Pest and disease identification method and system based on AI vision model

Abstract

The application provides a pest and disease identification method and system based on an AI vision model. First, the vision collection data of different growth parts of a plant to be identified at different collection times is obtained. Then, a pest and disease feature dynamic correlation chain containing the core vision feature units, correlation strength parameters and cross-stage feature evolution path information of multiple pests and diseases at different development stages is constructed. Then, the AI vision model is called to perform multi-dimensional feature interaction processing on the plant vision data set based on the correlation chain, generate feature interaction results and preliminary pest and disease identification results, and perform multi-round iteration optimization on the interaction matching parameters based on historical feedback data to obtain optimized parameters and corrected identification results. Finally, an identification report containing pest and disease type identification, development stage evaluation and feature matching instructions is generated and sent to the target terminal, thereby improving the accuracy, comprehensiveness and adaptability of pest and disease identification.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of computer vision, in particular to a pest and disease identification method and system based on an AI vision model. BACKGROUND

[0002] In the field of agriculture, accurate identification and timely prevention of plant pests and diseases are crucial for ensuring crop yield and quality. Traditional methods of plant pest and disease identification mainly rely on manual observation and experience-based judgment. These methods are not only inefficient, but also susceptible to the professional level and subjective factors of the observer, resulting in low identification accuracy. Especially in the early stages of pest and disease occurrence, symptoms are often not obvious, making manual identification more difficult and delaying the best prevention opportunity.

[0003] With the development of computer vision technology, some pest and disease identification methods based on image recognition have emerged. These methods usually involve collecting plant images, extracting features using image processing algorithms, and then comparing them with a pre-established pest and disease feature library to achieve identification. However, existing pest and disease identification methods based on image recognition have many limitations. On the one hand, they often only consider the image features of a plant at a certain moment or a certain part, ignoring the dynamic changes of pests and diseases in different growth parts and different development stages, resulting in inaccurate and incomplete identification results. On the other hand, these methods lack in-depth exploration and analysis of the relationship between features, making it difficult to effectively handle complex and variable pest and disease characteristics. When faced with new pests and diseases or similar pests and diseases, the recognition ability is significantly insufficient. In addition, existing methods lack a self-optimization and improvement mechanism in the identification process, making it difficult to continuously improve recognition accuracy based on feedback data from actual applications. SUMMARY

[0004] In view of the above-mentioned problems, in combination with the first aspect of the present application, the embodiments of the present application provide a pest and disease identification method based on an AI vision model, which comprises:

[0005] Obtaining a set of plant visual data to be identified, which contains visual collection data of different growth parts of the plant to be identified at different collection times;

[0006] Constructing a dynamic correlation chain of pest and disease characteristics, which contains core visual feature units of various pests and diseases at different development stages, correlation strength parameters between each core visual feature unit, and cross-stage feature evolution path information;

[0007] The AI vision model is called to perform multi-dimensional feature interaction processing on the plant visual data set based on the dynamic association chain of pest and disease characteristics, generate a feature interaction result and a preliminary pest and disease identification result corresponding to the to-be-identified plant visual data set, and the feature interaction result includes the interaction matching parameters of each growth part visual feature and the core visual feature unit in the association chain.

[0008] Based on the parameter adjustment rule in the historical pest and disease identification feedback data, the interaction matching parameters in the feature interaction result are processed by multiple rounds of iteration optimization to obtain optimized interaction matching parameters and a corrected pest and disease identification result.

[0009] Based on the corrected pest and disease identification result and the optimized interaction matching parameters, a pest and disease identification report containing pest and disease type identification, development stage evaluation and feature matching description is generated, and the pest and disease identification report is sent to the target terminal.

[0010] In another aspect, the embodiments of the present application also provide a pest and disease identification system based on an AI vision model, comprising:

[0011] A processor; a machine-readable storage medium for storing machine-executable instructions of the processor; wherein the processor is configured to execute the above-mentioned pest and disease identification method based on an AI vision model by executing the machine-executable instructions.

