Characteristic fusion rice variety identification method based on improved hyperspectral technology

Abstract

The invention discloses a feature fusion rice variety identification method based on an improved hyperspectral technology, and relates to the technical field of agricultural information.The identification method comprises the specific steps that S100, multi-view-angle data acquisition is conducted, specifically, in a rice planting area, representative rice plants are selected, a plurality of hyperspectral sensors with adjustable angles are arranged around the representative rice plants, and the hyperspectral sensors are used for identifying the representative rice plants; by arranging a plurality of angle-adjustable hyperspectral sensors, spectral information of rice plants can be captured in an omnibearing and three-dimensional manner, information loss or misjudgment possibly caused by a single view angle is effectively avoided, meanwhile, feature extraction is performed by using a multi-scale spectral transformation function, the richness and diversity of features are improved, and the accuracy of the feature extraction is improved. The identification capability of the model on fine differences of rice varieties is enhanced, the accuracy and reliability of rice variety identification are improved, more comprehensive and accurate data support is provided for subsequent spectral analysis and model training, and the development process of agriculture intelligence and precision is promoted.

Description

Technical Field

[0001] The present invention relates to the field of agricultural information technology, and particularly to a method for identifying rice varieties by feature fusion based on improved hyperspectral technology. Background Art

[0002] With the rapid development of modern agricultural technology, the accurate identification of rice varieties has become one of the key factors in improving crop yield and quality. As an advanced remote sensing detection method, hyperspectral technology can finely analyze the spectral information reflected or emitted by the target object to obtain detailed information on its physical, chemical, and biological characteristics. In recent years, hyperspectral technology has received extensive attention and application in the fields of crop monitoring, pest and disease diagnosis, and variety identification. By utilizing the rich information in hyperspectral data, scientists can achieve real-time monitoring and precise management of crop growth status, providing strong support for improving agricultural production efficiency and quality.

[0003] Although hyperspectral technology has significant advantages in the field of crop identification, there are still some deficiencies in traditional hyperspectral data processing methods. On the one hand, traditional methods often rely on the acquisition of hyperspectral data from a single perspective, which may lead to one-sidedness and incompleteness of information, thus affecting the accuracy and reliability of identification. On the other hand, in the process of feature extraction and dimensionality reduction, traditional methods often ignore the complementarity of spectral information at different scales, making it difficult for the extracted feature vectors to comprehensively reflect the inherent characteristics of rice varieties. In addition, when traditional identification methods implement feature fusion, they often lack effective strategies and methods, resulting in the difficulty of fully utilizing the complementary advantages of various types of information in the fused feature vectors, thereby affecting the performance of the identification model.

[0004] Therefore, the development of a method for identifying rice varieties by feature fusion based on improved hyperspectral technology will provide strong support for the rapid and accurate identification of rice varieties, and at the same time provide new technical ideas and solutions for the fields of variety identification and pest and disease diagnosis of other crops. Summary of the Invention

[0005] The purpose of the present invention is to make up for the deficiencies of the prior art, and provide a method for identifying rice varieties by feature fusion based on improved hyperspectral technology. This method realizes the rapid and accurate identification of rice varieties through steps such as multi-perspective data acquisition, multi-scale feature extraction, spectral feature processing and dimensionality reduction, model application and continuous optimization, and practical application and continuous monitoring. This method not only improves the identification accuracy, but also reduces the data dimension and calculation complexity. By fusing spectral features, image features, and physiological features, the identification ability of the model is further enhanced.

