Identification of wheat kernel varieties based on hyperspectral imaging technology and grouped convolutional neural network with feature intervals

doi:10.21203/rs.3.rs-1951327/v1

Download PDF

Research Article

Identification of wheat kernel varieties based on hyperspectral imaging technology and grouped convolutional neural network with feature intervals

https://doi.org/10.21203/rs.3.rs-1951327/v1

This work is licensed under a CC BY 4.0 License

Version 1

posted

You are reading this latest preprint version

Variety identification of seeds is essential for evaluating seed purity and ensuring crop yields. This study used hyperspectral imaging to determine 8 high-quality wheat kernel varieties in China combined with a deep learning network and spectral wavelength interval selection methods. The hyperspectral images with a spectral range of 400-1000 nm were firstly processed by three kinds of spectral wavelength interval selection methods including interval continuum removal (iCR), backward interval partial least squares (BiPLS), and interval random frog (iRF) for exploring optimal wavelength intervals. Then the convolutional neural network (CNN) classification models in combination with the processed hyperspectral images were developed. In order to extract effective features, two different structures with full convolution (FC) and grouped convolution (GC) were used. The performance of the proposed models was validated with 8000 samples from the 8 varieties of wheat kernels. The results showed that the combination of feature wavelength and grouping convolution enabled classification models effectively to obtain better classification accuracy and improve the inference speed of the model. The iCR-GC CNN model had the highest classification accuracy with a 4.4% increase and the fastest inference speed with a 44.21% reduction by comparing with the reference CNN classification model.

Hyperspectral imaging

Convolutional neural network

wheat kernel

spectral wavelength interval selection

Wheat is one of the most important food crops in temperate countries, widely used as livestock feed and staple food (Tadesse et al., 2019). With the development of crop science and breeding technology, an increasing number of wheat varieties have flooded the market. At the same time, the purity of wheat kernels has also attracted great attention from breeders, planters, and consumer (Koistinen and Hanhineva, 2017). However, wheat kernels of different varieties have an almost consistent visual appearance, which makes it difficult to distinguish with the naked eye. Traditional methods used to identify wheat varieties, such as high-performance liquid chromatography (HPLC) and gas chromatography-mass spectrometer (GC-MS) (Qiu et al., 2018), usually consume a huge amount of time and manpower, meanwhile causing damage to the tested samples. A series of recent advances in machine learning, spectroscopy, and computer graphics have led to a variety of rapid and non-destructive methods for the identification of crop seed varieties.

Hyperspectral imaging is an emerging technology that can fuse visible/near-infrared (VNIR) spectra into an image. The hyperspectral image is defined as a data cube with a shape of w×h×d where the w×h is the width and height of the image that contains the information in spatial dimensions and d is the bands of the image that contains the information in spectral dimensions. It can be regarded as a spectrum with an image on each band or an image with a spectrum on each pixel. Due to the above characteristics of the hyperspectral image, many rapid and nondestructive methods for the identification of crop seed varieties based on machine learning have been proposed (Ge et al., 2019; Xu et al., 2021). Huang et al. (2016) reported a 92.19% prediction accuracy in classifying 17 varieties of maize seed using the morphological, texture, and spectrum features of the hyperspectral image combined with the feature transformation/reduction by developing a least squares support vector machines model. Yang et al. (2015) extracted five morphological features and eight texture features to develop two classification models of four waxy corn seed varieties. The classification accuracies were 77.4%-98.2%. Wang et al. (2015) used the chalkiness degree and shape feature of rice and obtained 88.09%-94.45% accuracy in classifying three rice varieties with a backpropagation neural network classifier. However, all of these studies are at the core of the traditional machine learning algorithms built on a series of features extracted from the hyperspectral images. The process of extracting features is complicated, and the results of the research will depend on the method of extracting features.

