An intelligent crop seed vitality detection method and system

Abstract

The present invention proposes an intelligent method and system for detecting the vitality of crop seeds. The method includes: collecting and preprocessing seed data; extracting the appearance, spectrum, hardness and internal quality characteristics of the seeds to generate a feature set. The feature set is input into a vitality prediction model to obtain a vitality prediction value, and the classification threshold is dynamically adjusted to classify seed vitality. The present invention integrates four features to construct a high-dimensional feature set, and accurately characterizes the physiological activity and mechanical integrity of seeds through nanoindentation and X-ray CT technology. Error feedback and optimization methods are introduced to respond to data changes in real time and improve classification accuracy. Micron-level quantification is achieved through X-ray CT tomography, which improves the accuracy of embryo development assessment. A time series neural network is constructed to capture the dynamic relationship between features, predict the downward trend of seed vitality in advance, and have better long-term prediction stability.

Description

Technical Field

[0001] The present invention relates to the technical field of seed vigor detection, and in particular to an intelligent crop seed vigor detection method and system. Background Art

[0002] In agricultural production, seed vigor directly affects crop yield and quality, so rapid and accurate seed vigor detection is a key link in ensuring agricultural benefits. Traditional detection methods mainly rely on manual observation, germination tests, or physical and chemical analysis of a single indicator, which has the problems of low efficiency, strong subjectivity, high destructiveness, and difficulty in comprehensively evaluating the multi-dimensional characteristics of seeds. For example, conventional hardness tests use mechanical crushing methods, which can easily damage seeds and can only provide local data; internal quality inspections often rely on anatomical observations or low-resolution imaging, which cannot obtain three-dimensional structural information non-destructively; vitality prediction models are often based on a single spectral or morphological feature, ignoring the correlation between multimodal data, resulting in limited prediction accuracy. In addition, the classification threshold mostly uses a fixed threshold, which is difficult to adapt to the dynamic characteristics of different varieties and batches of seeds, and is prone to misjudgment due to environmental or sample differences.

[0003] In recent years, with the development of intelligent detection technology, some studies have attempted to improve detection accuracy through multi-source data fusion. For example, spectral analysis can reflect the chemical composition of seeds, but is limited to single-wavelength information. While X-ray imaging can capture internal morphology, two-dimensional images make it difficult to quantify the degree of embryo development. Machine learning models have made some progress in feature fusion, but static threshold classification still cannot address data drift. Existing technologies lack a systematic solution for comprehensively assessing seed viability from multiple dimensions, including spectroscopy, morphology, mechanics, and internal structure. Furthermore, the technical bottleneck of adaptive adjustment of dynamic classification thresholds has not been resolved, resulting in low stability and accuracy of detection results. Summary of the Invention

[0004] The present invention aims to at least solve the technical problem of low stability and accuracy of detection results in the prior art, and particularly innovatively proposes an intelligent crop seed vitality detection method and system.

[0005] In order to achieve the above-mentioned object of the present invention, the present invention provides an intelligent crop seed vitality detection method, which comprises:

[0006] S1, collecting seed data and preprocessing the seed data; the seed data includes seed spectrum data, seed 360-degree image data, seed hardness data and seed internal quality image;

[0007] S2. performing feature extraction on the seed data to obtain appearance features, spectral features, hardness features, and internal quality features of the seeds, and generating a seed feature set;

[0008] S3, inputting the feature set into a trained vitality prediction model to obtain a vitality prediction value;

[0009] S4. Dynamically adjust the classification threshold and classify the vitality of the seeds based on the vitality prediction value.

[0010] As an optional embodiment of the present invention, optionally, collecting seed hardness data in step S1 includes:

[0011] S101. Select seed samples, remove surface impurities, and store them in a constant temperature box for more than 24 hours;

[0012] S102, using a nanoindenter and an indenter, setting parameters, selecting a flat area of the seed without cracks or scratches, and performing hardness tests at different points for at least three times to obtain elastic recovery data for each point;

[0013] S103: Aggregate the elastic recovery data to obtain seed hardness data.

[0014] As an optional embodiment of the present invention, optionally, collecting the seed internal quality image in step S1 includes:

[0015] S106. Select seed samples of uniform size, remove surface impurities, and store in a constant temperature box for more than 24 hours;

[0016] S107, placing the seed sample on a rotating polytetrafluoroethylene tray, and performing a tomographic scan on the seed using an X-ray CT scanner to obtain an internal tomographic image of the seed;

[0017] S108, performing image enhancement processing on the internal tomographic image;

[0018] S109 , performing three-dimensional reconstruction on the enhanced internal tomographic image to obtain a three-dimensional image of the internal quality of the seed.

[0019] As an optional embodiment of the present invention, optionally, dynamically adjusting the classification threshold in step S4 includes:

[0020] S401. Determine an initial classification threshold based on historical classification accuracy;

[0021] S402, after each classification, calculate the current classification accuracy. If the current classification accuracy is higher than the historical classification accuracy, keep the current classification threshold unchanged;

[0022] S403: If the current classification accuracy is lower than the historical classification accuracy, the classification threshold is dynamically adjusted according to the classification error until the current classification accuracy is higher than or equal to the historical classification accuracy, or the preset upper limit of the number of adjustments is reached.