[0006] To solve the above technical problems, the present invention provides the following technical solution: A method for identifying rice varieties by feature fusion based on improved hyperspectral technology, and the specific steps of the identification method are as follows: S100, Multi-view data acquisition: In the rice planting area, select representative rice plants, arrange multiple adjustable-angle hyperspectral sensors around them, synchronously collect rice hyperspectral data from different perspectives, form a multi-view hyperspectral data set and store it; S200, Multi-scale feature extraction: Receive the collected hyperspectral data set, perform cleaning and standardization, use the multi-scale spectral transformation function for transformation, and extract feature vectors of different scales from the transformation results; S300, Spectral feature processing and dimensionality reduction: Determine the number of samples according to domain knowledge and analysis objectives, calculate the correlation index at each wavelength, screen out the bands closely related to rice varieties, form a sample feature vector, and perform dimensionality reduction on the features; S400, Model application and continuous optimization: Use an agricultural image acquisition instrument to obtain rice images and extract global feature vectors, use a SPAD-502 chlorophyll meter to measure rice physiological indicators to form physiological feature vectors, hierarchically fuse them with the feature vectors after spectral dimensionality reduction through a hierarchical multi-feature fusion formula, create a classification model based on the fused feature vectors and rice variety labels, and use sample data to train and optimize the model; S500, Practical application and continuous monitoring: Deploy the trained and optimized model to the actual agricultural production scenario, track and record the identification results, collect new sample data or sample data with differences, and regularly feedback to optimize the model.

[0007] Further, the specific steps of performing multi-scale transformation through the multi-scale spectral transformation function in S200, multi-scale feature extraction are as follows: (1) Preset the value range and step size of the scale factor; (2) Determine the wavelength offset according to the spectral resolution and analysis requirements; (3) Determine the number of wavelength offsets by analyzing the data characteristics; (4) Determine the weight coefficient according to the data in the relevant field; (5) Based on the set parameters, use the multi-scale spectral transformation function to transform the hyperspectral data. As the scale factor takes values in sequence, obtain the transformation results at different scales, and extract relevant features from the transformation results.

[0008] Even further, in S200, multi-scale feature extraction, the wavelength offset is dynamically adjusted according to the spectral resolution of the hyperspectral sensor. The higher the spectral resolution, the smaller the wavelength offset. The specific adjustment rule is: Let the spectral resolution of the hyperspectral sensor be , define the basic wavelength offset as , the actual wavelength offset has the following calculation formula: , where is an adjustment coefficient used to dynamically calibrate the wavelength offset, and its value range is [0.8, 1.2].

[0009] Further, in the S200, during multi-scale feature extraction, a multi-scale spectral transformation function is used to transform the hyperspectral data. Assume the collected hyperspectral data is , and the scale factor is defined as , which is used to control the granularity of the spectral transformation. The transformation function is constructed, and its calculation formula is: , where is the selected number of wavelength offsets, which is determined by the characteristics of the spectral data, is the corresponding weight coefficient, reflecting the importance of different offsets, is the wavelength offset dynamically adjusted according to the spectral resolution. By changing the value, the transformation results at different scales are obtained .

[0010] Further, in the S300, for spectral feature processing and dimensionality reduction, the correlation index is used to screen the bands closely related to the rice variety. The calculation formula is: , where is the spectral value of different samples at the wavelength , is the average value of all samples at this wavelength, is the number of samples. Select greater than the threshold as the key bands. The threshold is determined through statistical analysis of historical data.

[0011] Further, in the S300, during spectral feature processing and dimensionality reduction, the projection matrix is obtained by solving an optimization problem to reduce the dimensionality of the eigenvectors. The projection matrix is defined as , and the calculation formula is: , where is the sample eigenvector composed of the key bands, is the number of sample eigenvectors, is the candidate solution of the projection matrix, which is used to optimize the dimensionality reduction process. The projection matrix is obtained by solving, and the sample eigenvector is projected into a low-dimensional space to obtain the dimensionality-reduced eigenvector .

[0012] Further, in the S400, model application and continuous optimization, feature fusion is performed through a hierarchical multi-feature fusion formula. Let the image features of rice be extracted as , and the physiological features be extracted as vector . After spectral dimensionality reduction, the feature vector , the global image feature vector , and the physiological feature vector are fused through a hierarchical multi-feature fusion formula. The formula is: , where , , are weight coefficients, and the optimal values are determined through cross-validation.