Deep learning (DL) is one of the most recent emerging research focuses of data analysis and has been successfully applied to hyperspectral image processing (Li and He, 2020; Vaddi and Manoharan, 2020). DL adopts a neuron structure similar to the biological brain to learn the deep features of the data automatically, which is different from traditional machine learning. This neuron structure can extract various features of the original data layer by layer, and gradually convert them into some high-level abstract features for mining the deep information of the data (Zhao and Du, 2016). The convolutional neural network (CNN) is one of the important networks in deep learning, which has been applied to the identification of crop seed varieties. Yu et al. (2021) introduced CNN to identify hybrid okra seeds and indicated the effectiveness of CNN with a 93.79% prediction accuracy in classifying 17 varieties of hybrid okra seeds where CNN had the highest accuracy by comparison with extreme learning machine (ELM) and stacked sparse auto-encoder (SSAE). Singh et al. (2021) employed the near-infrared hyperspectral imaging coupled with the CNN to identify the barley seeds variety and demonstrated that the CNN model with a 98% prediction accuracy in classifying 35 Indian barley varieties has higher accuracy than the traditional machine learning. Chatnuntawech et al. (2018) showed better performance of CNN with 11.9% absolute improvement in the mean classification accuracy compared with the support vector machine (SVM) for identifying rice seed varieties. These satisfactory results show that this CNN classification model, which can automatically extract features, is expert in processing hyperspectral images and capable of classifying crop seeds. Unlike the traditional method, the high-dimensional hyperspectral data cube can be processed by CNN directly without feature extraction.

Although the high-dimensional hyperspectral data cube is beneficial to the application of remote sensing and crop seed recognition, there are still some problems. The hyperspectral images are usually highly redundant in the spectral dimensions (Zhang et al., 2012), which will lead to an increase in computational complexity, storage space, and communication bandwidth (Sui et al., 2014). This redundancy will not only lengthen the training and inference time of CNN which would have taken plenty of time, but also cause the CNN classification model to fail to focus on the most valuable information of hyperspectral images the first time. Thus, prior knowledge is needed for constraining the hyperspectral images in the spectral dimensions. In this study, the feature wavelength interval selection is introduced as prior knowledge of spectroscopy to constrain the spectral dimension of the hyperspectral images. The interval continuum removal (iCR) (Itoh and Iwasaki, 2013), backward interval partial least squares (BiPLS) (Huang et al., 2012), and interval random frog (iRF) (Yun et al., 2013) are used to extract the feature wavelength interval. The different wavelength intervals represent the characteristics of different components in wheat kernels. When constructing the CNN model, a grouped convolution (GC) in the first layer can independently extract the features of different information intervals, it is better to effectively characterize the hyperspectral images. Then the features are merged in the next convolutional layer and full connect layer. In order to verify the general applicability of the proposed method, experiments are conducted using eight varieties of wheat seed with 1000 samples from each variety, and the modeling performance between the proposed CNN and the traditional CNN was compared. The specific objectives of this study are to: (1) propose a hybrid approach combined wavelength intervals and grouped convolution for effectively extracting hyperspectral features; (2) develop the appropriate lightweight CNN of hyperspectral images for wheat kernels classification;

The remainder of this paper is organized as follows. Section 2 presents the experimental data collection, the hyperspectral imaging system, and image preprocessing. Section 3 describes the wavelength interval selection method briefly, then introduces the proposed CNN models. Section 4 discusses the results and performance analysis. Finally, Section 5 summarizes the conclusions.

2.1 Samples Collection and Preparation

Eight varieties of wheat kernels harvested in 2020 were collected from SINOCHEM AGRICULTURE HOLDINGS. The 1000 unbroken wheat kernels were collectively chosen from samples constituting each variety, thus the total number of wheat kernels included in the study was 8000. In order to minimize the interference from the external environment, each variety of wheat kernels was placed in sealed plastic bags and stored in a dark and dry environment (sealed jars with desiccant) at room temperature (20–25°C). The wheat variety, place of production, and genetic relationship are listed in Table 1, where the genetic relationship is worthy of hybrid male and female parents of this type of wheat.

