Original papers
Feasibility study on identifying seed viability of Sophora japonica with optimized deep neural network and hyperspectral imaging

https://doi.org/10.1016/j.compag.2021.106426Get rights and content

Highlights

  • Hyperspectral imaging was used to achieve testing of Sophora japonica seed viability.

  • Spectral data collection was performed with different viability and swelling states.

  • Compare the changes of components in the seed before and after aging and imbibition.

  • Use particle swarm optimization algorithm to optimize CNN and LSTM.

  • Based on the established multivariate data set model to identify the seed vitality.

Abstract

Aiming at the rapid identification of seed viability of Sophora japonica, different aging grades and imbibition states were proposed as the research focus. The hyperspectral data of the four viability grade seeds in the state of swelling at 0 h and 10 h were collected. The effects of swelling and aging on the content of water, protein, starch, fat and total sugar were compared, and then correlation analysis was established. Among them, fat was the most sensitive to changes in spectrum and viability. In the machine learning scale, naive bayes (NB), self-organizing feature mapping algorithm (SOFM), and support vector machine (SVM) were used. On the scale of deep learning, convolutional neural networks (CNN) and long short-term memory (LSTM) were applied to model on the one-dimensional spectral and two-dimensional image. The image-based deep learning model performed best among the three models, and the recognition accuracy was above 90%. A particle swarm optimization algorithm (PSO) was proposed to optimize the hyperparameters. It can increase the recognition rate, up to 99.73%, and increase the convergence speed. This study proved that deep learning combined with spectral or image has great potential in identifying the seed viability.

Introduction

Seed germination refers to a series of orderly physiological processes and morphogenesis processes starting from imbibition (Xie et al., 2014). The first stage is called imbibition, including rapid water absorption, which will cause the destruction of the cell membrane structure. This process lasts for about 5 h, and the hydrophilic substance inside attracts water molecules, making the seed volume increase rapidly (sometimes more than doubled) (Ion et al., 2012). The second stage is to continuously absorb water, which is characterized by repairing specific molecules, such as DNA and protein synthesis. In the third stage, the cells divide and enlarge, and the nutrients stored in the seeds begin to be consumed until the embryo breaks through the seed coat. Therefore, imbibition seems to be the triggering event that initiates the germination biochemical process and plays a vital role. Most seeds will experience a natural decline in vigor during storage, which is also called seed aging. This process will be accelerated under high temperature and humidity conditions, so artificial aging (Kandpal et al., 2016) is often used in experimental research to simulate the natural aging of seeds. It seems that the water binding within seeds and on seed cells is the main factor affecting seed aging, followed by the protein and lipid modification process caused by oxidation (Agelet and Hurburgh, 2014).

Sophora japonica (Sophora japonica Linn.), a plant of the bean order Papilionaceae, is widely planted in China, especially in the north. In addition to having ornamental value, it is also windproof and sand-fixing. Its flowers and pods are used as medicine to stop bleeding and lower blood pressure; the leaves and root bark have heat-clearing and detoxifying effects; the kernels contain starch, which can be used for winemaking or as paste and feed. Moreover, the seeds can also be squeezed for industrial use. Because of the high vigorous Sophora japonica seeds with rich practical value, they are widely promoted and applied. The hard seed coat is ubiquitous in legumes (Rolston, 1978, Baskin et al., 2000), and Sophora japonica seed is no exception, which makes it difficult to identify seed vigor quickly.

It is difficult, or even impossible, to screen and distinguish viable and non-viable seeds with the naked eye or similar basic methods. Traditional seed vigor detection methods, such as germination test and tetrazolium test (Kusumaningrum et al., 2018), can intuitively obtain seed viability, but they have the disadvantages of expensive, complicated procedures, destructive, and time-consuming. Some emerging fast vitality detection methods based on physics have also been developed and applied, including X-ray (Al-Hammad et al., 2017), nuclear magnetic resonance spectroscopy (Krishnan et al., 2004), Fourier spectroscopy, and Raman spectroscopy (Ambrose et al., 2016a, Ambrose et al., 2016b). But they also have the problems of low efficiency, complex operation and unable to be detected in batch.

Different from the above methods, the hyperspectral imaging (HSI) technology simultaneously obtains the spatial and spectral information of samples in an extensive spectral range, and provides imaging data simultaneously (Li et al., 2020). It is a non-invasive, non-contact, and rapid detection technology. Because different components have different spectrum absorption, the image will have an obvious reflection on the defect at a specific wavelength. In addition, spectral data are helpful to analyze the internal physical structure and chemical composition of samples (Liu et al., 2017). Many researchers have used these characteristics to conduct a series of studies based on hyperspectral imaging with seeds as the object. For example, it has been widely used in soybean, hybrid okra and other seed variety identification (Zhang et al., 2018, Zhu et al., 2020), as well as rice seed infection detection (Wu et al., 2020). In terms of rapid seed vigor detection, although not as adequate as the previous two studies, great progress has also been made. Li et al. (2019) combined support vector machine (SVM) and three variable selection methods to identify the presence or absence of viability soybean seeds with an accuracy of 100%. The wheat seeds were aged to different degrees, and then the three-day, five-day and non-germinated seed spectrum data were collected, respectively. Finally, 8 models for distinguishing wheat grain vigor were established with accuracy rates higher than 84.0% (Fan et al., 2020).

