A method and system for identifying ripe leaves of grape based on a machine learning model

Abstract

The application discloses a grape mature leaf recognition method and system based on a machine learning model, and relates to the technical field of leaf feature recognition. The original image of grape leaves is collected, and a standard grape leaf image is obtained after data preprocessing. Pixel-level contour segmentation is performed based on a visual processing software, inherent shape features are extracted, and a leaf maturity feature set is constructed. The feature set is sequentially subjected to elimination, screening and normalization processing to obtain normalized shape features. The normalized shape features are input into a trained lightweight double-branch recognition model to obtain a maturity recognition result. After division, a leaf maturity grading result and a corresponding confidence label are obtained. The application realizes accurate and automatic recognition of grape leaf maturity and senescence grade, effectively eliminates artificial recognition deviation, improves the stability and accuracy of leaf recognition in complex field scenes, and promotes the development of the grape planting industry in the direction of refinement, intelligence and industrialization.

Description

Technical Field

[0001] This invention relates to the field of leaf feature recognition technology, and in particular to a method and system for identifying ripe grape leaves based on a machine learning model. Background Technology

[0002] Grapes are a high-value-added cash crop, occupying an important position in the fruit and vegetable planting industry. Their yield and fruit quality directly affect the economic benefits of cultivation. Unlike conventional crops, grape cultivation is highly dependent on manual labor, requiring meticulous management throughout the entire process. Numerous field management procedures are involved, including pruning, inflorescence shaping, shoot pinching, and leaf thinning. The timing and precision of each procedure play a decisive role in grape growth, flower bud differentiation, fruit set, and fruit quality. Shoot pinching and leaf thinning are core management procedures in grape cultivation. Shoot pinching must be based on the number of new leaves and the degree of leaf expansion on the new shoots to accurately determine the pinching points for the main and lateral shoots, thus balancing vegetative and reproductive growth and ensuring the quality of flower bud differentiation and fruit set. Leaf thinning requires determining the pruning cycle based on the degree of leaf senescence. Aged leaves have significantly reduced photosynthetic capacity, consuming only tree nutrients and unable to effectively participate in photosynthesis. Timely removal of severely aged leaves optimizes ventilation and light penetration on the trellis, improving light energy utilization efficiency, which is crucial for achieving high-quality and high-yield grapes.

[0003] Currently, in domestic grape cultivation, the differentiation between new and old leaves, the determination of leaf senescence levels, the timing of topping, and the determination of the pruning cycle for old leaves mainly rely on the manual experience of growers. This traditional method of judgment is greatly influenced by the subjective factors of the operators. Different workers have significantly different planting experience and judgment standards, which easily leads to misjudgments of leaf age and inaccuracies in the determination of senescence levels. This, in turn, causes a series of production problems such as topping too early or too late, and delaying the pruning of old leaves. These problems directly lead to an imbalance between vegetative and reproductive growth in grapevines, a decline in the quality of flower bud differentiation, and also cause canopy closure, poor ventilation and light penetration, and reduced light energy utilization efficiency. This not only restricts the normal growth and development of grapes but also significantly reduces grape yield and fruit quality. Standardized and uniform pruning operations are difficult to achieve, severely limiting the development of large-scale, industrialized, and high-quality grape cultivation.

[0004] To address the numerous drawbacks of manual identification, the industry has gradually introduced automated leaf recognition technology for grape cultivation management. However, existing automated recognition technologies have significant shortcomings in adaptability and accuracy, making it difficult to adapt to the complex and ever-changing actual planting scenarios in the field. They suffer from insufficient recognition accuracy and a lack of quantification capabilities, failing to accurately quantify the growth stage of grape shoots, accurately classify leaf senescence levels, and standardize the topping and thinning operations across the entire vineyard. Consequently, they cannot meet the needs of refined and standardized grape cultivation and production.

