Intelligent identification and classification method for characteristics of sugarcane buds

Abstract

The invention discloses an intelligent identification and classification method for characteristics of sugarcane buds. The method comprises the following steps: marking sugarcane seeds; collecting microscope static sugarcane bud characteristic data; 360-degree dynamic sugarcane bud characteristic data are collected; the static sugarcane bud characteristics of the microscope and the 360-degree dynamic scanning sugarcane bud characteristics of the camera are fused; performing a sugarcane field planting experiment on the sugarcane seeds to obtain the actual germination rate of sugarcane buds; performing regression feature analysis on the basis of the budding data of the planted sugarcane seeds, extracting data influencing the budding features of sugarcane buds, obtaining the relationship between the budding features and the budding rate, and predicting the budding rate; establishing standards of good and bad buds of the sugarcane seeds according to characteristics of the sugarcane buds and the bud ratio, and carrying out combined modeling on classification of the good and bad buds and classification of the varieties of the sugarcane seeds; and identifying and detecting the sugarcane buds by using the model to finish classification of good and bad sugarcane buds and classification of different varieties of sugarcane seeds. According to the method, high-quality sugarcane buds can be identified more accurately, field planting can be guided, and the sugarcane seed screening efficiency is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of agricultural intelligence, and particularly relates to a method for intelligent recognition and classification of sugarcane bud characteristics. Background Art

[0002] At present, there are generally two methods for selecting sugarcane seeds and their ages. One is sugarcane seeds with an age of 6 - 8 months. Such sugarcane buds are healthy and plump, and the quality from the sugarcane tip to the root is not much different, and all can be used as sugarcane seeds. The other is sugarcane seeds with an age of about one year. The healthy sugarcane buds in the upper 1 / 3 section of this kind of sugarcane are larger than those in the middle 1 / 3 section of the sugarcane, and the healthy sugarcane buds in the middle 1 / 3 section of the sugarcane are larger than those in the root 1 / 3 section of the sugarcane. At present, the selection of sugarcane seeds mainly relies on manual identification and screening, with a large workload. If the screening and planting of the cut sugarcane seeds are not completed within the optimal time period (3 days), the sugarcane seeds will show phenomena such as dehydration, nutrient consumption, bud mutation, being fragile and easy to break, and pest and disease infection, which will affect the subsequent growth and yield of sugarcane seedlings. Sugarcane is asexual reproduction. At present, there are many and miscellaneous sugarcane varieties, and most sugarcane body appearances are similar. It is very difficult to distinguish the specific variety only by the detailed skills mastered by experience. Different sugarcane varieties have different growth advantages. For example: drought resistance, lodging resistance, pest and disease resistance, etc., or one variety has multiple growth advantages at the same time. Whether the seed selection is appropriate directly affects the sugarcane yield.

[0003] The information disclosed in this background art section is only intended to increase the understanding of the overall background of the present invention, and should not be regarded as an admission or any form of suggestion that this information constitutes the prior art already known to those of ordinary skill in the art. Summary of the Invention

[0004] The purpose of the present invention is to provide a method for intelligent recognition and classification of sugarcane bud characteristics, so as to overcome the disadvantages such as large labor intensity, low efficiency and low accuracy in sugarcane seed selection.

[0005] To achieve the above purpose, the present invention provides a method for intelligent recognition and classification of sugarcane bud characteristics, including the following steps:

[0006] (1) Label the sugarcane seeds;

[0007] (2) Collect static microscope sugarcane bud characteristic data;

[0008] (3) Collect 360° dynamic sugarcane bud characteristic data;

[0009] (4) Integrate the static microscope sugarcane bud characteristics and the 360° dynamic scanning sugarcane bud characteristics of the camera;

[0010] (5) Conduct a sugarcane field planting experiment on the sugarcane seeds to obtain the actual germination data of the sugarcane buds, that is, the germination situation;

[0011] (6) Based on the germination data of planted sugarcane seeds, perform regression feature analysis, extract the data of the characteristics affecting sugarcane bud germination, obtain the relationship between germination characteristics and germination rate, and predict the germination rate.

[0012] (7) According to the sugarcane bud characteristics and germination rate, establish the criteria for good and bad sugarcane buds, and jointly model the classification of good and bad buds and the classification of sugarcane seed varieties.

[0013] (8) Use the model to identify and detect sugarcane buds, complete the classification of good and bad sugarcane buds, and the classification of different varieties of sugarcane seeds.

[0014] Preferably, in the above technical solution, the sugarcane bud characteristic data obtained in step (2) includes the surface texture of the sugarcane bud, the health of the bud tip structure, and the microscopic color characteristics.

[0015] The method for obtaining the surface texture of the sugarcane bud includes extracting the texture parameters of the sugarcane bud through the Gray Level Co-occurrence Matrix (GLCM) to reflect the surface integrity of the sugarcane bud.

[0016] The method for obtaining the health of the bud tip structure includes: using an edge detection algorithm to extract the characteristics of the bud tip area and judge whether there are abnormalities such as fractures and corruptions.

[0017] The method for obtaining the microscopic color characteristics includes: analyzing the color distribution through the HSV color space to detect whether the bud tissue is degenerated or diseased.

[0018] Preferably, in the above technical solution, the 360° dynamic sugarcane bud characteristic data collected in step (3) includes: the camera captures the 360° rolling video of the sugarcane seeds, extracts each video frame frame by frame, captures the characteristics of different positions of the sugarcane seeds for real-time analysis and sugarcane bud recognition; among them, the dynamic sugarcane bud characteristics include the area and shape of the sugarcane bud (the image scanned dynamically is the image of the entire double-bud segment).

[0019] Preferably, in the above technical solution, in step (4), the image of the microscope and the video frame of the camera are processed to have consistent resolution and size, and then the image of the microscope is spliced onto the video frame for fusion. The fused image contains the high-resolution details of the microscope and the dynamic information of the video frame.

[0020] Preferably, in the above technical solution, after the sugarcane is planted, marks are made above the soil covering the sugarcane segments. The emergence situation of each sugarcane bud is counted every day. After 15 days, the sugarcane buds are classified into good, medium, and bad categories according to the emergence time of the sugarcane buds, and their corresponding label names are recorded; according to the experimental results, the pictures collected before cutting the seeds are classified and archived according to the emergence labels for dataset fusion and annotation; among them, the sugarcane bud classification criteria and the germination data of the sugarcane seeds include: good buds emerge within 10 days and are healthy, medium-grade buds emerge between 10 and 15 days, and bad buds emerge after 15 days or do not emerge.