Table 1

Information on Wheat Varieties
wheat variety	place of production	genetic relationship
bainong4199	Henan	bainong gaoguang3709F2/bainongAK58
jimai44	Shandong	ji954072/jinan17
zhoumai33	Henan	zhenmai366/bainongAK58
weilong169	Henan	shanmai94/xinong822
Shiluan02-1	Hebei	9411/9430
bainongAK58	Henan	zhoumai11/wenmai6/zhenzhou8960
xinmai26	Henan	9408E1/jinan17
jimai22	Shangdong	935024/935106

2.2 The Hyperspectral Imaging System

The hyperspectral images of wheat kernels were acquired by using a push broom (line scan) hyperspectral imaging system, and the schematic diagram of the system is shown in Fig. 1. The system consisted of an image acquisition module, halogen light source, and linear displacement platform.

The image acquisition module was mainly composed of a hyperspectral image spectrograph (1003A-10140 HyperspcTM VNIR C-Series, Headwall Photonics Inc. USA), CCD camera (pixelfly qe IC * 285AL, Cooke, USA), and lens (10004A-21226 Lens, F/1.4 FL23 mm, Standard Barrel, C -Mount., USA). The parallel light was irradiated on the linear displacement platform with samples by the light source through the collimator lens linked by the optical fiber. The linear displacement platform was driven with the aid of a stepper motor was used for a constant scanning speed through Hyperspec-M Rev A.2.1.6 software. All system components except the halogen light source are placed in a dark box to prevent ambient light from affecting the system. In this study, the power of the light source, the moving speed of the conveyer belt, the exposure time of the camera, and the distance from the lens to the plate were set as 60 w, 10 mm/s, 250 ms, and 28 cm, respectively. The band interval during spectrum sampling is 6.38nm, a total of 94 bands with the spectral range from 400 to 1,000 nm, the scan step is 100µm, the total scan length is 100mm (with 1,000 scanning lines for each hyperspectral image), and the scan width is 139.2mm. A hyperspectral image data cube of each batch of wheat kernels with the shape of 1000×1392×94 was obtained ultimately.

The kernels were placed on a counting board with small grooves for hyperspectral image acquisition. The counting board was made of special materials with a very low reflectance in the visible/near-infrared range. A total of 8,000 wheat samples were collected from 8 varieties (1000 samples for each variety).

2.3 Hyperspectral Image Preprocessing

Due to the uneven distribution of the light source in each band, the dark current in the camera, and the influence of the surrounding environment, the collected hyperspectral images will contain a certain amount of noise. To the extent possible to reduce these noises, black and white calibration was performed after every 5 times of hyperspectral image data collection. The calibration formula is as shown in Eq. (1).

$${I_{cablibration}}=\frac{{I - {I_{Black}}}}{{{I_{White}} - {I_{Black}}}}$$

Where ${I_{calibration}}$ is the hyperspectral image after calibration, is the raw hyperspectral image, ${I_{Black}}$ is the dark reference obtained by covering the lens with an opaque cap, ${I_{White}}$ is the white reference obtained by scanning a standard whiteboard.

Firstly, the original hyperspectral image was converted to tif format using ENVI5.1 (ITT Visual Information Solutions, Boulder, UT). Figure 2 shows the procedures of spectral data extraction. The pseudo-RGB image composed of the reflectance of HIS bands 1 (402.53 nm), 17 (505.60 nm), and 35 (621.54 nm) was converted to the YCbCr color space. The CB channel component of YCbCr images which contained the density offset component of blue in the pseudo-RGB image was taken as a grayscale image. The Otsu binarization method (Otsu, 1979) was adopted to binarize the grayscale image to divide it into the wheat region and another region preliminarily. Morphological filtering was performed on the binary image to obtain the mask of the wheat image. Finally, use the mask to extract the region of interest (ROI) of the image i.e. the wheat region. 0 was added around each wheat image to ensure that the size of each wheat image was 60×80×94.

In this study, iCR, BiPLS, and iRF were introduced as prior knowledge of spectroscopy to constrain the spectral dimension of hyperspectral images. The proposed CNN classification models with different structures are used to classify wheat.