[0005] In summary, current grape cultivation suffers from several problems: reliance on manual experience for leaf maturity and identification, strong subjective interference, inconsistent pruning standards, and poor adaptability and accuracy of automated identification technologies. These issues lead to problems such as deviations in topping timing, delayed thinning of old leaves, low-quality flower bud differentiation, and poor light energy utilization, which have long constrained the improvement of grape yield and quality, hindering the development of refined, intelligent, and industrialized grape cultivation. Therefore, there is an urgent need to develop a highly accurate method for identifying mature grape leaves that is adaptable to complex field conditions and can be standardized for implementation, in order to address the many shortcomings of existing technologies. Summary of the Invention

[0006] In view of this, the present invention provides a method and system for identifying mature grape leaves based on a machine learning model, which realizes automatic and accurate identification and quantitative determination of the maturity status of grape leaves, extracts leaf shape feature indicators, and classifies and scores the leaf growth status and senescence level by combining machine learning models, thereby solving the problems of low efficiency and poor consistency of manual identification, as well as the insufficient accuracy and weak generalization ability of traditional single identification methods in field scenarios.

[0007] To achieve the above objectives, the present invention adopts the following technical solution: A method for identifying ripe grape leaves based on a machine learning model includes the following steps: S1 Acquisition Steps: Acquire raw images of grape leaves in the target area; S2 processing steps: Perform data preprocessing operations on the original image to obtain a standard grape leaf image; S3 Extraction Steps: Based on visual processing software, perform pixel-level contour segmentation on standard grape leaf images, extract multidimensional inherent shape features of grape leaves, and construct a leaf maturity feature set; S4 screening steps: Based on the variance expansion factor and mutual information criterion, redundant features are removed and important features are screened from the leaf maturity feature set. After normalization, the normalized shape features of grape leaves in the field are obtained. S5 discrimination step: Input the normalized shape features into the trained lightweight two-branch discrimination model to obtain the maturity discrimination results of grape leaves in the field; S6 Output Steps: Based on the maturity level of grape leaves, the maturity identification results are divided to obtain the leaf maturity grading results and corresponding confidence labels.

[0008] Optionally, the content of S1 in the above method is as follows: using a field mobile inspection device equipped with a high-definition industrial camera or a handheld plant protection image acquisition terminal, sampling is performed in the target area under natural lighting conditions, according to the preset shooting distance and preset shooting angle, to obtain the original image of the grape leaves in the field.

[0009] The above method, optionally, includes data preprocessing in S2, which specifically includes performing ROI region cropping, illumination normalization, pest and disease masking, and distortion region removal on the original image in sequence.

[0010] The above method, optionally, includes cropping the ROI region by taking the main leaf region in the original image of the grape leaves in the field as the core cropping target, removing background interference regions, and retaining the complete grape leaf region. Illumination normalization uses a combination of Gamma correction and histogram equalization algorithm to unify the image brightness and contrast of the original image under different illumination intensities. The specific steps of pest and disease masking and distortion removal include: identifying leaf holes, mold spots, yellowing spots and wrinkled distortion areas through image thresholding and edge detection algorithms, and generating a binary mask matrix to mask and fill abnormal areas.

[0011] The above method can optionally include the multidimensional inherent shape characteristics of grape leaves in S3, specifically including leaf perimeter, leaf area, leaf length-to-width ratio, leaf margin curvature, leaf base angle, leaf outline rectangularity, and leaf compactness.

[0012] Optionally, in the above method, S4, based on the variance inflation factor and mutual information criterion, the process of removing redundant features and selecting important features from the leaf maturity feature set specifically includes: removing multicollinear redundant features in the leaf maturity feature set that are greater than the preset VIF threshold according to the preset VIF threshold and the preset mutual information threshold, and retaining highly correlated features that are greater than the preset mutual information threshold, thereby obtaining a preset number of core leaf maturity features.

[0013] Optionally, the normalization process in S4 of the above method includes: using a minimum-maximum normalization algorithm to map the core maturity characteristics of the leaf to a preset numerical range.

[0014] Optionally, the lightweight dual-branch discrimination model in S5 includes a parallel feature parsing branch and a maturity association inference branch. Among them, the feature parsing branch integrates the KNN model and the KKNN model to explore the nonlinear coupling relationship between the normalized shape features of grape leaves in the field, and captures the subtle morphological differences of leaves by parsing the local spatial distribution features of the normalized shape features. The maturity correlation inference branch integrates the Logit model and the GBM model. The Logit model is used to realize the basic classification of leaf maturity, while the GBM model is used to iteratively optimize the feature-maturity mapping function based on the physiological evolution law of leaf shape during the grape growing season and according to the preset number of iterations, so as to correct the deviation of the basic classification of leaf maturity.

