Hyperspectral imaging method and system based on machine learning and application of hyperspectral imaging method and system in cold resistance evaluation of pennisetum alopecuroides

Abstract

The invention belongs to the technical field of plant cold resistance detection, and particularly relates to a hyperspectral imaging method and system based on machine learning and application of the hyperspectral imaging method and system in cold resistance evaluation of pennisetum alopecuroides. The invention discloses a method and a system for predicting cold resistance of pennisetum alopecuroides based on a hyperspectral imaging technology of machine learning. The prediction method comprises the following steps: acquiring a pennisetum alopecuroides leaf hyperspectral image by using a hyperspectral imaging system, determining a region of interest, and correcting the region of interest to obtain standard hyperspectral data; obtaining the value of the low-temperature induction component of the leaf, calculating the weight ratio of each physiological index, and determining that the proline content is the physiological index with the maximum weight ratio; preprocessing the average hyperspectral data and screening characteristic wave bands; and constructing an estimated remote sensing monitoring image under the low-temperature stress of the pennisetum alopecuroides by utilizing the characteristic wave bands and different models. According to the method, the cold resistance of the pennisetum alopecuroides can be nondestructively and effectively detected, so that feasibility is provided for low-temperature stress diagnosis and breeding of the pennisetum alopecuroides.

Description

Technical Field

[0001] This invention belongs to the field of plant cold resistance detection technology, specifically involving a hyperspectral imaging method and system based on machine learning and its application in evaluating the cold resistance of Napier grass. Background Technology

[0002] In plant production practices, *Pennisetum alopecuroides* is susceptible to low-temperature stress, and low temperature has become one of the major meteorological hazards restricting its growth, development, and production stability. Developing and promoting cold-resistant cultivation techniques is an effective way to address this challenge, but the prerequisite for implementing this technique is the ability to quickly and accurately assess the cold resistance of *Pennisetum alopecuroides*. Therefore, how to establish an efficient and reliable method for evaluating cold resistance has become a key issue that urgently needs to be addressed.

[0003] Traditionally, the assessment of plant cold resistance has relied primarily on detecting physiological and biochemical damage indicators, such as measuring electrolyte permeability, malondialdehyde content, or antioxidant enzyme activity to determine the degree of cell damage. While these methods can reflect the physiological state of plants, they typically have significant limitations: the detection process is cumbersome and time-consuming, requires destructive sampling, is inefficient, and demands a high level of expertise from both the equipment and the operators, making them relatively expensive and unsuitable for large-scale, real-time, and rapid on-site screening applications.

[0004] In recent years, hyperspectral imaging technology, as an emerging non-destructive testing method, has demonstrated significant advantages in plant phenotypic analysis and stress diagnosis. This technology can simultaneously acquire spatial and continuous spectral information of samples, and is characterized by its speed, efficiency, non-destructive nature, and real-time, high-throughput data acquisition. It provides a powerful tool for accurately and dynamically monitoring changes in the physiological and biochemical components within leaves, thus opening new possibilities for the objective and rapid assessment of crop cold resistance. Currently, research on the inversion of plant physicochemical parameters and stress monitoring using hyperspectral imaging technology has made some progress. For example, existing studies have established a spectral regression model for the degree of frost damage to tea trees based on hyperspectral data, successfully achieving non-destructive monitoring and inversion of the degree of frost damage in tea trees. These works have verified the feasibility of applying hyperspectral technology to the diagnosis of plant cold damage.

[0005] However, despite the great potential of hyperspectral imaging technology in crop monitoring, there are few reports on its application specifically for the accurate and rapid assessment of the cold tolerance of Napier grass. Existing technologies are mostly focused on economic crops such as tea trees, and there is a lack of models that establish a correlation between hyperspectral characteristics and cold tolerance for the specific species Napier grass. There is also a lack of reliable methods and model systems specifically applicable to the rapid and non-destructive detection of the cold tolerance of Napier grass.

[0006] Therefore, this study utilizes hyperspectral imaging technology to obtain hyperspectral images and spectral data of Napier grass leaves, analyzes and screens characteristic spectral information highly correlated with cold resistance, and constructs an efficient and robust prediction model to achieve rapid, non-destructive, and accurate diagnosis of Napier grass's cold resistance. This has important theoretical and practical significance for early monitoring of low-temperature stress in Napier grass, breeding of cold-resistant varieties, and field cultivation management.