[0012] In another aspect, the embodiments of the present application also provide a computer program product, which comprises machine-executable instructions stored in a computer-readable storage medium, and a processor of a computer device reads the machine-executable instructions from the computer-readable storage medium, and the processor executes the machine-executable instructions, so that the computer device executes the above-mentioned pest and disease identification method based on an AI vision model.

[0013] Based on the above aspects, by acquiring visual collection data of different growth parts of the plant to be identified at different collection time points, and then constructing a disease and pest feature dynamic correlation chain, the core visual feature units, correlation strength parameters and cross-stage feature evolution path information of multiple diseases and pests at different development stages are integrated, which can deeply depict the dynamic change rule and internal correlation of the disease and pest features, and effectively solve the problem of ignoring the dynamic characteristics and feature correlation of the existing method. The AI vision model is called to perform multi-dimensional feature interaction processing on the plant visual data set based on the correlation chain, and the generated feature interaction result and preliminary disease and pest identification result fully consider the matching of the plant visual features and the core features of the disease and pest, which improves the accuracy and comprehensiveness of the identification. Based on the parameter adjustment rule in the historical disease and pest identification feedback data, the interaction matching parameters in the feature interaction result are iteratively optimized for multiple rounds, so that the model can continuously adjust and improve itself according to the feedback in the actual application, further improving the accuracy and adaptability of the identification. Finally, the disease and pest identification report containing the disease and pest type identification, development stage evaluation and feature matching description is generated, which provides detailed and accurate disease and pest information for agricultural practitioners, and helps them to take effective prevention and control measures in time, to ensure the healthy growth of crops and improve the efficiency of agricultural production. BRIEF DESCRIPTION OF DRAWINGS

[0014] Figure 1 is the execution flow diagram of the disease and pest identification method based on the AI vision model provided by an embodiment of the present application.

[0015] Figure 2 is a schematic diagram of exemplary hardware and software components of the disease and pest identification system based on the AI vision model provided by an embodiment of the present application. DETAILED DESCRIPTION

[0016] The present application will be described in detail below with reference to the accompanying drawings of the specification, Figure 1 is a flow diagram of the disease and pest identification method based on the AI vision model provided by an embodiment of the present application, and the disease and pest identification method based on the AI vision model will be described in detail below.

[0017] Step S110: acquiring a plant visual data set to be identified, the plant visual data set to be identified containing visual collection data of different growth parts of the plant to be identified at different collection time points.

[0018] In this embodiment, the identification of pests and diseases in tomato plants within a greenhouse is taken as an example. The plant to be identified is a tomato plant. When acquiring the visual data set of the plant to be identified, systematic visual data collection is required for different growth parts of the tomato plant. The growth parts of the tomato plant mainly include leaves (including new leaves, old leaves, and tender leaves), stems (including main stems and lateral branches), fruits (including green fruits, ripe fruits, and fruits changing color), and flowers. To comprehensively reflect the growth status of the tomato plant and the characteristics of possible pests and diseases, visual data collection needs to be conducted at different times, such as morning, noon, and evening each day, while also covering different weather conditions, such as sunny days, cloudy days, and overcast days, to avoid interference from lighting conditions on visual characteristics. The acquisition equipment uses an industrial camera with a fixed focal length and resolution. The camera is mounted on a movable track support, enabling it to capture images of different growth parts of the tomato plant from different angles (such as front, side, and top views). During the data collection process, it is essential to ensure that the image clarity of each growth part meets the requirements for subsequent feature extraction. Leaves must fully display their upper, lower, and edge surfaces; stems must be photographed at different heights to capture surface texture; and fruits must be photographed to capture their overall shape and surface details. For flowers, targeted data collection should be conducted during the flowering period.

[0019] Furthermore, when collecting privacy-sensitive data, such as the location information of greenhouses and the personal information of growers, privacy protection measures must be implemented. Specific technical measures include removing geolocation tags from image data, anonymizing metadata containing personal information, and using data encryption technology to encrypt data during transmission and storage to ensure that privacy is not leaked during data collection, transmission, and storage.