[0021] Preferably, in the above technical solution, in step (6), a feature matrix is constructed by combining the microscopic image texture feature GLCM, the edge feature Sobel, and the color feature HSV, and a model is established by combining the germination rate data from field experiments; data preprocessing includes normalization, data filling, and enhancement to ensure the stability and generalization ability of classification and regression tasks; the CatBoost regression model is used, and the performance is evaluated by MSE, MAE, and R 2 to evaluate the performance, and the influence of different features on the germination rate is quantified by combining feature importance analysis and SHAP; Bayesian optimization is used to tune the parameters, and finally the microscopic features are input into the model to predict the germination rate, realizing sugarcane bud classification and interpretive modeling.

[0022] Preferably, in the above technical solution, constructing the feature matrix includes:

[0023] 1) Process the microscopic image data, introduce CNN to extract microscopic image features, and extract the sugarcane bud texture features by calculating the gray-level co-occurrence matrix GLCM, including contrast, entropy, homogeneity, and energy, to describe the texture characteristics of the sugarcane bud surface;

[0024] 2) Use the edge detection algorithm Sobel to extract the sugarcane bud edge features, and calculate geometric features such as edge continuity, breakage ratio, and edge complexity to reflect the integrity of the bud tip structure;

[0025] 3) Convert the sugarcane bud image to the HSV color space and extract color features, including hue mean, saturation mean and standard deviation, and lightness mean and standard deviation, to capture the microscopic color characteristics of the sugarcane bud tissue;

[0026] 4) Combine the feature data with the germination rate data from field experiments to construct the feature matrix X = [GLCM feature, edge feature, color feature] and Y = germination rate.

[0027] Preferably, in the above technical solution, in step (6), the CatBoost regression model is used, and the performance is evaluated by MSE, MAE, and R 2 to evaluate the performance, and the influence of different features on the germination rate is quantified by combining feature importance analysis and SHAP, including:

[0028] 1) In the model construction stage, select the CatBoost gradient boosting tree method, the target is the regression task, and the optimization metrics are the mean square error MSE, the mean absolute error MAE, and the coefficient of determination R 2Hyperparameter settings include: The default value of the learning rate is 0.1, which can be adjusted to 0.01 - 0.3; the maximum depth of the tree max_depth is 5 - 10 to control the model complexity; the subsample ratio is set to 0.7 - 0.9 to prevent overfitting; the feature sampling ratio is 0.6 - 0.8 to control the number of features used in each tree; the number of trees is selected from 300 - 1000 and further optimized according to the performance on the validation set. During the training process, the training set is used to fit the model Y = f(X). The gradient boosting tree automatically evaluates the contribution of each feature to the budding rate by learning the complex relationships between features, providing a basis for subsequent feature importance analysis.

[0029] 2) In the model validation and tuning stage, evaluate the model performance on the test set, using the mean squared error MSE, mean absolute error MAE, and coefficient of determination R 2 as the main metrics;

[0030] 3) Utilize the built-in feature importance evaluation mechanism of the gradient boosting tree to analyze the contribution of each feature to the good, medium, and bad budding rates, and determine which features have a greater impact on the model prediction.

[0031] 4) Feature impact quantification: Extract the importance scores of GLCM features, edge features, and color features.

[0032] GLCM feature quantification:

[0033] Budding rate = a1 * Contrast + a2 * Entropy + a3 * Homogeneity + c

[0034] Edge feature quantification:

[0035] Budding rate = b1 * Edge continuity + b2 * (1 / Proportion of breakage) + b3 * (1 / Edge complexity) + d

[0036] Color feature quantification:

[0037] Budding rate = c1 * Hue mean + c2 * Saturation mean - c3 * Lightness standard deviation + e

[0038] After quantifying the importance of the features, combine the experimental data to form an explanatory model M for the impact of features on the budding rate.

[0039] 5) Use the feature importance analysis tool SHAP to visually present the impact of features on the prediction results. Take microscopic features as the input and use the trained model to predict the budding rate, where 0.8 < Y is defined as good buds; 0.5 ≤ Y ≤ 0.8 is medium buds; Y < 0.5 is bad buds.

[0040] Preferably, in the above technical solution, the model constructed in step (7) adopts a two-stage modeling strategy: in the first stage, the germination rate is predicted based on microscopic features, and good and bad buds are classified according to a threshold. In the second stage, multi-task joint modeling of good and bad bud classification and sugarcane variety classification is carried out. A shared feature extraction layer and task branches are used, and the cross-entropy loss is combined to optimize the two tasks. The model is trained through an optimizer, and the evaluation includes regression and classification metrics. SHAP analysis is used for visualization to analyze the importance of microscopic features, display the classification results and confusion matrix, and ensure the model performance and interpretability.

[0041] Preferably, in the above technical solution, in step (7), based on the sugarcane bud characteristics and germination rate, a standard for good and bad sugarcane buds is established, and joint modeling of good and bad bud classification and sugarcane variety classification is carried out, including:

[0042] 1) In the data preparation stage, the input data of the model includes microscopic images, video frame images, and fused images. The task labels are designed into task one for good and bad bud classification and task two for sugarcane variety classification. For task one of good and bad bud classification, classification labels are generated based on the predicted value of the germination rate Y. Among them, good buds are defined as the germination rate where 0.8 > Y, secondary buds are 0.5 ≤ Y ≤ 0.8, and bad buds are Y < 0.5. For task two of sugarcane variety classification, corresponding sugarcane variety labels are provided for each sample according to the experimental data.

[0043] 2) The model design adopts a two-stage modeling strategy: in the first stage, the trained germination rate prediction model M is used to predict the germination rate Y based on microscopic features, and Y is divided into good buds Y > 0.8, secondary buds 0.5 ≤ Y ≤ 0.8, and bad buds Y < 0.5 categories through threshold segmentation, thereby generating the classification labels for task one. In the second stage, based on the microscopic feature vector, multi-task joint modeling is carried out in combination with the labels of task one and task two. The model architecture includes a shared feature extraction layer and two task branches. First, the deep information of microscopic features is extracted through the shared layer CNN, and then the task one branch outputs the probability distribution of good and bad bud classification through a fully connected Dense layer, and the Softmax activation function is used to classify the three categories of sugarcane bad buds. The task two branch outputs the probability distribution of sugarcane variety classification through multiple Dense layers, and the Softmax activation function is also used.

[0044] 3) A joint loss function is used to optimize the two tasks, in the form of L = α·Ltask1 + β·Ltask2, where α and β are weight parameters used to balance the learning of the two tasks and can be tuned through the validation set. Cross-Entropy Loss is used for task one, and cross-entropy loss is also used for task two.

