A method and system for grape landmark recognition based on hyperspectral ensemble learning

Abstract

This invention discloses a hyperspectral ensemble learning-based method and system for grape landmark identification, relating to the field of grape landmark identification and monitoring technology. The hyperspectral ensemble learning-based grape landmark identification method includes the following steps: spectral information extraction and monitoring; dataset construction and splitting; feature set construction; and prediction accuracy evaluation. This invention extracts and monitors spectral information from grape landmark samples, then constructs and splits the dataset, followed by feature recognition and feature set construction to obtain a feature set. Finally, based on the feature training set, it constructs an SSC prediction model and performs Bayesian global hyperparameter tuning, achieving improved accuracy in SSC prediction and landmark identification for Korque Terrek grapes. This solves the problem in existing technologies where differences in planting adjustments lead to high deviations in SSC prediction and landmark identification results for Korque Terrek grapes.

Description

Technical Field

[0001] This invention relates to the field of grape landmark identification and monitoring technology, and in particular to a method and system for grape landmark identification using hyperspectral integrated learning. Background Technology

[0002] Kokterek grapes are a geographical indication agricultural product of Xinjiang. The most important indicator of the quality and flavor of Kokterek grapes is SSC (Soluble Solids Content). The SSC of Kokterek grapes is usually higher than 18% (Brix degree), far exceeding that of ordinary grape varieties (12%-15%). However, due to the significant difference in SSC between geographical indication and non-geographical indication products of Kokterek grapes, the nutritional quality of Kokterek grapes is significantly affected. Therefore, it is necessary to use detection technology to predict the SSC of Kokterek grapes and accurately identify the geographical indication.

[0003] Existing methods mainly use saccharimeters for measurement, and local identification usually uses isotope and elemental analysis methods. Hyperspectral imaging technology is used for fruit quality detection. However, existing methods mainly focus on single-task analysis of hyperspectral imaging technology, i.e., single regression or classification tasks, and rarely perform multi-task analysis.

[0004] For example, the invention patent application CN118111955A discloses a method and system for detecting soluble solids content in grapes based on hyperspectral imaging. This includes: acquiring hyperspectral images of grape bunches using a hyperspectral imager and scanning a standard white board for black-and-white correction to obtain the grape hyperspectral image; constructing a grape instance segmentation model and performing instance segmentation processing on the grape hyperspectral image; constructing a soluble solids content prediction model for different grape varieties based on the grape hyperspectral image; predicting the soluble solids content of the segmented grape hyperspectral image using the soluble solids content prediction model, obtaining the soluble solids content cutting force prediction of the grape, comparing it with experimental data, and analyzing the causes of cutting force errors in conjunction with the model establishment process.

[0005] For example, the invention patent announcement CN106872396B discloses a method for converting glucose concentration measurement models using different near-infrared instruments, which includes: 1) acquiring spectral data from two near-infrared instruments; 2) mathematical conversion between spectral data; 3) screening for common wavelengths in the data from the two instruments; 4) calculating the converted spectral set; and 5) constructing the converted model.

[0006] However, in the process of implementing the inventive technical solution in the embodiments of this application, it was found that the above-mentioned technology has at least the following technical problems:

[0007] In the existing technology, due to the significant differences in SSC between geographical indication and non-geographical indication products of Korque Terrek grapes, and the large fluctuations in SSC of non-geographical indication products caused by planting conditions, the prediction of SSC of Korque Terrek grapes and the accurate identification of geographical indication are affected by the differences. The differences in planting adjustment also lead to a high degree of deviation in the results of prediction of SSC of Korque Terrek grapes and accurate identification of geographical indication. Summary of the Invention

[0008] This application provides a hyperspectral integrated learning method and system for grape landmark identification, which solves the problem of high deviation in the results of SSC prediction and landmark accuracy identification of Korque Terrek grapes caused by differences in planting regulation in the prior art, and improves the accuracy of SSC prediction and landmark accuracy identification of Korque Terrek grapes.

[0009] This application provides a hyperspectral integrated learning method for grape landmark identification, comprising the following steps: S1, extracting and monitoring spectral information from acquired grape landmark samples to obtain hyperspectral images and measuring SSCs, the grape landmark samples including Xinjiang Kok Terek landmark and non-landmark grape samples; S2, constructing a spectral dataset based on the acquired spectral dataset, splitting the spectral dataset to obtain a spectral set, the spectral set including a spectral training set and a spectral test set, the spectral dataset construction data including hyperspectral data, SSCs and landmark categories; S3, performing feature recognition to obtain shared features, constructing a feature set based on the shared features, the feature set including... The process includes a feature training set and a feature test set. Feature recognition represents the identification of features corresponding to SSC and landmark categories through BA. The feature training set is used to improve the accuracy of SSC prediction model construction, and the feature test set is used to predict and evaluate the XGBoost ensemble learning model. The sample size of the feature test set is not less than the preset total number of samples. S4 involves building an SSC prediction model based on the feature training set and performing Bayesian global tuning of hyperparameters to improve the prediction accuracy of SSC and landmark categories. The SSC prediction model construction is used to obtain the XGBoost ensemble learning model, coefficient of determination, and root mean square error. The Bayesian global tuning of hyperparameters is used to optimize the XGBoost ensemble learning model to improve model performance.

[0010] This application provides a hyperspectral integrated learning-based grape landmark recognition system, including a spectral information extraction and monitoring module, a dataset construction and splitting module, a feature set construction module, and a prediction accuracy evaluation module. The spectral information extraction and monitoring module is used to extract and monitor spectral information from acquired grape landmark samples to obtain hyperspectral images and determine the SSC (Spectral Score). Grape landmark samples include Xinjiang Kok Terek landmark and non-landmark grape samples. The dataset construction and splitting module is used to construct a spectral dataset based on the acquired spectral dataset, and then split the dataset to obtain a spectral set, which includes a spectral training set and a spectral test set. The spectral dataset construction data includes hyperspectral data, SSC, and landmark categories. The feature set construction module is used to perform feature recognition to obtain shared features. The feature set is constructed based on shared features, including a feature training set and a feature test set. Feature recognition means that the features corresponding to SSC and landmark categories are identified through BA. The feature training set is used to improve the accuracy of the SSC prediction model construction, and the feature test set is used to predict and evaluate the XGBoost ensemble learning model. The sample size of the feature test set is not less than the preset total number of samples. The prediction accuracy evaluation module is used to construct the SSC prediction model based on the feature training set and to globally tune the hyperparameters using Bayesian to improve the prediction accuracy of SSC and landmark categories. The SSC prediction model construction is used to obtain the XGBoost ensemble learning model, the coefficient of determination, and the root mean square error. The globally tuned hyperparameters using Bayesian are used to optimize the XGBoost ensemble learning model to improve model performance.