[0013] Further, in the S400, model application and continuous optimization, for the creation of the classification model, let the fused feature vector be , and the corresponding rice variety label be represented as . Let the mapping function of the model be: , where is the weight vector corresponding to the feature , and is the bias term.

[0014] Further, the specific steps for training and optimizing the classification model in the S400, model application and continuous optimization are as follows: (1) Collect rice samples. For each sample, obtain its fused feature vector , and at the same time, numerically encode the corresponding rice variety label to obtain . Organize the feature vectors and variety labels of all samples into a training dataset , where is the number of samples, and divide the dataset into a training set and a test set according to a ratio; (2) Initialize the weight vector and the bias term in the model; (3) Take out a batch of samples from the training set. Let the batch size be . For each sample in the batch, input its fused feature vector into the model. The formula is: , and calculate the predicted value ; (4) According to the calculated predicted value and the actual variety label of the sample, use the loss function to calculate the loss value of this batch of samples. The formula is: , which measures the difference between the model prediction result and the actual label; (5) Calculate the loss function through the backpropagation algorithm for the weight vector and the bias term gradient; (6) According to the calculated gradient, use the stochastic gradient descent algorithm to update the weight vector and the bias term , the update formula is: ; , where is the learning rate, a preset hyperparameter that controls the step size of parameter update; (7) Repeat (3) to (6) to traverse the entire training set, complete one training cycle, and perform training cycles, is the preset number of training cycles, dynamically adjusted according to the model convergence situation; (8) After completing all training cycles, use the test set to evaluate the trained model, and retrain the model according to the evaluation results.

[0015] Compared with the prior art, the method for identifying rice varieties by feature fusion based on improved hyperspectral technology has the following beneficial effects: First, by arranging multiple hyperspectral sensors with adjustable angles, the present invention can capture the spectral information of rice plants omnidirectionally and three-dimensionally, effectively avoiding information loss or misjudgment caused by a single perspective. At the same time, using the multi-scale spectral transformation function for feature extraction not only improves the richness and diversity of features but also enhances the model's ability to identify subtle differences in rice varieties, not only improving the accuracy and reliability of rice variety identification but also providing more comprehensive and accurate data support for subsequent spectral analysis and model training, which helps to promote the development process of agricultural intelligence and precision.

[0016] Second, by fusing feature information from different sources and different dimensions, the present invention constructs a comprehensive and detailed rice variety identification model. This fusion strategy not only makes full use of the complementarity of various feature information but also improves the generalization ability and adaptability of the model. In addition, the present invention also uses techniques such as cross-validation, loss function calculation, backpropagation algorithm, and stochastic gradient descent algorithm to train and optimize the model, ensuring the stability and accuracy of the model in practical applications, not only improving the efficiency and accuracy of rice variety identification but also providing useful reference and inspiration for the fields of variety identification and pest and disease diagnosis of other crops.

[0017] Other advantages, objects, and features of the present invention will be set forth in part in the following description, and in part will be obvious to those skilled in the art from a study of the following, or may be learned from the practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the drawings in the following description are only some embodiments of the present invention, and for those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on these drawings.

[0019] Figure 1 It is a flowchart of a method for identifying rice varieties by feature fusion based on improved hyperspectral technology; Figure 2 It is a framework diagram of a method for identifying rice varieties by feature fusion based on improved hyperspectral technology. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0020] To further elaborate on the technical means and effects adopted by the present invention to achieve the intended invention purpose, the following will, in conjunction with the accompanying drawings and preferred embodiments, describe in detail the specific embodiments, structures, features, and their effects of the present invention as follows.

[0021] Embodiment 1: Variety identification in a large-scale rice planting base.

[0022] Multi-perspective data acquisition: In a large-scale planting base with a large area and multiple rice varieties planted, representative rice plants are carefully selected according to the distribution and growth trends of different rice varieties. To comprehensively obtain the spectral information of rice from various angles, multiple adjustable-angle hyperspectral sensors are evenly arranged around each rice plant. During the critical growth periods of rice, such as the tillering stage and the filling stage, ensure that each sensor synchronously collects the hyperspectral data of rice from different perspectives. These data are promptly transmitted and stored in the special data storage system of the base, providing a rich data basis for subsequent analysis.