[0020] Step S120: Construct a dynamic association chain of pest and disease features, which includes core visual feature units of various pests and diseases at different development stages, association strength parameters between core visual feature units, and cross-stage feature evolution path information.

[0021] Step S121: Obtain full-cycle growth monitoring data for various pests and diseases. The full-cycle growth monitoring data includes continuous visual acquisition data of each pest and disease from the initial occurrence stage, the spread stage to the severe damage stage, and corresponding growth environment parameter records.

[0022] For example, common diseases and pests affecting tomatoes include early blight, late blight, gray mold, leaf mold, aphids, and whiteflies. For each of these diseases and pests, a dedicated monitoring database needs to be established to collect continuous visual data from the initial occurrence stage, the spread stage, to the severe damage stage. Taking early blight as an example, the initial stage typically manifests as pinhead-sized brown spots on the leaves. As the disease progresses into the spread stage, the spots gradually enlarge, becoming circular or irregular in shape with distinct concentric rings. In the severe damage stage, leaves turn yellow and fall off in large quantities, and brown hard spots appear on the fruit. Throughout the monitoring period, the continuity of visual data collection must be ensured, with collections at least three times per week until the disease or pest develops into a severe damage stage or is effectively controlled. Simultaneously, corresponding environmental parameters need to be recorded, including air temperature, relative humidity, soil temperature, soil moisture, light duration, light intensity, and CO2 concentration. These environmental parameters are collected in real time using sensors installed inside the greenhouse. The sensors must be evenly distributed to ensure the representativeness of the collected data. For the acquired full-cycle growth monitoring data, a data quality assessment is required to remove blurry, out-of-focus, overexposed, or underexposed image data, as well as abnormal environmental parameter records, in order to ensure the reliability of the data.

[0023] Step S122: Extract visual feature units for each development stage from the full-cycle growth monitoring data of each pest and disease to obtain a multi-stage feature unit set corresponding to each pest and disease. The visual feature units include color feature units, morphological feature units, texture feature units, and lesion area distribution feature units.

[0024] Step S1221: The full-cycle growth monitoring data of each pest and disease is segmented in chronological order. The segmentation interval is determined according to the growth rate and morphological change range of the pest and disease. The continuous monitoring data is divided into multiple time periods, and each time period corresponds to a development stage.

[0025] For example, taking early blight of tomatoes as an example, the segmentation interval is determined based on its growth rate and the magnitude of morphological changes. In the initial stage, because the development of the disease and pests is relatively slow and the morphological changes are small, the segmentation interval can be set to a longer interval, such as every three days. After entering the spread stage, the development rate of the disease and pests accelerates and the morphological changes increase, so the segmentation interval can be shortened to every two days. In the severe damage stage, the disease develops rapidly, and the segmentation interval is further shortened to one day. Through the above segmentation method, the continuous monitoring data of early blight of tomatoes is divided into multiple time periods, each corresponding to a clear development stage. During the segmentation process, the start and end times of each time period need to be accurately marked to ensure clear time boundaries between each stage. For different types of diseases and pests, the segmentation interval needs to be adjusted according to their respective growth characteristics. For example, the spread rate of gray mold of tomatoes may be faster than that of early blight of tomatoes, so its segmentation interval needs to be shortened accordingly.

[0026] Step S1222: Preprocess the visual acquisition data within each time period, extract color features from the preprocessed visual data of the lesion area, and use a color space conversion method to convert the visual data from RGB color space to HSV color space, extracting the hue mean, saturation variance and brightness range of the lesion area as color feature units.