[0007] Furthermore, on the one hand, there are differences in understanding among those skilled in the art; on the other hand, the inventors studied a large number of documents and patents when making this invention, but due to space limitations, not all details and contents were listed in detail. However, this does not mean that the present invention does not possess the features of these prior art. On the contrary, the present invention already possesses all the features of the prior art, and the applicant reserves the right to add relevant prior art to the background art. Summary of the Invention

[0008] This invention belongs to the field of plant cold resistance detection technology, specifically involving a hyperspectral imaging method and system based on machine learning and its application in evaluating the cold resistance of Napier grass.

[0009] To address the aforementioned technical problems, one objective of this invention is to provide a hyperspectral imaging method based on machine learning, which includes the following steps: S1 Obtain leaf samples of Napier grass under low temperature stress, acquire hyperspectral images of the leaves using a hyperspectral imaging system, determine the region of interest of the leaves, obtain a standard hyperspectral image of reflectance, and measure the low temperature-induced physiological and biochemical indicators of each leaf sample, and calculate the weight ratio of each physiological indicator. S2 obtains the average spectral data from the standard hyperspectral images of each sample obtained in step S1, and uses a preprocessing and feature wavelength selection algorithm to extract a subset of feature wavelengths that are significantly related to proline content; S3 uses the reflectance data of the feature wavelength subset obtained from S2 as the independent variable and the physiological and biochemical indicators with the largest weight in S1 as the dependent variable. A regression prediction model is trained using a machine learning algorithm and then analyzed using R... 2 The best predictive model is selected through validation and RMSE.

[0010] According to a preferred embodiment, the feature band selection algorithm includes the SPA algorithm and the CARS algorithm.

[0011] According to a preferred embodiment, in S3, before the characteristic wavelength selection algorithm, the spectral data is further preprocessed. The preprocessing algorithms include SG, MSC-SG, SNV-SG, and 1 st Der-SG or 2 nd Der-SG algorithm.

[0012] According to a preferred embodiment, the characteristic wavelength subset includes 397.66 nm, 418.62 nm, 421.24 nm, 423.86 nm, 426.49 nm, 431.74 nm, 434.37 nm, 439.63 nm, 442.26 nm, 447.52 nm, 452.79 nm, 455.43 nm, 458.06 nm, 463.34 nm, 468.62 nm, 473.9 nm, 479.18 nm, 481.83 nm, 487.12 nm, 492.42 nm, 503.02 nm, 510.98 nm, 518.95 nm, 526.93 nm, 534.91 nm, 548.24 nm, 550.91 nm, 561.59 nm, and 572.29 nm. nm, 585.68 nm, 591.04 nm, 599.1 nm, 607.16 nm, 615.23 nm, 623.3 nm, 628.69 nm, 639.48 nm, 644.88 nm, 652.99 nm, 666.52 nm, 671.94nm, 677.36 nm, 685.5 nm, 699.09 nm, 712.7 nm, 731.79 nm, 745.45 nm, 753.66 nm, 761.88nm, 767.36 nm, 775.6 nm, 792.08 nm, 797.59 nm, 803.1 nm, 808.61 nm, 816.88 nm, 825.16nm, 833.45 nm, 841.75 nm, 847.29 One or more of the following wavelengths: nm, 852.83 nm, 855.6 nm, 863.92 nm, 869.47 nm, 872.25 nm, 880.58 nm, 886.15 nm, 891.71 nm, 897.28 nm, 902.86 nm, 908.43 nm, 914.02 nm, 919.6 nm, 927.98 nm, 930.78 nm, 933.58 nm, 939.18 nm, 941.98 nm, 947.58 nm, 950.38 nm, 955.99 nm, 958.8 nm, 961.6 nm, 964.41 nm, 967.22 nm, 970.03 nm, 972.84 nm, 975.65 nm, 978.46 nm, and 981.27 nm.

[0013] According to a preferred embodiment, the malondialdehyde content, soluble sugar content, soluble protein content, proline content, and chlorophyll content of *Phragmites australis* were collected at temperatures of 10°C and below, and the weighting percentage of each physiological indicator was calculated based on the following formula: (1); (2); (3) (4); Among them, in formula (1) It is the kth principal component, in formula (1) These are the standardized original variables, in formula (1) It is the loading of the p-th index on the k-th principal component, in formula (1) It is the loading of the first index on the kth principal component, in formula (1) It is the loading of the second index on the k-th principal component. In formula (2) It is the eigenvalue of the k-th principal component. It is the feature value of the i-th indicator. It is the variance contribution rate of the k-th principal component. In formula (3) It is the loading of the i-th variable on the k-th principal component. In formula (4) It is the weight of the i-th indicator. In the formula, m and p both represent the number of indicators, i.e., the number of original variables; MAD: malondialdehyde content, in nmol / mg; SS: soluble sugar content, in mg / ml; Pro: proline content, in nmol / mg; SP: soluble protein content, in mg / ml; Ch1: chlorophyll content, in μg / mL.