[0015] Optionally, the lightweight bi-branch discrimination model in S5 is iteratively trained using K-fold cross-validation based on manually labeled field grape leaf image samples. The data collection criteria for manually labeled field grape leaf image samples include grape variety, growth stage, lighting conditions, and senescence level.

[0016] A grape ripe leaf identification system based on a machine learning model is used to implement a grape ripe leaf identification method based on a machine learning model as described in any of the preceding claims, comprising a data acquisition module, a processing module, an extraction module, a filtering module, an identification module, and an output module connected in sequence. The acquisition module is used to acquire raw images of grape leaves in the target area. The processing module is used to perform data preprocessing operations on the original image to obtain a standard grape leaf image; The extraction module is used to perform pixel-level contour segmentation on standard grape leaf images based on visual processing software, extract multi-dimensional inherent shape features of grape leaves, and construct a leaf maturity feature set. The screening module is used to remove redundant features and screen important features from the leaf maturity feature set based on the variance expansion factor and mutual information criterion. After normalization, the normalized shape features of grape leaves in the field are obtained. The discrimination module is used to input normalized shape features into the trained lightweight two-branch discrimination model to obtain the maturity discrimination results of grape leaves in the field. The output module is used to classify the maturity identification results according to the maturity level of grape leaves, and obtain the leaf maturity grading results and corresponding confidence labels.

[0017] As can be seen from the above technical solution, compared with the prior art, the present invention provides a method and system for identifying ripe grape leaves based on a machine learning model, which has the following beneficial effects: (1) End-to-end automated closed loop: From field image acquisition, standardized preprocessing, feature extraction and screening to model reasoning and graded result output, the entire process is automated, completely replacing manual experience judgment, significantly reducing the labor cost of field cultivation and management, and avoiding the risk of human misjudgment; (2) High resistance to interference exclusive feature recognition: Through light normalization, pest and disease masking and distortion area removal, the image is accurately corrected and the inherent shape features of the leaves that are not affected by light, pigments and field noise are extracted, which can stably adapt to the complex and ever-changing planting scenarios in the field. (3) Bidirectional precise feature optimization: The dual criteria of variance expansion factor multicollinearity detection and mutual information nonlinear correlation judgment are integrated to filter redundant features in both directions, retaining only the core features that are strongly correlated with leaf lignification and photosynthetic maturity, taking into account both recognition accuracy and computational efficiency, and avoiding recognition deviation caused by feature redundancy; (4) Multi-model fusion robust inference: A dual-branch multi-model complementary architecture is adopted, relying on KNN and KKNN models to adapt to the local morphological differences of leaves, combined with Logit basic classification and GBM physiological law iterative correction, which effectively makes up for the defects of single model recognition and greatly improves the recognition stability and accuracy under complex leaf morphology. (5) Standardized, quantitative, and controllable grading: A unified standard for judging leaf maturity is constructed, and a confidence-based identification label is output. This achieves a unified standard for judging the maturity status of leaves throughout the garden, solves the problem of chaotic manual judgment standards, and provides precise data support for standardized topping, thinning, and refined cultivation of grapes. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0019] Figure 1 This is a flowchart illustrating a method for identifying mature grape leaves based on a machine learning model, as disclosed in this invention. Figure 2 This is a schematic diagram of the operation flow of a method for identifying mature grape leaves based on a machine learning model, as disclosed in an embodiment of the present invention. Figure 3 This is an example diagram of the morphological indicators of grape leaves disclosed in an embodiment of the present invention. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] In this application, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. The terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0022] See Figure 1 As shown, this invention discloses a method for identifying mature grape leaves based on a machine learning model, comprising the following steps: S1 Acquisition Steps: Acquire raw images of grape leaves in the target area; S2 processing steps: Perform data preprocessing operations on the original image to obtain a standard grape leaf image; S3 Extraction Steps: Based on visual processing software, perform pixel-level contour segmentation on standard grape leaf images, extract multidimensional inherent shape features of grape leaves, and construct a leaf maturity feature set; S4 screening steps: Based on the variance expansion factor and mutual information criterion, redundant features are removed and important features are screened from the leaf maturity feature set. After normalization, the normalized shape features of grape leaves in the field are obtained. S5 discrimination step: Input the normalized shape features into the trained lightweight two-branch discrimination model to obtain the maturity discrimination results of grape leaves in the field; S6 Output Steps: Based on the maturity level of grape leaves, the maturity identification results are divided to obtain the leaf maturity grading results and corresponding confidence labels.