[0023] Multi-scale Feature Extraction: After the data storage system receives the collected hyperspectral dataset, it starts the cleaning and normalization process. This process aims to remove outliers and noise in the data to make the data more accurate and reliable. Then, technicians determine a series of key parameters based on the spectral resolution of the hyperspectral sensor, the actual situation of the planting base, and the data analysis experience accumulated over the years. These parameters include the value range and step size of the scale factor, the wavelength offset, and the number of wavelength offsets. The wavelength offset is dynamically adjusted according to the spectral resolution of the hyperspectral sensor. The higher the spectral resolution, the smaller the wavelength offset. The specific adjustment rule is: Let the spectral resolution of the hyperspectral sensor be , define the basic wavelength offset as , and the formula for the actual wavelength offset is: , where is the adjustment coefficient, and its value range is [0.8, 1.2]. The number of wavelength offsets is determined by analyzing the data characteristics. At the same time, the weight coefficient is determined in combination with the data in related fields. Using these set parameters, the hyperspectral data is transformed by means of a multi-scale spectral transformation function. Let the collected hyperspectral data be , define the scale factor as , and construct the transformation function , and the formula is: , where is the selected number of wavelength offsets, is the corresponding weight coefficient, is the wavelength offset dynamically adjusted according to the spectral resolution. By changing the value of , the transformation results at different scales are obtained, and the feature vectors are accurately extracted from the transformation results at different scales.

[0024] Spectral Feature Processing and Dimensionality Reduction: Technicians comprehensively consider the historical data of the planting base, the characteristic differences of different rice varieties, and the current research needs to determine the appropriate number of samples. By calculating the correlation index at each wavelength, the bands closely related to the rice variety are selected. The formula is: , where is the spectral value of different samples at wavelength , is the average value of all samples at this wavelength, is the number of samples, and select greater than the threshold as the key band. The threshold is determined by statistical analysis of historical data. These key bands are used to form the sample feature vector, and then the projection matrix is obtained by solving a specific optimization problem. The formula is: , where, is the sample feature vector composed of key bands, is the number of samples, and the projection matrix is obtained by solving , project the sample feature vector into a low-dimensional space to obtain the feature vector after dimensionality reduction , while retaining the key information, reduce the data dimension and improve the efficiency of subsequent processing.

[0025] Model application and continuous optimization: Obtain the image information of rice with the help of an agricultural image acquisition instrument, and extract the global feature vector from it ; Use the SPAD-502 chlorophyll meter to measure the physiological indicators of rice, and then form the physiological feature vector , the feature vector after spectral dimensionality reduction , the image global feature vector and the physiological feature vector are fused hierarchically, and the calculation formula is: , where , , are weight coefficients, and the optimal values are determined by the method of cross-validation to achieve the best fusion effect. Based on the fused feature vector and the rice variety label, a classification model is constructed. Let the fused feature vector be , and the corresponding rice variety label can be numerically encoded as . Let the mapping function of the model be: , where is the weight vector corresponding to the feature , is the bias term. To continuously optimize the model performance, continuously collect the rice sample data in different regions and at different growth stages in the base, and train and optimize the model.

[0026] Practical application and continuous monitoring: The trained and optimized model is deployed to the intelligent management system of the planting base. During the daily production process, hyperspectral data, image data, and physiological data of rice are regularly collected, and these data are input into the model for variety identification. The management system will record the identification results of each time in detail, and at the same time collect the newly emerging sample data and the sample data that is different from the existing data. Every once in a while, these data will be fed back to the model for further optimization, so that the model can better adapt to the changes in the environment in the planting base and the update of rice varieties.