[0014] Preferably, the proline content parameter of *Pennisetum alopecuroides* is collected to predict its cold resistance. Preferably, physiological and biochemical indicators of *Pennisetum alopecuroides* are collected on days 1, 3, 5, and 7.

[0015] According to a preferred embodiment, R 2 RMSE and RMSE are calculated using formulas (5) and (6) respectively: (5), (6), in, and Let represent the predicted value and the measured value of the i-th sample in the sample set, respectively. denoted as the average of the measured values ​​in the sample set, and n represents the number of samples in the sample set.

[0016] According to a preferred embodiment, in S3, the screening criteria are: Test the R-squared of the regression prediction model 2 The larger; or R 2 The larger the value, the smaller the RMSE.

[0017] According to a preferred embodiment, the machine learning algorithm is a PLSR, SVR, or RF neural network.

[0018] According to a preferred embodiment, the hyperspectral imaging method further includes: S5. Collect hyperspectral images of Napier grass leaves, extract the characteristic wavelengths of the spectral data of the images in the same way as in step S4, and input them into the best prediction model to calculate the predicted proline content of each plant, thereby generating a visual image reflecting the spatial distribution of cold resistance.

[0019] One of the objectives of this invention is to provide a system for predicting the cold hardiness of Napier grass using the machine learning-based hyperspectral imaging method described above, characterized in that the system comprises: The data acquisition module is used to acquire hyperspectral image data of Napier grass leaves; The processing module communicates with the data acquisition module and is used to obtain the current cold resistance assessment results of Napier grass based on hyperspectral image data.

[0020] According to a preferred embodiment, the data acquisition module includes a hyperspectral camera. The hyperspectral camera has an image resolution of 640×640 px, a spectral resolution of 8 nm, a 280 W halogen lamp as the light source, and a lens height of 32 cm above the sample. Preferably, the acquisition method involves using the hyperspectral camera to capture images from a top-down view of a horizontally positioned Napier grass leaf.

[0021] According to a preferred embodiment, the specific steps of S2 are as follows: The average hyperspectral data of standard hyperspectral images were collected, and the region of interest was determined by thresholding using the environmental visualization program ENVI. The average band values ​​of the leaves of Napier grass in the region of interest were then extracted. Within the environmental visualization program, algorithms are used to denoise the average band values ​​in S1, reducing noise in the average hyperspectral data and improving the availability of effective information.

[0022] According to a preferred embodiment, the specific steps in S3 for training the regression prediction model using a machine learning algorithm are as follows: When building a regression prediction model, all datasets are divided into 80% training set and 20% test set, using the coefficient of determination R0. 2 The root mean square error (RMSE) is used to evaluate the performance of the regression prediction model and select the best model. Physiological and biochemical indicators of Pennisetum under low temperature stress were fitted with hyperspectral data to construct a regression prediction model for Pennisetum cold resistance. To evaluate the accuracy of the regression prediction models described above, the estimates of all regression prediction models are compared to verify the stability of the regression prediction models.

[0023] One of the objectives of this invention is to provide an application of the optimal prediction model constructed based on the above-mentioned machine learning-based hyperspectral imaging method in the evaluation of the cold resistance of Napier grass.

[0024] The beneficial effects of this technical solution are: This invention creatively combines hyperspectral imaging technology with machine learning algorithms to establish a non-destructive and rapid method for predicting the cold resistance of Napier grass. The beneficial effects of this technical solution are reflected in: This method achieves technological integration and innovation. By selecting physiological indicators with high weighting and utilizing machine learning models, it accurately establishes a quantitative inversion relationship between the score and hyperspectral features, realizing for the first time a direct and stable estimation of comprehensive cold resistance from spectral information. Simultaneously, it achieves a significant methodological breakthrough. This method is entirely based on spectral analysis, enabling dynamic and continuous monitoring of the cold resistance physiological state of living leaves without damaging plant tissue. Combined with feature band screening technology, it significantly improves the model's efficiency and interpretability. Furthermore, it has significant advantages in application value. This method is not only highly efficient and low-cost, but more importantly, it can generate remote sensing monitoring images reflecting the spatial distribution of cold resistance, achieving a leap from "point" measurement to "area" assessment. It provides a revolutionary technical tool for large-scale phenotypic screening in the field, cold resistance breeding screening, and precise diagnosis of low-temperature stress. Attached Figure Description