[0023] Optionally, S1 specifically involves using a mobile field inspection device equipped with a high-definition industrial camera or a handheld plant protection image acquisition terminal to perform full-time, full-coverage sampling in the target area under natural lighting conditions, based on preset shooting distance and preset shooting angle, to obtain the original image of the grape leaves in the field.

[0024] refer to Figure 2As shown, this invention uses an active field-based image acquisition method to obtain leaf images. In the preparation work for identifying grape leaves in the target area, this invention uses a field-based mobile acquisition device equipped with a high-definition industrial camera, or other photographic devices such as handheld plant protection imaging equipment, to collect original images of grape leaves at different growth stages, under different lighting conditions, and in different growth states in a natural field setting. This achieves full coverage sampling of leaves from different shoots and canopy positions, covering sunny days, cloudy days, and low-light morning and evening hours, thus obtaining original images of grape leaves in multiple scenes and states.

[0025] Optionally, data preprocessing in S2 includes sequentially performing ROI region cropping, illumination normalization, pest and disease masking, and distortion region removal on the original image.

[0026] Specifically, ROI cropping involves using the main leaf area in the original image of grape leaves in the field as the core cropping target, removing background interference areas, and retaining the complete grape leaf area. Illumination normalization uses a combination of Gamma correction and histogram equalization algorithm to unify the image brightness and contrast of the original image under different illumination intensities. The specific steps of pest and disease masking and distortion removal include: identifying leaf holes, mold spots, yellowing spots and wrinkled distortion areas through image thresholding and edge detection algorithms, and generating a binary mask matrix to mask and fill abnormal areas.

[0027] Combination Figure 2 As shown, after acquiring the original images of grape leaves, the image data is transmitted to computers and other terminal devices for image processing, providing accurate and reliable standardized image data support for subsequent grape leaf identification. Although the acquisition of the original images standardized the shooting distance and angle, the sheer number of original grape leaf images and the common interference problems such as uneven lighting, background clutter, insect holes and mold spots on the leaves, and wrinkles and distortions in the complex field environment make them unsuitable for direct use in leaf feature extraction and maturity determination. Therefore, a complete set of standardized preprocessing operations, including ROI region cropping, lighting normalization, pest and disease masking, and distortion region removal, is used to optimize and correct the original images, remove invalid interference areas, and standardize image imaging standards to obtain clean and standardized grape leaf target images. This achieves a smooth transition from image acquisition to feature extraction, ensuring the accuracy and stability of subsequent leaf maturity identification work.

[0028] Optionally, the multidimensional inherent shape characteristics of grape leaves in S3 include leaf perimeter, leaf area, leaf length-to-width ratio, leaf margin curvature, leaf base angle, leaf outline rectangularity, and leaf compactness.

[0029] like Figure 3 As shown in the figure, some features of grape leaves were measured according to the illustration, resulting in the grape leaf morphological index table shown in Table 1 below. This invention uses grape leaf morphological indices as the core extraction basis. Based on the visual processing software ImageJ, pixel-level contour segmentation and refined morphological feature extraction are performed on standard grape leaf images. Referring to Table 1, the leaf morphological indices specifically include key geometric parameters such as leaf area, leaf length, leaf width, vein length, the angle between the petiole and the upper edge of the leaf, and the angle between the petiole and the lower edge of the leaf. Shape features are the geometric details of the region of interest in the grape leaf. They are a regional feature description method used for accurate expression and difference matching of leaf shape. Compared to image features such as color and texture, which are easily affected by environmental interference, leaf geometric shape features have extremely high stability and are not affected by the intensity of field light or changes in leaf pigmentation. They can truly and objectively reflect the physiological growth and ripening state of grape leaves, and are the core foundation for constructing a high-precision leaf maturity identification model.