[0045] 4) During the model training stage, input microscopic features and their corresponding two-task labels batch by batch, and optimize the weights of the shared feature layer and the two-task branches simultaneously; during the verification process, monitor the loss changes of Task 1 and Task 2 to ensure the balance of the training process;

[0046] 5) Model evaluation includes the independent performance and overall performance of the two tasks. The evaluation of Task 1 includes regression metrics MSE, MAE, R 2 , to measure the accuracy of the germination rate prediction, and classification metrics, including accuracy and F1 score, to evaluate the classification effect of good buds, substandard buds, and bad buds; the evaluation of Task 2 analyzes the sugarcane variety classification performance through a confusion matrix, classification report, and classification accuracy, where the classification includes precision, recall, and F1 score; the overall performance is evaluated through the joint loss function value to judge the learning effect and balance of the model when considering both tasks;

[0047] 6) In terms of visualization and interpretation, use SHAP for feature importance analysis to reveal the contributions of microscopic features to the classification of good and bad buds and sugarcane variety classification; in terms of result display, output the prediction distribution of good and bad bud classification, including the number of good buds, substandard buds, and bad buds, and display the confusion matrix and the distribution of each category of the sugarcane variety classification result.

[0048] Compared with the prior art, the present invention has the following beneficial effects:

[0049] (1) The intelligent identification and classification method for sugarcane bud characteristics of the present invention can capture the fine-grained microscopic characteristics and macroscopic structural characteristics of sugarcane buds through the fusion of microscopic static feature collection and 360° dynamic scanning data. This combination of multi-modal data improves the comprehensive understanding and accurate prediction ability of sugarcane bud characteristics. Associate the actual germination rate data of field planting experiments with the laboratory feature collection results, and extract the key features affecting germination through regression analysis. This method not only considers the actual planting conditions but also combines high-precision experimental data, providing higher reliability for the prediction model. Based on the germination rate prediction and feature analysis, establish the criteria for good and bad sugarcane buds, and realize the joint modeling of good and bad bud classification and sugarcane variety classification. Compared with the prior art, it can more accurately identify high-quality sugarcane buds and guide field planting, improving the efficiency of sugarcane seed selection.

[0050] (2) The intelligent identification and classification method for sugarcane bud characteristics of the present invention provides a more scientific basis for the classification task by predicting the germination rate, realizing the transition from quantitative to qualitative analysis. Compared with traditional single classification methods, using the predicted germination rate to dynamically divide the categories of good and bad buds improves the scientific nature of the classification criteria. Description of the Drawings

[0051] Figure 1It is the overall flowchart of the intelligent recognition and classification method for sugarcane bud characteristics according to the present invention;

[0052] Figure 2 It is the flowchart of the combined modeling steps for good and bad bud classification and sugarcane variety classification according to the present invention;

[0053] Figure 3 It is the column statistical table of the budding category results of different sugarcane nodes in the first batch of sugarcane field planting experiments of the present invention;

[0054] Figure 4 It is the pie chart of the budding category results of sugarcane buds in the first batch of sugarcane field planting experiments of the present invention. Specific Embodiments

[0055] The following combines the accompanying drawings to describe the specific embodiments of the present invention in detail, but it should be understood that the protection scope of the present invention is not limited by the specific embodiments.

[0056] Unless otherwise clearly stated, throughout the specification and claims, the term "comprising" or its variations such as "including" or "having" etc. will be understood to include the stated elements or components, without excluding other elements or other components.

[0057] As Figures 1 to 2 shown, an intelligent recognition and classification method for sugarcane bud characteristics according to a specific embodiment of the present invention includes the following steps:

[0058] Step 1: Label the sugarcane seeds.

[0059] Paper labels need to be pasted on each sugarcane from the tip to the tail. The information to be labeled includes: sugarcane variety, sugarcane root number, and each sugarcane node number, so that each sugarcane bud has a unique number.

[0060] The sugarcane seeds used for sample experiments include Guitang 42, Zhongzhe 9, Zhongzhe 14, and Zhongzhe 15. Select ratoon sugarcane with uniform sugarcane node length and diameter and about 25 sugarcane nodes to ensure more accurate budding rate in single-factor experiments. Use a marker pen to number and mark each sugarcane node from the sugarcane tip to the sugarcane root. Then, cut the sugarcane seeds into double-bud segments in sequence, and paste paper labels on each sugarcane node. The label information includes the sugarcane variety, the number of this sugarcane (for example: which sugarcane), and the number of each sugarcane node on this sugarcane (for example: which sugarcane node).

[0061] Step 2: Collect static sugarcane bud characteristic data under the microscope.

[0062] An industrial stereo microscope with an ultra-clear measurement CCD camera is used to clearly observe the entire area of the sugarcane buds and truly restore the color of the sugarcane buds. The marked double-bud sugarcane seeds are placed on the working table of the industrial microscope with the front surface of the sugarcane buds facing up. At a high resolution of 3840*2160, the sugarcane buds are magnified 30 times and photographed and recorded for preservation.

[0063] A high-resolution microscope is used to obtain microscopic images of the sugarcane buds, capturing key features of the health status of the sugarcane buds at the microscopic level. The obtained sugarcane bud feature data includes the surface texture of the sugarcane buds, the health of the bud tip structure, and the microscopic color characteristics.

[0064] Surface texture of sugarcane buds: Texture parameters are extracted through the gray-level co-occurrence matrix (GLCM) to reflect the integrity of the sugarcane bud surface. Health of the bud tip structure: Edge detection algorithms are used to extract the characteristics of the bud tip area to determine whether there are abnormalities such as breaks and spoilage. Microscopic color characteristics: The color distribution is analyzed through the HSV color space to detect whether the bud tissue has degenerated or become diseased.

[0065] Step 3: Collect 360° dynamic sugarcane bud feature data.

[0066] The equipment used is a 360° intelligent sugarcane bud screening machine. By flipping the sugarcane segments 360°, the sugarcane buds on both sides of the sugarcane seeds can be observed, and the sugarcane buds are identified and screened comprehensively, and the inferior seeds are removed. The Chinese patent application number of this screening machine is 2023110903066, and the name is a 360° flipping sugarcane bud identification and screening equipment for sugarcane seeds and its control method.

[0067] The 360° intelligent sugarcane bud screening machine consists of a bracket, a sugarcane seed conveyor belt, a black box, a camera, a cylinder, a sugarcane seed baffle, etc. The conveyor belt is designed with a slope and is divided into an uphill and a downhill. The camera is located above the front of the descending section of the slope. The sugarcane seeds are put into the entrance of the 360° intelligent sugarcane bud screening machine. The sugarcane seeds start to roll 360° from the highest point of the conveyor belt. The camera records the rolling video and collects the data characteristics of the double buds on the opposite side of the sugarcane seeds. Among them, the dynamic sugarcane bud characteristics include the area and shape of the sugarcane buds (the image scanned dynamically is the image of the entire double-bud segment).

