Wheat seed viability nondestructive testing method based on hyperspectral technology

Abstract

The invention discloses a wheat seed viability nondestructive testing method and system based on a near-infrared hyperspectral imaging technology. According to the method, near-infrared hyperspectral image data (the wavelength range is 936-1720 nm) of wheat seeds are collected, standardized preprocessing is carried out, and viability prediction is carried out through a constructed seed spectrum converter (SST) deep learning model. The SST model combines a convolutional neural network (CNN) and a self-attention mechanism, can effectively extract local and global features of spectral data, and enhances the training stability and generalization ability of the model through residual connection and layer normalization. Model hyper-parameters are optimized through a scatter diagram and a box plot, optimal configuration is determined through a grid search method or a Bayesian optimization method, and the accuracy and stability of the model are ensured. In addition, the detection accuracy of the model on unknown varieties is evaluated through a cyclic cross validation method, and the final accuracy rate reaches 85.53% or above. Experimental results show that the method can complete detection within several seconds, does not damage the physical structure of the seeds, is suitable for non-destructive detection of the vitality of various wheat varieties, and meets the requirements of the modern seed industry for rapid, non-destructive and high-throughput detection.

Description

Technical Field

[0001] The present invention relates to the technical field of agricultural seed detection, and specifically relates to a method and system for non-destructively detecting the viability of wheat seeds by using near-infrared hyperspectral imaging technology combined with a deep learning model. This method can complete the detection within a few seconds without damaging the physical structure of the seeds, and is applicable to the quality control and breeding screening of wheat seeds. Background Art

[0002] The viability of wheat seeds is an important indicator for measuring the quality and germination potential of wheat seeds. Traditional methods for detecting seed viability, such as germination tests and TTC staining methods, have problems of long time consumption, complex operation, and destructiveness. For example, germination tests usually take 7 - 14 days, and the results are greatly affected by environmental conditions; although the TTC staining method is faster, it will damage the seeds and cannot be used for subsequent planting. These methods are difficult to meet the requirements of the modern seed industry for rapid, non-destructive, and high-throughput detection. In recent years, near-infrared hyperspectral imaging technology has received attention because it can quickly obtain the spectral information of seeds. However, most of the existing technologies are limited to modeling of single or a small number of varieties, resulting in poor performance of the model in cross-variety detection and being difficult to be generalized to different varieties of wheat seeds. Therefore, it is of great practical significance to develop a non-destructive detection method for the viability of multiple wheat varieties. Summary of the Invention

[0003] The present invention aims to provide a method and system for non-destructively detecting the viability of wheat seeds based on near-infrared hyperspectral imaging technology. By constructing a "hyperspectral-viability" dataset containing multiple varieties and combining a deep learning model, rapid and non-destructive detection of the viability of multiple wheat seeds is achieved.

[0004] Technical Solution: Data Acquisition: Use a near-infrared hyperspectral imaging device (such as SPECIM SisuCHEMA) to scan wheat seeds to obtain spectral image data of the seeds, with a wavelength range of 936 - 1720 nm.

[0005] Data Preprocessing: Perform standardization preprocessing (such as Z-score standardization) on the collected spectral data to convert the data into a distribution with a mean of 0 and a variance of 1, so as to enhance the difference between samples and improve the model training efficiency.

[0006] Model Construction: Propose a deep learning model of a Seed Spectral Transformer (SST) based on the Transformer framework. This model combines a convolutional neural network (CNN) to extract local features and a self-attention mechanism to integrate global features to achieve comprehensive feature extraction of seed viability.

[0007] Model Optimization: Analyze the mean and variance of the model accuracy through scatter plots and box plots, and use grid search or Bayesian optimization to tune hyperparameters to determine the optimal hyperparameter configuration.

[0008] Generalization Ability Evaluation: Evaluate the detection accuracy of the model for unknown varieties through cyclic cross-validation methods (such as 5-fold cross-validation) to prove its good generalization ability.

[0009] System Composition: Hyperspectral Imaging Module: Used to collect spectral image data of seeds. The SPECIMSisuCHEMA device is adopted, which can collect high-resolution spectral images in the wavelength range of 936 - 1720 nm.

[0010] Data Processing Module: Used to preprocess and extract features from the collected data, including operations such as standardization and noise reduction.