[0011] One or more technical solutions provided in the embodiments of this application have at least the following technical effects or advantages:

[0012] 1. Spectral information is extracted and monitored using grape landmark samples. Then, a spectral dataset is constructed based on the spectral dataset, and the dataset is split to obtain a spectral set. Next, feature recognition is performed to obtain shared features, and a feature set is constructed. Finally, an SSC prediction model is built based on the feature training set, and the hyperparameters are globally tuned using Bayesian. This improves the effectiveness of SSC and landmark category prediction results, thereby improving the accuracy of SSC prediction and landmark identification for Korque Terrek grapes. This effectively solves the problem of high deviation in SSC prediction and landmark identification results for Korque Terrek grapes caused by differences in planting regulation in existing technologies.

[0013] 2. By combining hyperspectral extraction data with coupling processing, an extraction accuracy quantization value is obtained. Then, it is determined whether the extraction accuracy quantization value meets the extraction accuracy conditions. When the extraction accuracy quantization value meets the extraction accuracy conditions, the corresponding hyperspectral data is acquired. When the extraction accuracy quantization value does not meet the extraction accuracy conditions, the extraction accuracy is optimized. This achieves accurate evaluation of the extraction accuracy of hyperspectral data from the equatorial center image of the fruit, thereby improving the accuracy of acquiring hyperspectral data.

[0014] 3. By processing the difference between the actual and predicted SSC values ​​of grapes, the grape SSC difference value is obtained. Then, the difference between the actual and mean actual SSC values ​​is processed to obtain the average grape SSC difference value. Finally, after performing a convergence analysis on the grape SSC difference value and the average grape SSC difference value, the deviation of the fit is processed to obtain the coefficient of determination. This achieves a precise assessment of the fitting accuracy of the regression model, thereby improving the fitting accuracy of the regression model. Attached Figure Description

[0015] Figure 1 A flowchart illustrating a hyperspectral ensemble learning-based grape landmark recognition method provided in this application embodiment;

[0016] Figure 2 Hyperspectral image of Kökterjek grape provided in an embodiment of this application;

[0017] Figure 3 The average spectrum of Cockterjek grapes, both local and non-local, provided for embodiments of this application;

[0018] Figure 4 SSC and landmark sharing features provided in the embodiments of this application;

[0019] Figure 5 This is a scatter plot comparing SSC content provided in an embodiment of this application.

[0020] Figure 6 The Kokterjek grape landmark identification confusion matrix provided in the embodiments of this application. Detailed Implementation

[0021] This application provides a hyperspectral integrated learning method and system for grape landmark identification, which solves the problem of high deviation in SSC prediction and landmark accuracy identification of Korque Terrek grapes due to differences in planting regulation in the prior art. The method involves extracting and monitoring spectral information from grape landmark samples to obtain hyperspectral images and measuring SSCs. Then, a spectral dataset is constructed based on the spectral dataset. The dataset is then split to obtain a spectral set. Next, shared features are identified, and a feature set is constructed based on these shared features. Finally, an SSC prediction model is built based on the feature training set, and Bayesian global hyperparameter tuning is performed to improve the accuracy of SSC and landmark category predictions. This improves the accuracy of SSC prediction and landmark accuracy identification for Korque Terrek grapes.

[0022] The technical solution in this application aims to address the problem of high deviation in SSC prediction and accurate identification of geographical features for Korque Terrek grapes caused by differences in planting adjustments. The overall approach is as follows:

[0023] By extracting and monitoring spectral information from grape landmark samples, and then constructing a spectral dataset based on the spectral dataset, and splitting the dataset to obtain a spectral set, then performing feature recognition to obtain shared features and constructing a feature set, and finally constructing an SSC prediction model based on the feature training set and globally tuning hyperparameters using Bayesian, the accuracy of SSC prediction and landmark identification results for Kok Terek grapes was improved.

[0024] To better understand the above technical solutions, the following will provide a detailed explanation of the technical solutions in conjunction with the accompanying drawings and specific implementation methods.

[0025] like Figure 1The flowchart shown is a hyperspectral integrated learning method for grape landmark recognition provided in this application embodiment. The method includes the following steps: S1, spectral information extraction and monitoring: spectral information extraction and monitoring are performed based on the acquired grape landmark samples to obtain hyperspectral images and determine the SSC. The grape landmark samples include Xinjiang Kok Terek landmark and non-landmark grape samples; S2, dataset construction and splitting: a spectral dataset is constructed based on the acquired spectral dataset. Simultaneously, the extraction accuracy quantification value is used to quantify the hyperspectral data extraction accuracy of the fruit equatorial center image and determine whether to optimize the extraction accuracy. The dataset is split based on the spectral dataset to obtain a spectral set, which includes a spectral training set and a spectral test set (e.g., a spectral training set of 290 samples and a spectral test set of 72 samples). The spectral dataset construction data includes hyperspectral data, SSC, and landmark categories; S3, feature set construction: shared features are obtained through feature recognition. A feature set is constructed based on the shared features. The process includes a feature training set and a feature test set (the sample size of the feature test set is no less than 20% of the total sample size). Feature recognition means identifying the features corresponding to SSC and landmark categories through BA (Base Accuracy). The feature training set is used to improve the accuracy of the SSC prediction model construction, and the feature test set is used to predict and evaluate the XGBoost ensemble learning model. The sample size of the feature test set is no less than the preset total sample size. S4, Prediction Accuracy Evaluation: Based on the feature training set, the SSC prediction model is constructed and Bayesian global tuning of hyperparameters is performed to improve the prediction accuracy of SSC and landmark categories. The SSC prediction model construction is used to obtain the XGBoost ensemble learning model, the coefficient of determination, and the root mean square error. Bayesian global tuning of hyperparameters is used to optimize the XGBoost ensemble learning model to improve model performance. Specifically, the XGBoost ensemble learning model is constructed using the feature training set. In the Bayesian algorithm, the Gaussian process is used as a surrogate model, the expected improvement is the objective function, the maximum number of iterations is 50, and the hyperparameters are globally optimized. The input is SSC and landmark categories, and the output is the model's prediction of SSC and landmark categories.

[0026] Among them, such as Figure 2 The image shown is a hyperspectral image of a Korque Terrek grape provided in an embodiment of this application; as shown... Figure 3 The image shows the average spectrum of Coq Terrek grapes, both local and non-local, provided in an embodiment of this application; as shown... Figure 4 As shown, this application provides SSC and landmark sharing features in its embodiments; Figure 5 As shown, this is a scatter plot comparing SSC content provided in the embodiments of this application. Figure 2-5 It can be seen that the local and non-local spectra of Cocterjek grapes differ.

[0027] The prediction accuracy was R² = 0.943, RMSE = 0.302, and the hyperparameters were mtry = 3, min_n = 6, tree_depth = 7, learn_rate = 0.00191, loss_reduction = 0.001, and sample_prop = 0.967.