[0068] The double-bud sugarcane seeds are flipped 360° in the 360° sugarcane bud identification and screening equipment. The camera in the black box collects the video data of the 360° rolling of the sugarcane seeds and sends the video data of the sugarcane seeds to the controller. After the controller obtains the sugarcane seed video data transmitted by the camera, it uses OpenCV to open the video stream and extracts video frames frame by frame. During the rolling of the sugarcane seeds, the camera captures the dynamic video data of the sugarcane buds in real time. When the sugarcane seeds are flipped 360°, the camera can observe the sugarcane buds from multiple angles and capture their characteristics at different positions. These data will be used for real-time analysis and sugarcane bud identification.

[0069] Step Four: Integrate the static sugarcane bud features observed under the microscope and the 360° dynamic scanning sugarcane bud features captured by the camera.

[0070] Process the microscope images and the video frames of the camera to have consistent resolution and size, and then splice the microscope images onto the video frames for integration. The integrated images contain the high-resolution details of the microscope and the dynamic information of the video frames.

[0071] Specifically, the microscope images have a high resolution and can clearly show the details of the sugarcane buds. Since the microscope images only cover a small part of the sugarcane buds, and the dynamic video of the camera can capture the 360° tumbling of the entire sugarcane seed, it is necessary to ensure that the two are as temporally matched as possible. Since the microscope image acquisition is completed before the camera video acquisition, the microscope images are matched with the video frames through serial numbers to ensure their spatial and temporal correlation.

[0072] Microscopic image (A): Provides high-resolution local features (microscopic structure, texture, morphology of sugarcane buds). Video frame (B): Completely shows the overall appearance information of the sugarcane segment (such as shape, color, overall features). The two are complementary: the microscopic image captures microscopic features, and the video frame shows macroscopic features. The combined input can significantly improve the prediction accuracy.

[0073] Import and read the high-resolution microscope images, crop or scale the images so that they have the same resolution and size as the video frames of the camera for easy integration. Use OpenCV to open the dynamic video stream of the camera and extract video frames frame by frame. Each video frame also needs to be resized and normalized to have the same size and format as the microscope images. Splice the microscope images to the upper right corner of the video frames. The final integrated images will contain the high-resolution details from the microscope and the dynamic information from the video frames.

[0074] Step Five: Conduct sugarcane field planting experiments on sugarcane seeds to obtain the actual bud emergence situation of the sugarcane buds.

[0075] Field planting experiments provide standardized conditions for the experiments. The standard steps for sugarcane planting are as follows: First, the seedling is treated by cutting it into double-bud segments and then soaking and disinfecting it. The sowing method is double-row top-connecting strip sowing. The row spacing of double-row sowing is greater than 10 cm. 7 cm of loose soil is laid under the bottom of the ditch, and 5 cm of loose soil is covered on the top. The soil is taken from both sides of the ditch for covering, forming a state of deep ditch and shallow planting to facilitate water retention. The row spacing is set at 1.4 m per ridge, and 5-6 double-bud segments are placed at the bottom of each meter of the ditch to ensure the planting density. When digging the ditch, the bottom width of the ditch is required to be 20-25 cm, the top width of the ditch is 40-50 cm, and the depth is 20-30 cm. When placing the seeds, the seedlings need to be placed flat in the center of the ditch bottom, with the bud eyes facing both sides, ensuring that the seed stems are in close contact with the soil. After covering the soil, for easy identification, plastic tags corresponding to paper tags are inserted above the covered soil of the sugarcane segments as marks for subsequent observation and statistics of the germination rate data.

[0076] After 15 days, the sugarcane buds are divided into three categories: good, medium, and bad according to the germination time of the sugarcane buds, and the corresponding tag names are recorded. According to the experimental results, the pictures collected before cutting the seeds are classified and archived according to the good, medium, and bad tags for use in dataset fusion and annotation. The classification criteria for sugarcane buds are as follows: The germination data of sugarcane seeds include good buds (the sugarcane buds emerge from the soil and are healthy within 10 days), medium-grade buds (emerge from the soil within 10 to 15 days), and bad buds (the sugarcane buds that emerge from the soil after 15 days or do not emerge).

[0077] Step 6: Based on the germination data of the planted sugarcane seeds, perform regression feature analysis, extract the characteristic data affecting the germination of sugarcane buds, obtain the relationship between the germination characteristics and the germination rate, and predict the germination rate.

[0078] By combining the microscopic image texture feature GLCM, edge feature Sobel, and color feature HSV to construct a feature matrix, and combining with the germination rate data of the field experiment for modeling; data preprocessing includes normalization, data filling, and enhancement to ensure the stability and generalization ability of classification and regression tasks; using the CatBoost regression model, through MSE, MAE, and R 2 evaluate the performance, and combine feature importance analysis and SHAP to quantify the influence of different features on the germination rate; use Bayesian optimization to tune the parameters, and finally input the microscopic features into the model to predict the germination rate, realizing the classification of sugarcane buds (good buds, medium buds, bad buds) and interpretive modeling. Specifically, it includes the following:

[0079] (1) Data preparation

[0080] 1) Process the microscopic image data, introduce CNN to extract microscopic image features, and first extract the texture features of sugarcane buds by calculating the gray-level co-occurrence matrix (GLCM). It includes contrast, entropy, homogeneity, and energy to describe the texture characteristics of the sugarcane bud surface;

[0081] The surface texture features of sugarcane buds reflect the fineness and regularity of the surface texture of sugarcane buds. Among them, the contrast reflects the difference in gray values. The higher the value, the stronger the surface inhomogeneity; the entropy represents the complexity of the texture. The higher the value, the more disordered the surface; the homogeneity measures the degree of uniformity of the gray distribution. The higher the value, the smoother the surface; the energy reflects the repeatability of the image. The higher the value, the more regular the texture.

[0082] Relationship between features and sugarcane buds mapping:

[0083] The surface of healthy buds is usually relatively flat, with high homogeneity and low entropy. The contrast and energy also show medium levels. Contrast: Medium (smooth but regular cell texture). Entropy: Low (high surface regularity). Homogeneity: High (no cracks or damage on the surface).

[0084] 2) Use the edge detection algorithm (Sobel) to extract the edge features of sugarcane buds, and calculate geometric features such as edge continuity, breakage ratio, and edge complexity to reflect the integrity of the bud tip structure.

[0085] Edge continuity: Proportion of continuous edges (the proportion of edges is large when the integrity is high). Edge complexity: Calculate the mean and standard deviation of the edge curvature to reflect abnormal cusps or curved surfaces. Breakage ratio: Calculate the proportion of the damaged area in the entire bud tip area. Healthy bud tip: Smooth edge, high continuity, small curvature change, low breakage ratio. The germination rate is positively correlated with edge continuity and negatively correlated with the breakage ratio and edge complexity.