[0027] For example, preprocessing steps include image denoising, image enhancement, and contrast adjustment. Image denoising uses Gaussian filtering to remove Gaussian noise from the image; image enhancement uses histogram equalization to improve image contrast, making lesion areas more prominent. After preprocessing, color features are extracted from the visual data of the lesion area. First, a color space conversion method is used to convert the image data from the RGB color space to the HSV color space. In the HSV color space, the hue (H) component more intuitively reflects the type of color, the saturation (S) component reflects the purity of the color, and the lightness (V) component reflects the brightness of the color. For the lesion area, the hue mean is extracted, which represents the overall color tendency of the lesion area; the saturation variance is extracted, which reflects the uniformity of color purity in the lesion area; and the lightness range is extracted, which represents the range of brightness variation in the lesion area. The hue mean, saturation variance, and lightness range are combined to form a color feature unit. For example, in the initial stage of early blight in tomatoes, the mean hue of the lesion area may be within a certain range, with a small saturation variance and a narrow brightness range; as the disease progresses, the mean hue will change, the saturation variance will increase, and the brightness range will widen.

[0028] Step S1223: Extract morphological features from the preprocessed visual data of the lesion area, obtain the contour information of the lesion area through the edge detection algorithm, and calculate the perimeter, area and roundness of the contour as morphological feature units.

[0029] In this embodiment, an edge detection algorithm, such as the Canny edge detection algorithm, is first used to detect the edges of the lesion area and obtain its contour information. During edge detection, a threshold parameter needs to be set appropriately to ensure accurate detection of the contour edges of the lesion area while avoiding misdetecting background noise as edges. After obtaining the contour information, the perimeter of the lesion area is calculated based on the contour, where the perimeter represents the total length of the lesion area's contour; the area enclosed by the contour is calculated, where the area reflects the size of the lesion area; and the roundness is calculated using the formula (4×π×area) / (perimeter×perimeter), which describes how close the lesion area's contour is to a circle. The perimeter, area, and roundness are used as morphological feature units. For example, the initial lesion area of ​​early blight in tomatoes may have a relatively regular contour, high roundness, and small perimeter and area; as it progresses into the spread stage, the contour gradually becomes irregular, the roundness decreases, and the perimeter and area increase.

[0030] Step S1224: Extract texture features from the preprocessed visual data of the lesion area, and use the gray-level co-occurrence matrix method to calculate the energy, entropy, and contrast of the lesion area as texture feature units.

[0031] For example, first, the color image is converted to a grayscale image, and then the gray-level co-occurrence matrix (GLCM) method is used to calculate texture features. The GLCM is a matrix constructed by statistically analyzing the probability of occurrence of different grayscale pixel pairs in a specific direction and distance. Based on this matrix, the energy, entropy, and contrast of the lesion region are calculated. Energy reflects the uniformity and regularity of the image texture; the higher the energy value, the more uniform and regular the texture. Entropy reflects the complexity of the image texture; the higher the entropy value, the more complex the texture. Contrast reflects the clarity and depth of the grooves in the image texture; the higher the contrast value, the deeper the grooves. These three factors—energy, entropy, and contrast—are used as texture feature units. For example, the texture of the lesion region of tomato leaf mold may be more complex, with a higher entropy value and higher contrast; while the texture of the lesion region of tomato powdery mildew may be more uniform, with a higher energy value and lower entropy value.

[0032] Step S1225: Extract the distribution features of the lesion area from the preprocessed visual data of the lesion area, and mark the location coordinates, distribution density, and distance of the lesion area in the overall visual data of the plant as the distribution feature unit of the lesion area.

[0033] For example, firstly, the location coordinates of each lesion area are marked in the overall visual data of the plant. The coordinate system is based on the top left corner of the image as the origin, with the horizontal direction as the X-axis and the vertical direction as the Y-axis. The location of each lesion area is represented by the coordinates of the top left and bottom right corners of its bounding rectangle. Then, the distribution density of the lesion areas is calculated. Distribution density is measured by the number of lesion areas per unit area, i.e., the ratio of the total number of lesion areas in the image to the total area of ​​the image. Finally, the distance between adjacent lesion areas is calculated. For each lesion area in the image, the nearest other lesion area is found, the Euclidean distance between them is calculated, and the mean and variance of all distances between adjacent lesion areas are calculated. The combination of the location coordinates, distribution density, and distances between adjacent lesion areas is used as the lesion area distribution feature unit. For example, lesions from tomato aphid disease may be relatively scattered with a low distribution density and large distances between adjacent lesion areas; while lesions from tomato late blight may show a clustered distribution with a high distribution density and small distances between adjacent lesion areas.