[0023] As an optional embodiment of the present invention, optionally, in step S403, if the current classification accuracy is lower than the historical classification accuracy, dynamically adjusting the classification threshold according to the classification error includes:

[0024] S4031. If the current classification accuracy fluctuates in the short term, the classification threshold is adjusted using the error feedback method;

[0025] S4032. If the current classification accuracy fluctuates over a long period of time, recalibrate the classification threshold using the ROC curve optimization method;

[0026] S4033. If the current classification accuracy is abnormal, switch to the adaptive window method and dynamically adjust the classification threshold;

[0027] S4034: Verify the adjusted classification threshold.

[0028] On the other hand, the present invention also provides an intelligent crop seed vitality detection system, the system including an intelligent crop seed vitality detection method; the system also includes:

[0029] A collection module, used for collecting seed data of seeds;

[0030] A processing module, connected to the acquisition module, for preprocessing and feature extraction of the collected seed data;

[0031] A prediction module, connected to the processing module, for obtaining a predicted vitality value of the seed using a vitality prediction model based on the extracted features;

[0032] The classification module is connected to the prediction module and is used to classify seeds according to the vitality prediction value and dynamically adjust the classification threshold according to the historical classification accuracy.

[0033] The present invention integrates four key features: spectral, morphological, mechanical, and internal structural, to construct a high-dimensional feature set. For example, mechanical parameters such as elastic modulus and hardness acquired through nanoindentation, combined with embryo volume and porosity data from X-ray CT 3D reconstruction, can accurately characterize seed physiological activity and mechanical integrity. The present invention incorporates error feedback, receiver operating characteristic (ROC) curve optimization, and an adaptive windowing method to respond in real time to changes in data distribution. Dynamic threshold adjustment reduces the fluctuation range of classification accuracy from ±8% to ±3% when faced with seed batch differences or fluctuations in storage environment. For example, in a sudden low-temperature stress experiment, dynamic thresholding reduced viability classification error by 45%, effectively improving the stability of test results. Through X-ray CT tomography and 3D reconstruction, the present invention achieves micron-level quantification of parameters such as embryo volume and porosity. Compared with traditional 2D imaging, the accuracy of embryo development assessment is improved, enabling precise identification of underdeveloped seeds (e.g., samples with an embryo volume fraction <75%), significantly enhancing the accuracy of vigor detection. A temporal neural network with a contextual layer is constructed to capture the dynamic coupling relationships between features. For example, through the correlation analysis of the decay law of spectral wavelength over time and hardness parameters, the downward trend of seed vitality can be predicted in advance, which has better long-term prediction stability.

[0034] Additional aspects and advantages of the present invention will be set forth in part in the description which follows and, in part, will be obvious from the description which follows, or may be learned by practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the following description of the embodiments with reference to the accompanying drawings, in which:

[0036] Figure 1 This is a flow chart of an intelligent crop seed vitality detection method according to embodiment 1 of the present invention;

[0037] Figure 2 This is a structural diagram of an intelligent crop seed vitality detection system according to embodiment 2 of the present invention. DETAILED DESCRIPTION

[0038] The following describes embodiments of the present invention in detail. Examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are intended only to explain the present invention and are not to be construed as limiting the present invention.

[0039] Example 1

[0040] like Figure 1 As shown, an intelligent crop seed vitality detection method comprises:

[0041] S1, collecting seed data and preprocessing the seed data; the seed data includes seed spectrum data, seed 360-degree image data, seed hardness data and seed internal quality image;

[0042] It should be noted that seed spectral data was collected using a portable near-infrared spectrometer (such as the ASD FieldSpec 4 series) with a wavelength range of 900–1700 nm (suitable for seed composition analysis), a spectral resolution of ≤3 nm, and a signal-to-noise ratio of ≥1000:1. The spectrometer was equipped with a reflectance probe (including a built-in whiteboard calibration module), a fiber optic transmission unit, and a sample cup holder. The seeds were inserted vertically into the fixture groove and adjusted so that the seed equatorial plane was parallel to the rotating stage plane. The rotating stage was set to rotate in 10° increments (a total of 36 angles), triggering the camera to capture images at each angle. The light source intensity was adjusted to ensure uniform brightness across the seed surface (grayscale fluctuation <5%). Oversampling (30% overlap between images) was used to ensure texture continuity. Multi-view images were stitched together into a panoramic image using feature point matching (SIFT algorithm) and normalized to a resolution of 1024×1024 pixels. Blurred images were automatically removed (based on the Laplacian operator clarity score, with a threshold of ≥0.8). Ten percent of the stitched images were manually reviewed to ensure that the embryo region was free of obstructions or deformation.