[0086] 3) Convert the sugarcane bud image into the HSV color space and extract color features, including hue mean, saturation mean and standard deviation, and lightness mean and standard deviation, to capture the microscopic color characteristics of sugarcane tissues.

[0087] Hue mean: Healthy buds are usually in a specific hue range (such as green or yellow-green). Saturation mean and standard deviation: A decrease in saturation may indicate tissue degradation. Lightness mean and standard deviation: Dull buds may be diseased. Normalize these features to eliminate the influence of image illumination differences. Hue is greenish or uniform: Related to health and high germination rate. High saturation and stable distribution: An indicator of healthy buds. Moderate lightness: Too bright or too dark may reflect health problems.

[0088] 4) Combine the feature data with the germination rate data of field experiments to construct the feature matrix X = [GLCM features, edge features, color features] and Y = germination rate.

[0089] (2) Data preprocessing

[0090] To ensure the modeling effect, the feature data is normalized to avoid the impact of feature magnitude differences on model training. At the same time, missing values are filled in by mean filling. If there are many missing values, consider removing some missing samples or increasing data collection. In addition, the dataset is divided into a training set, a validation set, and a test set according to the ratio of 7:1.5:1.5, and the cross-validation method is used to improve the model stability.

[0091] (3) Dataset annotation

[0092] Fifteen days later, the sugarcane buds are classified into three categories: good, medium, and bad according to the emergence time of the sugarcane buds. Locate the microscopic images archived before planting according to the geographical tags to ensure that each picture corresponds to a unique sugarcane bud number, and organize these images into an initial dataset. Use the professional image annotation tool LabelImg to perform sugarcane variety annotation and good / bad bud classification annotation on the fused images respectively. Mark the whole segment of Zhongzhe 9 as 9, the whole segment of Zhongzhe 14 as 14, and the whole segment of Zhongzhe 15 as 15; mark the good buds as 2, the medium buds as 2, and the bad buds as 0. Manually review to ensure the accuracy and consistency of the annotation. Divide the annotated dataset into a training set, a validation set, and a test set with a ratio of 70% (training set), 15% (validation set), and 15% (test set) to ensure the uniform distribution of samples of different categories in each set and avoid under-sampling and over-sampling. Data augmentation: Perform data augmentation operations on the fused sugarcane bud images, including: rotation, translation, scaling, brightness adjustment, flipping, and noise addition, to generate more diverse image samples and improve the generalization ability of the model under different conditions.

[0093] (4) Model construction

[0094] In the model construction stage, the CatBoost gradient boosting tree method is selected. The target is a regression task, and the optimization metrics are mean squared error (MSE), mean absolute error (MAE), and coefficient of determination R2. The hyperparameter settings include: the default value of the learning rate (LearningRate) is 0.1, which can be adjusted to 0.01 - 0.3; the maximum depth of the tree (max_depth) is 5 - 10 to control the model complexity; the subsample ratio (subsample) is set to 0.7 - 0.9 to prevent overfitting; the feature sampling ratio (colsample_bytree) is 0.6 - 0.8 to control the number of features used by each tree; the number of trees (n_estimators) is selected from 300 - 1000 and further optimized according to the performance of the validation set. During the training process, use the training set to fit the model Y = f(X). The gradient boosting tree automatically evaluates the contribution of each feature to the budding rate by learning the complex relationships between features, providing a basis for subsequent feature importance analysis.

[0095] (5) Model validation and tuning

[0096] In the model validation and tuning phase, first evaluate the model performance on the test set, using the mean squared error (MSE), mean absolute error (MAE), and coefficient of determination R2 as the main metrics. The formula for MSE is:

[0097]

[0098] where MSE: Mean Squared Error, is a metric that measures the difference between the predicted value and the actual value. The smaller the value, the closer the predicted value is to the actual value.

[0099] n: The number of samples, i.e., the total number of data points.

[0100] The predicted value of the i-th sample, calculated by the model.

[0101] y i : The actual value of the i-th sample, the true value obtained from experiments or observations.

[0102] The formula for R2 is:

[0103]

[0104] where R 2 : Coefficient of determination, which represents the goodness of fit of the model to the data, and its value ranges from 0 to 1.

[0105] R 2 The closer it is to 1, the stronger the explanatory power of the model.

[0106] Y: Observed value (actual value), which is the true data obtained from experiments or measurements.

[0107] Predicted value, the value calculated or predicted by the model.

[0108] The formula for the mean absolute error (MAE) is: The formula for is:

[0109]

[0110] where n: The number of samples, i.e., the total number of data points.

[0111] y i : The true value of the i-th sample.

[0112] The predicted value of the i-th sample.

[0113] The absolute value of the prediction error of the i-th sample.

[0114] (6) Utilize the feature importance evaluation mechanism built into the gradient boosting tree to analyze the contribution of each feature to the good, medium, and bad budding rates, which helps us understand which features have a greater impact on model prediction.

[0115] Feature impact quantification: Extract the importance scores of GLCM features, edge features, and color features.

[0116] Budding rate = a1 * Contrast + a2 * Entropy + a3 * Homogeneity + c.

[0117] Among them, the budding rate: the target variable, representing the germination proportion of sugarcane buds, usually expressed as a percentage (for example, 80% means 80 out of 100 buds germinate successfully).

[0118] Contrast: One of the image texture features, reflecting the degree of difference between image gray levels. The greater the contrast, the more obvious the difference between the bright and dark regions in the image. It is used to measure the texture clarity of sugarcane buds in the sugarcane bud experiment.

[0119] Entropy: Measures the randomness or complexity of the image gray distribution, reflecting the amount of information in the image. The greater the entropy, the more complex the pixel value changes in the image. In sugarcane bud detection, it is used to evaluate the surface complexity or distribution characteristics of sugarcane buds.

[0120] Homogeneity: Describes the degree of similarity of pixels in the image. The larger the value, the closer the gray levels of adjacent pixels in the image. It is related to the uniformity or health of the sugarcane bud surface.

[0121] a1, a2, a3: Coefficients of the regression model, representing the influence weights of each feature on the budding rate. C is the intercept term, representing the baseline budding rate when the contrast, entropy, and homogeneity are all zero.

[0122] This is a multiple linear regression model for predicting the budding rate. The coefficient weights of each feature (contrast, entropy, homogeneity) can be determined by fitting experimental data (features extracted from image processing and the actual budding rate). The larger the weight, the more significant the influence of the feature on the budding rate.

[0123] Edge feature quantification:

[0124] Budding rate = b1 * Edge continuity + b2 * (1 / Proportion of breakage) + b3 * (1 / Edge complexity) + d

[0125] Among them, edge continuity: A feature in the image, representing the coherence and integrity degree of the sugarcane bud edge. The higher the edge continuity, usually the more complete the shape of the sugarcane bud and the healthier the physical state.