[0027] In summary, this embodiment utilizes improved hyperspectral technology to achieve accurate variety identification in large-scale rice planting bases. By collecting multi-perspective data to obtain rich spectral information, extracting multi-scale features to mine features at different scales, processing spectral features and reducing dimensions to screen key information, fusing multi-features and optimizing models to improve identification accuracy, and finally deploying to actual production scenarios and continuously monitoring and optimizing. This method provides strong support for planting bases in variety management and yield prediction, helps improve planting efficiency, optimize resource allocation, promotes the development of smart agriculture, and has significant application value and economic benefits.

[0028] Embodiment 2: New variety screening in rice research institutions.

[0029] Multi-perspective data collection: In the experimental fields of rice research institutions, there are various newly cultivated rice varieties and common varieties as controls. To deeply study the characteristics of different varieties, researchers select rice plants that can represent the typical characteristics of each variety. Around these plants, multiple adjustable-angle hyperspectral sensors are reasonably arranged. At multiple important growth stages of rice, such as the seedling stage, heading stage, and maturity stage, hyperspectral data is synchronously collected from different perspectives. The collected data is orderly stored in the professional database of the research institution, providing comprehensive data support for subsequent research and analysis.

[0030] Multi-scale feature extraction: After the database receives the collected hyperspectral dataset, it first cleans and standardizes the data to ensure data quality. Researchers determine the value range and step size of the scale factor, the wavelength offset and the number of wavelength offsets according to the environmental characteristics of the experimental field, the spectral resolution, and the research objective. The wavelength offset is dynamically adjusted according to the spectral resolution of the hyperspectral sensor. The higher the spectral resolution, the smaller the wavelength offset. The specific adjustment rule is: Let the spectral resolution of the hyperspectral sensor be , define the basic wavelength offset as , and the actual wavelength offset is calculated by the formula: , where is the adjustment coefficient, and its value range is [0.8, 1.2]. The number of wavelength offsets is determined by analyzing the data characteristics. The weight coefficient is determined based on the research results and personal experience in related fields. The multi-scale spectral transformation function is used to transform the hyperspectral data, and feature vectors at different scales are extracted from the transformation results. Let the collected hyperspectral data be , define the scale factor as , and construct the transformation function , and the calculation formula is: , where, is the selected number of wavelength offsets, is the corresponding weight coefficient, is the wavelength offset dynamically adjusted according to the spectral resolution. By changing value, transformation results at different scales are obtained , laying a foundation for accurately analyzing varietal differences subsequently.

[0031] Spectral feature processing and dimensionality reduction: According to the previous research accumulation and the purpose of this experiment, researchers determine an appropriate number of samples. By calculating the correlation index at each wavelength, key bands that are significant for differentiating different rice varieties are screened out. The calculation formula is: , select greater than the threshold as the key bands. These key bands are composed into a sample feature vector, and a projection matrix is obtained by solving an optimization problem. The calculation formula is: . By solving the above optimization problem, the projection matrix is obtained. The sample feature vector is projected into a low-dimensional space to obtain a dimensionality-reduced feature vector , realizing the dimensionality reduction of the feature vector. This process helps to more clearly analyze and compare the characteristic differences between different varieties.

[0032] Model application and continuous optimization: Use an agricultural image acquisition device to obtain rice images and extract the global feature vectors therein; use a SPAD-502 chlorophyll meter to measure the physiological indicators of rice and form physiological feature vectors. The spectral dimensionality-reduced feature vector , the image global feature vector and the physiological feature vector are fused hierarchically. The calculation formula is: , where , , are weight coefficients to achieve the best fusion effect. Based on the fused feature vector and the rice variety labels, a classification model is constructed. Let the fused feature vector be , and the corresponding rice variety labels can be numerically encoded as . Let the mapping function of the model be: . To improve the accuracy and reliability of the model in new variety screening, researchers continuously collect data of rice samples planted in different batches in the experimental field and train and optimize the model.

[0033] Practical applications and continuous monitoring: The optimized trained model is applied to the new variety screening work in scientific research institutions. In each new variety cultivation experiment, scientific researchers collect relevant data of the newly cultivated rice, use the model to identify the differences from existing varieties, and assist in judging the characteristics and potential of the new varieties. Scientific researchers closely monitor the identification results of the model, collect new data, and regularly optimize the model to adapt to the changing research needs and the characteristics of newly emerging rice varieties.