3.1 Spectral wavelength interval selection methods

3.1.1 Interval continuum removal

Continuum removal may be defined as a spectral processing method that is used to enhance and standardize the specific absorption characteristics of the biochemical components of the sample spectrum (Malenovský et al., 2013). In the literature (Huang et al., 2004), the continuum tends to be used to refer to a series of straight-line segments fitted over the top of a spectrum that connect local spectral maxima and represents the “background absorption” onto which other absorption features are superimposed. Continuum removal is done by dividing the reflectance of each point in the absorption feature by the reflectance of the continuum at the corresponding wavelength. The raw profile and continuum profile are shown in Fig. 3. Different varieties of wheat all have absorption valleys at the same wavelengths, but the extreme points of these absorption valleys are not completely consistent. Only using the combination of a single waveband as the input of the simplified model may lead to the inability to accurately represent these absorption valleys. The characteristic absorption valley and its nearby wavelengths will be selected as the interval that would be used to develop simplified classification models. In this study, the interval region is set to 10.

3.1.2 Backward interval partial least squares

BiPLS (Leardi and Nørgaard, 2004) can be loosely described as a fast automatic way to remove non-relevant spectral regions prior which is widely used for spectral wavelength interval selection. BiPLS divides the spectrum into multiple intervals and finds the optimal interval combination by backward elimination whose removal results in the lowest root-mean-square error of cross-validation of PLS. The maximum number of PLS components is 4 and the data sample is split into 5 groups for cross-validation. The number of sub-intervals depends on the characteristic curve of wheat kernels.

3.1.3 Interval random frog

iRF (Yun et al., 2013) is a PLS-based heuristic group evolution algorithm for spectral wavelength interval selection. A fixed-width moving window is introduced to divide the spectra into sub-intervals. Random frog coupled with PLS is applied to select these intervals. There are six parameters to be initialized of iRF. The interval width w depends on the chemical bandwidth. According to experience, the number of iterations N is set to 1000. In order to make the initial set contain more intervals the number of variables, Q is set to 20. According to the reference, the remaining three parameters θ, ω and η have no significant effect on the results are set to 0.3, 3, and 0.1. And the maximum number of PLS components is 10.

3.2 Convolutional neural network

CNN is a deep learning algorithm used widely which can extract abstract features from the sample data automatically. The architecture of CNN is mainly constituted by convolutional layers, batch normalization layers, activation layers, pooling layers, flattening layers, and full connection layers. The convolutional layer with a series of local receptive fields convolution kernels is generally employed as an excellent feature extractor (Weng et al., 2021). The activation layer with a non-linear activation function is typically applied after the convolutional layer to provide nonlinearity for the network. The pooling layer is a nonlinear downsampling method used to reduce the spatial size or dimension of the output feature map generated by the convolutional layer. The output feature map generated by the convolutional layer is flattened into 1-dimensional data through the flattening layer and input to the full connection layer for classification. In the reference (Santurkar et al., 2018), the author pointed out that the batch normalization layer may make the optimization landscape significantly smoother which will bring prevention of exploding or vanishing gradients, robustness to different settings of hyperparameters such as learning rate and initialization scheme, and keeping most of the activations away from saturation regions of non-linearities. In this study, two CNN classification models with different structures are used for wheat classification and shown in Fig. 4–5. Three convolutional modules were applied in the first CNN classification model in Fig. 4. Each convolutional module includes a convolutional layer, a batch normalization (BN) layer, a ReLU activation layer, and a maxpooling layer. The number of convolutional filters for the first convolution layer is set to 160 with a kernel size of 3×3, a stride of 1, and a padding mode of ‘same’. The second and third convolution layer both have 32 convolutional filters with a kernel size of 3×3, a stride of 1, and a padding mode of ‘same’. HSI of the sample is extracted into a feature map through these convolution models and input to the fully connected layer to achieve the classification effect.

In order to extract the spectral features of different intervals better, a CNN structure combining group convolution is introduced. The second CNN classification model proposed in this article, as shown in Fig. 5, has the same structures as the first one which is shown in Fig. 4 except for the first convolutional layer. The total number of convolutional filters for the first convolution layer is set to 160 which is the same as the first CNN classification model proposed in this article. After the spectral wavelength interval selection, the convolution filters in the first convolution layer are divided into multiple groups with the same number of intervals. Then the output feature maps generated by these convolution filter groups will be spliced into one feature map in the depth dimension by the depth connection layer. Since the spectral information contained in different intervals is generally of different substances, this grouped convolution structure enables the network to extract the information of different substances independently in the first convolutional layer and then merge them in the second and third convolutional layers. This can significantly improve the accuracy of the classification and reduce the network's inference time and parameters.