[0034] Step S1226: Integrate the color feature units, morphological feature units, texture feature units and lesion area distribution feature units extracted in each time period to form the stage feature unit group corresponding to that time period; arrange the stage feature unit groups corresponding to all time periods of each disease and pest in chronological order to form a multi-stage feature unit set corresponding to each disease and pest.

[0035] In this embodiment, each feature unit is represented as a feature vector. For example, a color feature unit can be represented as a three-dimensional vector containing the mean hue, saturation variance, and brightness range; a morphological feature unit can be represented as a three-dimensional vector containing perimeter, area, and roundness; a texture feature unit can be represented as a three-dimensional vector containing energy, entropy, and contrast; and a lesion area distribution feature unit can be represented as a multi-dimensional vector containing location coordinates (multiple coordinate points), distribution density, and adjacent distances (mean and variance). These feature vectors are concatenated sequentially to form a stage feature unit group corresponding to the given time period. The stage feature unit group contains comprehensive visual feature information of the lesion area within that time period. For each pest or disease, all its corresponding stage feature unit groups are arranged chronologically, starting from the first time period of the initial occurrence stage and ending at the last time period of the severe damage stage, forming a multi-stage feature unit set corresponding to that pest or disease. The multi-stage feature unit set completely records the changes in visual features throughout the entire process of the pest or disease's development from occurrence to maturity.

[0036] Step S123: Perform core feature screening on the multi-stage feature unit set of each disease and pest, and retain the visual feature unit that appears more frequently in any development stage of the disease and pest than in other stages and can uniquely identify that stage as the core visual feature unit, forming a subset of core visual feature units for each stage of each disease and pest.

[0037] For example, taking the multi-stage characteristic unit set of tomato early blight as an example, for each development stage (initial occurrence stage, spread stage, severe damage stage), the frequency of occurrence of each visual characteristic unit in the stage characteristic unit group is statistically analyzed and compared with the frequency of occurrence in other stages. For example, in the initial occurrence stage, if a certain hue mean range occurs more frequently than in the spread and severe damage stages, and this hue mean range only appears in the initial occurrence stage, uniquely identifying this stage, then the color characteristic unit corresponding to this hue mean range is retained as the core visual characteristic unit of the initial occurrence stage. For morphological characteristic units, if a certain roundness range occurs most frequently in the spread stage and is significantly different from the roundness range in other stages, uniquely identifying the spread stage, then it is taken as the core morphological characteristic unit of the spread stage. Similarly, a similar screening is performed on texture characteristic units and lesion area distribution characteristic units. Through the above screening process, core visual characteristic units are determined for each development stage of each disease and pest, and these core visual characteristic units together constitute the core visual characteristic unit subset of that stage. There should be clear differences between the subsets of core visual feature units at each stage to ensure that different development stages can be accurately distinguished through the core visual feature units.

[0038] Step S124: Analyze the correlation between core visual feature units in adjacent development stages of each disease and pest, and calculate the co-occurrence frequency, morphological similarity, and evolutionary dependence of core visual feature units in adjacent stages respectively; normalize the co-occurrence frequency, morphological similarity, and evolutionary dependence to the same numerical range, and determine the correlation strength parameter between each core visual feature unit based on the weighted calculation results of the normalized three.

[0039] When analyzing the correlation between core visual feature units in adjacent developmental stages of each pest or disease, the core visual feature units of two adjacent developmental stages, such as the initial occurrence stage and the spread stage, are extracted first. The co-occurrence frequency is calculated, i.e., the frequency with which core visual feature units in two adjacent stages appear simultaneously; a higher co-occurrence frequency indicates a stronger correlation between the two. Morphological similarity is calculated by comparing various feature parameters of the core visual feature units, such as the mean hue of color feature units and the area of ​​morphological feature units, using methods such as cosine similarity or Euclidean distance. Higher morphological similarity indicates a stronger morphological correlation between the two. Evolutionary dependency is calculated, i.e., the degree to which the appearance of a later-stage core visual feature unit depends on the earlier-stage core visual feature unit. This can be determined by analyzing the causal relationship between the two over time; a higher evolutionary dependency indicates that the earlier-stage core visual feature unit is an important foundation for the evolution of the later-stage core visual feature unit.

