Due to the high quality and the resistance to multiple stressors, “Jingke 968” maize is extensively cultivated in eastern and northern China. Therefore, it was selected as the experiment sample in this study. To ensure the accuracy of the experiment, seeds with uniform size and shape were manually selected, then all seeds were disinfected by soaking them in a 0.5% sodium hypochlorite solution for 5 minutes, followed by rinsing with distilled water five times, and air-dried under natural conditions.

To simulate the natural aging process of seeds, the experiment samples were artificially aged. All seeds were exposed to high temperature and high humidity conditions (45 °C and a relative humidity of 95%) and stirred twice a day to ensure uniform exposure (Zhang et al., 2020). 150 maize seeds were taken out randomly at aging 2, 4, 6, and 8 days, respectively. Additionally, 150 untreated seeds were selected as the control group (CK). Therefore, a total of 750 maize seeds within five aging stages were obtained and used for subsequent experimentation.

Two HSI systems, the Vis-SWNIR and LWNIR, have been built in the Intelligent Detection Laboratory of the China Agricultural Equipment Technology Research Center (Fan et al., 2018). The Vis-SWNIR system is capable of acquiring hyperspectral images within the wavelength range of 327-1098 nm, encompassing 1000 spectral variables, while the LWNIR system can capture images within the range of 930-2548 nm, containing 256 spectral variables. The Vis-SWNIR system includes an imaging spectrometer, an electron-multiplying charge-coupled device camera with a resolution of 502×500, a camera lens, and a spectraCube data acquisition software. Similarly, the LWNIR system includes an imaging spectrometer, a charge-coupled device camera with a resolution of 320×256, a camera lens, and a spectral acquisition software (Tian et al., 2021). And the acquisition software of both systems was developed using LabVIEW (National Instruments Inc., Austin, TX, USA) to facilitate the acquisition of spectral images, as well as to manage the camera and motor operations. Both systems share two 300-watt halogen lamps to provide stable illumination. In addition, an electrically operated moving platform and a computer are available for sample placement (Capable of accommodating up to 96 samples simultaneously) and hyperspectral image acquisition (Figure 1A) (Liu et al., 2022).

To ensure the accuracy and reliability of the hyperspectral images (E_raw ), calibration operation is essential to eliminate the effects of uneven illumination of the light source and camera dark current changes (An et al., 2022). The calibration operation involved using a white reflection board (with a reflectance of 99%) (E_w ) to acquire a standard white reference image in the same sampling environment as the sample, while turning off the light source and covering the lens to obtain a black reference image (with a reflectance of 0%) (E_d ). The calibrated image can be calculated using the following formula:

After calibration, in the Vis-SWNIR region, a subset of 347 spectral variables within the 420-1000 nm range was selected for further analysis, considering the abundance of spectral data and the presence of duplicate information in adjacent spectra. On the other hand, in the near-infrared region, due to the limited number of available bands, all spectral variables (256) were directly included in the analysis. To separate maize seeds from the background, a mask was applied to segment the hyperspectral image. The gray-scale images at 801 nm and 1098 nm were selected as the mask images for the Vis-SWNIR and LWNIR bands, respectively. The average spectral curves were obtained by calculating the mean reflectance under the mask. Lastly, in order to eliminate the influence of the instrument, the Savitzky-Golay (SG) and Standard Normal Variate (SNV) methods were utilized to preprocess the spectra.

A transparent box measuring 25cm×25cm was used as a germination chamber, and 75 seeds were placed in each box. A total of 10 boxes were used in the experiment. Prior to the germination test, the germination boxes were sterilized with 75% ethanol (Suksungworn et al., 2021), and three layers of gauze were placed in each germination box to provide continuous moisture for the seeds. A black gauze was placed on the top layers as the background for photography (Figure 2A). An equal amount of distilled water was added to each box, and the temperature was set to 25°C with 12-hour intervals of light and dark (Figure 2B). Throughout the 7-day germination experiment (Long et al., 2022), the germination progress of maize seeds was monitored daily at specific time intervals. According to the ISTA standard, the germination rate was determined (Wang et al., 2022c).

RGB images of maize seeds were captured using BASLER industrial cameras (acA1920-25um/uc, BASLER AG, Germany, 2.4 MP,100 fps) during germination test (Figure 2D) (Shen et al., 2023). An adjustable camera platform was built to ensure consistency of the images and prevent camera shake. The position of the germination box relative to the lens was kept fixed during each image capture. Indoor lighting was turned on and curtains were drawn for each capture. After placing the seeds into the boxes (Day 0), images of each box were immediately captured. Subsequently, images were captured every 15 hours for 7 consecutive days (Figure 2C). The dataset used in this study consisted of a total of 3000 maize seeds (All the captured RGB images collectively contain 3000 seeds). Among them, 2250 seeds were designated as training samples, while the remaining seeds were allocated to the test set.