[0043] To capture 360-degree seed image data, an industrial line scan camera (resolution ≥5000×3000, frame rate ≥80 fps) was used in conjunction with a ring light source (color temperature 5500K, CRI ≥90) and a rotation stage (0.01° graduation, ±0.05° repeatability) for 360-degree seed rotation control. A black background was used to eliminate ambient reflections. A custom seed fixture (3D-printed, matte plastic) was used to ensure that the seed's geometric center coincided with the rotation axis. The seed was inserted vertically into the fixture's groove and adjusted so that the seed's equatorial plane was parallel to the plane of the rotation stage. The rotation stage was set to rotate in 10° increments (a total of 36 angles), triggering the camera to capture at each angle. The light intensity was adjusted to ensure uniform brightness across the seed surface (grayscale fluctuation <5%). Oversampling (30% overlap between images) was used to ensure texture continuity. Multi-view images were stitched into a panoramic image using feature point matching (SIFT algorithm) and normalized to a resolution of 1024×1024 pixels. 10% of the stitching results were manually reviewed to ensure that the embryo area was not blocked or deformed.

[0044] Preprocessing of seed data involves data cleaning and removing outliers and missing values to ensure data integrity and accuracy. Spectral data requires smoothing filtering to reduce noise, while image data requires image enhancement and denoising to improve image quality. Furthermore, seed data requires standardization, converting data of varying dimensions to a common scale to facilitate subsequent feature extraction and analysis.

[0045] S2. performing feature extraction on the seed data to obtain appearance features, spectral features, hardness features, and internal quality features of the seeds, and generating a seed feature set;

[0046] It should be noted that in this embodiment, appearance feature extraction utilizes deep learning algorithms, such as convolutional neural networks (CNNs), to process 360-degree seed image data, automatically identifying and extracting appearance features such as seed shape, color, and texture. Spectral data from seeds is acquired using a near-infrared spectrometer, and dimensionality reduction processing, such as principal component analysis (PCA), is performed to extract key spectral features that reflect the chemical composition and physiological state of the seeds.

[0047] S3, inputting the feature set into a trained vitality prediction model to obtain a vitality prediction value;

[0048] It should be noted that the vitality prediction model of this embodiment adopts a time-aware neural network (Time-Aware Neural Network) with a context layer, which is specially designed to process the multimodal time series features of seeds. The input layer receives the fused four-dimensional feature vector (spectral, morphological, mechanical, internal structure) and inputs it after standardization. The hidden layer contains a bidirectional LSTM unit to capture the dynamic correlation between features (such as the coupling effect of spectral changes and hardness attenuation). The context layer stores historical hidden states to enhance the model's sensitivity to long-term dependencies. The output layer, a fully connected layer, outputs the seed vitality value (range 0-1, the higher the value, the stronger the vitality).

[0049] When training the vitality prediction model:

[0050] Training data utilizes data from multiple batches of seeds throughout their lifecycle (including labels such as germination rate and storage time). Stratified sampling ensures balanced distribution of samples across different vigor levels. Spectral reflectance curves, 360° image texture features, nanoindentation mechanical parameters, and CT 3D embryo structure data are mapped to the same dimensional space through maximum and minimum value normalization. A self-attention mechanism dynamically weights key features (such as the spectral red edge shift during the critical germination period). Seeds are regularly inspected during storage to construct a time series feature matrix to capture vigor decay trends.

[0051] The process of generating predicted values is as follows:

[0052] The input feature vector passes through LSTM units sequentially, and the context layer updates the historical state. Finally, the hidden layer state is mapped to the vitality prediction value through a fully connected layer. Monte Carlo dropout is used to perform multiple forward propagations during the test phase to generate a distribution range for the prediction value and evaluate the model confidence.

[0053] Model validation and optimization are:

[0054] We combine mean squared error (MSE) with smoothed L1 loss to balance outlier robustness and gradient stability. We use a cosine annealing learning rate schedule with an initial learning rate of 0.001 and a batch size of 128. We use early stopping to prevent overfitting, and terminate training if the validation set loss does not decrease for five consecutive epochs.

[0055] On an independent test set, the coefficient of determination (R²) was calculated to be >0.92, and the correlation coefficient between predicted and actual germination rates was >0.95. Vigor classification accuracy was analyzed using a confusion matrix to ensure identification rates of high-vigor seeds (>0.8) exceeding 98%. Network weights were fine-tuned regularly (e.g., quarterly) using new batches of data to adapt to seasonal variations in seed physiological characteristics. Predicted value distribution drift was monitored, and if the confidence level of consecutive sample predictions fell below a threshold, a model retraining process was triggered.

[0056] Through the above process, the model can robustly output high-precision vitality prediction values, and combined with the dynamic threshold classification strategy, it can significantly improve the reliability and practicality of seed vitality assessment.

[0057] S4. Dynamically adjust the classification threshold and classify the vitality of the seeds based on the vitality prediction value.