[0011] Deep Learning Model Module: Predict the viability of seeds based on the SST model, and combine CNN and self-attention mechanisms to extract features.

[0012] Result Display Module: Used to display the detection results of wheat seed viability and provide visual reports and data analysis. Specific Implementation Manner

[0013] Select 17 wheat varieties with different colors, oil contents, and protein contents, and about 900 seeds of each variety as experimental materials. Conduct artificial aging treatment on the seeds, simulate the natural aging process in a high-temperature aging oven at 50°C, set the aging time from 0 to 40 days, with a total of 9 gradients to cover seeds with different aging degrees. Use a near-infrared hyperspectral imaging device (SPECIM SisuCHEMA) to scan the seeds to obtain spectral image data. Preprocess the collected spectral data by standardization to enhance the differences between samples. Construct an SST model, and combine CNN and self-attention mechanisms to extract features. Optimize the model hyperparameters through scatter plots and box plots to determine the optimal configuration. Use the Adam optimizer for model training, set the learning rate to 0.001, and the number of training epochs to 100. Evaluate the detection accuracy of the model for unknown varieties through cyclic cross-validation methods. The experimental results show that the final accuracy of the present invention reaches 85.53%, and the accuracy is increased by more than 10% compared with traditional methods. Description of the Drawings

[0014] Figure 1 : Schematic diagram of the SST model structure.

[0015] Figure 2 : Cyclic cross-validation scheme.

Claims

1. A nondestructive detection method for wheat seed vitality based on near-infrared hyperspectral imaging technology, characterized in that: The method comprises the following steps: a. using a near-infrared hyperspectral imaging device to collect spectral image data of wheat seeds, wherein the wavelength range of the spectral image data is 936-1720 nm; b. performing standardization (SS) preprocessing on the collected spectral data to enhance the differences between samples; c. constructing a wheat seed spectral transformer (SST) deep learning model based on the Transformer framework, wherein the model combines a convolutional neural network (CNN) to extract local features and a self-attention mechanism to integrate global features; d. analyzing the mean and variance of the model accuracy through scatter plots and box plots to determine the optimal hyperparameter configuration; e. using the optimized SST model to predict the vitality of wheat seeds; f. The detection accuracy of the model for unknown varieties was evaluated through cyclic cross-validation, and the final accuracy rate reached more than 85.53%.

2. The detection method according to claim 1, characterized in that: The SST model includes the following structures: a. a convolutional layer for extracting local features in spectral data; b. Embedding layer, which is used to convert local features into higher-dimensional vector representations and embed location information; c. Self-attention layer, used to capture global features in spectral data; d. Residual connections and layer normalization layers are used to enhance the training stability and generalization ability of the model.

3. The detection method according to claim 1, characterized in that: The standardization (SS) preprocessing step involves converting the feature values ​​of the spectral data into standardized data with zero mean and unit variance to accelerate the model training process and improve prediction accuracy.

4. The detection method according to claim 1, characterized in that: The cyclic cross-validation method comprises the following steps: a. dividing seed data of different varieties into training sets and test sets; b. gradually increasing the number of varieties in the training set to evaluate the detection accuracy of the model for unknown varieties in the test set; c. determining the generalization ability of the model for unknown varieties through multiple cyclic validations.

5. The detection method according to claim 1, characterized in that: The near-infrared hyperspectral imaging device is a SPECIMSisuCHEMA device, which can collect high-resolution spectral images within the wavelength range of 936-1720 nm.

6. The detection method according to claim 1, characterized in that: The model optimization step uses a grid search method or a Bayesian optimization method to perform hyperparameter tuning to determine the optimal hyperparameter configuration.

7. The detection method according to claim 1, characterized in that: The SST model was trained using the Adam optimizer, with the learning rate set to 0.001 and the number of training rounds being 100.

8. The detection method according to claim 1, characterized in that: The wheat seeds include 17 varieties with different colors, oil contents and protein contents, with about 900 seeds of each variety.

9. The detection method according to claim 1, characterized in that: The wheat seeds are artificially aged at a temperature of 50° C. and an aging time ranging from 0 to 40 days, with a total of 9 gradients.

10. The detection method according to claim 1, characterized in that: The detection method is applicable to non-destructive detection of the vitality of various wheat varieties, and meets the needs of the modern seed industry for rapid, non-destructive, and high-throughput detection.

Images