To improve the credibility of model evaluation, k-fold cross-validation was used to evaluate the CNN classification model. The specific operation process shows as follows: firstly, all samples are randomly and equally divided into k subsets. Take one of the subsets as the validation set during each training, and the remaining subsets as the training set. The cross-validation will be repeated k times to ensure each subset has been verified once. The average performance of k times will be accepted as the result. By employing k-fold cross-validation, more observations can be obtained to achieve a reliable accuracy estimate and the dependency among accuracy estimates can also be reduced (Wong and Yeh, 2019). In this study, 10-fold cross-validation was used to evaluate the CNN classification model.

4.1 Spectral wavelength interval selection results

The raw spectral of the eight types of wheat shown in Fig. 6 (a) is relatively smooth and lacks obvious absorption and reflection characteristics which will lead to the difficulty of directly analyzing whether there are absorption and reflection characteristics in the spectral curve. The spectrum processed by the continuum removal, with obvious absorption characteristics, are similar spectral trends with differences in the magnitude throughout these spectral regions. The primary structural composition and molecules were revealed from the spectral data enhanced by continuum removal. Specifically, the absorption peaks near 421 nm and 486 nm are related to the protein content of wheat kernels (Huang et al., 2016). The absorption peak is identified near 627 nm, and it is mainly related to the absorption of oil and starch in the wheat kernels (Bao et al., 2019). The absorption peak near 911 nm is related to the water content of wheat kernels (Bruning et al., 2019) and the absorption peak near 982 nm is related to the interaction between protein and the O-H bond in water (Singh et al., 2010).

It can be seen from Fig. 6 that the extreme point shift of the same absorption valley of different varieties of wheat kernels generally does not exceed 30nm. In order to ensure as much as possible that each interval selected by BiPLS contains the absorption bands of the same substance in different varieties of wheat, the spectrum is divided into 19 intervals in this study. The width of the absorption valley, that is, the chemical bandwidth of wheat on VNIR is about 60 nm, so the interval width w is set to 10. In this study, iCR, BiPLS, and iRF were introduced to identify characteristic wavelengths that would be employed to develop simplified CNN classification models. The results of the spectral wavelength interval selection procedure are visualized in Fig. 7. The continuum method found that the important bands lie in the range of 402–460 nm, 466–524 nm, 608–666 nm, 879–937 nm, and 943–1000 nm. The BiPLS method selected a range of 402–438 nm, 466–492 nm, 499–534 nm, 531–557 nm, 627–653 nm, 692–718 nm, 724–750 nm, 885–911 nm, 917–943 nm and 950–975 nm. The iRF methods produced high importance scores in the range of 408–466 nm, 479–537 nm, 576–634 nm, 776–834 nm, and 930–988 nm. The intervals found by the above methods all contain parts or all of the absorption properties mentioned in the previous section. Therefore, as a simplification of full wavelengths, these wavelength intervals contained important information about different physical and chemical characteristics and could be combined with CNN classifiers to perform further discriminant analysis.

4.2 Classification results and analysis

The effects of feature selectors based on iCR, BiPLS, and iRF were compared. The classification accuracies with different feature extraction algorithms are summarized in Table 2. The following observations can be found: (1) In structure 1, the overall performances (accuracy rate and inference time) of the selected wavelength intervals are better than that of the full wavelength which means that the introduced prior knowledge can well exclude wavelength with low relevance to classification to maintain a high accuracy rate and reduce inference time. Especially in inference time, after introducing prior knowledge, the average inference time is significantly reduced by 47.56%. (2) After adopting structure 2 there is an average of 3.75% improvement in accuracy compared to structure 1, and an average of 3.87% improvement in accuracy compared with the full wavelength model. At the same time, the reasoning time is further reduced by 4.81% based on the introduction of prior knowledge. This indicates that by introducing the group structure to deal with the selected wavelength intervals, the model could focus on the absorption peaks of physical and chemical characteristics contained in each interval at the first time which leads to an increase in classification accuracy. Grouped convolution can reduce the parameter in the layer employing this structure to 1/n of the original where n is the number of groups.