[0025] Figure 1 This is a flowchart of a method for predicting the cold resistance of Napier grass based on machine learning hyperspectral imaging technology according to the present invention; Figure 2 Spectral curves for different preprocessing methods; Figure 3The feature bands selected for competitive adaptive reweighted sampling (CARS) are denoised based on the average hyperspectral data of all samples using a built-in environment visualization program. The figure shows the spectral intensity after denoising based on the original data (RAW), Savitzky-Golay smoothing (SG), spectral multivariate scattering correction-Savitzky-Golay smoothing (MSC-SG), standard normal variable transformation-Savitzky-Golay smoothing (SNV-SG), first derivative-Savitzky-Golay smoothing (1st Der-SG), and second derivative-Savitzky-Golay smoothing (2nd Der-SG). Figure 4 The denoising process of the average hyperspectral data of all samples for the feature bands selected by the Continuous Projection Algorithm (SPA) is based on the built-in environment visualization program. The figure shows the spectral intensity after denoising based on the original data (RAW), Savitzky-Golay smoothing (SG), spectral multivariate scattering correction-Savitzky-Golay smoothing (MSC-SG), standard normal variable transformation-Savitzky-Golay smoothing (SNV-SG), first derivative-Savitzky-Golay smoothing (1st Der-SG), and second derivative-Savitzky-Golay smoothing (2nd Der-SG). Figure 5 The figure shows the validation results based on the PLSR model; Figure 6 The graph shows the validation results based on the RF model. Figure 7 The image shows the verification results based on the SVR model. Detailed Implementation

[0026] In the description of this invention, terminology is used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly defined.

[0027] Unless otherwise specified, the experimental methods used in the following examples are all conventional methods; the materials, reagents or instruments used, unless otherwise specified by the manufacturer, are all commercially available reagents and materials; the conditions not specified in the examples are all carried out according to conventional conditions or conditions recommended by the manufacturer. At the same time, the present invention does not limit the source of the raw materials used. Unless otherwise specified, the raw materials used in the present invention are all commercially available products in this technical field.

[0028] RAW: Raw data.

[0029] SG: SG filtering smoothing process (Savitzky-Golay).

[0030] MSC-SG: Multiplicative scatter correction - Savitzky - Golay.

[0031] SNV-SG: Standard Normal Variate-Savitzky-Golay.

[0032] 1st Der-SG: First derivative-Savitzky-Golay filtering smoothing.

[0033] 2nd Der-SG: Second Derivative-Savitzky-Golay filtering smoothing process.

[0034] SPA: Successive projections algorithm.

[0035] CARS: Competitive adaptive reweighted sampling.

[0036] The data acquisition apparatus described in the following embodiments includes a hyperspectral camera, a darkroom, and a low-temperature index detection device. The hyperspectral camera is used to acquire hyperspectral images of *Pennisetum alopecuroides* leaves. The darkroom is equipped with a fixed light source for housing the hyperspectral camera to acquire spectral images of the *Pennisetum alopecuroides* leaves. The low-temperature index detection device (e.g., a reagent kit) is used to determine the values ​​of low-temperature induced components in the *Pennisetum alopecuroides* leaves.

[0037] In the following embodiments, R 2 The measured values ​​involved in the RMSE screening model are the leaf proline content actually measured in the laboratory using traditional chemical methods (such as reagent kits); the predicted values ​​are the proline content values ​​calculated and output by the model after inputting the spectral data of the same leaf into the constructed regression model.

[0038] To find the optimal model, this technical solution employs cross-validation. The method involves dividing all sample data into a training set (80%) and a validation set (20%). Different models (PLSR, SVR, RF) are "trained" using the training set data (including spectral X and measured proline Y), allowing the models to learn the relationship between X and Y. The trained models are then used to predict the proline content of samples in the validation set (which the models have never seen before), yielding "predicted values." Then, the coefficient of determination R between the "predicted values" and "measured values" on the validation set is calculated using formulas (5) and (6). 2 And the root mean square error RMSE. 2 The larger the value (closer to 1), the better the model's predictions match the actual values, and the stronger the model's interpretability. The smaller the RMSE, the smaller the model's prediction error and the higher its accuracy.