[0028] like Figure 6 The image shows the confusion matrix for Kokterek grape landmark recognition provided in this embodiment of the application, where the recall rate is 92.4% and the optimal hyperparameters are mtry = 3; min_n = 12; tree_depth = 2; learn_rate = 0.0107; loss_reduction = 0.171; sample_prop = 0.922.

[0029] This application provides a hyperspectral integrated learning-based grape landmark recognition system, comprising a spectral information extraction and monitoring module, a dataset construction and splitting module, a feature set construction module, and a prediction accuracy evaluation module. The spectral information extraction and monitoring module is used to extract and monitor spectral information from acquired grape landmark samples to obtain hyperspectral images and determine the SSC (Spectral Characteristic Score). Grape landmark samples include Xinjiang Kok Terek landmark and non-landmark grape samples. The dataset construction and splitting module is used to construct a spectral dataset based on the acquired spectral dataset, and to split the dataset to obtain a spectral set, which includes a spectral training set and a spectral test set. The spectral dataset construction data includes hyperspectral data, SSC, and landmark categories. The feature set construction module is used to perform feature recognition to obtain shared features. The feature set is constructed based on shared features, including a feature training set and a feature test set. Feature recognition means that the features corresponding to SSC and landmark categories are identified through BA. The feature training set is used to improve the accuracy of the SSC prediction model construction, and the feature test set is used to predict and evaluate the XGBoost ensemble learning model. The sample size of the feature test set is not less than the preset total number of samples. The prediction accuracy evaluation module is used to construct the SSC prediction model based on the feature training set and to globally tune the hyperparameters using Bayesian to improve the prediction accuracy of SSC and landmark categories. The SSC prediction model construction is used to obtain the XGBoost ensemble learning model, the coefficient of determination, and the root mean square error. The globally tuned hyperparameters using Bayesian are used to optimize the XGBoost ensemble learning model to improve model performance.

[0030] In this embodiment, hyperspectral images are acquired and SSCs are measured through spectral information extraction and monitoring; hyperspectral data meeting the extraction accuracy conditions are obtained through dataset construction and splitting; simultaneously, the acquired spectral dataset is used to construct a spectral dataset and a spectral set; a feature set is obtained through feature set construction to predict and evaluate the XGBoost ensemble learning model; the XGBoost ensemble learning model, coefficient of determination, and root mean square error are obtained through prediction accuracy evaluation, and hyperparameters are globally tuned using Bayesian to improve the prediction accuracy of SSCs and landmark categories; the synergistic effect of spectral information extraction and monitoring, dataset construction and splitting, feature set construction, and prediction accuracy evaluation helps to provide more accurate hyperspectral data, thereby improving the accuracy of SSC prediction and landmark identification results for the Koc Terrek grape.

[0031] Furthermore, the hyperspectral image is acquired by setting the spectral image of the marked grape landmark sample information using a preset hyperspectral imaging device; the marked grape landmark sample information represents grape landmark sample information of no less than a preset sample quantity (generally 100) obtained from the database; the spectral image settings include band range settings and resolution settings; the band range setting indicates that the band corresponding to the grape landmark sample information set by the preset personnel is within a preset band range (generally the visible to near-infrared band of 400 to 2500 nm); the resolution setting indicates that the spectral resolution corresponding to the grape landmark sample information set by the preset personnel is within a preset resolution range (generally 2-5 nm).

[0032] It should be added that the spectral dataset is constructed based on the acquired spectral dataset. The specific process is as follows: The equatorial center image of the fruit is acquired using a preset hyperspectral imaging device. This involves whiteboard calibration using the preset hyperspectral imaging device and a whiteboard, turning off the halogen lamp, and using a blackboard to correct for ambient light sources to eliminate errors during imaging scans of each sample. Hyperspectral data is obtained by extracting data from the fruit equatorial center image. This extraction is performed using ENVI software from a preset image region at the fruit equatorial center (the number of preset image regions is greater than or equal to 3). The SSC (Spectral Signature Sample) of the labeled grape is measured using a saccharimeter (the measurement results are independently repeated at least twice). Landmark categories are then determined based on the SSC. The spectral dataset is constructed by combining the spectral dataset construction data into a single set (e.g., integrating spectral data, SSC, landmark categories, and acquisition time into a unified storage format). The acquired spectral dataset is then smoothed using Savitzky-Golay smoothing to eliminate noise in the spectral dataset.

[0033] In this embodiment, by setting the spectral image, it is helpful to avoid feature loss due to fixed parameters (such as failure to capture key absorption peaks) and improve the correlation between spectral data and chemical composition. Spectral data is extracted from a preset region of the fruit equatorial center image using ENVI software to ensure synergistic optimization of spatial resolution and spectral resolution. The Savitzky-Golay smoothing algorithm is used to eliminate random noise in the spectral data (such as instrument noise and ambient light interference), which helps to preserve key spectral features and improve the image quality of the fruit equatorial center image. This, in turn, improves the accuracy of SSC prediction and landmark identification results for Kock Terjek grapes.

[0034] Furthermore, based on the acquired spectral dataset, a dataset is constructed to obtain a spectral dataset, which also includes quantifying the extraction accuracy. The specific process for quantifying the extraction accuracy is as follows: Hyperspectral extraction accuracy parameters are obtained during data extraction from the fruit's equatorial center image; based on the acquired hyperspectral extraction accuracy parameters and preset hyperspectral extraction accuracy parameters, the extraction accuracy of the hyperspectral data from the fruit's equatorial center image is quantified to obtain a quantized extraction accuracy value; the specific process for obtaining the quantized extraction accuracy value is as follows:

[0035] AA1, after analyzing the proportion of preset spectral resolution deviation and spectral resolution deviation, a weighted operation is performed using the resolution deviation-extraction accuracy factor to obtain the resolution deviation-extraction accuracy value. This value reflects the effect of spectral resolution deviation on the accuracy of hyperspectral data extraction from the fruit equatorial center image. Specifically, the expression for the resolution deviation-extraction accuracy value is as follows: F = 1, 2, ..., J, where F represents the number of the fruit equatorial center image, J represents the total number of fruit equatorial center images, GGP1(F) represents the resolution deviation of the F-th fruit equatorial center image minus the extracted accurate value, ΔGPP(F) represents the average spectral resolution deviation of the F-th fruit equatorial center image, ΔGPP(0) represents the preset average spectral resolution deviation, and M... 1 The resolution deviation is represented by the extraction accuracy factor, which is the absolute value of the difference between the resolution at a preset wavelength of the fruit equatorial center image and the preset spectral resolution (pre-set by preset personnel) monitored by preset hyperspectral imaging equipment (such as spectrometers, atomic emission spectrometers, high-resolution cameras, etc.), and the average value of these differences is taken as the average spectral resolution deviation.