Table 2

The classification accuracies and Inference time achieved by different feature selectors and classifiers.
Convolutional structure	Variable number	Interval selection method	Acc.^[b]	Inference time^[b] (s)
Structure 1^[a]	94	-	90.96%±1.50%	0.380
	50	iCR	91.83%±0.86%	0.172
	50	BiPLS	90.46%±0.68%	0.199
	50	iRF	90.95%±0.93%	0.170
Structure 2^[a]	50	iCR	95.36%±0.85%	0.168
	50	BiPLS	94.02%±1.24%	0.178
	50	iRF	95.11%±0.80%	0.169
^[a]Structure 1 in convolutional structure represents the structure shown in Fig. 4, and structure 2 is shown in Fig. 5.
^[b]Acc represents the classification accuracy of the 10-fold cross-validation described in section 2.5.5. Inference time(s) is the time that takes for the classification model to judge the varieties of 100 samples.

The relationship between iteration and training accuracy were shown in Fig. 8. The classification model that adopts the grouped convolution structure has a faster convergence speed which converges to 100% classification accuracy around 6000 iterations. The convergence speed of the classification model without grouped convolution structure is slower, and it converges to 100% classification accuracy around 10,000 iterations. The descent of the curve around 5500 iterations and 8250 iterations is caused by the stochastic gradient descent algorithm jumping out of the local optimum.

The overall results demonstrated that it was feasible to combine CNN especially group CNN and spectral prior knowledge for varieties identification of wheat kernels based on hyperspectral images and such methods generally obtained good classification performance. Compared with the performance of original full bands and the optimal band intervals selected based on the prior information, the result illustrates that optimal intervals contain enough information for classification and the structure of grouped convolution can urge the classifier to extract the information distributed in the optimal interval. Trained in an end-to-end manner, CNN combined with spectral prior knowledge can be used as a reliable model for varieties identification of wheat kernels.

This paper aimed to propose a novel approach of using hyperspectral images combined with CNN models for varieties identification of wheat kernels. The classification result of 8 varieties of wheat kernels by the full wavelength CNN model achieved more than 90%, which showed the hyperspectral images with the CNN classification model were feasible for wheat kernels. In order to simplify the CNN model, three spectral wavelength interval selection methods (iCR, BiPLS and iRF) were as prior knowledge first to reduce the input dimension. The optimal interval contained some crucial information about varieties discrimination, and the interval CNN model based on the spectral prior information has a similar classification accuracy to the full wavelength CNN model. Then, two convolutional structures (FC and GC) of CNN models were introduced to extract features. The GC models outperformed all the FC models, confirming that the ability of GC combined with spectral prior knowledge in representing the spectral difference of different varieties of wheat kernels is excellent. It was more conducive to extracting the information distributed in the hyperspectral images, in which the convolution kernel extracted the different features contained in different intervals.

In summary, the GC-CNN models with spectral wavelength interval provided accurate and rapid identification of wheat kernel variety which would help to develop an online system for quality detection of wheat kernels as well as other crop seeds. The combination of hyperspectral imaging, CNN, and spectral prior knowledge will become a powerful tool for large-scale crop seed detection in the modern agriculture industry.

Ethics approval and consent to participate Not applicable

Consent for publication Not applicable

Competing interests The authors have no competing interests to declare that are relevant to the content of this article.

Author contributions Haotian Que and Xin Zhao wrote the main manuscript text, Xiulan Sun worked on formal analysis and investigation, Qibing Zhu and Min Huang completed the analysis and interpretation of data. All authors reviewed the manuscript.

Funding This work was supported by the Fundamental Research Funds for the Central Universities under Grant JUSRP121017, High level project of Jiangsu Province under Grant JSSCBS20210858 and National Nature Science Foundation of China under Grant 61775086.

Availability of data and materials The datasets analysed during the current study are available from the corresponding author on reasonable request.