[0040] After calculating co-occurrence frequency, morphological similarity, and evolutionary dependency, all three need to be normalized to the same numerical range, for example, to between 0 and 1. The normalization process uses the min-max normalization method: for a given parameter, its actual value is subtracted from its minimum value across all samples, and then divided by the difference between its maximum and minimum values ​​to obtain the normalized parameter value. After normalization, different weight coefficients are assigned to each parameter, determined based on their importance to the association relationship. For example, the weight coefficient for co-occurrence frequency can be set higher, followed by morphological similarity, and then evolutionary dependency. A weighted calculation is performed based on the normalized parameter values ​​and their corresponding weight coefficients. The weighted results are then summed to obtain the association strength parameter between each core visual feature unit. The association strength parameter ranges from 0 to 1; a higher value indicates a stronger association between core visual feature units at adjacent stages.

[0041] Step S125: Track the changes in the core visual feature units of each disease and pest from the initial occurrence stage to the severe damage stage, record the order of appearance, morphological transition mode and duration of each core visual feature unit, and form cross-stage feature evolution path information for each disease and pest.

[0042] Step S1251: Assign time labels to each development stage in the full-cycle growth monitoring data of each pest and disease, marking the start and end times of each development stage.

[0043] For example, taking tomato late blight as an example, its full-cycle growth monitoring data includes the initial occurrence stage, the spread stage, and the severe damage stage. A unique time label is assigned to each stage, such as using a stage number combined with the start and end times, like "Stage 1 - Start Time T1 - End Time T2", "Stage 2 - Start Time T2 - End Time T3", etc.

[0044] Step S1252: Extract the occurrence time of each core visual feature unit from the subset of core visual feature units of each stage of each disease and pest, wherein the occurrence time refers to the moment when the core visual feature unit first appears in the visual acquisition data.

[0045] For example, for each core visual feature unit, within the time period corresponding to its development stage, find the image frame in the visual acquisition data where the core visual feature unit first appears. The timestamp of this image frame is the appearance time of the core visual feature unit. For example, a core morphological feature unit in the tomato late blight spread stage first appears in an image frame within the second time period of that stage; the timestamp of this image frame is the appearance time of the core visual feature unit. The extraction of the appearance time needs to be accurate to the second to ensure the accuracy of the time series. For core visual feature units that span multiple time periods, only the time of their first appearance is recorded.

[0046] Step S1253: Sort all core visual feature units of each disease and pest according to the order of their appearance to form a temporal sequence of core visual feature units.

[0047] For example, the occurrence times of all core visual feature units can be compared and arranged in ascending order to form a chronological sequence of core visual feature units. In this sequence, each core visual feature unit is arranged according to its chronological order of appearance in the development of the pest or disease; for example, core visual feature units in the initial stage are listed first, followed by those in the spread stage, and finally those in the severe damage stage. For multiple core visual feature units appearing within the same developmental stage, they are arranged in their corresponding positions within the chronological sequence according to their occurrence time. This chronological sequence of core visual feature units clearly demonstrates the order in which core visual features appear during the development of pests and diseases, reflecting the evolutionary process of the pests and diseases.

[0048] Step S1254: Analyze the morphological differences between two adjacent core visual feature units in the time sequence of core visual feature units, and determine the morphological transition mode between them. The morphological transition mode includes gradual transition, abrupt transition and superposition transition. Gradual transition refers to the gradual change of morphological parameters, abrupt transition refers to the sudden change of morphological parameters, and superposition transition refers to the subsequent core visual feature unit adding morphological parameters on the basis of the previous core visual feature unit.