[0036] AA2, after analyzing the proportion of the average spectral signal signal-to-noise ratio (SNR) and the preset average spectral signal SNR, a weighted operation based on the SNR-extraction accuracy factor is performed to obtain the SNR-extraction accuracy value. This value reflects the effect of the average spectral signal SNR on the accuracy of hyperspectral data extraction from the fruit equatorial center image. Specifically, the expression for the SNR-extraction accuracy value is as follows: GGP2(F) represents the signal-to-noise ratio (SNR) of the Fth fruit equatorial center image - the accurate extraction value, ΔGPXZ(F) represents the average spectral signal SNR of the Fth fruit equatorial center image, ΔGPXZ(0) represents the preset average spectral signal SNR, M 2 The signal-to-noise ratio (SNR) is defined as the ratio of the power to the noise power of the preset spectral signal in the fruit's equatorial center image, which is monitored using a preset hyperspectral imaging device and a power meter. The average value of these ratios is used as the average spectral signal SNR.

[0037] AA3, after analyzing the proportion of the preset average wavelength shift and average wavelength shift, combines the wavelength shift-extraction accuracy factor for weighted calculation to obtain the wavelength shift-extraction accuracy value. This value reflects the effect of the average wavelength shift on the accuracy of hyperspectral data extraction from the fruit equatorial center image. Specifically, the expression for the wavelength shift-extraction accuracy value is as follows: GGP3(F) represents the wavelength offset of the equatorial center image of the Fth fruit - the accurate extraction value, ΔBC(F) represents the average wavelength offset of the equatorial center image of the Fth fruit, ΔBC(0) represents the preset average wavelength offset, M 3 The wavelength offset is extracted by using a preset hyperspectral imaging device to monitor the absolute value of the difference between the preset wavelength and the preset wavelength (pre-set by preset personnel) of the fruit equatorial center image, and the average value of these values ​​is used as the average wavelength offset.

[0038] AA4, combined with hyperspectral extraction data, is coupled to obtain the extraction accuracy quantification value. The extraction accuracy quantification value is used to reflect the combined effect of the hyperspectral extraction accuracy parameter and the preset hyperspectral extraction accuracy parameter on the extraction accuracy of hyperspectral data of the fruit equatorial center image. The hyperspectral extraction data includes resolution deviation-extraction accuracy value, signal-noise ratio-extraction accuracy value and wavelength shift-extraction accuracy value, and all of the hyperspectral extraction data are greater than 0.

[0039] The accuracy quantification value was obtained through the following method:

[0040] GGP(F)=GGP1(F)+GGP2(F)+GGP3(F);

[0041] In the formula, GGP(F) represents the quantification value of the extraction accuracy of the Fth fruit equatorial center image.

[0042] In this application, a database storing various preset data is established before designing the grape landmark recognition method based on hyperspectral integrated learning. This database includes, but is not limited to, preset average spectral resolution deviation, preset average spectral signal-to-noise ratio, and preset average wavelength shift, with each value directly set by a technician. The accurate hyperspectral extraction parameters include average spectral resolution deviation, average spectral signal-to-noise ratio, and average wavelength shift. The preset accurate hyperspectral extraction parameters are represented by the average value of the accurate hyperspectral extraction parameters over a historical time period. The units for average spectral resolution deviation and preset average spectral resolution deviation are nanometers. The units for average spectral signal-to-noise ratio and preset average spectral signal-to-noise ratio are decibels. The units for average wavelength shift and preset average wavelength shift are nanometers.

[0043] In this embodiment, the extraction accuracy quantification value is further obtained by analyzing the hyperspectral extraction data. A larger resolution deviation-extraction accuracy value means a stronger effect of spectral resolution deviation on the extraction accuracy of hyperspectral data from the fruit's equatorial center image, resulting in a larger extraction accuracy quantification value. A larger signal-to-noise ratio-extraction accuracy value means a stronger effect of the average spectral signal-to-noise ratio on the extraction accuracy of hyperspectral data from the fruit's equatorial center image, resulting in a larger extraction accuracy quantification value. A larger wavelength shift-extraction accuracy value means a stronger effect of the average wavelength shift on the extraction accuracy of hyperspectral data from the fruit's equatorial center image, resulting in a larger extraction accuracy quantification value. In summary, in this embodiment, the hyperspectral extraction data and the extraction accuracy quantification value are positively correlated.

[0044] In this embodiment, the monitored hyperspectral extraction parameters are not isolated but interconnected, requiring correlation analysis to describe their combined effects. An increase in average spectral resolution deviation indicates greater inhomogeneity in the wavelength intervals of the spectral data, potentially leading to distorted spectral signals and a lower average signal-to-noise ratio (SNR). Average spectral resolution deviation can also affect the accuracy of wavelength measurements; a larger deviation means greater error in wavelength measurement by the spectrometer, potentially resulting in inaccurate wavelength measurements and a larger average wavelength shift. Conversely, a lower average SNR indicates more noise in the spectral signal, which may interfere with accurate wavelength measurement and further increase the average wavelength shift. By analyzing the combined effects of these parameters, a precise assessment of the accuracy of hyperspectral data extraction from the equatorial center image of the fruit was achieved, thereby improving the accuracy of SSC prediction and landmark identification for Korque Terrek grapes.

[0045] Furthermore, the specific acquisition process of hyperspectral data is as follows: It is determined whether the extraction accuracy quantization value meets the extraction accuracy condition; when the extraction accuracy quantization value meets the extraction accuracy condition, an extraction qualification prompt is sent, and the corresponding hyperspectral data is acquired; when the extraction accuracy quantization value does not meet the extraction accuracy condition, an extraction failure prompt is sent, and extraction accuracy optimization is performed; extraction accuracy optimization is used to improve the extraction accuracy of hyperspectral data from the fruit equatorial center image; the extraction accuracy condition indicates that the extraction accuracy quantization value is not lower than the preset extraction accuracy quantization value obtained from the database.

[0046] The specific steps involved in optimizing extraction accuracy are as follows:

[0047] BB1, during the data extraction process of the fruit equatorial center image, optimized and qualified parameters were obtained; after analyzing the proportion of the light source intensity homogenization ratio and the preset light source intensity homogenization ratio, a weighted calculation was performed using the first light source homogenization factor to obtain the first light source homogenization value, which reflects the effect of the light source intensity homogenization ratio on the light source homogenization during the hyperspectral data extraction process of the fruit equatorial center image. Specifically, the expression for the first light source homogenization value is as follows: F = 1, 2, ..., J, where F represents the number of the fruit equatorial center image, J represents the total number of fruit equatorial center images, JYX1(F) represents the first light source uniformity value of the F-th fruit equatorial center image, JYB(F) represents the light source intensity uniformity ratio of the F-th fruit equatorial center image, JYB(0) represents the preset light source intensity uniformity ratio, E 1 The first light source uniformity factor is defined as the ratio of the maximum and minimum light source intensity of the preset illumination area of ​​the fruit equatorial center image monitored by a preset hyperspectral imaging device and a light intensity sensor, and the average value of the ratio is used as the light source intensity uniformity ratio.