Bao Y, Mi C, Wu N, et al. (2019) Rapid Classification of wheat grain varieties using hyperspectral imaging and chemometrics. Applied Sciences, 9(19): 4119. https://doi.org/10.3390/app9194119
Bruning B, Liu H, Brien C, et al. (2019) The development of hyperspectral distribution maps to predict the content and distribution of nitrogen and water in wheat (Triticum aestivum). Frontiers in plant science, 10: 1380. https://doi.org/10.3389/fpls.2019.01380
Chatnuntawech I, Tantisantisom K, Khanchaitit P, et al. (2018) Rice classification using spatio-spectral deep convolutional neural network. arXiv preprint, 1805.11491. https://doi.org/10.48550/arXiv.1805.11491
Ge X, Sun J, Lu B, et al. (2019) Classification of oolong tea varieties based on hyperspectral imaging technology and BOSS‐LightGBM model. Journal of Food Process Engineering, 42(8): e13289. https://doi.org/10.1111/jfpe.13289
Huang M, He C, Zhu Q, et al. (2016) Maize seed variety classification using the integration of spectral and image features combined with feature transformation based on hyperspectral imaging. Applied Sciences, 6(6): 183. https://doi.org/10.3390/app6060183
Huang X, Qian M, Xu F. (2012) Nondestructive detection of dried apricots quality based on machine vision and near-infrared diffuse reflectance spectroscopy technology. Transactions of the Chinese Society of Agricultural Engineering, 28(7): 260-265. https://doi.org/10.3969/j.issn.1002-6819.2012.07.043
Huang Z, Turner B J, Dury S J, et al. (2004) Estimating foliage nitrogen concentration from HYMAP data using continuum removal analysis. Remote Sensing of Environment, 93(1-2): 18-29. https://doi.org/10.1016/j.rse.2004.06.008
Itoh Y, Iwasaki A. (2013) Enhancement of hyperspectral unmixing using continuum removal. IEEE International Geoscience and Remote Sensing Symposium, 2013: 445-448. https://doi.org/10.1109/igarss.2013.6721188
Koistinen V M, Hanhineva K. (2017) Mass spectrometry-based analysis of whole-grain phytochemicals. Critical reviews in food science and nutrition, 57(8): 1688-1709. https://doi.org/10.1080/10408398.2015.1016477
Leardi R, Nørgaard L. (2004) Sequential application of backward interval partial least squares and genetic algorithms for the selection of relevant spectral regions. Journal of Chemometrics: A Journal of the Chemometrics Society, 18(11): 486-497. https://doi.org/10.1002/cem.893
Li Y, He L. (2020) An improved hybrid CNN for hyperspectral image classification. Eleventh International Conference on Graphics and Image Processing. International Society for Optics and Photonics, 11373: 113731R. https://doi.org/10.1117/12.2557384
Malenovský Z, Homolová L, Zurita-Milla R, et al. (2013) Retrieval of spruce leaf chlorophyll content from airborne image data using continuum removal and radiative transfer. Remote Sensing of Environment, 131: 85-102. https://doi.org/10.1016/j.rse.2012.12.015
Otsu N. (1979) A threshold selection method from gray-level histograms. IEEE transactions on systems, man, and cybernetics, 9(1): 62-66. https://doi.org/10.1109/tsmc.1979.4310076
Qiu Z, Chen J, Zhao Y, et al. (2018) Variety identification of single rice seed using hyperspectral imaging combined with convolutional neural network. Applied Sciences, 8(2): 212. https://doi.org/10.3390/app8020212
Santurkar S, Tsipras D, Ilyas A, et al. (2018) How does batch normalization help optimization. Proceedings of the 32nd international conference on neural information processing systems, 2018: 2488-2498. https://dl.acm.org/doi/10.5555/3327144.3327174
Singh C B, Jayas D S, Paliwal, J, et al. (2010) Identification of insect-damaged wheat kernels using short-wave near-infrared hyperspectral and digital colour imaging. Computers electronics in agriculture, 73(2): 118-125. https://doi.org/10.1016/j.compag.2010.06.001