[0030] Table 1. Morphological Indicators of Grape Leaves

[0031] The S3 extraction step uses ImageJ software to perform precise contour segmentation on standardized, interference-free grape leaf images, accurately quantifies and extracts various stable inherent geometric morphological features, discards easily distorted shallow image features, and retains only multidimensional inherent shape features strongly correlated with the physiological maturity characteristics of grape leaves. Based on multiple accurate and stable morphological indicators, a dedicated shape feature set adapted for grape leaf maturity recognition is integrated and constructed, providing high-precision and high-stability raw feature data support for subsequent feature selection and accurate maturity inference.

[0032] Optional, in the S4 filtering step

[0033] Based on the variance inflation factor and mutual information criterion, the process of eliminating redundant features and selecting important features in the leaf maturity feature set includes: removing multicollinear redundant features in the leaf maturity feature set that are greater than the preset VIF threshold and the preset mutual information threshold, and retaining highly correlated features that are greater than the preset mutual information threshold, so as to obtain the leaf core maturity features that meet the preset number. The normalization process specifically includes: using a min-max normalization algorithm to map the core maturity characteristics of the leaf to a preset numerical range.

[0034] In the current embodiment of the present invention, the variance inflation factor (VIF) is used to quantify the degree of multicollinearity among various shape features of grape leaves. Using all specific shape features as the calculation sample, the larger the VIF value of a single feature, the stronger the linear correlation between the feature and the other features. The mutual information criterion is used to quantify the nonlinear correlation between a single shape feature and the discrete label of leaf maturity. The higher the mutual information value (MI), the stronger the contribution of the feature to maturity classification. The two work together to achieve bidirectional and accurate filtering of redundant features.

[0035] Based on the above explanation, S4 specifically includes the following steps: S41 Redundancy Filtering Step: Traverse all features in the dedicated shape feature set, and calculate the inflation factor of each feature by substituting it into the VIF formula. Using VIF=10 as a preset threshold, remove highly collinear redundant features with VIF>10, eliminate linear overlap interference between features, and retain a subset of effective features with high independence. That is, calculate the inflation factor of each feature in the leaf maturity feature set according to the VIF formula, remove the corresponding features with inflation factors greater than the preset threshold, and obtain a subset of effective features. S42 Invalid Removal Step: Calculate the mutual information value between the filtered feature subset and the leaf maturity label dimension by dimension. Use MI=0.15 as the preset mutual information threshold. Features with less than the preset mutual information threshold are identified as low-association invalid features. After removal and screening, a high-quality feature set with strong correlation to leaf lignification and photosynthetic maturity is obtained. S43 sorting and selection steps: Sort the high-quality feature set from high to low according to the mutual information association weight, and select the top 7 features by weight as the core maturity features of the leaf according to the preset number. S44 Normalization Processing Steps: Perform min-max normalization operations on the 7 core leaf maturity features obtained in S43 dimension by dimension to unify the numerical range and dimensions of all features, eliminate the differences in numerical scale of different morphological features, and finally output normalized shape features that are adapted to the input of the dual-branch model in S5, effectively improving the model training convergence speed and recognition accuracy.

[0036] Optionally, the lightweight dual-branch discrimination model in S5 includes a parallel feature parsing branch and a maturity association inference branch; Among them, the feature parsing branch integrates the KNN model and the KKNN model to explore the nonlinear coupling relationship between the normalized shape features of grape leaves in the field, and captures the subtle morphological differences of leaves by parsing the local spatial distribution features of the normalized shape features. The maturity correlation inference branch integrates the Logit model and the GBM model. The Logit model is used to realize the basic classification of leaf maturity, while the GBM model is used to iteratively optimize the feature-maturity mapping function based on the physiological evolution law of leaf shape during the grape growing season and according to the preset number of iterations, so as to correct the deviation of the basic classification of leaf maturity.