[0048] BB2, after analyzing the proportion of the average light source intensity and the preset average light source intensity, combines the second light source uniformity factor for weighted calculation to obtain the second light source uniformity value. This value reflects the effect of the average light source intensity on the uniformity of the light source during the extraction of hyperspectral data from the fruit equatorial center image. Specifically, the expression for the second light source uniformity value is as follows: JYX2(F) represents the uniform value of the second light source in the equatorial center image of the Fth fruit, GQP(F) represents the average light source intensity in the equatorial center image of the Fth fruit, GQP(0) represents the preset average light source intensity, E 2 The second light source uniformity factor is represented by the light source intensity of the preset illumination area points of the fruit equatorial center image monitored by a preset hyperspectral imaging device and a light intensity sensor, and the average value of these points is taken as the average light source intensity.

[0049] BB3, after analyzing the proportion of the preset average image reflectance and average image reflectance, combines the third light source uniformity factor for weighted calculation to obtain the third light source uniformity value. This value reflects the effect of average image reflectance on the light source uniformity during the extraction of hyperspectral data from the fruit equatorial center image. Specifically, the expression for the third light source uniformity value is as follows: JYX3(F) represents the uniform value of the third light source in the equatorial center image of the Fth fruit, FS(F) represents the average image reflectance of the equatorial center image of the Fth fruit, FS(0) represents the preset average image reflectance, E 3 The uniformity factor of the third light source is represented by the average reflectance of the pixels corresponding to the preset illumination area points of the fruit equatorial center image, which is monitored by a preset hyperspectral imaging device.

[0050] BB4, after coupling processing with the extraction accuracy optimization data, yields the extraction accuracy optimization value. The extraction accuracy optimization value reflects the combined effect of the optimized qualified parameters and the preset optimized qualified parameters on the uniformity of the light source during the extraction of hyperspectral data of the fruit equatorial center image. The extraction accuracy optimization data includes the uniform values ​​of the first light source, the second light source, and the third light source, and all of the extraction accuracy optimization data are greater than 0.

[0051] The accuracy optimization value was obtained through the following method:

[0052] JYX(F)=JYX1(F)+JYX2(F)+JYX3(F);

[0053] In the formula, JYX(F) represents the optimized value for the extraction accuracy of the Fth fruit equatorial center image.

[0054] In summary, the optimized qualified parameters include the light source intensity homogenization ratio, average light source intensity, and average image reflectivity. The preset optimized qualified parameters include the preset light source intensity homogenization ratio, preset average light source intensity, and preset average image reflectivity. Among them, the preset optimized qualified parameters are represented by the average value of the optimized qualified parameters over a historical time period. The light source intensity homogenization ratio and the preset light source intensity homogenization ratio are both unitless. The units of the preset average light source intensity and the average light source intensity are both watts per square meter. The preset average image reflectivity and the average image reflectivity are both unitless.

[0055] An extraction adjustment factor is obtained by performing a ratio analysis between the optimized extraction accuracy value and the preset optimized extraction accuracy value obtained from the database. The preset optimized extraction accuracy value is represented by the average of the optimized extraction accuracy values ​​over a historical time period. The extraction adjustment factor is represented by the ratio of the optimized extraction accuracy value to the preset optimized extraction accuracy value and is used to adjust the extraction accuracy of hyperspectral data from the fruit's equatorial center image. A light source intensity setting is implemented, which sends a prompt to preset personnel to gradually increase the light source intensity according to the magnitude corresponding to the extraction adjustment factor. A working distance setting is also implemented, which sends a prompt to preset personnel to gradually decrease the working distance of the preset hyperspectral imaging device according to the magnitude corresponding to the extraction adjustment factor. When the working distance of the preset hyperspectral imaging device decreases to the preset minimum working distance or the light source intensity increases to the preset maximum light source intensity, if the extraction accuracy quantification value does not meet the extraction accuracy conditions, an alarm is sent. The working distance of the preset hyperspectral imaging device and the preset maximum light source intensity are preset by preset personnel.

[0056] In this embodiment, when a light source intensity setting prompt is detected, the light source intensity is gradually increased by extracting the amplitude corresponding to the adjustment factor to compensate for the light attenuation in the equatorial region. When a working distance setting prompt is detected, the working distance of the preset hyperspectral imaging device is gradually decreased by extracting the amplitude corresponding to the adjustment factor to enhance the ability to capture spectral details. By dynamically adjusting the light source intensity and working distance, the performance and reliability of the hyperspectral imaging device in grape SSC detection are improved.

[0057] This embodiment further obtains the optimized extraction accuracy value by analyzing the data. A larger first light source uniformity value means a stronger effect of the light source intensity homogenization ratio on the uniformity of the light source during the extraction of hyperspectral data from the fruit's equatorial center image, resulting in a larger optimized extraction accuracy value. A larger second light source uniformity value means a stronger effect of the average light source intensity on the uniformity of the light source during the extraction of hyperspectral data from the fruit's equatorial center image, resulting in a larger optimized extraction accuracy value. A larger third light source uniformity value means a stronger effect of the average image reflectance on the uniformity of the light source during the extraction of hyperspectral data from the fruit's equatorial center image, resulting in a larger optimized extraction accuracy value. In summary, in this embodiment, the optimized extraction accuracy data and the optimized extraction accuracy value are positively correlated.

[0058] In this embodiment, the monitored optimized parameters are not isolated but interconnected, requiring correlation analysis to describe their combined effects. A higher light source intensity homogenization ratio indicates a more uniform intensity distribution of the light source within the irradiated area, which helps maintain the stability and effectiveness of the average light source intensity, thus leading to a higher average light source intensity. Conversely, a lower light source intensity homogenization ratio indicates an uneven intensity distribution of the light source within the preset irradiated area, resulting in relatively uneven reflected light distribution and a lower average image reflectance. A higher average image reflectance may lead to overexposure of the fruit's equatorial center image, causing loss of detail in the preset area and reducing the light source homogenization in the preset area, thus resulting in a lower light source intensity homogenization ratio. By analyzing the comprehensive influence of these parameters, accurate assessment of light source homogenization during the hyperspectral data extraction process of the fruit's equatorial center image is achieved, thereby improving the accuracy of SSC prediction and landmark identification results for Kock Terrek grapes.

[0059] Furthermore, the specific process of splitting the spectral dataset to obtain the spectral set based on the spectral dataset is as follows: Step 1, set the split ratio, which means sending a prompt to the preset personnel to set the split ratio of the spectral dataset; Step 2, stratified random sampling, which means splitting the spectral dataset using a stratified random sampling method; stratified random sampling is used to ensure that the landmark categories and SSCs in the spectral set are normally distributed.

