GCT-BCLN 水稻种子活力双向闭环高光谱无损检测网络(Frontiers in Plant Science 2026-08-17)

Abstract

Rice seed vigor is a key determinant of germination performance and final crop yield, making its rapid and non-destructive assessment essential for seed quality evaluation. Conventional vigor detection methods are often destructive, labor-intensive, and time-consuming. Hyperspectral imaging provides a promising non-destructive alternative, but hyperspectral data are typically high-dimensional, redundant, and susceptible to noise and scattering interference. Moreover, existing models still have limited ability to discriminate subtle spectral differences among seed vigor levels. To address these challenges, this study proposes a gated recurrent unit (GRU)-guided closed-loop CNN-Transformer network (GCT-BCLN) for accurate, non-destructive identification of rice seed vigor. The model establishes bidirectional information flow between CNN and Transformer via the GRU, enabling dynamic and synergistic optimization of local spectral features and global spectral representations. In addition, a combined preprocessing strategy integrating adaptive iteratively reweighted penalized least squares (AirPLS), Savitzky–Golay (SG) smoothing, and multiplicative scatter correction (MSC) was adopted to improve spectral quality. Experimental results showed that GCT-BCLN achieved a test accuracy of 0.9795 for hybrid indica rice, outperforming the CNN-Transformer fusion model by 1.37%. The model also achieved accuracies of 0.9793 and 0.9758 on conventional japonica rice and glutinous japonica rice, respectively, showing consistent performance across the three evaluated variety-specific datasets under the controlled experimental protocol. These results support the feasibility of GCT-BCLN for laboratory-scale, non-destructive discrimination of aging-induced rice seed categories under controlled conditions, while practical application requires further external validation.

1 Introduction

Rice is one of the world’s most important staple crops, and its yield and grain quality are closely related to global food security (; ). In crop production, seed quality is a key determinant of field performance, among which seed vigor is a crucial physiological indicator affecting germination, seedling establishment, and ultimately crop productivity (). High-vigor seeds, characterized by stronger stress tolerance and growth potential, contribute to stable and high yields (). Therefore, accurate and efficient assessment of rice seed vigor is of great significance for seed breeding, processing, and storage management. Conventional methods, including standard germination tests, electrical conductivity measurements, and tetrazolium staining, can differentiate seed vigor to some extent, but they are generally limited by laborious procedures, long detection cycles, and destructive sampling requirements (). With the development of intelligent and precision agriculture, there is an increasing demand for non-destructive and accurate methods for rice seed vigor assessment.

Recent advances in hyperspectral imaging (HSI) have provided a promising non-destructive technique for seed vigor evaluation (). By combining spatial information, such as seed morphology and texture, with spectral signatures associated with molecular vibrations of internal biochemical constituents, HSI enables seed quality assessment without physical damage (). Its potential has been demonstrated in several seed vigor studies. For example, applied HSI to detect rice seed vigor under sample-imbalanced conditions. used HSI to predict the vigor of naturally aged Xishuangbanna cucumber seeds, and achieved non-destructive detection of maize seed vigor using a data-fusion-driven HSI approach. Nevertheless, hyperspectral data are typically characterized by high dimensionality (), substantial redundancy (), susceptibility to noise (), and sample-specific variability, all of which hinder robust feature extraction and modeling. Conventional pipelines combining spectral preprocessing methods, such as first derivative (FD) (), Savitzky-Golay (SG) smoothing (), and standard normal variate (SNV) transformation (), with traditional machine learning models, including partial least squares regression (PLSR) (), support vector machine (SVM) (), and Random Forest (RF) (), can improve seed vigor prediction to some extent. However, these methods usually depend on manual feature engineering and often fail to capture the complex nonlinear relationships between spectral information and seed vigor, resulting in limited generalization and reduced sensitivity to subtle vigor differences.