Deep learning is a machine learning technology based on representation learning, which automatically learns and discovers the features required for classification by processing multiple layers of input data (LeCun et al., 2015). Deep Convolutional Neural Networks (DCNN) can extract high-level spatial features and show strong robustness and effectiveness in image classification. Long short-term memory (LSTM) (Ma et al., 2020), a variant of cyclic neural network (RNN), is suitable for processing and predicting important events with long intervals, and satisfactory results have been obtained in time series data clustering. In the field of hyperspectral image analysis and classification, DCNN was first used to classify hyperspectral remote sensing data, and it has been well developed in recent years (Yu et al., 2017). In terms of spectral analysis, deep learning models have also been gradually applied to wheat Fusarium head blight disease detection (Jin et al., 2018), black goji berry chemical composition detection (Zhang et al., 2020) and rice variety identification (Weng et al., 2020). Therefore, it is a good attempt to identify hyperspectral information of Sophora japonica seeds with different vigor by using multiple depth models.

Previous studies on seed viability often focused on crops but less on forestry crops. And they often used hyperspectral imaging to study seed aging, but few seed imbibition. The purpose of this study is to use a variety of deep neural networks and their optimized forms to quickly detect the viability of Sophora japonica seeds in different aging and swelling states. The specific objectives are as follows: (1) to observe the spectral representation of different aging and imbibition seeds and analyze the component changes, and then establish the correlation between spectra and components; (2) to compare the feasibility of different pretreatment methods and typical machine learning methods in identifying seed viability; (3) to study the results of particle swarm optimization (PSO) for optimizing the depth neural network and outputting the optimal hyperparameters; (4) to use multiple deep learning networks to establish seed viability recognition models based on spectra and images.

Section snippets

Seed sample preparation

The seed samples of Sophora japonica used in this study were purchased at the Beijing seed market in October 2019. The seed coat structure of Sophora japonica has the typical characteristics of legumes plants, which seed coat has a palisade layer with tightly integrated cells and has poor water permeability. Therefore, seeds were soaked in water with an initial temperature of 85-90℃ to destroy the seed epidermis. This process continued until the water temperature returned to normal and then

Spectral characteristics and qualitative analysis of PCA

Fig. 4a and 4b show the average spectra of four grades of seeds in the state of imbibition at 0 h and 10 h, respectively. The spectral range used to analyze spectral characteristics was 400–1000 nm. In general, no matter which state, the spectral curves of four viability seeds had the same wave pattern and the peaks and troughs of similar positions, but their reflectance values were different. The internal composition of the seeds was consistent so that they had the same absorption peak in the

Conclusion

In cultivating agricultural and forestry crops, it is very important to determine seed viability quickly and accurately. This paper used hyperspectral imaging and deep learning algorithm optimized by particle swarms to identify different viability Sophora japonica seeds. The spectral and image data of four seed vigor levels under two imbibition states were collected to judge the vigor of these two states. In the spectrum performance, the four vigor levels had a similar trend, but there were

Funding

This study was supported by National Natural Science Foundation of China (Grant No. 31770769), National Natural Science Foundation of China (Grant No. 32171742).

CRediT authorship contribution statement

Lei Pang: Investigation, Methodology, Software, Writing – original draft. Lianming Wang: Data curation, Validation. Peng Yuan: Formal analysis, Visualization. Lei Yan: Funding acquisition, Supervision. Qing Yang: Visualization, Writing – review & editing. Jiang Xiao: Conceptualization.

Declaration of Competing Interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

References (49)

  • H. Shimodaira

    Improving predictive inference under covariate shift by weighting the log-likelihood function

    J. Stat. Plan. Inference

    (2000)
  • G.P. Silva et al.

    Biochemical and physiological changes in Dipteryx alata Vog. seeds during germination and accelerated aging

    S. Afr. J. Bot.

    (2020)
  • C. Wakholi et al.

    Rapid assessment of corn seed viability using short wave infrared line-scan hyperspectral imaging and chemometrics

    Sensor Actuat. B-Chem.

    (2018)
  • S. Weng et al.

    Hyperspectral imaging for accurate determination of rice variety using a deep learning network with multi-feature fusion

    Spectrochim. Acta A

    (2020)
  • N.a. Wu et al.

    Practicability investigation of using near-infrared hyperspectral imaging to detect rice kernels infected with rice false smut in different condition

    Sensor Actuat. B-Chem.

    (2020)
  • Q. Xia et al.

    Integrating proteomics and enzymatic profiling to decipher seed metabolism affected by temperature in seed dormancy and germination

    Plant Sci.

    (2018)
  • M. Xu et al.

    Physiological and proteomic analyses of coix seed aging during storage

    Food Chem.

    (2018)
  • S. Yu et al.

    Convolutional neural networks for hyperspectral image classification

    Neurocomputing

    (2017)
  • C. Zhang et al.

    Developing deep learning based regression approaches for determination of chemical compositions in dry black goji berries (Lycium ruthenicum Murr.) using near-infrared hyperspectral imaging

    Food Chem.

    (2020)
  • L.E. Agelet et al.

    Limitations and current application of near infrared spectroscopy for single seed analysis

    Talanta

    (2014)
  • I. Baek et al.

    Rapid measurement of soybean seed viability using kernel-based multispectral image analysis

    Sensors

    (2019)
  • J.M. Baskin et al.

    Taxonomy, anatomy and evolution of physical dormancy in seeds

    Plant Spec. Biol.

    (2000)
  • H. Chen et al.

    Combination optimization of multiple scatter correction and Savitzky-Golay smoothing modes applied to the near infrared spectroscopy analysis of soil organic matter

    Comput. Appl. Chem.

    (2011)
  • X. Chen et al.

    Visualization model of big data based on self-organizing feature map neural network and graphic theory for smart cities

    Cluster Comput.

    (2019)
  • Cited by (0)

    View full text