[0034] In summary, in the new variety screening work of rice scientific research institutions, this method based on improved hyperspectral technology plays an important role. From data collection from multiple perspectives, to multi-scale feature extraction, spectral feature processing and dimensionality reduction, to multi-feature fusion for model construction and continuous optimization, it provides a scientific and efficient means of variety identification for scientific researchers. With the help of this method, scientific researchers can more accurately analyze the characteristics of new varieties, judge their potential, accelerate the cultivation process of new varieties, and contribute to more achievements in rice scientific research. It is of great significance for improving the rice breeding level in China and ensuring food security.

[0035] The above description is only a preferred embodiment of the present invention and does not impose any form of limitation on the present invention. Although the present invention has been disclosed above with the preferred embodiment, it is not intended to limit the present invention. Any person skilled in the art can make some changes or modifications to equivalent embodiments by using the disclosed technical content within the scope of the technical solution of the present invention. However, as long as it does not depart from the content of the technical solution of the present invention, any brief modification, equivalent change and modification made to the above embodiments based on the technical essence of the present invention still fall within the scope of the technical solution of the present invention.

Claims

1. A method for identifying rice varieties based on feature fusion of improved hyperspectral technology, characterized in that: The specific steps of this identification method are: S100, multi-view data acquisition: In the rice-growing area, select representative rice plants, arrange multiple hyperspectral sensors with adjustable angles around them, and synchronously collect rice hyperspectral data from different perspectives to form and store multi-view hyperspectral data sets; S200, multi-scale feature extraction: receiving the collected hyperspectral data set, cleaning and standardizing it, transforming it using a multi-scale spectral transformation function, and extracting feature vectors of different scales from the transformation results; S300, spectral feature processing and dimension reduction: determine the number of samples based on domain knowledge and analysis objectives, calculate the correlation index at each wavelength, screen out the bands closely related to rice varieties, form sample feature vectors, and reduce the dimension of the features; S400, model application and continuous optimization: Use agricultural image acquisition instrument to obtain rice images to extract global feature vectors, use SPAD-502 chlorophyll meter to measure rice physiological indicators to form physiological feature vectors, and use hierarchical multi-feature fusion formula to hierarchically fuse them with the feature vectors after spectral dimension reduction. Create a classification model based on the fused feature vectors and rice variety labels, and use sample data to train and optimize the model; S500, practical application and continuous monitoring: deploy the trained and optimized model to the actual agricultural production scenario, track and record the identification results, collect new sample data or different sample data, and provide regular feedback to optimize the model.

2. The method for identifying rice varieties based on feature fusion of improved hyperspectral technology according to claim 1, characterized in that: The specific steps of performing multi-scale transformation by using a multi-scale spectral transformation function in the multi-scale feature extraction in S200 are as follows: (1) Predetermine the range and step size of the scale factor; (2) Determine the wavelength offset based on spectral resolution and analysis requirements; (3) Analyze the characteristics of the data and determine the amount of wavelength shift; (4) Determine the weight coefficient based on relevant field data; (5) Based on the set parameters, the hyperspectral data is transformed using a multi-scale spectral transformation function. As the scale factors are taken in turn, the transformation results at different scales are obtained, and relevant features are extracted from the transformation results.

3. The method for identifying rice varieties based on feature fusion of improved hyperspectral technology according to claim 2, characterized in that: In S200, the wavelength offset in the multi-scale feature extraction is dynamically adjusted according to the spectral resolution of the hyperspectral sensor. The higher the spectral resolution, the smaller the wavelength offset. The specific adjustment rule is: assuming that the spectral resolution of the hyperspectral sensor is , the basic wavelength offset is defined as , the actual wavelength shift The calculation formula is: ,in It is an adjustment coefficient used to dynamically calibrate the wavelength offset, and its value range is [0.8, 1.2].