[0058] It should be noted that in step S4, the seed vitality level threshold is divided into three types: first-level seeds: predicted value ≥ 0.85 (high vitality, suitable for long-term storage or precision sowing); second-level seeds: 0.7 ≤ predicted value < 0.85 (medium vitality, requiring short-term storage or for general planting); third-level seeds: predicted value < 0.7 (low vitality, recommended for elimination or use in scientific research experiments);

[0059] In the process of dynamically adjusting the classification threshold, the system will make intelligent judgments based on the historical classification accuracy and the characteristics of the current batch of seeds. For example, when the system detects that the overall vitality level of the current batch of seeds is high, it will automatically increase the classification threshold to ensure that high-vigor seeds can be accurately identified and retained. Conversely, if the system detects that the seed vitality is generally low, it will lower the classification threshold accordingly to avoid mistakenly classifying seeds with potential as low-vigor seeds. This dynamic adjustment mechanism enables the seed vitality detection system of the present invention to adapt to the characteristics of different batches of seeds, improving the accuracy and practicality of classification.

[0060] In summary, the intelligent crop seed vigor detection method of this embodiment integrates four key features: spectral, morphological, mechanical, and internal structural, to construct a high-dimensional feature set. For example, mechanical parameters such as elastic modulus and hardness acquired through nanoindentation, combined with embryo volume and porosity data from X-ray CT 3D reconstruction, can accurately characterize seed physiological activity and mechanical integrity. This method incorporates error feedback, receiver operating characteristic (ROC) curve optimization, and an adaptive windowing method to respond in real time to changes in data distribution. Dynamic threshold adjustment reduces the fluctuation range of classification accuracy from ±8% to ±3% when seed batches vary or storage environments fluctuate. For example, in a sudden low-temperature stress experiment, dynamic thresholding reduced vigor classification error by 45%, effectively improving the stability of test results. This method utilizes X-ray CT tomography and 3D reconstruction to achieve micron-level quantification of parameters such as embryo volume and porosity. Compared to traditional 2D imaging, the accuracy of embryo development assessment is improved, enabling precise identification of underdeveloped seeds (e.g., samples with an embryo volume fraction <75%), significantly enhancing the accuracy of vigor detection. A temporal neural network with a contextual layer is constructed to capture the dynamic coupling relationships between features. For example, by analyzing the correlation between the decay of spectral wavelength over time and hardness parameters, it is possible to predict the decline in seed vigor in advance. Field tests have shown that this model improves the accuracy of aging seed identification by 18% compared to static models and exhibits greater long-term prediction stability.

[0061] As an optional embodiment of the present invention, optionally, collecting seed hardness data in step S1 includes:

[0062] S101. Select seed samples, remove surface impurities, and store them in a constant temperature box for more than 24 hours;

[0063] It should be noted that when selecting seed samples in step S101, it is important to ensure that the samples are representative and reflect the overall characteristics of the seed batch. Surface impurities are removed to prevent them from interfering with the hardness measurement results. The seed samples are stored in a constant temperature chamber for at least 24 hours to allow the seeds to reach temperature equilibrium, eliminating any temperature effects on the hardness measurement results.

[0064] S102, using a nanoindenter and an indenter, setting parameters, selecting a flat area of the seed without cracks or scratches, and performing hardness tests at different points for at least three times to obtain elastic recovery data for each point;

[0065] It should be noted that, when adopting nanoindenter to carry out hardness test in step S102, need to ensure that indenter is clean and wear-free, to avoid producing error to test result. Parameters such as indenter indentation depth, loading rate are set, and the selection of these parameters should be reasonably set according to the characteristic and test requirement of seed. Select the smooth area without crack and scratch on seed surface to test, in order to ensure the accuracy and reliability of test result. Carry out at least three hardness tests on each selected point, in order to obtain more accurate average hardness value, reduce the error that single test may bring. By recording the elastic recovery data of each point, mechanical property and elastic recovery situation of seed can be further analyzed, for the vitality assessment of seed provides more comprehensive data support.

[0066] S103: Aggregate the elastic recovery data to obtain seed hardness data.

[0067] It should be noted that in step S103, the elastic recovery data obtained from multiple hardness tests are aggregated and processed to calculate an average hardness value, reflecting the overall hardness characteristics of the seed. Hardness data, as an important component of seed internal quality characteristics, can reflect the mechanical strength and tolerance of the seed, providing key information for subsequent vigor prediction. By accurately collecting and processing hardness data, the accuracy and reliability of seed vigor detection can be further improved.

[0068] As an optional embodiment of the present invention, optionally, the expression for calculating the hardness characteristic in step S2 is:

[0069] ,

[0070] ,

[0071] ,

[0072] ;

[0073] in, Indicates hardness characteristics, Represents a splicing operation, represents the elastic modulus, Indicates hardness. represents the resonant frequency, represents the unloading stiffness, represents the geometric correction coefficient of the pressure head, represents the contact area, Indicates the maximum indentation load, represents the equivalent stiffness, Indicates seed quality, represents the natural frequency of undamped free vibration.