Singh T, Garg N M, Iyengar S R S. (2021) Nondestructive identification of barley seeds variety using near‐infrared hyperspectral imaging coupled with convolutional neural network. Journal of Food Process Engineering, 44(10): e13821. https://doi.org/10.1111/jfpe.13821
Sui C, Tian Y, Xu Y, et al. (2014) Unsupervised band selection by integrating the overall accuracy and redundancy. IEEE Geoscience and Remote Sensing Letters, 12(1): 185-189. https://doi.org/10.1109/lgrs.2014.2331674
Tadesse W, Sanchez-Garcia M, Assefa S G, et al. (2019) Genetic gains in wheat breeding and its role in feeding the world. Crop Breed. Genet. Genom, 1: e190005. https://doi.org/10.20900/cbgg20190005
Vaddi R, Manoharan P. (2020) Hyperspectral image classification using CNN with spectral and spatial features integration. Infrared Physics and Technology, 107: 103296. https://doi.org/10.1016/j.infrared.2020.103296
Wang L, Liu D, Pu H, et al. (2015) Use of hyperspectral imaging to discriminate the variety and quality of rice. Food analytical methods, 8(2): 515-523. https://doi.org/10.1007/s12161-014-9916-5
Weng S, Han K, Chu Z, et al. (2021) Reflectance images of effective wavelengths from hyperspectral imaging for identification of Fusarium head blight-infected wheat kernels combined with a residual attention convolution neural network. Computers and Electronics in Agriculture, 190: 106483. https://doi.org/10.1016/j.compag.2021.106483
Wong T T, Yeh P Y. (2019) Reliable accuracy estimates from k-fold cross validation. IEEE Transactions on Knowledge and Data Engineering, 32(8): 1586-1594. https://doi.org/10.1109/tkde.2019.2912815
Xu M, Sun J, Zhou X, et al. (2021) Research on nondestructive identification of grape varieties based on EEMD‐DWT and hyperspectral image. Journal of Food Science, 86(5): 2011-2023. https://doi.org/10.1111/1750-3841.15715
Yang X, Hong H, You Z, et al. (2015) Spectral and image integrated analysis of hyperspectral data for waxy corn seed variety classification. Sensors, 15(7): 15578-15594. https://doi.org/10.3390/s150715578
Yu Z, Fang H, Zhangjin Q, et al. (2021) Hyperspectral imaging technology combined with deep learning for hybrid okra seed identification. Biosystems Engineering, 212: 46-61. https://doi.org/10.1016/j.biosystemseng.2021.09.010
Yun Y, Li H, Wood L R E, et al. (2013) An efficient method of wavelength interval selection based on random frog for multivariate spectral calibration. Spectrochimica Acta Part A: Molecular and Biomolecular Spectroscopy, 111: 31-36. https://doi.org/10.1016/j.saa.2013.03.083
Zhang L, Zhang L, Tao D, et al. (2012) Tensor discriminative locality alignment for hyperspectral image spectral–spatial feature extraction. IEEE Transactions on Geoscience and Remote Sensing, 51(1): 242-256. https://doi.org/10.1109/tgrs.2012.2197860
Zhao W, Du S. (2016) Spectral-spatial feature extraction for hyperspectral image classification: A dimension reduction and deep learning approach. IEEE Transactions on Geoscience and Remote Sensing, 54(8): 4544-4554. https://doi.org/10.1109/tgrs.2016.2543748

No competing interests reported.

Download PDF

Version 1

posted

You are reading this latest preprint version

Identification of wheat kernel varieties based on hyperspectral imaging technology and grouped convolutional neural network with feature intervals

Status:

Version 1

Abstract

Figures

1. Introduction

2. Materials And Methods

2.1 Samples Collection and Preparation

2.2 The Hyperspectral Imaging System

2.3 Hyperspectral Image Preprocessing

3. Data Analysis Methods

3.1 Spectral wavelength interval selection methods

3.1.1 Interval continuum removal

3.1.2 Backward interval partial least squares

3.1.3 Interval random frog

3.2 Convolutional neural network

4. Results And Discussion

4.1 Spectral wavelength interval selection results

4.2 Classification results and analysis

5. Conclusions

Declarations

References

Additional Declarations

Status:

Version 1