[0037] In the dual-branch machine learning model, the feature parsing branch and the maturity association inference branch are operated in parallel and independently. This parallel branch serves as a result fusion and correction architecture, sharing the same normalized shape feature input with no sequential dependency. Feature computation and classification inference are performed synchronously, and the outputs of the two branches are ultimately fused to obtain a high-precision grape leaf maturity identification result. The specific computation process is as follows: (1) Parallel computation process of feature parsing branch. The feature parsing branch integrates the KNN model and the KKNN model. The core is used to mine the spatial distribution pattern and nonlinear coupling relationship of core shape features, and adapt to the local morphological differences of different grape leaves.

[0038] After the normalized shape features obtained in S4 are input into the dual-branch machine learning model, the feature vectors corresponding to the normalized shape features are simultaneously input into the KNN model and the KKNN model. The KNN model calculates the Euclidean distance between the feature vector to be identified and the feature vectors of the labeled samples in the training set, selects K nearest neighbor samples, and initially matches clusters of samples with similar leaf morphology to achieve global morphological feature comparison. The feature vectors of the labeled samples in the training set are the manually labeled dataset used in the training process of the dual-branch machine learning model, which will be explained in detail in subsequent embodiments.

[0039] The KKNN model, based on the KNN nearest neighbor selection, introduces a kernel function weighting mechanism to assign differentiated weights to nearest neighbor samples at different distances for each feature vector. This strengthens the feature contribution of closely similar samples and weakens the influence of distant interference samples, accurately capturing subtle morphological differences such as leaf margin curvature, leaf base angle, and local compactness. It outputs a leaf morphology feature coupling correlation matrix, completing the refined analysis and representation of complex leaf shape features, and providing fine-grained feature support for maturity classification.

[0040] (2) Parallel computation process of maturity correlation inference branch. The maturity correlation inference branch integrates the Logit model and the GBM model, which are closely aligned with the physiological evolution of grape leaves, and realize maturity grading and deviation correction.

[0041] After the normalized shape features are input, the maturity association inference branch and the feature parsing branch simultaneously receive the normalized core shape features. In the branch, the Logit model first performs a linear weighted operation on the input normalized core shape features based on the regression coefficients that have been trained. According to the preset four-level maturity classification threshold, it completes the basic probability classification of young leaves, transitional mature leaves, fully mature leaves, and senescent leaves, and outputs the initial maturity classification results and basic confidence scores.

[0042] Meanwhile, the GBM model uses the physiological evolution of leaf shape at different growth stages of grapes as an iterative constraint. Through multiple rounds of decision tree iteration, it fits the nonlinear mapping relationship between features and leaf lignification and photosynthetic maturity, corrects the morphological adaptation bias and growth stage adaptation bias in the basic classification of the Logit model round by round, dynamically optimizes the feature-maturity association function, and outputs the corrected maturity classification results and the corrected maturity inference results with corresponding confidence levels.

[0043] (3) Dual-branch result fusion decision process. The model finally performs weighted fusion of the refined morphological representation results of the feature analysis branch and the hierarchical correction results of the maturity association inference branch. Combining the operational confidence weights of the two models, the model comprehensively determines the leaf maturity level and outputs the final identification label with global confidence. This effectively avoids the problems of low accuracy and poor stability of a single model in identifying irregular leaves and transitional mature leaves, and greatly improves the robustness and accuracy of leaf maturity identification in complex field scenarios.

[0044] Furthermore, the lightweight bi-branch discrimination model in S5 is iteratively trained using K-fold cross-validation based on manually labeled field grape leaf image samples; the data collection criteria for manually labeled field grape leaf image samples include grape variety, growth stage, lighting conditions, and aging level.

[0045] First, field leaf image samples covering different grape varieties, different growth stages, different light conditions, and different aging levels were collected. After image preprocessing and feature filtering processes such as S1-S3, a standardized feature dataset was constructed, and four levels of leaf maturity were manually and accurately labeled.

[0046] Subsequently, the training set, validation set, and test set are divided according to a preset ratio. The dual-branch model is jointly iteratively trained using K-fold cross-validation. The standardized feature dataset with precise manual annotation in the divided training set is suitable for the calculation and recognition of the KNN model in the feature parsing branch.