[0074] As an optional embodiment of the present invention, optionally, collecting the seed internal quality image in step S1 includes:

[0075] S106. Select seed samples of uniform size, remove surface impurities, and store in a constant temperature box for more than 24 hours;

[0076] S107, placing the seed sample on a rotating polytetrafluoroethylene tray, and performing a tomographic scan on the seed using an X-ray CT scanner to obtain an internal tomographic image of the seed;

[0077] It should be noted that when adopting X-ray CT scanning equipment to carry out tomographic scanning of seeds in step S107, it is necessary to ensure that the scanning parameters are reasonably set, such as tube voltage, tube current, scanning speed, etc., to obtain clear internal tomographic images. Placing the seed sample on a polytetrafluoroethylene tray with a rotating function can ensure that the seed can rotate evenly during the scanning process, thereby obtaining a full range of internal image information. The seed internal tomographic image obtained by tomographic scanning can clearly show the internal structure of the seed, such as the position, form, size and internal pore distribution of the embryo, etc., providing an important basis for subsequent internal quality feature extraction. After obtaining the internal tomographic image, image reconstruction and processing are also required to improve image quality and resolution, so as to facilitate subsequent feature extraction and analysis.

[0078] S108, performing image enhancement processing on the internal tomographic image;

[0079] It should be noted that the internal tomographic image is enhanced in step S108 to improve the visualization and detail of the image, and to provide clearer and more accurate image information for subsequent feature extraction and analysis. Image enhancement processing may include operations such as contrast adjustment, noise suppression, and edge sharpening, which can be reasonably selected and combined according to the specific circumstances and needs of the image. By adjusting the contrast, the brightness difference between different objects or areas in the image can be increased, making the image clearer and easier to observe. Noise suppression can remove random noise and interference in the image, and improve the purity and signal-to-noise ratio of the image. Edge sharpening can enhance the edges and contours of objects in the image, making the image more three-dimensional and vivid.

[0080] S109 , performing three-dimensional reconstruction on the enhanced internal tomographic image to obtain a three-dimensional image of the internal quality of the seed.

[0081] It should be noted that, in step S109, the internal tomographic image after the enhanced processing is carried out three-dimensional reconstruction. Through advanced image reconstruction algorithm, two-dimensional tomographic images are superimposed and fused to form the internal three-dimensional structure of seed. This step not only requires a high degree of image processing ability, but also needs to have a deep understanding of the internal structure of seed, to ensure the accuracy and authenticity of the reconstruction result. The three-dimensional image after the reconstruction can intuitively show the internal structure of seed, such as the three-dimensional form of embryo, internal pore network etc., for subsequent feature extraction provides more intuitive and comprehensive information. After obtaining the internal quality three-dimensional image, further analysis and processing are also needed to extract the key features relevant to seed vigor.

[0082] As an optional embodiment of the present invention, optionally, the expression for extracting the internal quality feature in step S2 is:

[0083] ,

[0084] ,

[0085] ,

[0086] ;

[0087] in, Represents internal quality characteristics, Represents a splicing operation, represents the internal porosity of the seed, represents the actual embryo volume, represents the theoretical maximum embryo volume, represents the sound wave attenuation coefficient, represents the propagation distance, represents the logarithmic function, represents the initial amplitude, Indicates the received amplitude, represents the proportion of embryo area, represents the pixel area of the embryo region, Indicates the total pixel area of the seed.

[0088] As an optional embodiment of the present invention, optionally, in step S3, the expression for obtaining the vitality prediction value using the vitality prediction model is:

[0089] ,

[0090] ,

[0091] ,

[0092] ;

[0093] in, Indicates that at time step When the output layer The predicted output value of each neuron, represents the total number of neurons, represents the hidden layer neurons to the output layer The connection weights of neurons, represents the hidden layer neurons at time step The output, represents the activation function, Represents the hidden layer neurons The net input, represents the input feature dimension, Represents the input layer neurons to the hidden layer The connection weights of neurons, Represents the input layer neurons at time step The input value of Represents the context layer neurons to the hidden layer The feedback weight of each neuron, Represents the context layer The neuron at the previous time step The feedback value of Indicates that at time step After that, the context layer neurons The state of the current hidden layer neuron is updated The output of is used as feedback for the next time step.

[0094] As an optional embodiment of the present invention, optionally, dynamically adjusting the classification threshold in step S4 includes:

[0095] S401. Determine an initial classification threshold based on historical classification accuracy;

[0096] It should be noted that the initial classification threshold value set in step S401 is based on a large amount of historical data and classification results, and is derived through statistical analysis. A reasonable initial classification threshold value can improve the accuracy and stability of classification. The setting of the initial classification threshold value requires consideration of various factors, such as seed type, storage conditions, and testing environment, to ensure its adaptability and accuracy.

[0097] S402, after each classification, calculate the current classification accuracy. If the current classification accuracy is higher than the historical classification accuracy, keep the current classification threshold unchanged;

[0098] It should be noted that, in step S402, by dynamically adjusting the classification threshold, the classification result can be continuously optimized to make it more in line with the actual situation. If the current classification accuracy is lower than the historical classification accuracy, it means that the current classification threshold may not be suitable for the current test environment or seed characteristics, and the classification threshold needs to be adjusted at this time. The adjustment method can be gradual fine-tuning, or it can be adjusted according to specific rules or algorithms. During the adjustment process, it is necessary to continuously monitor the changes in the classification accuracy to ensure the effectiveness and accuracy of the adjustment. Through continuous adjustment and optimization, the accuracy and stability of the classification can be gradually improved.