[0039] Select the best performing (R) on the validation set. 2 The model with the largest RMSE (smallest RMSE) is selected as the final best prediction model. According to Table 6, the R-value of the 2nd Der-SG-SPA-SVR model on the validation set... 2 It has the highest value (0.877) and the lowest RMSE (29.086), so it was selected as the optimal model.

[0040] The technical solution involved in this invention involves using principal component analysis to screen the most representative proline from multiple traditional biochemical indicators as the target, and then using machine learning algorithms to establish a quantitative relationship model between proline and hyperspectral data. Ultimately, the application prospect of this optimal model is that by performing a single hyperspectral scan on unknown Napier grass leaves and inputting the spectral data, the predicted value of its proline content can be quickly and non-destructively output, and its cold resistance can be assessed based on the proline content.

[0041] Example 1 A method for predicting the cold resistance of Napier grass based on machine learning hyperspectral imaging technology.

[0042] The samples for the test set and training set are recorded in the germplasm resource information table for the tested materials (Table 1).

[0043] Table 1

[0044] See flowchart Figure 1 It includes the following steps: Step 1: Acquisition of hyperspectral images, including the following specific steps. 1.1 Ensure the leaf to be tested is placed horizontally and set the parameters of the hyperspectral camera. Use a hyperspectral camera (Gaiafieldpro) Hyperspectral images of Napier grass leaves were acquired using a Dualix Spectral Imaging (V10, China) camera. The camera's field of view was 29°, and the object distance was 32 cm. The camera resolution was set to 640 × 640 px, with a spectral resolution of 8 nm. A total of 224 bands were used in the hyperspectral camera, covering the range of 400–1000 nm.

[0045] 1.2 The acquired hyperspectral images were preprocessed and corrected using SpecView software (V1.0, Dualix Spectral Imaging, China).

[0046] Step 2: Obtaining Physiological and Biochemical Values ​​of Napier Grass Leaves The physiological and biochemical values ​​of various low-temperature induced components in mature leaves were measured from the collected hyperspectral images.

[0047] The collected Napier grass leaves were flash-frozen in liquid nitrogen and then ground into a fine powder for subsequent physiological and biochemical index determination. The contents of soluble sugar (SS), soluble protein (SP), proline (Pro), chlorophyll (Chl), and malondialdehyde (MDA) were determined using kits purchased from Jiangsu Aidisheng Biological Technology Co., Ltd., strictly following the instructions and employing the microplate method. The corresponding product codes for each kit are: SS (ADS-W-TDX039), SP (ADS-W-SP001), Pro (ADS-W-AJS004), Chl (ADS-W-GH001), and MDA (ADS-W-YH002).

[0048] Step 3: Calculate the weighting percentage of physiological indicators of low-temperature stress in *Phragmites australis*. To quantify the physiological response of *Phragmites australis* under low-temperature stress, biological software was used to calculate the weight of each physiological indicator. First, five key low-temperature response physiological indicators (SS, SP, Pro, Chl, and MDA) were standardized and analyzed. Then, based on principal component analysis, the contribution coefficient of each original variable to each principal component was calculated using formulas (1), (2), (3), and (4). ) and variance contribution rate ( ), and finally the weight of the indicator is measured ( ).

[0049] (1) (2) (3) (4); Among them, in formula (1) It is the kth principal component, in formula (1) These are the standardized original variables, in formula (1) It is the loading of the p-th index on the k-th principal component, in formula (1) It is the loading of the first index on the kth principal component, in formula (1) It is the loading of the second index on the k-th principal component. In formula (2) It is the eigenvalue of the k-th principal component. It is the feature value of the i-th indicator. It is the variance contribution rate of the k-th principal component. In formula (3) It is the loading of the i-th variable on the k-th principal component. In formula (4) It is the weight of the i-th indicator. In the formula, m and p both represent the number of indicators, that is, the number of original variables.

[0050] It should be noted that, These are the standardized raw physiological indicator data. In this embodiment, p=5, and they correspond to the contents of malondialdehyde (MDA), soluble sugar (SS), proline (Pro), soluble protein (SP), and chlorophyll (Chl), respectively. Standardization is used to eliminate the influence of differences in units and magnitudes among different indicators. This is the "loading coefficient" of the p-th original index (e.g., malondialdehyde) on the k-th principal component. The loading coefficient is directly obtained by performing principal component analysis on the standardized original data.

[0051] Formula (1) defines the principal components. It is a linear combination of the original variables X, and the combination coefficient is the load v.