[0047] The feature parsing branch learns the subtle differences in leaf morphology and the coupling rules of features through KNN and KKNN models, while the maturity association inference branch completes the basic classification fitting through the Logit model. It also relies on the GBM model combined with the physiological evolution characteristics of grape leaves to continuously iterate and optimize the mapping relationship between features and maturity and correct classification bias. The accuracy and generalization error are monitored in real time during the training process, and training is terminated in a timely manner to avoid overfitting and underfitting. Finally, the model weights that converge to the optimal value are saved to obtain a lightweight dual-branch discrimination model that can be adapted to complex field scenarios.

[0048] Optionally, in the S6 output step, the leaf maturity grading results include young leaves, transitional mature leaves, fully mature leaves, and senescent leaves.

[0049] Specifically, based on the trained lightweight bi-branch discrimination model, the maturity discrimination results of grape leaves in the field are obtained by identifying the normalized shape feature set. The maturity level of the leaves is divided into four levels: young leaves, transitional mature leaves, fully mature leaves, and senescent leaves. The discrimination labels with confidence are output to complete the automatic identification of mature grape leaves.

[0050] The leaf maturity grading results are used to guide refined pruning in grape fields. Based on the distribution ratio of young leaves, transitional mature leaves, fully mature leaves, and senescent leaves, the timing of new shoot pinching and the cycle of old leaf thinning are standardized throughout the vineyard. This effectively optimizes the ventilation and light penetration environment of the grape trellis, scientifically regulates the nutrient distribution ratio of the tree, balances the vegetative and reproductive growth of the plant, improves the quality of flower bud differentiation and light energy utilization efficiency, and fundamentally standardizes the cultivation and management standards of grape fields. This solves the technical pain points of traditional manual pruning, such as strong subjectivity, inconsistent standards, and delayed timing, effectively improving the standardization, refinement, and intelligence level of grape cultivation, and ensuring a steady increase in grape yield and fruit quality.

[0051] This invention also discloses a grape ripe leaf identification system based on a machine learning model, used to achieve, for example... Figure 1 The method for identifying mature grape leaves based on a machine learning model includes a data acquisition module, a processing module, an extraction module, a filtering module, an identification module, and an output module connected in sequence. The acquisition module is used to acquire raw images of grape leaves in the target area. The processing module is used to perform data preprocessing operations on the original image to obtain a standard grape leaf image; The extraction module is used to perform pixel-level contour segmentation on standard grape leaf images based on visual processing software, extract multi-dimensional inherent shape features of grape leaves, and construct a leaf maturity feature set. The screening module is used to remove redundant features and screen important features from the leaf maturity feature set based on the variance expansion factor and mutual information criterion. After normalization, the normalized shape features of grape leaves in the field are obtained. The discrimination module is used to input normalized shape features into the trained lightweight two-branch discrimination model to obtain the maturity discrimination results of grape leaves in the field. The output module is used to classify the maturity identification results according to the maturity level of grape leaves, and obtain the leaf maturity grading results and corresponding confidence labels.

[0052] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for identifying mature grape leaves based on a machine learning model, characterized in that, Includes the following steps: S1 Acquisition Steps: Acquire raw images of grape leaves in the target area; S2 processing steps: Perform data preprocessing operations on the original image to obtain a standard grape leaf image; S3 Extraction Steps: Based on visual processing software, perform pixel-level contour segmentation on standard grape leaf images, extract multidimensional inherent shape features of grape leaves, and construct a leaf maturity feature set; S4 screening steps: Based on the variance expansion factor and mutual information criterion, redundant features are removed and important features are screened from the leaf maturity feature set. After normalization, the normalized shape features of grape leaves in the field are obtained. S5 discrimination step: Input the normalized shape features into the trained lightweight two-branch discrimination model to obtain the maturity discrimination results of grape leaves in the field; S6 Output Steps: Based on the maturity level of grape leaves, the maturity identification results are divided to obtain the leaf maturity grading results and corresponding confidence labels.

2. The method for identifying ripe grape leaves based on a machine learning model according to claim 1, characterized in that, The specific content of S1 is as follows: using a field mobile inspection device equipped with a high-definition industrial camera or a handheld plant protection image acquisition terminal, sampling is carried out in the target area under natural lighting conditions, according to the preset shooting distance and preset shooting angle, to obtain the original image of grape leaves in the field.

