differences in oil content. It includes
comprehensive information about the differences in material composition within all samples. Therefore, it finally decided to classify haploid
and diploid using a qualitative analysis method based on joint modeling
of embryo and non-embryo. It studies whether the identification accuracy of the model meets the identification requirements.
When selecting samples, it is necessary to note that each kernel
included both the data of embryo and non-embryo. Therefore, in selecting the training set, we must simultaneously select embryo and nonembryo data of the same kernel, and use the method of joint modeling
of embryo and non-embryo of maize kernels. Similarly, one kernel can
only exist in the form of embryo or non-embryo, therefore the embryo
and the non-embryo data were tested respectively. The sample number
of the training set and test set was still distributed according to 2:1.
Haploid identification models based on the joint modeling method,
and for Zhengdan 958 the test set of embryo of haploid and diploid
achieved the CAR of 99.09% and 99.70%, the FAR of 0% respectively,
for non-embryo, yielded CAR as 99.4% and 100%, the FAR of 0%. At
the same time, we can see from the experimental results of Nongda 616
that the test set of embryo and non-embryo of haploid and diploid
achieved yielded CAR of 100%, the FAR of 0% (see Table 3).
The above results showed that the haploid identification model
based on the joint modeling of embryo and non-embryo is of good
classification accuracy for untrained samples, so as to ensure the principle that there is sufficient rejection ability for diploid in mixed kernels
and recognize the haploid in the mixed kernels as much as possible.
Therefore, the haploid can be effectively recognized through use of
hyperspectral imaging technology.
To further verify the effectiveness of hyperspectral imaging
Table 1
Classification accuracy of spectral data of embryo of two varieties.
Variety
Zhengdan 958
Nongda 616
Test sample
Haploid
Diploid
Haploid
Diploid
Training set
Test set
CAR (%)
FAR (%)
CAR (%)
FAR (%)
100
100
100
100
0
0
0
0
99.68
99.34
100
100
0.61
0
0
0
Table 2
Classification accuracy of spectral data of non-embryo of two varieties.
Variety
Zhengdan 958
Nongda 616
Test sample
Haploid
Diploid
Haploid
Diploid
Training set
Test set
CAR (%)
FAR (%)
CAR (%)
FAR (%)
97.87
100
100
100
0
0
0.1587
0
98.79
99.39
98.79
100
0
0
0
0
respectively, and R = 0.5659 in the original space but R = 3.0622 in
the feature space, which indicated that there were significant differences between haploid and diploid by feature extraction. The nonembryonic data of two varieties were tested respectively, and the results of the test were shown in Table 2.
Table 2 shows the final results of the classification of haploid and
diploid using the non-embryonic data of maize kernels. As shown in
Table 2, for Zhengdan 958, the test set of haploid and diploid achieved
the CAR of 98.79% and 99.39%, respectively, the FAR of 0%. For
Nongda 616, the test set of haploid and diploid achieved the CAR of
98.79% and 100%, respectively, the FAR of 0%. It is indicated that the
haploid and diploid can be classified accurately by the non-embryo data
of maize kernels. It is proved that the hyperspectral imaging can also
obtain enough information of embryo through the starch layer of nonembryo for classification.
Through analysis of the representative samples which Zhengdan
958 and Nongda 616, the classification accuracy of haploid and diploid
were above 98% for the test set, whether through embryo or non-embryo of maize kernel. The results showed that hyperspectral imaging
method can effectively distinguish haploid and diploid by means of
embryo and non-embryo, and it proves that the feasibility to recognize
haploid and diploid by means of hyperspectral imaging technology.
Table 3
Classification results of the test set of the embryo and non-embryo based on the
joint modeling.
Variety
Zhengdan 958
3.4. Qualitative identification of haploid based on joint modeling
Nongda 616
It has been proved that hyperspectral imaging method can
194
Test sample
Haploid
Diploid
Haploid
Diploid
Embryo
Non-embryo
CAR (%)
FAR (%)
CAR (%)
FAR (%)
99.09
99.70
100
100
0
0
0
0
99.40
100
100
100
0
0
0
0
Computers and Electronics in Agriculture 153 (2018) 188–195
Y. Wang et al.
technology in maize haploid discrimination, this paper was done on the
accuracy of 32 haploid and 33 diploid maize kernels determined by
hyperspectral imaging technology through field planting detection
method. The results showed that the outcome of the field planting detection method and the hyperspectral imaging detection method were
same, which indicated that the hyperspectral imaging detection method
can effectively identify the maize haploid, which has a broad application prospects.
Agron. Sinica 29, 587–590.
De La Fuente, G.N., Carstensen, J.M., Edberg, M.A., Lübberstedt, T., Léon, J., 2017.
Discrimination of haploid and diploid maize kernels via multispectral imaging. Plant
Breeding 136, 50–60.
Dwivedi, S.L., Britt, A.B., Tripathi, L., Sharma, S., Upadhyaya, H.D., Ortiz, R., 2015.
Haploids: constraints and opportunities in plant breeding. Biotechnol. Adv. 33,
812–829.
Huang, M., Tang, J., Yang, B., Zhu, Q., 2016. Classification of maize seeds of different
years based on hyperspectral imaging and model updating. Comput. Electron. Agric.
122, 139–145.
Jia, S., Liu, Z., Li, S., Li, L., Ma, Q., Zhang, X., Zhu, D., Yan, Y., An, D., 2013. Study on
method of maize hybrid purity identification based on hyperspectral image technology. Spectroscopy Spectral Anal. 33, 2847–2852.