Deep learning, with its end-to-end feature learning capability, has shown strong potential for modeling high-dimensional and highly correlated hyperspectral data. In seed vigor detection, proposed a multi-scale spectral attention residual network (MSARN) for common bean seed vigor assessment, achieving accuracy, precision, recall, and F1-score values of 98.75%, 98.97%, 98.80%, and 98.81%, respectively. combined CNNs with a long short-term memory (LSTM) network to model the vigor of Styphnolobium japonicum seeds, achieving recognition accuracy above 90%. Despite these advances, conventional CNNs are limited by their kernel-size-dependent receptive fields and tend to emphasize local spectral patterns while underrepresenting long-range spectral relationships. This limitation restricts their ability to model complex spectral signatures associated with seed aging and biochemical heterogeneity. To jointly capture local and global information, recent studies have introduced Transformer architectures, which model long-range dependencies through self-attention. developed a CNN-Transformer fusion model for rice variety classification using reconstructed hyperspectral images, achieving 99.413% accuracy and markedly outperforming a standalone CNN. proposed a dual-feature transformation fusion transformer (DFTFT), which explicitly models global spectral associations for wheat and maize leaf disease and quality grade classification, improving classification accuracy by 5.2%–7.8% compared with single-CNN baselines. designed an interpretable lightweight CNN-Transformer hybrid model for rapid rice leaf disease identification, achieving an accuracy of up to 97%.

Although CNN-Transformer hybrid models have shown advantages over conventional CNNs in crop hyperspectral recognition and classification, rice seed vigor detection remains challenging. First, spectral differences among seeds at different aging stages are often subtle, requiring models with strong sensitivity to fine inter-class variations. Second, the curved surface morphology of rice seeds induces scattering heterogeneity, causing spectra from seeds with the same vigor level to be affected by sample-specific physical differences. Third, varietal diversity leads to differences in spectral baselines and characteristic peaks, further complicating feature learning. Most existing CNN-Transformer fusion models adopt cascaded serial architectures (), in which features are transmitted unidirectionally and statically. In such designs, local feature extraction is weakly guided by global spectral representations, while global modeling provides limited feedback for refining local features. This restricts model robustness under subtle vigor differences and spectral interference. Therefore, dynamic and synergistic optimization of local and global features is needed to improve the adaptability of hyperspectral models for rice seed vigor assessment. Bidirectional feedback has been shown to enhance feature interaction in other fields. proposed a lightweight bidirectional feedback network for image super-resolution, reducing parameters while improving reconstruction quality. incorporated encoder–decoder bidirectional transmission with dense connections into U-Net (BiDFDC-Net), improving skin lesion segmentation. In remote sensing, proposed a dual-stream spectral interactive network with multi-scale enhancement, cross-attention complementarity, and cross-layer interaction for salient object detection. However, such bidirectional feedback mechanisms remain rarely explored in agricultural hyperspectral analysis, particularly for seed vigor assessment. Collectively, these studies indicate that reciprocal feature transmission can improve representation refinement, but existing approaches have not yet provided a stateful feedback mechanism for iterative interaction between local and global representations of wavelength-ordered hyperspectral signals.

To address this limitation, this study proposes a GRU-guided bidirectional closed-loop CNN-Transformer network (GCT-BCLN) for rice seed vigor identification. In this framework, the term bidirectional describes the two-way information flow between the CNN and Transformer: CNN-derived local spectral features are propagated forward for global dependency modeling, whereas Transformer-derived global representations are fed back to refine CNN-based local feature extraction. By contrast, the term closed-loop refers to the recurrent use of this two-way pathway, whereby feedback-refined local features re-enter the GRU-Transformer pathway in subsequent iterations until the prediction distribution stabilizes. The GRU is adopted as the feedback carrier because hyperspectral bands form an ordered sequence, and its gated hidden state can retain wavelength-dependent contextual information while selectively integrating the current global representation with information from previous feedback iterations. Through this iterative local-global interaction, GCT-BCLN is designed to enhance the representation of subtle vigor-related spectral differences and improve robustness to spectral interference. The proposed framework provides a useful methodological basis for non-destructive rice seed vigor assessment using hyperspectral data.