4. The method for identifying rice varieties based on feature fusion of improved hyperspectral technology according to claim 2, characterized in that: In the S200, the multi-scale feature extraction uses a multi-scale spectral transformation function to transform the hyperspectral data. Suppose the collected hyperspectral data is , define the scale factor as , used to control the granularity of spectral transformation and construct transformation function , the calculation formula is: ,in, is the selected wavelength shift amount, which is determined by the characteristics of the spectral data. is the corresponding weight coefficient, reflecting the importance of different offsets, To dynamically adjust the wavelength offset according to the spectral resolution, by changing The value of , get the transformation results at different scales .

5. The method for identifying rice varieties based on feature fusion of improved hyperspectral technology according to claim 1, characterized in that: S300, spectral feature processing and dimensionality reduction through correlation index To screen the bands closely related to rice varieties, the calculation formula is: ,in, The wavelength of different samples The spectral value at is the average value of all samples at this wavelength, is the sample size, select Greater than threshold As the key band, the threshold Determined through statistical analysis of historical data.

6. The method for identifying rice varieties based on feature fusion of improved hyperspectral technology according to claim 1, characterized in that: In the S300, in the spectral feature processing and dimensionality reduction, the projection matrix is ​​obtained by solving the optimization problem to reduce the dimension of the feature vector, and the projection matrix is ​​defined as: , the calculation formula is: ,in, is the sample feature vector composed of key bands, is the number of sample feature vectors, It is a candidate solution of the projection matrix, which is used to optimize the dimensionality reduction process. The projection matrix is ​​obtained by solving , project the sample feature vector into the low-dimensional space to obtain the feature vector after dimensionality reduction .

7. The method for identifying rice varieties based on feature fusion of improved hyperspectral technology according to claim 1, characterized in that: In the S400, in the model application and continuous optimization, feature fusion is performed through a hierarchical multi-feature fusion formula, and the image feature of the extracted rice is assumed to be , extract physiological features as vectors , the eigenvector after spectrum dimension reduction , image global feature vector and physiological feature vector Feature fusion is performed through the hierarchical multi-feature fusion formula, the formula is: ,in , , is the weight coefficient, and the optimal value is determined by cross-validation method.

8. The method for identifying rice varieties based on feature fusion of improved hyperspectral technology according to claim 1, characterized in that: In S400, the creation of a classification model in model application and continuous optimization, the fused feature vector is assumed to be , the corresponding rice variety label is expressed as , let the mapping function of the model be: ,in It is a feature The corresponding weight vector, is the bias term.

9. The method for identifying rice varieties based on feature fusion of improved hyperspectral technology according to claim 1, characterized in that: The specific steps of classification model training and optimization in S400, model application and continuous optimization are: (1) Collect rice samples and obtain the fusion feature vector for each sample , and the corresponding rice variety labels are numerically encoded to obtain , organize the feature vectors and variety labels of all samples into a training data set ,in is the number of samples, and the data set is divided into training set and test set in proportion; (2) Weight vector in the model and the bias term Initialize; (3) Take a batch of samples from the training set and set the batch size to , for each sample in the batch , and fuse the feature vector Input into the model, the formula is: , calculate the predicted value ; (4) Based on the calculated predicted value And the actual species label of the sample , use the loss function to calculate the loss value of the batch of samples, the formula is: , which measures the difference between the model prediction and the actual label; (5) Calculate the loss function through the back propagation algorithm For the weight vector and the bias term The gradient of (6) Based on the calculated gradient, the weight vector is updated using the stochastic gradient descent algorithm and the bias term , the update formula is: ; ,in is the learning rate, which is a pre-set hyperparameter that controls the step size of parameter updates; (7) Repeat (3) to (6) to traverse the entire training set and complete a training cycle. training cycles, The preset number of training cycles is dynamically adjusted according to the model convergence; (8) After completing all training cycles, use the test set to evaluate the trained model and retrain the model based on the evaluation results.

Images