[0126] The breakage ratio refers to the ratio of the damaged area of the sugarcane buds to the total area.

[0127] 1 / breakage ratio: It indicates that the smaller the breakage ratio (the more complete the sugarcane buds), the greater the contribution of this item to the germination rate.

[0128] The edge complexity refers to the degree of irregularity of the edges of the sugarcane buds, which is calculated by the ratio of the perimeter to the area. 1 / edge complexity: It indicates that the lower the complexity (the more regular the edges), the greater the contribution to the germination rate.

[0129] b1, b2, b3: The weight coefficients in the regression model, indicating the influence degree of each feature on the germination rate.

[0130] d: The intercept term, indicating the baseline germination rate when all feature terms are zero.

[0131] Color feature quantization:

[0132] Germination rate = c1 * Mean Hue + c2 * Mean Saturation - c3 * Standard Deviation of Lightness + e

[0133] Among them, Mean Hue: Hue is a component in the HSV color model, indicating the type of color (such as red, green, blue, etc.). The Mean Hue is the average color of the sugarcane bud area, reflecting the overall color characteristics. In sugarcane bud detection, different Mean Hue values may correspond to the differences between healthy buds, sub-optimal buds, and bad buds.

[0134] Mean Saturation: Saturation represents the purity or intensity of the color. The Mean Saturation is the average saturation of the colors in the sugarcane bud area. A higher saturation may indicate bright colors (characteristics of healthy buds), while a lower saturation may indicate sub-optimal buds or bad buds.

[0135] Standard Deviation of Lightness: Lightness represents the brightness of the color. The Standard Deviation of Lightness reflects the degree of change in brightness within the sugarcane bud area. A larger Standard Deviation of Lightness may indicate uneven color distribution on the bud surface, usually associated with poor health conditions.

[0136] c1, c2, c3: The regression coefficients in the model, indicating the influence weights of each feature on the germination rate.

[0137] e: The intercept term, indicating the baseline germination rate when the weight terms of all features are zero.

[0138] After quantifying the importance of features and combining experimental data, an explanatory model M for the impact of features on the budding rate is formed. The feature importance analysis tool SHAP is used to visually present the impact of features on the prediction results. Using microscopic features as input, the trained model is used to predict the budding rate, where a budding rate of 0.8 < Y is defined as good buds; 0.5 ≤ Y ≤ 0.8 is defined as sub-good buds; and Y < 0.5 is defined as bad buds.

[0139] Step 7: Based on the characteristics of sugarcane buds and the budding rate, establish criteria for good and bad sugarcane seeds, and jointly model the classification of good and bad buds and the classification of sugarcane seed varieties.

[0140] The constructed model adopts a two-stage modeling strategy: In the first stage, the budding rate is predicted through microscopic features and good and bad buds are classified according to thresholds. In the second stage, multi-task joint modeling of the classification of good and bad buds and the classification of sugarcane seed varieties is carried out. A shared feature extraction layer and task branches are used, and the cross-entropy loss is combined to optimize the two tasks; model training is carried out through an optimizer, and the evaluation includes regression and classification metrics; visualization uses SHAP to analyze the importance of microscopic features, display the classification results and confusion matrix, and ensure the performance and interpretability of the model. Specifically, it includes:

[0141] (1) In the data preparation stage, the input data of the model includes: microscopic images (microscopic images of a single sugarcane bud, containing detailed features), video frame images (dynamic feature images of the entire sugarcane), and fused images formed by splicing the two (the microscopic image is embedded in the upper right corner of the video frame to form a unified input). For the design of task labels, Task 1 (classification of good and bad buds): Classification labels are generated based on the predicted value of the budding rate Y: where good buds are defined as the budding rate

[0142] (0.8 > Y: class 2), sub-good buds are (0.5 ≤ Y ≤ 0.8: class 1), and bad buds are (Y < 0.5: class 0); Task 2 (classification of sugarcane seed varieties): Provide corresponding sugarcane seed variety labels for each sample based on experimental data. In addition, data augmentation operations are performed on microscopic images, video frames, and fused images, specifically including image rotation, translation, scaling, noise addition, brightness adjustment, and horizontal or vertical flipping to improve the robustness and generalization ability of the model.

[0143] (2) The model design adopts a two-stage modeling strategy: In the first stage (the germination rate prediction model M), the pre-trained model M is used to predict the germination rate Y based on microscopic features. Through threshold segmentation, Y is divided into good buds (Y > 0.8), secondary buds (0.5 ≤ Y ≤ 0.8), and bad buds (Y < 0.5) categories, thus generating the classification labels for Task 1. In the second stage (the multi-task classification model), based on the microscopic feature vector, multi-task joint modeling is carried out by combining the labels of Task 1 and Task 2. The model architecture includes a shared feature extraction layer and two task branches. First, the deep information of microscopic features is extracted through the shared layer CNN. Then, the task 1 branch (good and bad bud classification) outputs the probability distribution of good and bad bud classification through a fully connected Dense layer, and the Softmax activation function is used to classify the three categories (good buds, secondary buds, bad buds). The task 2 branch (sugarcane seed variety classification) outputs the probability distribution of sugarcane seed variety classification through multiple Dense layers, and the Softmax activation function is also used. Such a design can efficiently complete the joint modeling of the two tasks and make full use of the microscopic feature information.

[0144] (3) The loss function design uses a joint loss function to optimize the two tasks, in the form of L = α·Ltask1 + β·Ltask2, where α and β are weight parameters used to balance the learning of the two tasks and can be tuned through the validation set. Task 1 (good and bad bud classification) uses the cross-entropy loss, and Task 2 (sugarcane seed variety classification) also uses the cross-entropy loss. By jointly optimizing the loss function, the model can improve the performance of the two tasks simultaneously.

[0145] In the model training stage, the microscopic features and the corresponding two-task labels are input batch by batch, and the weights of the shared feature layer and the two task branches are optimized simultaneously. The optimizer uses Adam, the initial learning rate is set to 0.001, and it is dynamically adjusted in combination with the learning rate scheduler. During the validation process, the loss changes of Task 1 and Task 2 are monitored to ensure the balance of the training process. In the validation and testing stages, Task 1 is evaluated through classification accuracy, F1 score, precision, recall, and confusion matrix, while Task 2 is evaluated through indicators such as classification accuracy, precision, recall, and F1 score.

[0146] (4) The model evaluation includes the independent performance and overall performance of the two tasks. The evaluation of Task 1 includes regression metrics (MSE, MAE, R 2 ) to measure the accuracy of germination rate prediction, and classification metrics (accuracy, F1 score) to evaluate the classification effect of good buds, secondary buds, and bad buds. The evaluation of Task 2 analyzes the sugarcane seed variety classification performance through the confusion matrix, classification report (precision, recall, F1 score), and classification accuracy. In addition, the overall performance is evaluated through the joint loss function value to judge the learning effect and balance of the model when considering both tasks.