[0049] For example, consider two adjacent core visual feature units A and B, where A is the preceding core visual feature unit and B is the following core visual feature unit. Compare their morphological parameters, such as perimeter, area, and roundness. If the morphological parameters of B show a gradual change relative to those of A, meaning that the parameters change continuously and smoothly over time without any abrupt changes, then the morphological transition between the two is determined to be a gradual transition. If the morphological parameters of B change significantly and abruptly relative to those of A within a short period of time, showing a large difference from those of A, then it is determined to be an abrupt transition. If the morphological parameters of B include all the morphological parameters of A, and new morphological parameters are added on top of those of A, such as adding special contour shape features, then it is determined to be an overlay transition. Through the above analysis, the morphological transition mode is determined for each pair of adjacent core visual feature units in the time series and recorded in the transition mode list.

[0050] Step S1255: Calculate the time span from the appearance time to the disappearance time of each core visual feature unit, and take the time span as the duration of the core visual feature unit. If the core visual feature unit continues until the end of the severe hazard stage, then the duration of the core visual feature unit is the time span from the appearance time to the end of the severe hazard stage.

[0051] In this embodiment, the disappearance time refers to the moment when the core visual feature unit last appeared in the visual acquisition data. By traversing the visual acquisition data of the development stage to which the core visual feature unit belongs and subsequent stages, the image frame in which the core visual feature unit no longer appears is found, and the moment before the image frame is the disappearance time. If the core visual feature unit continues to appear until the end of the severe damage stage, then its disappearance time is the end time of the severe damage stage. The duration is obtained by subtracting the appearance time from the disappearance time, which is the time span experienced by the core visual feature unit from its first appearance to its final disappearance. The duration reflects the existence time of the core visual feature unit in the development process of the pest and disease.

[0052] Step S1256: Integrate the temporal sequence of core visual feature units, the morphological transition mode of each adjacent core visual feature unit, and the duration of each core visual feature unit to form basic information of cross-stage feature evolution path; draw a cross-stage feature evolution path diagram for each disease and pest based on the basic information of cross-stage feature evolution path, and mark the position, occurrence time, duration, and morphological transition mode of each core visual feature unit; supplement the cross-stage feature evolution path diagram with the changing trend of the correlation strength parameter corresponding to each core visual feature unit, and mark the curve of the correlation strength parameter changing with time.

[0053] In this embodiment, basic information is stored in the form of a data structure, including a time-series linked list, a transition mode array, and a duration dictionary. Based on the above basic information, a cross-stage feature evolution path diagram is drawn. The path diagram uses time as the horizontal axis and the core visual feature units as the vertical axis. The position of each core visual feature unit on the horizontal axis corresponds to its occurrence time and duration, forming a time interval. In the path diagram, different graphics or colors are used to identify the type of core visual feature units, such as circles for color feature units and squares for morphological feature units. Adjacent core visual feature units are connected by directed line segments. The type or color of the line segment indicates the morphological transition mode, such as solid lines for gradual transitions, dashed lines for abrupt transitions, and dotted lines for superposition transitions. The correlation strength parameter value is labeled next to the line segment. In addition, the correlation strength parameter change trend is added to each core visual feature unit in the cross-stage feature evolution path diagram. By selecting multiple time points on the time axis, the association strength parameters between the core visual feature unit and its adjacent core visual feature units at different time points are calculated. The above parameter values ​​are connected to form a curve showing the change of association strength parameters over time. The curve can intuitively show the dynamic changes of association strength.

[0054] Step S1257: Integrate the cross-stage feature evolution path diagram and the corresponding textual description information to form cross-stage feature evolution path information for each disease and pest.

[0055] In this embodiment, the textual description information includes detailed descriptions of each core visual feature unit in the cross-stage feature evolution path diagram, such as the type of feature unit, specific feature parameters, occurrence time, and duration; the specific manifestations of each morphological transition mode, such as the change process of morphological parameters in gradual transition and the specific change magnitude in abrupt transition; and the analysis of the trend of association strength parameters, such as when the association strength increases, when it decreases, and the possible reasons. Integrating the cross-stage feature evolution path diagram with the above textual description information forms complete cross-stage feature evolution path information for each pest and disease. The cross-stage feature evolution path information comprehensively records the evolution process of core visual feature units of pests and diseases from the initial occurrence stage to the severe damage stage.