[0060] The specific process for obtaining shared features through feature recognition is as follows: SS1, Generate BA, which is obtained through random forest ensemble; SS2, Reproduce the code, which means reproducing the feature using R code; SS3, Perform feature-level fusion, which means identifying the features corresponding to SSC and landmark categories through BA and combining them to obtain fused features; SS4, Perform average importance assessment, which means calculating the average importance score based on the impurity of random forest nodes; SS5, Set the importance threshold, which means sending prompts to preset personnel to set an importance threshold (generally p = 0.6); SS6, Perform feature filtering, which means filtering fused features based on the importance threshold to obtain shared features.

[0061] In this embodiment, a stratified random sampling method is used to split the spectral dataset. For example, by pre-setting the split ratio of the spectral training set and the spectral test set to be greater than 4:1, it is ensured that the landmark categories and SSCs corresponding to the spectral sets are normally distributed. The BA of the spectral dataset is obtained through the random forest ensemble algorithm, which helps to evaluate the effect of feature selection on the prediction performance. The random forest algorithm is reproduced through R language, which helps to ensure the repeatability of the results and optimize the feature recognition effect. Based on the features identified by the random forest, the features corresponding to the SSCs and landmark categories are identified through BA, which helps to remove redundant features and enhances the prediction ability of the Kok Terrek grape SSC and landmarks. Based on the reduction of impurity of random forest nodes, the average importance score of each fused feature is calculated, which provides a quantitative basis for feature selection. Thus, the accuracy of the Kok Terrek grape SSC prediction and landmark accurate recognition results are improved.

[0062] Furthermore, the specific process for obtaining the coefficient of determination is as follows: CC1, the grape SSC difference value is obtained by performing difference processing on the actual and predicted grape SSC values; the grape SSC difference value reflects the relative deviation between the actual and predicted grape SSC values, and their combined effect on the fitting accuracy of the regression model. The grape SSC difference value is represented by the sum of squared errors between the actual and predicted grape SSC values. Specifically, the expression for the grape SSC difference value is: y obs This represents the actual value of grape SSC, y pre denoted as SSC predicted value for grapes, i represents the sample number, and n represents the sample size.

[0063] CC2 calculates the average grape SSC difference value by processing the difference between the actual grape SSC values ​​and the mean of the actual grape SSC values. The average grape SSC difference value reflects the relative deviation between the actual grape SSC values ​​and the mean of the actual grape SSC values, and their combined impact on the accuracy of the regression model's fit. The average grape SSC difference value is represented by the sum of squared errors between the actual grape SSC values ​​and the mean of the actual grape SSC values. Specifically, the expression for the average grape SSC difference value is as follows: y ave This represents the average actual value of the grape SSC.

[0064] CC3 is a coefficient of determination obtained by performing a convergence analysis on the grape SSC difference value and the average grape SSC difference value, followed by a deviation adjustment for the goodness of fit. The coefficient of determination is used to reflect the combined effect of grape evaluation parameters and grape SSC predicted values ​​on the accuracy of the regression model. Grape evaluation parameters include the actual grape SSC value and the mean of the actual grape SSC value.

[0065] The coefficient of determination is obtained using the following method:

[0066]

[0067] Where R 2 S represents the coefficient of determination, S1 represents the grape SSC difference value, and S2 represents the average grape SSC difference value.

[0068] The specific process for obtaining the root mean square error is as follows:

[0069] DD1, obtained by analyzing the proportion of variance in grape SSC values ​​and sample size, yields the root mean square error (RMSE). The RMSE is used to assess the impact of the actual grape SSC values ​​on the deviation from the accuracy of the grape SSC. Specifically, the expression for the RMSE is: y obs This represents the actual value of grape SSC, y pre This represents the predicted SSC value for grapes.

[0070] DD2 is obtained by taking the square root of the root mean square error. The root mean square error is used to reflect the combined effect of the actual and predicted values ​​of grape SSC on the deviation of the accuracy of grape SSC.

[0071] The root mean square error is obtained through the following method:

[0072]

[0073] In the formula, RMSE represents the root mean square error, and H represents the root mean square error reflection value.

[0074] It should be added that the construction of the SSC prediction model based on the feature training set and the global tuning of hyperparameters by Bayesian to improve the accuracy of SSC and landmark category prediction also include the evaluation of classification model accuracy; the evaluation of classification model accuracy means obtaining accuracy and recall based on the confusion matrix, which is used to evaluate the classification accuracy of the XGBoost ensemble learning model.

[0075] Specifically, accuracy = (TP+TN) / (TP+TN+FP+FN); recall = TP / (TP+FN); where TP represents positive correct samples; TN represents negative correct samples; FP represents positive false samples; and FN represents negative false samples.

[0076] In this embodiment, a larger grape SSC difference value means that the relative deviation between the actual grape SSC value and the predicted grape SSC value has a stronger effect on the fitting accuracy of the regression model, resulting in a smaller coefficient of determination. A larger average grape SSC difference value means that the relative deviation between the actual grape SSC value and the mean of the actual grape SSC value has a stronger effect on the fitting accuracy of the regression model (the regression model corresponding to the XGBoost ensemble learning model), resulting in a larger coefficient of determination. A larger root mean square error value means that the actual grape SSC value has a stronger effect on the deviation of the grape SSC accuracy, resulting in a larger root mean square error.

[0077] Monitoring the coefficient of determination and root mean square error helps improve the model's predictive capabilities, thereby enhancing the accuracy of SSC predictions for the Kokterek grape and precise identification of landmarks.

[0078] This embodiment provides a specific comparative example 1, which uses NIR (Near-Infrared Spectroscopy) spectroscopy to construct a traditional PLS (Partial Least Squares) model to predict SSC and origin. The specific method is as follows:

[0079] 1) 131 local and 231 non-local grape fruits were collected from Xinjiang Kok Terek grapes. Each fruit was labeled and each sample was imaged and scanned using a pre-set hyperspectral imaging device. Hyperspectral data were extracted from three regions on the equatorial region of each sample surface using ENVI software. The SSC of each sample was then measured using a saccharimeter. The wavelength range was 400 to 1000 nm, and the soluble solids content was measured twice.

[0080] 2) Obtain the full wavelength dataset using full wavelength, SSC, and landmark categories, and split it into a training set of 290 samples and a test set of 72 samples;

[0081] 3) Train the traditional PLS model using the training set and make predictions using the prediction set.