2 Materials and methods

To systematically achieve non-destructive rice seed vigor detection, this study established an integrated workflow encompassing sample preparation, hyperspectral data acquisition and processing, spectral preprocessing, and model construction, as illustrated in Figure 1. Sample preparation generated seed samples with different vigor levels for subsequent data acquisition. Hyperspectral imaging and spectral preprocessing provided high-quality spectral inputs for model development, while standard germination tests were conducted to establish reference labels for vigor assessment. Finally, the proposed model was constructed to map spectral features to corresponding vigor levels. The operational procedures and principles of each stage are described in detail in the following sections.

Figure 1

2.1 Preparation of rice seed samples with different vigor levels

This study employed an artificial accelerated aging method () to prepare rice seed samples with aging-induced vigor categories under controlled laboratory conditions. Three rice varieties were used in this study, including hybrid indica rice, conventional japonica rice, and glutinous japonica rice. All seeds were harvested in 2024 and provided by the Rice Research Institute of the Anhui Academy of Agricultural Sciences, China. For each variety, diseased, damaged, and empty grains were removed, and 1,200 plump, healthy seeds with uniform morphology were selected. Therefore, a total of 3,600 individual seeds were included in the experiments. The hybrid indica rice dataset was used as the primary dataset for model development and comparative evaluation, whereas the conventional japonica rice and glutinous japonica rice datasets were used as additional variety-specific evaluation datasets. Before aging, the seeds were surface-sterilized by immersion in 1% sodium hypochlorite (NaClO) solution for 15 min, thoroughly rinsed with sterile deionized water at least five times, and air-dried at room temperature until no visible surface moisture remained. Accelerated aging was conducted under controlled high-temperature and high-humidity conditions (). Briefly, 500 mL of sterile water was added to a glass desiccator, which was then placed in a dark incubator maintained at 45°C. After the internal environment stabilized at 45°C and approximately 100% relative humidity, the seeds were placed in breathable nylon mesh bags, arranged flat on a perforated porcelain plate inside the desiccator, and exposed to the aging treatment. For each rice variety, the 1,200 seeds were equally divided into six aging-duration groups, with 200 seeds in each group. Samples were collected after 0 d (control), 1 d, 2 d, 3 d, 4 d, and 5 d of aging to generate seed lots with a controlled vigor deterioration gradient. Temperature and relative humidity inside the desiccator were continuously monitored to ensure treatment consistency. After aging, seeds from each treatment were transferred to room-temperature conditions for moisture equilibration, thereby minimizing the influence of moisture differences on subsequent hyperspectral measurements and germination testing. It should be emphasized that the accelerated-aging protocol was used to generate a controlled and reproducible vigor gradient for methodological development rather than to reproduce the complete deterioration process of naturally aged commercial seed lots. The six classification targets in this study were defined according to aging duration rather than commercially defined vigor grades. Therefore, the current dataset should be interpreted as a laboratory-scale dataset for methodological evaluation rather than as a direct representation of naturally aged commercial seed lots.