[0052] Formula (2) calculates the variance contribution rate of each principal component. This value represents the percentage of total information in the original data that the principal component can explain.

[0053] Formula (3) calculates the absolute loadings of the same index (e.g., proline) on all principal components. The weighted sum is calculated, and the weights are the variance contribution rates of the corresponding principal components. ).

[0054] Formula (4) is the comprehensive score obtained from Formula (3). Normalization is performed to make the sum equal to 1, thus obtaining the final weight value for each indicator. .

[0055] The principal component table of cold-resistance related traits of Napier grass is shown in Table 2.

[0056] MDA, SS, Pro, SP, and Chl can all be used as key cold-resistance physiological indicators characterizing the low-temperature stress response of *Phragmites australis*. Following the relevant biological software analysis workflow, to integrate multi-dimensional information and reduce data redundancy, this study conducted principal component analysis (PCA) on the above five indicators. The weights of each physiological indicator were calculated based on the variance contribution rate and loadings of each principal component. The results showed that proline content had the largest proportion among the five physiological indicators, reaching 0.3526, while malondialdehyde content had the smallest proportion, at 0.2643. These results indicate that among the selected five indicators, the change in proline content is most closely related to the overall physiological response of *Phragmites australis* under low-temperature stress (represented by the combined principal components). Based on the comprehensive evaluation of multiple indicators, this embodiment uses proline content as the core target variable (dependent variable) for subsequent model prediction.

[0057] The dependent variable (Y) is the "physiological and biochemical indicator with the largest weight," namely proline content. In model building, it is considered the "true value," what the model attempts to predict. The independent variable (X) is the "reflectance data of a subset of characteristic wavelengths." This data comes from hyperspectral images, representing the spectral information of Napier grass leaves at different specific wavelengths. The dependent variable (Y) and the independent variable (X) are not causally related, but rather statistically correlated. The model's logic is that low-temperature stress causes changes in the internal physiological and biochemical components of Napier grass leaves (especially proline), and these internal changes affect the leaf's reflectance and absorption characteristics at specific wavelengths. Therefore, a mathematical function Y=f(X) can be learned and established through machine learning algorithms to describe this correlation between "spectral reflectance (X)" and "proline content (Y)." Once this function (i.e., the model) is established and validated, in the future, it will be possible to quickly predict the proline content (Y) of unknown samples by simply inputting their spectral data (X), thus achieving non-destructive and rapid detection.

[0058] Therefore, proline content can serve as a key physiological indicator of cold resistance, providing accurate values ​​for subsequent construction of models to predict proline content.

[0059] Table 2

[0060] Step four, preprocessing and selection of characteristic bands, includes the following specific steps: (1) The average hyperspectral data of the corrected hyperspectral image was collected. The region of interest (ROI) was determined by thresholding using the environmental visualization program ENVI. The average band values ​​of the hyperspectral data of the collected Napier grass leaves were then extracted, such as... Figure 2 As shown.

[0061] (2) Using SG, MSC-SG, SNV-SG, 1 st Der-SG, 2 nd Multiple algorithms, including Der-SG, first denoise the average hyperspectral data of *Pennisetum alopecuroides* leaves, reducing noise and highlighting effective data. Then, the denoised average hyperspectral data is processed using the SPA and CARS algorithms to filter feature bands and extract characteristic data from the average band values ​​of these *Pennisetum alopecuroides* leaves. Figure 3 , 4 As shown in the figure, see Table 3.

[0062] Step 5: Establishing the regression model, including the following specific steps: (1) A regression prediction model was established using machine learning methods (PLSR, SVR, RF), and the best model was selected. PLSR integrates the advantages of multiple linear regression, principal component analysis and canonical correlation analysis; SVM is based on nonlinear mapping theory and has relatively low robustness; RF combines multiple weak classifiers, and the final result is obtained by voting or taking the average, so that the overall model has high accuracy and generalization performance.

[0063] (2) When building the model, all datasets are divided into 80% training set and 20% test set. The coefficient of determination (R²) is used to determine the training set. 2 The performance of the model is evaluated using R and the root mean square error (RMSE). 2 A larger R value indicates a smaller RMSE, signifying better model performance. 2 RMSE and RMSE are calculated using formulas (5) and (6) respectively: (5) (6) in, and Let represent the predicted value and the measured value of the i-th sample in the sample set, respectively. denoted as the average of the measured values ​​in the sample set, and n represents the number of samples in the sample set.