[0099] S403: If the current classification accuracy is lower than the historical classification accuracy, the classification threshold is dynamically adjusted according to the classification error until the current classification accuracy is higher than or equal to the historical classification accuracy, or the preset upper limit of the number of adjustments is reached.

[0100] It should be noted that, in step S403, when the classification accuracy rate is continuously lower than the historical classification accuracy rate, the system will automatically adjust the classification threshold value according to the size and direction of the classification error. This adjustment is dynamic and intelligent, and is intended to gradually reduce the classification error and improve the accuracy and stability of the classification. During the adjustment process, the system will take into account various factors, such as the distribution of the classification error, the change of the seed characteristics, etc., to ensure the rationality and effectiveness of the adjustment. If, after multiple adjustments, the current classification accuracy rate is still lower than the historical classification accuracy rate and has reached the preset upper limit of the number of adjustments, the system will prompt the user to recheck the test environment, seed sample or classification model, etc., to find out the reasons that may cause the classification accuracy rate to decline, and to make corresponding improvements and optimizations. Through this series of dynamic adjustments and optimizations, the intelligent agricultural crop seed vigor detection method and system of the present invention can achieve accurate and rapid evaluation of seed vigor.

[0101] As an optional embodiment of the present invention, optionally, in step S403, if the current classification accuracy is lower than the historical classification accuracy, dynamically adjusting the classification threshold according to the classification error includes:

[0102] S4031. If the current classification accuracy fluctuates in the short term, the classification threshold is adjusted using the error feedback method;

[0103] It should be noted that the error feedback method in step S4031 is a feedback adjustment mechanism based on classification error. When the system detects short-term fluctuations in the current classification accuracy, it assumes this is due to random factors or short-term environmental changes. At this point, the error feedback method is used for adjustment. The error feedback method fine-tunes the classification threshold based on the magnitude and direction of the classification error to quickly respond to such short-term fluctuations and quickly restore the classification accuracy to normal levels. Through adjustments made using the error feedback method, the system can maintain the stability and accuracy of the classification results and avoid the accumulation of classification errors caused by short-term fluctuations.

[0104] S4032. If the current classification accuracy fluctuates over a long period of time, recalibrate the classification threshold using the ROC curve optimization method;

[0105] It should be noted that in step S4032, if the current classification accuracy shows a long-term fluctuating trend, this may indicate significant changes in the test environment, seed characteristics, or the classification model itself. In this case, relying solely on error feedback for adjustment may not achieve the desired results. Therefore, the present invention employs a strategy of recalibrating the classification threshold in conjunction with ROC curve optimization. The ROC curve (Receiver Operating Characteristic Curve) is a tool for evaluating the performance of a classification model. By plotting the relationship between the true positive rate (TRR) and the false positive rate (FPR) at different classification thresholds, it can intuitively demonstrate the model's classification performance. Combined with the ROC curve optimization method, the system can analyze the distribution of classification errors and find the optimal classification threshold, thereby recalibrating the classification results. Through this strategy, the present invention can address long-term fluctuations in classification accuracy and ensure the accuracy and stability of classification results. After recalibrating the classification threshold, the system will continue to monitor changes in the classification accuracy and make further adjustments and optimizations as needed.

[0106] S4033. If the current classification accuracy is abnormal, switch to the adaptive window method and dynamically adjust the classification threshold;

[0107] It should be noted that if a sudden abnormality occurs in the current classification accuracy in step S4033, this may be due to an abnormal situation in the test environment, such as extreme weather, equipment failure, etc., or there may be special problems with the seed sample itself. At this time, the system needs to respond quickly and switch to the adaptive window method for adjustment. The adaptive window method is a dynamic adjustment strategy based on a sliding window. It automatically adjusts the size and position of the window according to the changing trend and fluctuation range of the classification accuracy to achieve dynamic adjustment of the classification threshold. Through the adjustment of the adaptive window method, the system can quickly adapt to changes in the test environment, reduce classification errors, and improve the accuracy and stability of classification. During the adjustment process, the system will continue to monitor the changes in the classification accuracy and make further adjustments and optimizations as needed. If the classification accuracy still cannot be restored to normal levels after adjustment using the adaptive window method, the system will prompt the user to make further checks and adjustments to ensure the normal operation of the system and the accuracy of the classification results.

[0108] S4034: Verify the adjusted classification threshold.

[0109] It should be noted that verifying the adjusted classification threshold in step S4034 is a key step in ensuring the effectiveness of the adjustment. The verification process mainly consists of two parts: first, testing the adjusted classification threshold using a validation dataset to evaluate its classification performance; second, comparing the verification results with historical classification accuracy to confirm the effectiveness of the adjustment.

[0110] During the verification process, the system uses methods such as cross-validation to ensure the reliability and stability of the verification results. By comparing the verification results with historical classification accuracy, the system can determine whether the adjusted classification threshold has achieved the desired effect. If the verification results show that the adjusted classification threshold can significantly improve classification accuracy, the system will officially adopt this classification threshold for subsequent seed viability assessments. If the verification results are not ideal, the system will continue to adjust the classification threshold until the optimal solution is found.