Li, H., Li, W., Qin, H., Chen, S., Liu, J., Li, W., 2016a. Classifying method of haploid and
diploid based on least square error. Trans. Chin. Soc. Agric. Mach. 47, 259–264.
Li, W., Liu, Y., Chen, S., Qin, H., Liu, J., Tian, Z., 2016b. Automatic separating system of
maize haploid based on machine vision. J. Agric. Mech. Res. 81–85.
Liu, J., Guo, T., Yang, P., Wang, H., Liu, L., Lu, Z., Xu, X., Hu, J., Huang, Q., Chen, S.,
2012. Development of automatic nuclear magnetic resonance screening system for
haploid kernels in maize. Trans. Chin. Soc. Agric. Eng. 28, 233–236.
Ma, J., Jiang, M., Liu, X., Wang, H., 2011. Discussions on haploid breeding in maize. J.
Northeast Agric. Univ. 42, 135–139.
Melchinger, A.E., Schipprack, W., Friedrich Utz, H., Mirdita, V., 2014. In Vivo haploid
induction in maize: identification of haploid seeds by their oil content. Crop Sci. 54,
1497.
Nanda, D., Chase, S., 1966. An embryo marker for detecting monoploids of maize (Zea
Mays L). Crop Sci. 6, 381–382.
Ning, X., Li, W., Li, H., Tang, B., He, H., 2018. BULDP: biomimetic uncorrelated locality
discriminant projection for feature extraction in face recognition. IEEE Trans. Image
Process. 27, 2575–2586.
Qin, H., Ma, J., Chen, S., Yan, Y., Li, W., Wang, P., Liu, J., 2016. Identification of haploid
maize kernel using NIR spectroscopy in reflectance and transmittance modes: a
comparative study. Spectroscopy Spectral Anal. 36, 292–297.
Serranti, S., Cesare, D., Bonifazi, G., 2013. The development of a hyperspectral imaging
method for the detection of Fusarium-damaged, yellow berry and vitreous Italian
durum wheat kernels. Biosyst. Eng. 115, 20–30.
Sun, F., Bai, C., Feng, Y., Zhao, R., Su, E., Zhang, L., Liu, Z., Li, Q., 2009. Haploid technique and its application in corn selection. Inner Mongolia Agric. Sci. Technol. 3,
90–91.
Tang, J., Huang, M., Zhu, Q.B., 2015. Purity detection model updata of maize seeds based
on active learing. Spectroscopy Spectral Anal. 35.
Wang, H., Liu, J., Xu, X., Huang, Q., Chen, S., Yang, P., Chen, S., Song, Y., 2016. Fullyautomated high-throughput NMR system for screening of haploid kernels of maize
(corn) by measurement of oil content. PLoS ONE 11, e0159444.
Wang, S., 2002. Bionic (topological) pattern recognition-a new model of pattern recognition theory and its applications. Acta Electron. Sinica 30, 1417–1420.
Wang, S., Wang, B., 2002. Analysis and theory of high-dimension space geometry for
artificial neural networks. Acta Electron. Sinica 30, 1–4.
Williams, P.J., Geladi, P., Britz, T.J., Manley, M., 2012. Investigation of fungal development in maize kernels using NIR hyperspectral imaging and multivariate data analysis. J. Cereal Sci. 55, 272–278.
Yan, Y., Zhao, L., Han, D., Yang, S., 2005. Basic and Application of Near Infrared
Spectroscopy Analysis. China Light Industry Press, Beijing, China.
Zhang, J., Wu, Z., Song, P., Li, W., Chen, S., Liu, J., 2013. Embryo feature extraction and
dynamic recognition method for maize haploid seeds. Trans. Chin. Soc. Agric. Mach.
29, 199–203.
Zhao, J., Chen, Q., Lin, H., 2011. Introduction Modern Imaging Technology and its
Application in the Detection of Food and Agricultural Products. China Machine Press,
Beijing, China.
4. Conclusions
In this study, the feasibility of haploid maize kernel identification
was explored by the near-infrared hyperspectral imaging technology.
Taking the haploid and diploid of two varieties as experimental materials, the average reflectivity spectra of maize kernels were extracted
from ROIs of hyperspectral images as classification characteristics in
the study. The haploid identification model was established combined
with pretreatment, BULDP feature extraction and BPR. In order to
identify the haploid without human involvement in the later automation equipment, the influence of embryonic orientation on identification model and the separability between the embryo and non-embryo
were analyzed. Finally, the qualitative identification of haploid was
determined by using the method of joint modeling of embryo and nonembryo. The experiments showed that qualitative analysis fully extracted the effective classification information, and it was not the only
indicator of oil. From the perspective of the classification accuracy of
the test set, this method achieved the highest CAR above of 99%, the
FAR being stabilized below 1%, with a high accuracy rate, achieving
more accurate screening of haploid. It can be concluded from the above
results that it is feasible to identify haploid with spectral dimension
information of hyperspectral imaging technology. Additionally, it can
provide the basis for haploid maize kernel identification technology.
Acknowledgements
The authors gratefully acknowledge the financial support from the
Maize Research Centre of Beijing Academy of Agriculture and Forestry
Sciences (Grant No. 10162130106232029) and the National Key
Research and Development Program of China during the 13th Five-Year
Plan Period (Grant No. 21196058) of the National Maize Improvement
Centre of China. The maize seed samples were provided by National
Maize Improvement Centre of China.
References
Chen, S., Song, T., 2003. Identification haploid with high oil xenia effect in maize. Acta
195
Документ
Категория
Без категории
Просмотров
1
Размер файла
1 808 Кб
Теги
012, 2018, compaq
1/--страниц
Пожаловаться на содержимое документа