[0064] (3) Physiological and biochemical indicators of Napier grass were fitted with hyperspectral data to establish a prediction model for Napier grass cold resistance. Tables 4, 5 and 6 show the effect analysis of the established SVR, PLSR and RF prediction models.

[0065] (4) To evaluate the inversion accuracy of each model, the predicted values ​​of all models were compared, and the model stability of the cold resistance of *Phragmites australis* was verified, such as... Figure 5 , 6 As shown in Figure 7 ( Figure 5, 6 7 corresponds to the values ​​in Tables 4, 5, and 6). When the R of the test model... 2 The larger the value and the smaller the RMSE, the better the stability of the model. As can be seen from the above examples, 2 nd Der-SG-SPA-SVR's R 2 The highest value and the lowest RMSE indicate that this model is best suited for predicting the cold hardiness of Napier grass.

[0066] The results show that using the characteristic bands in Table 6 to scan hyperspectral images, combined with 2 nd The Der-SG-SPA-SVR model is the most suitable method for predicting the cold hardiness of Pennisetum acutum.

[0067] Preferably, the characteristic band subset includes 397.66 nm, 418.62 nm, 421.24 nm, 423.86 nm, 426.49 nm, 431.74 nm, 434.37 nm, 439.63 nm, 442.26 nm, 447.52 nm, 452.79 nm, 455.43 nm, 458.06 nm, 463.34 nm, 468.62 nm, 473.9 nm, 479.18 nm, 481.83 nm, 487.12 nm, 492.42 nm, 503.02 nm, 510.98 nm, 518.95 nm, 526.93 nm, 534.91 nm, 548.24 nm, 550.91 nm, 561.59 nm, 572.29 nm, and 585.68 nm. nm, 591.04 nm, 599.1 nm, 607.16 nm, 615.23 nm, 623.3nm, 628.69 nm, 639.48 nm, 644.88 nm, 652.99 nm, 666.52 nm, 671.94 nm, 677.36 nm, 685.5 nm, 699.09 nm, 712.7 nm, 731.79 nm, 745.45 nm, 753.66 nm, 761.88 nm, 767.36nm, 775.6 nm, 792.08 nm, 797.59 nm, 803.1 nm, 808.61 nm, 816.88 nm, 825.16 nm, 833.45nm, 841.75 nm, 847.29 nm, 852.83 One or more of the following wavelengths: nm, 855.6 nm, 863.92 nm, 869.47 nm, 872.25 nm, 880.58 nm, 886.15 nm, 891.71 nm, 897.28 nm, 902.86 nm, 908.43 nm, 914.02 nm, 919.6 nm, 927.98 nm, 930.78 nm, 933.58 nm, 939.18 nm, 941.98 nm, 947.58 nm, 950.38 nm, 955.99 nm, 958.8 nm, 961.6 nm, 964.41 nm, 967.22 nm, 970.03 nm, 972.84 nm, 975.65 nm, 978.46 nm, and 981.27 nm.

[0068] Table 3. Sensitive characteristic wavelengths extracted from spectral samples

[0069] Table 4. Effect of PLSR on estimating proline content in Napier grass

[0070] Table 5. Analysis of the effect of RF on estimating proline content in Napier grass

[0071] Table 6. Analysis of the effect of SVR on estimating proline content in *Sophora japonica*

[0072] This invention presents a method for predicting the cold resistance of Napier grass using hyperspectral imaging technology based on machine learning. Traditional methods for determining low-temperature induced components rely on manual methods and experience-based judgment, which are prone to misjudgment and have low detection efficiency. Therefore, this invention combines hyperspectral imaging technology with machine learning methods to predict the cold resistance of Napier grass by applying the degree of low-temperature stress.

[0073] It should be noted that the specific embodiments described above are exemplary, and those skilled in the art can devise various solutions inspired by the disclosure of this invention. These solutions all fall within the scope of this invention and its protection. Those skilled in the art should understand that this specification and its accompanying drawings are illustrative and not intended to limit the scope of the claims. The scope of protection of this invention is defined by the claims and their equivalents.