[0111] As an optional embodiment of the present invention, optionally, in step S4031, the expression for adjusting the classification threshold using the error feedback method is:

[0112] ;

[0113] in, represents the updated classification threshold, Indicates the current classification threshold, represents the learning rate, represents the number of samples, Indicates the The true labels of samples, Indicates the The predicted labels of samples, Indicates the direction of sensitivity of the predicted value to the threshold;

[0114] In step S4032, the expression for recalibrating the classification threshold by combining the ROC curve optimization method is:

[0115] ;

[0116] in, represents the optimal classification threshold, Indicates that the threshold The area under the ROC curve;

[0117] The expression for dynamically adjusting the classification threshold in step S4033 is:

[0118] ,

[0119] ;

[0120] in, represents the mixed classification threshold, represents the smoothing coefficient, represents the dynamic classification threshold, represents the historical optimal threshold, Represents a sliding window The mean of the internal prediction probabilities, represents the adjustment coefficient, Represents a sliding window The standard deviation of the predicted probabilities.

[0121] Example 2

[0122] like Figure 2 As shown, an intelligent crop seed vitality detection system includes an intelligent crop seed vitality detection method; the system is used to implement the intelligent crop seed vitality detection method; the system also includes:

[0123] A collection module, used for collecting seed data of seeds;

[0124] It should be noted that the acquisition module is mainly responsible for collecting various data related to seed vitality assessment, including but not limited to key indicators such as seed appearance characteristics, size, weight, germination rate, and growth rate.

[0125] A processing module, connected to the acquisition module, for preprocessing and feature extraction of the collected seed data;

[0126] It should be noted that the processing module is primarily responsible for cleaning, filtering, and feature extraction of collected seed data. Preprocessing the raw data removes noise and outliers, improving data accuracy and reliability. Feature extraction extracts key features from the raw data that are crucial for seed viability assessment. The processing module utilizes advanced algorithms and technologies to efficiently process large amounts of data, ensuring the system's real-time performance and accuracy.

[0127] A prediction module, connected to the processing module, for obtaining a predicted vitality value of the seed using a vitality prediction model based on the extracted features;

[0128] It's important to note that the prediction module uses a pre-trained vigor prediction model to analyze and calculate the features extracted by the processing module to derive a predicted seed vigor value. This value reflects the seed's growth potential and germination ability, providing an important reference for subsequent seed screening and planting. The vigor prediction model used in the prediction module has been trained and validated with extensive data, resulting in high accuracy and stability, capable of handling a wide range of seed types and characteristics.

[0129] The classification module is connected to the prediction module and is used to classify seeds according to the vitality prediction value and dynamically adjust the classification threshold according to the historical classification accuracy.

[0130] It should be noted that the classification module is primarily responsible for classifying the vitality prediction values output by the prediction module to determine the seed vitality level. The classification process is based on preset classification criteria and thresholds. At the same time, the classification module also has the ability to dynamically adjust the classification threshold, automatically adjusting the classification threshold based on historical classification accuracy and current classification results to ensure classification accuracy and stability. Through continuous adjustment and optimization, the classification module can adapt to seeds of different types and characteristics and achieve accurate assessment of seed vitality. Through the collaborative work of various modules, the entire intelligent crop seed vitality detection system can achieve rapid and accurate assessment of seed vitality, providing strong technical support for agricultural production.

[0131] The intelligent crop seed vigor detection system works as follows: the collection module collects various seed data, the processing module preprocesses the data and extracts features, the prediction module uses a vigor prediction model to derive predicted vigor values, and finally, the classification module classifies the seeds based on the predicted vigor values. During the classification process, the system not only relies on preset classification criteria and thresholds but also has the ability to intelligently adjust the classification thresholds. This dynamic adjustment mechanism ensures accurate and stable classification, enabling the system to adapt to different seed types and characteristics. The coordinated operation of the entire system enables comprehensive, rapid, and accurate assessment of seed vigor.

[0132] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and alterations may be made to the embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the claims and their equivalents.