Hyperspectral data typically consists of numerous bands, and certain bands may exhibit high correlation or contain redundant information (Han et al., 2022).When training 3DCNN with full-band data, it will lead to a significant increase in the number of networks training parameters, resulting in a more complex model. This phenomenon is commonly referred to as the curse of dimensionality. (Köppen, 2000). However, band selection (Sun and Du, 2019) allows retaining spectral bands that are closely related to seed vigor assessment while removing irrelevant bands, thereby enhancing the feature extraction and discriminative capabilities of the model.

Additionally, the use of dimensionality reduction data sets can effectively reduce the complexity of the model, mitigating the risk of overfitting and enhancing the model’s generalization ability and stability (Aloupogianni et al., 2023). Moreover, fewer computing resources are required during model training and inference, leading to a significant improvement in the computational efficiency of the model (XingJia et al., 2022).

Successive projections algorithm is a classical band selection method that can map high-dimensional spectral data to a low-dimensional space through multiple projections(de Almeida et al., 2018). SPA is a forward iterative search method used for selecting spectral information with minimal redundancy to address collinearity issues. The steps of SPA are shown in Table 1. The SPA is widely used in hyperspectral image processing attributed to its advantages of fast computation speed and easy implementation (Chen et al., 2023b). Therefore, the SPA was used in this study to perform feature selection on the processed average spectra of Vis-SWNIR and LWNIR, in order to perform dimensionality reduction on the hyperspectral data.

Support Vector Machine (SVM) (Cortes and Vapnik, 1995) is a powerful algorithm for classification and regression that finds an optimal hyperplane to separate data points of different classes. It handles high-dimensional datasets, avoids overfitting, and can handle non-linear problems using kernel functions. K-Nearest Neighbor (KNN) (Zhang, 2022) is a basic algorithm that selects the K nearest samples based on their feature values and uses their labels as predictions. Subspace Discriminant Analysis (SDA) (Zhao and Phillips, 1999) is a pattern classification method that aims to find a low-dimensional subspace to maximize the separation between different classes. In this study, the aforementioned machine learning methods were used to classify the viability of maize seeds at different aging stages for optimal classification accuracy.

The CNN combines the concepts of convolutional filtering and neural networks by utilizing local receptive fields and weight sharing to reduce the number of network parameters and speed up model training (Ghaderizadeh et al., 2021). Compared to the widespread use of two-dimensional convolution, three-dimensional convolution is less commonly used in practice. However, HSI contain rich spectral information, and using two-dimensional convolution may make the interband correlations of HSIs underutilized (Ge et al., 2020). To address this issue, this study introduced a 3DCNN, which can thoroughly extract feature relationships across different feature channels (Figure 1C), thereby enabling it to concurrently extract integrated spectral and spatial features from hyperspectral imagery (Sun et al., 2022).

Before inputting hyperspectral images into the network, standardization is performed to ensure that the data is within the same scale and range, enabling the network to learn weights faster and converge more easily during training. Moreover, data standardization can help avoid the problems of gradient disappearance or explosion, and improve the stability and generalization ability of the network. To obtain multiple convolutional features of HSI, multi-scale convolution is employed in the same convolutional layer, which can acquire both global and local information. Four different convolution kernels of 3×3×3, 3×3×5, 3×5×5, and 5×5×5 were selected to extract feature information and fused on the channel. This method can enhance the classification accuracy of the model. As illustrated in Figure 1D, each convolution kernel in the first convolution module has 16 filters, each kernel in the second convolution module has 32 filters, and each kernel in the third convolution module has 64 filters. The activation function in the three-dimensional convolution module uses Rectified Linear Unit (RELU) and is compressed by the pooling layer to reduce the amount of data and parameters, as well as alleviate the overfitting phenomenon. To ensure that the features extracted by different convolution kernels in the same module can be effectively connected, different parameters need to be set according to different situations, such as stride and padding. Finally, the output is produced through 1 fully connected layer and 1 output layer, and the output layer employs the SoftMax activation function.

To extract features from hyperspectral images of maize seeds at a more microscopic level and increase the amount of data, a window size of 5×5 was selected for segmentation (Figure 1B). To eliminate the influence of background on classification, small blocks containing 0-pixel points were discarded. As the size of maize seeds varies, the number of blocks obtained from different segments of maize seeds is also inconsistent. To address this issue, this study employed a majority principle labeling aggregation method, as Table 2.

In this study, the germination experiment showed that 404 viable samples and 346 nonviable samples were collected from 750 seeds. Given that the hyperspectral images were collected in a sequential manner based on the aging gradients of the seeds, it was crucial to maintain a balanced distribution of germinated and non-germinated samples in the test set. Therefore, a representative test set was carefully selected, consisting of 75 seeds, including the first seed, the 10th seed, the 20th seed, and so on. The remaining 675 seeds were allocated for the training phase. Through this meticulous approach, it was ensured that the test set encompassed samples from diverse categories, enabling an accurate evaluation of the classification model’s performance.