[0082] This embodiment provides a specific comparative example 2, which uses NIR spectroscopy to construct a PLS model of BA characteristic wavelengths to predict SSC and origin. The specific method is as follows:

[0083] 1) 131 local and 231 non-local grape fruits were collected from Xinjiang Kok Terek grapes. Each fruit was labeled and each sample was imaged and scanned using a pre-set hyperspectral imaging device. Hyperspectral data were extracted from three regions on the equatorial region of each sample surface using ENVI software. The SSC of each sample was then measured using a saccharimeter. The wavelength range was 400 to 1000 nm, and the soluble solids content was measured twice.

[0084] 2) Obtain the full wavelength dataset using full wavelength, SSC, and landmark categories, and split it into a training set of 290 samples and a test set of 72 samples;

[0085] 3) Use the BA algorithm to filter the characteristic wavelengths of SSC and landmarks, and reconstruct the training set and prediction set.

[0086] 4) Construct a traditional PLS model using a reconstructed training set.

[0087] In summary, this method involves extracting and monitoring spectral information from grape landmark samples, constructing a spectral dataset based on the spectral dataset, splitting the dataset to obtain a spectral set, identifying shared features, constructing a feature set, and finally building an SSC prediction model and globally tuning hyperparameters using Bayesian based on the feature training set. This improves the effectiveness of SSC and landmark category prediction results, thereby enhancing the accuracy of SSC prediction and landmark identification for Korque Terrek grapes. It effectively solves the problem of high deviation in SSC prediction and landmark identification results for Korque Terrek grapes caused by differences in planting regulation in existing technologies.

[0088] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0089] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0090] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0091] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0092] Although the preferred embodiments of the present invention have been described, those skilled in the art may make additional changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the present invention.

[0093] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A grape landmark recognition method based on hyperspectral ensemble learning, characterized in that, Includes the following steps: S1, Spectral information is extracted and monitored based on the obtained grape landmark samples to obtain hyperspectral images and determine SSC. The grape landmark samples include Xinjiang Kok Terek landmark and non-landmark grape samples. S2, construct a spectral dataset based on the acquired spectral dataset, split the spectral dataset to obtain a spectral set, the spectral set includes a spectral training set and a spectral test set, and the spectral dataset construction data includes hyperspectral data, SSC and landmark categories; S3, perform feature recognition to obtain shared features, and construct a feature set based on the shared features to obtain a feature set, which includes a feature training set and a feature test set. The feature recognition means recognizing the features corresponding to SSC and landmark categories through BA. The feature training set is used to improve the accuracy of SSC prediction model construction, and the feature test set is used to predict and evaluate the XGBoost ensemble learning model. S4. Based on the feature training set, construct the SSC prediction model and perform Bayesian global tuning of hyperparameters to improve the prediction accuracy of SSC and landmark categories. The SSC prediction model construction is used to obtain the XGBoost ensemble learning model, the coefficient of determination, and the root mean square error. The Bayesian global tuning of hyperparameters is used to optimize the XGBoost ensemble learning model to improve model performance.

2. The grape landmark recognition method based on hyperspectral ensemble learning as described in claim 1, characterized in that, The hyperspectral image is obtained by setting the spectral image of the marked grape landmark sample information using a preset hyperspectral imaging device. The labeled grape landmark sample information represents grape landmark sample information of no less than a preset sample quantity obtained from the database; The spectral image settings include band range settings and resolution settings; The band range setting indicates that the band corresponding to the grape landmark sample information set by the preset personnel is within the preset band range; The resolution setting indicates that a prompt is sent to a preset person to set the spectral resolution corresponding to the grape landmark sample information to be within a preset resolution range.

3. The grape landmark recognition method based on hyperspectral ensemble learning as described in claim 1, characterized in that, The process of constructing the spectral dataset based on the acquired spectral dataset is as follows: An equatorial center image of the fruit is acquired by imaging and scanning using a preset hyperspectral imaging device; Hyperspectral data is obtained by extracting data from the image of the fruit's equatorial center. The data extraction refers to the extraction from a preset image region at the fruit's equatorial center using ENVI software. SSCs were obtained by measuring labeled grape landmark samples using a saccharimeter. Landmark categories are obtained by classifying landmarks according to SSC. The spectral dataset is constructed by building the data into a dataset. The dataset construction means summarizing the spectral dataset into a set. The acquired spectral dataset is smoothed. The smoothing process refers to Savitzky-Golay smoothing to eliminate noise in the spectral dataset.

4. The grape landmark recognition method based on hyperspectral ensemble learning as described in claim 3, characterized in that, The process of constructing a spectral dataset based on the acquired spectral dataset also includes quantifying the extraction accuracy. The specific process for quantifying the accuracy of extraction is as follows: Accurate parameters for hyperspectral extraction were obtained during the data extraction process of the fruit's equatorial center image. Based on the obtained hyperspectral extraction accuracy parameters and the preset hyperspectral extraction accuracy parameters, the extraction accuracy of the hyperspectral data of the fruit equatorial center image is quantified to obtain the extraction accuracy quantization value; The specific process for obtaining the quantitative value of the extraction accuracy is as follows: After analyzing the proportion of preset spectral resolution deviation and spectral resolution deviation, a weighted operation is performed by combining resolution deviation-extraction accuracy factor to obtain the resolution deviation-extraction accuracy value, which is used to reflect the effect of spectral resolution deviation on the accuracy of hyperspectral data extraction from fruit equatorial center image. After analyzing the proportion of the average spectral signal signal-to-noise ratio and the preset average spectral signal signal-to-noise ratio, the signal-to-noise ratio-extraction accuracy value is obtained by weighting the signal-to-noise ratio-extraction accuracy factor. This value reflects the role of the average spectral signal signal-to-noise ratio in the accuracy of extracting hyperspectral data from the fruit equatorial center image. After analyzing the proportion of the preset average wavelength offset and average wavelength offset, the wavelength offset-extraction accuracy value is obtained by weighting the wavelength offset-extraction accuracy factor. This value reflects the effect of the average wavelength offset on the accuracy of extracting hyperspectral data from the fruit equatorial center image. The extraction accuracy is quantified by coupling the hyperspectral extraction data with the data. The extraction accuracy quantification value is used to reflect the combined effect of the hyperspectral extraction accuracy parameters and the preset hyperspectral extraction accuracy parameters on the accuracy of hyperspectral data extraction from the fruit equatorial center image; The hyperspectral extraction data includes resolution deviation-extraction accuracy, signal-to-noise ratio-extraction accuracy, and wavelength shift-extraction accuracy, and all hyperspectral extraction data are greater than 0; The accurate parameters for hyperspectral extraction include average spectral resolution deviation, average spectral signal-to-noise ratio, and average wavelength offset.