Claims

1. An intelligent method for detecting the vitality of crop seeds, characterized in that: The method comprises: S1, collecting seed data and preprocessing the seed data; the seed data includes seed spectrum data, seed 360-degree image data, seed hardness data and seed internal quality image; S2. performing feature extraction on the seed data to obtain appearance features, spectral features, hardness features, and internal quality features of the seeds, and generating a seed feature set; S3, inputting the feature set into a trained vitality prediction model to obtain a vitality prediction value; S4. Dynamically adjust the classification threshold and classify the vitality of the seeds based on the vitality prediction value; Dynamically adjusting the classification threshold in step S4 includes: S401. Determine an initial classification threshold based on historical classification accuracy; S402, after each classification, calculate the current classification accuracy. If the current classification accuracy is higher than the historical classification accuracy, keep the current classification threshold unchanged; S403: If the current classification accuracy is lower than the historical classification accuracy, the classification threshold is dynamically adjusted according to the classification error until the current classification accuracy is higher than or equal to the historical classification accuracy, or the preset upper limit of the number of adjustments is reached; In step S403, if the current classification accuracy is lower than the historical classification accuracy, dynamically adjusting the classification threshold according to the classification error includes: S4031. If the current classification accuracy fluctuates in the short term, the classification threshold is adjusted using the error feedback method; The expression for adjusting the classification threshold using the error feedback method in step S4031 is: ； in, represents the updated classification threshold, Indicates the current classification threshold, represents the learning rate, represents the number of samples, Indicates the The true labels of samples, Indicates the The predicted labels of samples, Indicates the direction of sensitivity of the predicted value to the threshold; S4032. If the current classification accuracy fluctuates over a long period of time, recalibrate the classification threshold using the ROC curve optimization method; S4033. If the current classification accuracy is abnormal, switch to the adaptive window method and dynamically adjust the classification threshold; S4034: Verify the adjusted classification threshold.

2. The intelligent crop seed vitality detection method according to claim 1, characterized in that: Collecting seed hardness data in step S1 includes: S101. Select seed samples, remove surface impurities, and store them in a constant temperature box for more than 24 hours; S102, using a nanoindenter and an indenter, setting parameters, selecting a flat area of the seed without cracks or scratches, and performing hardness tests at different points for at least three times to obtain elastic recovery data for each point; S103: Aggregate the elastic recovery data to obtain seed hardness data.

3. The intelligent crop seed vitality detection method according to claim 2, characterized in that: The expression for calculating the hardness characteristic in step S2 is: ， ， ， ； in, Indicates hardness characteristics, Represents a splicing operation, represents the elastic modulus, Indicates hardness. represents the resonant frequency, represents the unloading stiffness, represents the geometric correction coefficient of the pressure head, represents the contact area, Indicates the maximum indentation load, represents the equivalent stiffness, Indicates seed quality, represents the natural frequency of undamped free vibration.

4. The intelligent crop seed vitality detection method according to claim 1, characterized in that: Acquiring the seed internal quality image in step S1 includes: S106. Select seed samples of uniform size, remove surface impurities, and store in a constant temperature box for more than 24 hours; S107, placing the seed sample on a rotating polytetrafluoroethylene tray, and performing a tomographic scan on the seed using an X-ray CT scanner to obtain an internal tomographic image of the seed; S108, performing image enhancement processing on the internal tomographic image; S109 , performing three-dimensional reconstruction on the enhanced internal tomographic image to obtain a three-dimensional image of the internal quality of the seed.

5. The intelligent crop seed vitality detection method according to claim 4, characterized in that: The expression for extracting the internal quality features in step S2 is: ， ， ， ； in, Represents internal quality characteristics, Represents a splicing operation, represents the internal porosity of the seed, represents the actual embryo volume, represents the theoretical maximum embryo volume, represents the sound wave attenuation coefficient, represents the propagation distance, represents the logarithmic function, represents the initial amplitude, Indicates the received amplitude, represents the proportion of embryo area, represents the pixel area of the embryo region, Indicates the total pixel area of the seed.

6. The intelligent crop seed vitality detection method according to claim 1, characterized in that: In step S3, the expression for obtaining the vitality prediction value using the vitality prediction model is: ， ， ， ； in, Indicates that at time step When the output layer The predicted output value of each neuron, represents the total number of neurons, represents the hidden layer neurons to the output layer The connection weights of neurons, represents the hidden layer neurons at time step The output, represents the activation function, Represents the hidden layer neurons The net input, represents the input feature dimension, Represents the input layer neurons to the hidden layer The connection weights of neurons, Represents the input layer neurons at time step The input value of Represents the context layer neurons to the hidden layer The feedback weight of each neuron, Represents the context layer The neuron at the previous time step The feedback value of Indicates that at time step After that, the context layer neurons The state of the current hidden layer neuron is updated The output of is used as feedback for the next time step.

7. The intelligent crop seed vitality detection method according to claim 1, characterized in that: In step S4032, the expression for recalibrating the classification threshold by combining the ROC curve optimization method is: ； in, represents the optimal classification threshold, Indicates that the threshold The area under the ROC curve; The expression for dynamically adjusting the classification threshold in step S4033 is: ， ； in, represents the mixed classification threshold, represents the smoothing coefficient, represents the dynamic classification threshold, represents the historical optimal threshold, Represents a sliding window The mean of the internal prediction probabilities, represents the adjustment coefficient, Represents a sliding window The standard deviation of the predicted probabilities.

8. An intelligent crop seed vitality detection system, characterized in that: The system includes the intelligent crop seed vitality detection method according to any one of claims 1 to 7; the system also includes: A collection module, used for collecting seed data of seeds; A processing module, connected to the acquisition module, for preprocessing and feature extraction of the collected seed data; A prediction module, connected to the processing module, for obtaining a predicted vitality value of the seed using a vitality prediction model based on the extracted features; The classification module is connected to the prediction module and is used to classify seeds according to the vitality prediction value and dynamically adjust the classification threshold according to the historical classification accuracy.

Images