3. The method for identifying mature grape leaves based on a machine learning model according to claim 1, characterized in that, Data preprocessing in S2 specifically includes performing ROI region cropping, illumination normalization, pest and disease masking, and distortion region removal on the original image in sequence.

4. The method for identifying ripe grape leaves based on a machine learning model according to claim 3, characterized in that, ROI cropping specifically involves using the main leaf area in the original image of grape leaves in the field as the core cropping target, removing background interference areas, and retaining the complete grape leaf area; Illumination normalization uses a combination of Gamma correction and histogram equalization algorithm to unify the image brightness and contrast of the original image under different illumination intensities. The specific steps of pest and disease masking and distortion removal include: identifying leaf holes, mold spots, yellowing spots and wrinkled distortion areas through image thresholding and edge detection algorithms, and generating a binary mask matrix to mask and fill the abnormal areas.

5. The method for identifying ripe grape leaves based on a machine learning model according to claim 1, characterized in that, The S3 grape leaf multidimensional inherent shape characteristics include leaf perimeter, leaf area, leaf length-to-width ratio, leaf margin curvature, leaf base angle, leaf outline rectangularity, and leaf compactness.

6. The method for identifying ripe grape leaves based on a machine learning model according to claim 1, characterized in that, In S4, based on the variance inflation factor and mutual information criterion, redundant feature removal and important feature selection of the leaf maturity feature set are specifically carried out as follows: according to the preset VIF threshold and the preset mutual information threshold, multicollinear redundant features in the leaf maturity feature set that are greater than the preset VIF threshold are removed, and highly correlated features that are greater than the preset mutual information threshold are retained to obtain the preset number of core leaf maturity features.

7. The method for identifying ripe grape leaves based on a machine learning model according to claim 6, characterized in that, The normalization process in S4 specifically includes: using a minimum-maximum normalization algorithm to map the core maturity characteristics of the leaf to a preset numerical range.

8. The method for identifying ripe grape leaves based on a machine learning model according to claim 1, characterized in that, The lightweight dual-branch discrimination model in S5 includes a parallel feature parsing branch and a maturity association inference branch; Among them, the feature parsing branch integrates the KNN model and the KKNN model to explore the nonlinear coupling relationship between the normalized shape features of grape leaves in the field, and captures the subtle morphological differences of leaves by parsing the local spatial distribution features of the normalized shape features. The maturity correlation inference branch integrates the Logit model and the GBM model. The Logit model is used to realize the basic classification of leaf maturity, while the GBM model is used to iteratively optimize the feature-maturity mapping function based on the physiological evolution law of leaf shape during the grape growing season and according to the preset number of iterations, so as to correct the deviation of the basic classification of leaf maturity.

9. The method for identifying ripe grape leaves based on a machine learning model according to claim 1, characterized in that, The lightweight bi-branch discrimination model in S5 is iteratively trained using K-fold cross-validation based on manually labeled field grape leaf image samples. The data collection criteria for manually labeled field grape leaf image samples include grape variety, growth stage, lighting conditions, and senescence level.

10. A grape ripe leaf identification system based on a machine learning model, characterized in that, A method for identifying mature grape leaves based on a machine learning model, as described in any one of claims 1-9, comprises a data acquisition module, a processing module, an extraction module, a filtering module, an identification module, and an output module connected in sequence. The acquisition module is used to acquire raw images of grape leaves in the target area. The processing module is used to perform data preprocessing operations on the original image to obtain a standard grape leaf image; The extraction module is used to perform pixel-level contour segmentation on standard grape leaf images based on visual processing software, extract multi-dimensional inherent shape features of grape leaves, and construct a leaf maturity feature set. The screening module is used to remove redundant features and screen important features from the leaf maturity feature set based on the variance expansion factor and mutual information criterion. After normalization, the normalized shape features of grape leaves in the field are obtained. The discrimination module is used to input normalized shape features into the trained lightweight two-branch discrimination model to obtain the maturity discrimination results of grape leaves in the field. The output module is used to classify the maturity identification results according to the maturity level of grape leaves, and obtain the leaf maturity grading results and corresponding confidence labels.

Images