5. The grape landmark recognition method based on hyperspectral ensemble learning as described in claim 4, characterized in that, The specific process for acquiring the hyperspectral data is as follows: Determine whether the quantitative value of extraction accuracy meets the extraction accuracy criteria; When the extraction accuracy quantification value meets the extraction accuracy condition, an extraction qualified prompt is sent and the corresponding hyperspectral data is obtained; When the extraction accuracy quantification value does not meet the extraction accuracy conditions, an extraction failure prompt is sent to optimize the extraction accuracy. The aforementioned accuracy optimization is used to improve the accuracy of extracting hyperspectral data from fruit equatorial center images; The extraction accuracy condition means that the extraction accuracy quantification value is not lower than the preset extraction accuracy quantification value obtained from the database.

6. The grape landmark recognition method based on hyperspectral ensemble learning as described in claim 5, characterized in that, The specific steps for optimizing extraction accuracy are as follows: Optimized and qualified parameters were obtained during the data extraction process of the fruit's equatorial center image; After analyzing the proportion of the light source intensity homogenization ratio and the preset light source intensity homogenization ratio, the first light source homogenization value is obtained by weighting the first light source homogenization factor. This value is used to reflect the role of the light source intensity homogenization ratio in the extraction of hyperspectral data of the fruit equatorial center image. After analyzing the proportion of average light source intensity and preset average light source intensity, the second light source uniformity value is obtained by weighting the second light source uniformity factor. This value is used to reflect the role of average light source intensity in the extraction of hyperspectral data of fruit equatorial center image. After analyzing the proportion of the preset average image reflectance and average image reflectance, the third light source uniformity value is obtained by weighting the third light source uniformity factor. This value is used to reflect the role of average image reflectance in the extraction of hyperspectral data of fruit equatorial center image. The optimized extraction accuracy value is obtained by coupling the data with the data for extraction accuracy optimization. The extraction accuracy optimization value is used to reflect the combined effect of the optimized qualified parameters and the preset optimized qualified parameters on the uniformity of the light source during the extraction of hyperspectral data from the fruit equatorial center image; The data for optimizing extraction accuracy includes the uniform values ​​of the first light source, the second light source, and the third light source, and all of the data for optimizing extraction accuracy are greater than 0. The optimized qualified parameters include the light source intensity homogenization ratio, average light source intensity, and average image reflectance; An extraction adjustment factor is obtained by performing a ratio analysis between the optimized extraction accuracy value and the preset optimized extraction accuracy value obtained from the database; Setting the light source intensity means sending a prompt to a preset person to gradually increase the light source intensity by extracting the amplitude corresponding to the adjustment factor. Setting the working distance means sending a prompt to a preset person to extract the amplitude corresponding to the adjustment factor and gradually reduce the working distance of the preset hyperspectral imaging device. If the working distance of the preset hyperspectral imaging device decreases to the preset minimum working distance or the light source intensity increases to the preset maximum light source intensity, and the extraction accuracy quantification value does not meet the extraction accuracy conditions, an alarm will be sent.

7. The grape landmark recognition method based on hyperspectral ensemble learning as described in claim 1, characterized in that, The specific process of splitting the spectral dataset to obtain the spectral set is as follows: Step 1: Set the split ratio. The setting of the split ratio means sending a prompt to a preset person to set the split ratio of the spectral dataset. Step 2, stratified random sampling, whereby the stratified random sampling means splitting the spectral dataset using a stratified random sampling method; The stratified random sampling is used to ensure that the landmark categories and SSCs in the spectral set are normally distributed.

8. The grape landmark recognition method based on hyperspectral ensemble learning as described in claim 1, characterized in that, The specific process of obtaining shared features through feature recognition is as follows: SS1, Generate BA, which is obtained through random forest ensemble; SS2, code reproduction, where code reproduction means reproduction using R code; SS3, perform feature-level fusion, which means identifying the features corresponding to SSC and landmark categories through BA and combining them to obtain fused features; SS4, perform average importance assessment, which means calculating the average importance score based on the reduction of node impurity in the random forest; SS5, set an importance threshold, which means sending a prompt to a preset person to set an importance threshold; SS6, perform feature filtering, which means filtering fused features based on importance thresholds to obtain shared features.

9. The grape landmark recognition method based on hyperspectral ensemble learning as described in claim 1, characterized in that, The specific process for obtaining the coefficient of determination is as follows: The difference value of grape SSC is obtained by processing the difference between the actual value and the predicted value of grape SSC. The grape SSC difference value is used to reflect the relative deviation between the actual grape SSC value and the predicted grape SSC value, and together they affect the fitting accuracy of the regression model. The average grape SSC difference value is obtained by processing the difference between the actual value of grape SSC and the mean of the actual value of grape SSC. The average grape SSC difference value is used to reflect the relative deviation between the actual grape SSC value and the mean of the actual grape SSC value, and together they affect the fitting accuracy of the regression model. The coefficient of determination was obtained by performing a convergence analysis on the grape SSC difference value and the average grape SSC difference value, followed by a deviation adjustment for the goodness of fit. The coefficient of determination is used to reflect the combined effect of grape evaluation parameters and grape SSC predicted values ​​on the accuracy of the regression model. The grape evaluation parameters include the actual SSC value of the grape and the average actual SSC value of the grape; The construction of the SSC prediction model based on the feature training set and the Bayesian global tuning of hyperparameters to improve the accuracy of SSC and landmark category prediction also include the evaluation of classification model accuracy. The accuracy evaluation of the classification model refers to obtaining the accuracy and recall based on the confusion matrix, which is used to evaluate the classification accuracy of the XGBoost ensemble learning model.

10. A grape landmark recognition system based on hyperspectral ensemble learning, characterized in that, It includes a spectral information extraction and monitoring module, a dataset construction and splitting module, a feature set construction module, and a prediction accuracy evaluation module. The spectral information extraction and monitoring module is used to extract and monitor spectral information based on the acquired grape landmark samples to obtain hyperspectral images and measure SSC. The grape landmark samples include Xinjiang Kok Terek landmark and non-landmark grape samples. The dataset construction and splitting module is used to construct a spectral dataset based on the acquired spectral dataset, and to split the spectral dataset to obtain a spectral set. The spectral set includes a spectral training set and a spectral test set. The spectral dataset construction data includes hyperspectral data, SSC, and landmark categories. The feature set construction module is used to perform feature recognition to obtain shared features, and to construct a feature set based on the shared features. The feature set includes a feature training set and a feature test set. The feature recognition means to identify the features corresponding to SSC and landmark categories through BA. The feature training set is used to improve the accuracy of SSC prediction model construction, and the feature test set is used to predict and evaluate the XGBoost ensemble learning model. The prediction accuracy evaluation module is used to construct an SSC prediction model based on the feature training set and perform Bayesian global tuning of hyperparameters to improve the prediction accuracy of SSC and landmark categories. The SSC prediction model construction is used to obtain the XGBoost ensemble learning model, the coefficient of determination, and the root mean square error. The Bayesian global tuning of hyperparameters is used to optimize the XGBoost ensemble learning model to improve model performance.

Images