2.2 Hyperspectral data acquisition

Hyperspectral images of aged rice seeds were acquired using a FigSpec FS-20 hyperspectral imaging system, as illustrated in Figure 2. The system acquired 300 spectral bands over the wavelength range of 400–1000 nm, corresponding to a nominal spectral sampling interval of approximately 2.0 nm. The spatial dimensions of each hyperspectral image were 1920 × 1920 pixels. For each aging treatment, 200 seeds were imaged in two batches. In each batch, 100 seeds were arranged without overlap in a 10 × 10 matrix on a black matte background to minimize background reflection and facilitate subsequent ROI extraction. Illumination was provided by two 100 W tungsten-halogen lamps positioned symmetrically on both sides of the sample platform at an angle of approximately 60°. The hyperspectral camera was mounted vertically above the samples, with a lens-to-sample working distance of 30 cm. The integration time was set to 20 ms. Before image acquisition, the lamps were warmed up for 15 min to ensure stable illumination, and all acquisition parameters were kept constant throughout the experiment. Before each acquisition batch, white- and dark-reference images were collected for radiometric calibration. The white-reference image was acquired using a polytetrafluoroethylene (PTFE) diffuse-reflectance panel with a nominal reflectance of approximately 99%. The reference panel was placed under the same imaging geometry and camera settings as the seed samples. The dark-reference image was acquired with the camera lens completely covered. To reduce the effects of dark current, non-uniform illumination, and sensor noise, the raw hyperspectral images were radiometrically calibrated using the manufacturer-provided software (). The corrected reflectance was calculated according to (Equation 1). After calibration, all subsequent data processing was performed using ENVI 5.6 software. For each seed, a fixed 10 × 10-pixel square region of interest (ROI) was manually positioned within the visually homogeneous central endosperm region, with its center located near the longitudinal axis of the seed. To reduce edge scattering and background interference, the ROI boundary was kept at least 3 pixels inside the visible seed contour. The embryo-bearing basal one-third of the seed, shadowed pixels, and visibly damaged areas were excluded. When positional adjustment was necessary, the ROI was translated within the central endosperm region without changing its dimensions. The mean reflectance spectrum of the 100 pixels within the ROI was used to represent the corresponding seed and was subsequently exported for spectral preprocessing and modeling. The same ROI-selection criteria were applied to all seeds and aging treatments.

Figure 2

(1)

$$
R = \frac{R_{\text{O}} - R_{\text{D}}}{R_{\text{W}} - R_{\text{D}}}
$$

(1)

where $R$ is the calibrated reflectance image, $R_{\text{O}}$ is the original hyperspectral image of the rice seeds, $R_{\text{D}}$ is the dark-reference image acquired with the camera lens completely covered, and $R_{\text{W}}$ is the white-reference image acquired from the PTFE reference panel under identical acquisition conditions.

2.3 Standard laboratory germination test

To quantify the physiological decline caused by accelerated aging, a standard laboratory germination test was conducted with reference to the principles of the International Rules for Seed Testing (). For each aging treatment, 150 seeds were tested in three replicates, with 50 seeds per replicate. Seeds were placed on moistened filter paper in 90-mm Petri dishes and incubated in a controlled-environment chamber at 30 ± 0.5°C and 60 ± 5% relative humidity (RH) for 14 days. Sterile distilled water was added daily to maintain adequate moisture, and germination status was recorded daily. A seed was considered germinated when the radicle protruded through the seed coat and reached a length equal to or greater than half of the seed length (). After 14 days of incubation, the germination percentage was calculated for each replicate, and the mean value of the three replicates was used as the final germination percentage for the corresponding aging treatment. The six class labels used for spectral modeling corresponded to the aging durations (0–5 d), whereas germination percentage was used as a group-level physiological indicator to verify the progressive decline in germination performance. Therefore, the classification results should be interpreted as discrimination among aging-induced vigor categories under controlled laboratory conditions, rather than as direct measurements of the absolute vigor of naturally aged commercial seeds.

2.4 Dataset partitioning and spectral preprocessing

For each rice variety, 1,200 seed-level spectra were obtained, corresponding to six aging-duration categories, with 200 spectra in each category. The hybrid indica rice dataset was used as the primary dataset for model development and comparative evaluation, whereas the conventional japonica rice and glutinous japonica rice datasets were used as additional variety-specific evaluation datasets. To ensure balanced class distribution and reduce evaluation bias, the 200 spectra from each aging-duration category were randomly divided into training, validation, and test subsets at a ratio of 7:2:1, yielding 140, 40, and 20 spectra per level, respectively. Consequently, each variety-specific dataset contained 840 training spectra, 240 validation spectra, and 120 test spectra. The global random seed was fixed at 42 during dataset partitioning and model weight initialization to improve experimental reproducibility. Five-fold cross-validation was conducted only on the training set of the primary hybrid indica rice dataset to support hyperparameter tuning and assess training stability. The validation set was used for model selection, early stopping, and ablation analysis, whereas the test set was strictly reserved for final performance evaluation. For the conventional japonica rice and glutinous japonica ri