Claims

1. A hyperspectral imaging method based on machine learning, characterized in that, Includes the following steps: S1 Obtain leaf samples of Napier grass under low temperature stress, acquire hyperspectral images of the samples using a hyperspectral imaging system, determine the region of interest of the leaf, obtain a standard hyperspectral image of reflectance, and measure the low temperature-induced physiological and biochemical indicators of each leaf sample, and calculate the weight ratio of each physiological indicator. S2 obtains the average spectral data from the standard hyperspectral images of each sample obtained in step S1, and uses a preprocessing and feature wavelength selection algorithm to extract a subset of feature wavelengths that are significantly related to proline content; S3 uses the reflectance data of the feature wavelength subset obtained from S2 as the independent variable and the physiological and biochemical indicators with the largest weight in S1 as the dependent variable. A regression prediction model is trained using a machine learning algorithm and then analyzed using R... 2 The best predictive model is selected through validation and RMSE.

2. The hyperspectral imaging method based on machine learning according to claim 1, characterized in that, The characteristic wavelength subset includes 397.66 nm, 418.62 nm, 421.24 nm, 423.86 nm, 426.49 nm, 431.74 nm, 434.37 nm, 439.63 nm, 442.26 nm, 447.52 nm, 452.79 nm, 455.43 nm, 458.06 nm, 463.34 nm, 468.62 nm, 473.9 nm, 479.18 nm, 481.83 nm, 487.12 nm, 492.42 nm, 503.02 nm, 510.98 nm, 518.95 nm, 526.93 nm, 534.91 nm, 548.24 nm, 550.91 nm, 561.59 nm, 572.29 nm, and 585.68 nm. nm, 591.04 nm, 599.1 nm, 607.16 nm, 615.23 nm, 623.3 nm, 628.69 nm, 639.48nm, 644.88 nm, 652.99 nm, 666.52 nm, 671.94 nm, 677.36 nm, 685.5 nm, 699.09 nm, 712.7nm, 731.79 nm, 745.45 nm, 753.66 nm, 761.88 nm, 767.36 nm, 775.6 nm, 792.08 nm, 797.59 nm, 803.1 nm, 808.61 nm, 816.88 nm, 825.16 nm, 833.45 nm, 841.75 nm、847.29nm、852.83 One or more of the following wavelengths: nm, 855.6 nm, 863.92 nm, 869.47 nm, 872.25 nm, 880.58 nm, 886.15 nm, 891.71 nm, 897.28 nm, 902.86 nm, 908.43 nm, 914.02 nm, 919.6 nm, 927.98 nm, 930.78 nm, 933.58 nm, 939.18 nm, 941.98 nm, 947.58 nm, 950.38 nm, 955.99 nm, 958.8 nm, 961.6 nm, 964.41 nm, 967.22 nm, 970.03 nm, 972.84 nm, 975.65 nm, 978.46 nm, and 981.27 nm.

3. The hyperspectral imaging method based on machine learning according to claim 1, characterized in that, In S3, the spectral data undergoes preprocessing before the characteristic wavelength selection algorithm. The preprocessing algorithms include SG, MSC-SG, SNV-SG, and 1... st Der-SG, 2 nd Der-SG algorithm.

4. The hyperspectral imaging method based on machine learning according to claim 1, characterized in that, Physiological and biochemical indicators of Napier grass were collected at temperatures of 10℃ and below. These indicators included malondialdehyde content, soluble sugar content, soluble protein content, proline content, or chlorophyll content.

5. The hyperspectral imaging method based on machine learning according to claim 1, characterized in that, Feature band selection algorithms include the SPA algorithm and the CARS algorithm.

6. The hyperspectral imaging method based on machine learning according to claim 1, characterized in that, R 2 RMSE and RMSE are calculated using formulas (5) and (6) respectively: （5）， （6）， in, and Let represent the predicted value and the measured value of the i-th sample in the sample set, respectively. denoted as the average of the measured values ​​in the sample set, and n represents the number of samples in the sample set.

7. The hyperspectral imaging method based on machine learning according to claim 6, characterized in that, In S3, the filtering criteria are: Test the R-squared of the regression prediction model 2 The larger; or R 2 The larger the value, the smaller the RMSE.

8. The hyperspectral imaging method based on machine learning according to claim 1, characterized in that, The machine learning algorithm is a PLSR, SVR, or RF neural network.

9. A system for predicting the cold hardiness of Napier grass using the machine learning-based hyperspectral imaging method as described in any one of claims 1-8, characterized in that, The system includes: The data acquisition module is used to acquire hyperspectral image data of Napier grass leaves; The processing module is communicatively connected to the data acquisition module and is used to obtain the current cold resistance assessment result of Napier grass based on the hyperspectral image data.

10. Application of the optimal prediction model constructed based on the machine learning-based hyperspectral imaging method as described in any one of claims 1-8 in the evaluation of the cold resistance of Napier grass.

Images