[0147] (5) In terms of visualization and interpretation, SHAP is used for feature importance analysis to reveal the contributions of microscopic features to the classification of good and bad buds and the classification of sugarcane varieties. In terms of result display, the predicted distribution of good and bad bud classification is output, including the numbers of good buds, secondary buds, and bad buds, and the confusion matrix of the sugarcane variety classification result and the distribution of each category are displayed.

[0148] Step Eight: Use the model to identify and detect sugarcane buds to complete the classification of good and bad sugarcane buds and the classification of different sugarcane varieties.

[0149] Use the model to identify and detect sugarcane buds. The variety label of the sugarcane is displayed on the terminal display screen when the sugarcane segment rolls 360°. At the same time, when the sugarcane segment rolls 360°, determine good, secondary, and bad buds of the sugarcane buds on the segment. Obtain the coordinate information of good, secondary, and bad buds through the target tracking algorithm, transmit the coordinate information to the single-chip microcomputer, and transport sugarcane buds of different grades to different channels to complete classification and screening.

[0150] Use the method of the present invention to conduct a sugarcane field planting experiment. The test variety is Guitang 42, with 14 roots, 1 - 25 sugarcane joints, and a total of 350 buds. Among them, the number of good buds is 162, the number of secondary buds is 69, and the number of bad buds is 119. Figure 3 and Figure 4 are the statistical results of good and bad buds in the first batch of sugarcane field planting experiments. The sugarcane experimental sample is Guitang 42, with 14 sugarcane roots, and the sugarcane buds are from the first joint to the 25th joint, with a total of 350 buds. The budding statistical results are: 162 good buds, 69 secondary-grade buds, and 119 bad buds. Figure 3 is the proportion of good, secondary, and bad buds in each sugarcane bud; Figure 4 is the proportion of good, secondary, and bad buds in the overall sample. The experimental results of this batch will be used for regression analysis of the extracted microscopic features of sugarcane buds to predict the budding rate model and to lay the foundation for sugarcane bud classification.

[0151] The foregoing description of the specific exemplary embodiments of the present invention is for purposes of illustration and exemplification. These descriptions are not intended to limit the present invention to the precise forms disclosed, and obviously, many changes and variations are possible in light of the above teachings. The purpose of selecting and describing the exemplary embodiments is to explain the specific principles of the present invention and its practical applications, so that those skilled in the art can implement and utilize various different exemplary embodiments of the present invention, as well as various different selections and changes. The scope of the present invention is intended to be defined by the claims and their equivalents.

Claims

1. A method for intelligent identification and classification of sugarcane bud features, characterized in that: The following steps are involved: (1) Marking of sugarcane varieties; (2) Collecting static sugarcane bud characteristic data under a microscope; (3) Collecting 360° dynamic sugarcane bud characteristic data; (4) Combining the static sugarcane bud features of the microscope with the 360° dynamic scanning sugarcane bud features of the camera; (5) Carrying out sugarcane field planting experiments on the sugarcane seeds to obtain the actual sprouting conditions of the sugarcane sprouts; (6) Based on the budding data of planted sugarcane varieties, regression feature analysis is performed to extract the data on the characteristics that affect sugarcane budding, derive the relationship between budding characteristics and budding rate, and predict the budding rate; (7) Based on the characteristics of sugarcane buds and germination rates, the standards for good and bad sugarcane buds are established, and the classification of good and bad buds and sugarcane variety classification are jointly modeled; (8) Use the model to identify and detect sugarcane buds, classify good and bad sugarcane buds, and classify different varieties of sugarcane.

2. The method for intelligent identification and classification of sugarcane bud features according to claim 1, characterized in that: The sugarcane bud characteristic data acquired in step (2) include sugarcane bud surface texture, bud tip structure health and microscopic color characteristics; The method for obtaining the surface texture of sugarcane buds includes extracting the texture parameters of sugarcane buds through gray level co-occurrence matrix GLCM to reflect the surface integrity of sugarcane buds; Methods for obtaining the health of the bud tip structure include: using edge detection algorithms to extract features of the bud tip region and determine whether there are abnormalities such as fractures and corruption; Methods for obtaining microscopic color characteristics include: analyzing color distribution through HSV color space and detecting whether bud tissue is degenerated or diseased.

3. The method for intelligent identification and classification of sugarcane bud features according to claim 1, characterized in that: Step (3) collecting 360° dynamic sugarcane bud feature data includes: the camera collects 360° rolling video of the sugarcane, extracts video frames frame by frame, captures features of different positions of the sugarcane, and uses them for real-time analysis and sugarcane bud identification; wherein the dynamic sugarcane bud features include the area and shape of the sugarcane bud.

4. The method for intelligent identification and classification of sugarcane bud characteristics according to claim 1, characterized in that: Step (4) processes the microscope image and the camera video frame to have consistent resolution and size, and then splices the microscope image onto the video frame for fusion. The fused image contains the high-resolution details of the microscope and the dynamic information of the video frame.

5. The method for intelligent identification and classification of sugarcane bud characteristics according to claim 1, characterized in that: Step (5) After planting sugarcane, mark the top of the sugarcane segment covered with soil, count the emergence of each sugarcane bud every day, and classify the sugarcane buds into three categories: good, inferior, and bad according to the emergence time after 15 days, and record their corresponding label names; according to the experimental results, archive the pictures collected before cutting according to the emergence label classification, and use them for data set fusion and annotation; wherein, the sugarcane bud classification standard and the emergence data of sugarcane varieties include: good buds are sugarcane buds that emerge within 10 days and are healthy, inferior buds are sugarcane buds that emerge between 10 and 15 days, and bad buds are sugarcane buds that emerge after 15 days or have not emerged.

6. The method for intelligent identification and classification of sugarcane bud characteristics according to claim 1, characterized in that: Step (6) constructs a feature matrix by combining the microscopic image texture feature GLCM, edge feature Sobel and color feature HSV, and models it with the field experiment germination rate data; data preprocessing includes normalization, data filling and enhancement to ensure the stability and generalization ability of classification and regression tasks; uses the CatBoost regression model to calculate the performance of the model through MSE, MAE and R 2 The performance was evaluated, and feature importance analysis and SHAP were combined to quantify the impact of different features on germination rate. Bayesian optimization was used to adjust parameters, and finally the microscopic features were input into the model to predict germination rate, thus achieving sugarcane bud classification and explanatory modeling.