In order to measure the length of maize seed bud, the Mask R-CNN (He et al., 2017) (With resnet50_fpn as backbone) model was utilized to segment the bud from single-seed image firstly, then a skeleton extraction algorithm was applied to extract the skeleton of the bud (Figure 3A). Next, the bud length detection algorithm was used to remove the branches in the skeleton for obtaining the central skeleton image. Finally, the actual bud length was calculated by converting pixels to actual length (Figure 3B).

Mask R-CNN is a deep learning model that combines object detection and instance segmentation. It extends Faster R-CNN by generating binary masks for each region of interest (ROI), achieving pixel-level segmentation. The network consists of three main components: a backbone network, a Region Proposal Network (RPN) responsible for generating candidate object regions, and two parallel branches dedicated to object detection and mask prediction. Mask R-CNN excels in instance segmentation, object detection, and keypoint detection, making significant contributions to computer vision advancements (Casado-García et al., 2019). The model employs a multi-task loss function, comprising classification loss (L_cls ), bounding box loss (L_bbox ), and predicted mask loss (L_mask ), as represented by equations (2) to (5) (Cong et al., 2023).

L_cls measures the deviation between predicted and actual values for overall accuracy assessment. L_bbox quantifies the disparity between predicted and actual position parameters, assessing the model’s accuracy in bud localization. L_mask evaluates the model’s confidence in pixel-level classification using binary cross-entropy. Combining these components into a multi-task loss function allows for comprehensive evaluation across multiple tasks, resulting in enhanced overall performance.

The skeleton extraction algorithm is a technique used to extract the central line or skeleton of an object in a binary image (Fu et al., 2023). By progressively shrinking connected regions within the object contour, the algorithm produces a concise contour that provides valuable information for image processing tasks like recognition and matching. Various algorithms, such as Zhang-Suen, Morphological Thinning, and Medial Axis Transform, can be employed for this purpose. The Medial Axis Transform (MAT) algorithm, specifically, extracts the object’s central line by iteratively dilating boundary pixels and identifying the nearest internal pixels as skeleton pixels. This process continues until the skeleton pixels stabilize, resulting in a stable and versatile representation suitable for subsequent image processing tasks. The MAT algorithm handles different object shapes and can process grayscale information within binary images. Seed germination images exhibit a wide range of shape features, such as bud length, curvature, and angle. However, traditional methods for measuring bud length rely on manual measurements, which are time-consuming and prone to significant subjective biases. The MAT (Medial Axis Transform) skeleton extraction algorithm was chosen to obtain the central line of buds. However, the resulting skeleton may contain branches that need to be eliminated to derive the center skeleton. The process of centerline skeleton extraction is illustrated in the following Figure 3B.

In this study, a transparent box with a side length of 250 mm was used as a reference to convert pixels to actual lengths in millimeters. The calculation formula is:

Here, L_box represents the side length of the transparent box, and 1164 is the number of pixels corresponding to the transparent box in the image. According to the calculation formula, it can be derived that one pixel corresponds to 0.215 mm.

The seed quality detection methods such as germination and staining techniques are time-consuming and rely heavily on human intervention, which may lead to inaccurate results due to human error. In order to develop an automated and standardized method for detecting seed germination that is efficient, accurate, and reliable, the YOLOv7 (Wang et al., 2022b) object detection algorithm was selected in this study, which is one of the most widely used algorithms for object detection since its release in 2015 (Dewi et al., 2023). YOLOv7 is a real-time object detection algorithm (Soeb et al., 2023), which has evolved from YOLOv5 and has faster inference speed, improved detection accuracy, and reduced computational complexity. The algorithm consists of three main parts: the input layer, backbone layer, and output layer (Tang et al., 2023), and uses either a loss function with or without an auxiliary training head (Zhou et al., 2023).

The loss function is used to update the gradient loss during the training process (Cai et al., 2023). The YOLOv7 algorithm is evaluated using various metrics such as precision, mAP, recall, and F1 score (Zhao et al., 2023), and curves such as the F1-Confidence curve, precision-confidence curve, recall-confidence curve, and precision-recall curve are used to optimize the algorithm’s performance and achieve the best balance between precision and recall.

This study utilized a self-built dataset of maize seeds, comprising images of seeds from various angles and sizes, each with corresponding labels in YOLO format. The data collection and preprocessing process was conducted using the same method as Mask R-CNN. The dataset used in this study consisted of a total of 7000 maize seeds. Among these, 4200 seeds were designated as training samples, 1400 seeds were allocated for the test sets, and the remaining seeds were assigned to the validation sets. To enhance the accuracy and robustness of the model, the YOLOv7.pt (https://github.com/WongKinYiu/yolov7) pretrained weights provided by the official website were employed for training. These weights were trained on a large-scale dataset, which can significantly reduce the training time while improving the training effect. The Adam optimizer, a widely used optimizer that can optimize at different learning rates, was used to update the model parameters during training. The parameters of the Adam optimizer were adjusted based on the size of the learning rate in the training process to achieve better training results. A batch size of 2 and a training iteration of 300 were used in this study.