7. The method for intelligent identification and classification of sugarcane bud characteristics according to claim 6, characterized in that: Constructing the feature matrix includes: 1) The microscopic image data was processed, CNN was introduced to extract the microscopic image features, and the texture features of the sugarcane buds were extracted by calculating the gray level co-occurrence matrix GLCM, including contrast, entropy, homogeneity and energy, to describe the texture characteristics of the sugarcane bud surface; 2) The edge detection algorithm Sobel was used to extract the edge features of the sugarcane buds, and the geometric features such as edge continuity, damage ratio and edge complexity were calculated to reflect the integrity of the bud tip structure; 3) Convert the sugarcane bud image into HSV color space and extract color features, including hue mean, saturation mean and standard deviation, and brightness mean and standard deviation, to capture the microscopic color characteristics of sugarcane bud tissue; 4) Combine the feature data with the germination rate data of the field experiment to construct the feature matrix X = [GLCM feature, edge feature, color feature] and Y = germination rate.

8. The method for intelligent identification and classification of sugarcane bud characteristics according to claim 6, characterized in that: Step (6) uses the CatBoost regression model and uses MSE, MAE and R 2 Evaluate performance and combine feature importance analysis and SHAP to quantify the impact of different features on germination rate, including: 1) In the model building stage, the CatBoost gradient boosting tree method is selected, the target is the regression task, and the optimization indicators are the mean square error MSE, the mean absolute error MAE and the determination coefficient R 2 ; Hyperparameter settings include: the default value of the learning rate is 0.1, which can be adjusted to 0.01-0.3; the maximum depth of the tree is 5-10, which is used to control the complexity of the model; the sub-sample ratio is set to 0.7-0.9 to prevent overfitting; the feature sampling ratio is 0.6-0.8 to control the number of features used by each tree; the number of trees is selected to be 300-1000, and further optimized according to the performance of the validation set; during the training process, the training set is used to fit the model Y=f(X), and the gradient boosting tree automatically evaluates the contribution of each feature to the budding rate by learning the complex relationship between the features, providing a basis for the subsequent feature importance analysis; 2) During the model validation and tuning phase, the model performance is evaluated on the test set using mean square error (MSE), mean absolute error (MAE), and coefficient of determination (R). 2 As the main indicator; 3) Using the built-in feature importance evaluation mechanism of the gradient boosting tree, we analyze the contribution of each feature to the good, bad, and poor budding rates, and determine which features have a greater impact on model predictions; 4) Feature impact quantification: extract the importance scores of GLCM features, edge features, and color features; GLCM feature quantification: Germination rate = a1*contrast+a2*entropy+a3*homogeneity+c Edge feature quantization: Germination rate = b1*edge continuity + b2*(1 / breakage ratio) + b3*(1 / edge complexity) + d Color feature quantization: Germination rate = c1* mean hue + c2* mean saturation - c3* standard deviation of brightness + e After quantifying the importance of the features, combined with the experimental data, an explanatory model M of the effects of the features on germination rate was formed; 5) Use the feature importance analysis tool SHAP to intuitively present the impact of features on the prediction results; use the microscopic features as input and use the trained model to predict the germination rate, where 0.8<Y is a good bud; 0.5≤Y≤0.8 is a bad bud; Y<0.5 is a bad bud.

9. The method for intelligent identification and classification of sugarcane bud characteristics according to claim 1, characterized in that: The model constructed in step (7) adopts a two-stage modeling strategy: the first stage predicts the budding rate through microscopic features and divides good and bad buds according to the threshold, and the second stage performs multi-task joint modeling of good and bad bud classification and sugarcane variety classification. A shared feature extraction layer and task branch are used to optimize the two tasks in combination with cross entropy loss; model training is performed through an optimizer, and evaluation includes regression and classification indicators; visualization uses SHAP to analyze the importance of microscopic features, display classification results and confusion matrix, and ensure model performance and interpretability.

10. The method for intelligent identification and classification of sugarcane bud characteristics according to claim 9, characterized in that: Step (7) establishes the standard of good and bad buds of sugarcane varieties according to the characteristics of sugarcane buds and the budding rate, and jointly models the classification of good and bad buds and the classification of sugarcane varieties, including: 1) In the data preparation stage, the input data of the model includes microscopic images, video frame images and fused images; the task label design is divided into task 1: good and bad bud classification and task 2: sugarcane variety classification. Task 1: good and bad bud classification generates classification labels based on the predicted value of germination rate Y, where good buds are defined as germination rate, 0.8>Y, inferior buds are 0.5≤Y≤0.8, and bad buds are Y<0.5; Task 2: sugarcane variety classification provides corresponding sugarcane variety labels for each sample based on experimental data; 2) The model design adopts a two-stage modeling strategy: In the first stage, the germination rate prediction model M that has been trained is used to predict the germination rate Y based on microscopic features. Y is divided into good buds Y>0.8, inferior buds 0.5≤Y≤0.8 and bad buds Y<0.5 categories through threshold segmentation, thereby generating the classification label of task one; in the second stage, multi-task joint modeling is performed based on the microscopic feature vector and the labels of task one and task two are combined; the model architecture includes a shared feature extraction layer and two task branches. First, the deep information of microscopic features is extracted through the shared layer CNN, and then the task one branch outputs the probability distribution of good and bad bud classification through the fully connected Dense layer, and the Softmax activation function is used to classify the three types of sugarcane varieties with bad buds; the task two branch outputs the probability distribution of sugarcane variety classification through a multi-layer Dense layer, and the Softmax activation function is also used; 3) A joint loss function is used to optimize the two tasks in the form of L = α·Ltask1+β·Ltask2, where α and β are weight parameters used to balance the learning of the two tasks and can be tuned through the validation set. Task 1 uses cross-entropy loss, and task 2 also uses cross-entropy loss; 4) During the model training phase, microscopic features and the corresponding two-task labels are input in batches, and the weights of the shared feature layer and the two-task branches are optimized at the same time; during the verification process, the loss changes of task one and task two are monitored to ensure the balance of the training process; 5) Model evaluation includes the independent performance and overall performance of the two tasks. The evaluation of task one includes regression indicators MSE, MAE, R 2 , to measure the accuracy of germination rate prediction, as well as classification indicators, including accuracy and F1 score, to evaluate the classification effect of good buds, inferior buds, and bad buds; the evaluation of task 2 analyzes the classification performance of sugarcane varieties through confusion matrix, classification report and classification accuracy, where classification includes precision, recall rate, and F1 score; the overall performance is evaluated by the joint loss function value to judge the learning effect and balance of the model when taking into account the two tasks; 6) In terms of visualization and interpretation, SHAP was used to perform feature importance analysis to reveal the contribution of microscopic features to the classification of good and bad buds and sugarcane varieties. In terms of result presentation, the predicted distribution of good and bad bud classification was output, including the number of good buds, inferior buds, and bad buds, and the confusion matrix of the sugarcane variety classification results and the distribution of each category were displayed.

Images