[0056] Step S126: Integrate the core visual feature unit subsets of each stage of each disease and pest, the corresponding association strength parameters, and the cross-stage feature evolution path information to form a feature association sub-chain for a single disease and pest.

[0057] For example, taking tomato powdery mildew as an example, its feature association subchain includes a subset of core visual feature units in the initial occurrence stage, a subset of core visual feature units in the spread stage, and a subset of core visual feature units in the severe damage stage. Each stage's core visual feature unit subset is connected to the core visual feature unit subsets of adjacent stages through an association strength parameter, which reflects the degree of association between core visual feature units at different stages. Cross-stage feature evolution path information describes the evolutionary sequence, transition mode, and duration of each core visual feature unit throughout the entire pest and disease development process. Through the above integration, the various parts of information are organically organized to form a complete and structured feature association subchain for a single pest and disease. This subchain clearly demonstrates the core characteristics of the pest and disease at different development stages and their interrelationships, serving as a fundamental component in constructing a dynamic feature association chain for pests and diseases.

[0058] Step S127: Collect all disease and pest feature association subchains for each single disease and pest, classify and sort them according to disease and pest type, associate and mark similar core visual feature units that exist between different disease and pest types, and supplement feature association reference information across disease and pest types.

[0059] Collect feature association sub-chains for all individual pests and diseases, such as feature association sub-chains for early blight, late blight, and gray mold in tomatoes. Classify and sort these sub-chains according to pest and disease type, such as fungal diseases, bacterial diseases, viral diseases, and insect pests. Each category contains corresponding feature association sub-chains. Based on the classification and sorting, associate and mark similar core visual feature units between different pest and disease types. For example, early blight and late blight in tomatoes may have similar color feature units at a certain developmental stage; their similarity is determined by comparing parameters such as the mean hue and variance of saturation. For similar core visual feature units, add association markers to each feature association sub-chain, indicating the core visual feature units of other similar pests and diseases and their respective pest and disease types. Simultaneously, supplement cross-pest and disease type feature association reference information, including similarity values ​​of similar feature units, analysis of possible causes of confusion, and key distinguishing points. This reference information helps avoid misjudgments due to feature similarity in subsequent pest and disease identification processes.

[0060] Step S128: Based on the feature association reference information across pest and disease types, adjust the association strength parameters of the core visual feature units in each pest and disease feature association sub-chain to ensure that the feature association logic between different pest and disease feature association sub-chains remains consistent.

[0061] In this embodiment, for different pests and diseases with similar core visual feature units, the differences in their association strength parameters are analyzed. If the association strength parameters of similar feature units differ significantly in different sub-chains, it may lead to inconsistencies in feature association logic. In this case, the association strength parameters need to be adjusted based on the similarity values ​​and distinguishing points in the reference information. For example, for core visual feature units with high similarity but belonging to different pests and diseases, their association strength parameters with adjacent feature units can be appropriately reduced to highlight their differences from other pests and diseases. During the adjustment process, it is necessary to ensure that the logic of the association strength parameters within each pest and disease feature association sub-chain is self-consistent, while maintaining consistency in the feature association logic between different sub-chains to avoid contradictory or conflicting association relationships. The adjusted association strength parameters need to be recalculated and updated in the feature association sub-chains.

[0062] Step S129: Integrate all the adjusted disease and pest feature association sub-chains to construct a dynamic disease and pest feature association chain containing multiple disease and pest feature association information, and store it in the feature association database of the AI ​​vision model.

[0063] In this embodiment, a unified indexing mechanism needs to be established during the integration process to uniquely identify the core visual feature units of all pests and diseases, ensuring accurate association between feature units in different sub-chains. The dynamic association chain of pest and disease features is stored in a mesh structure, where each node represents a core visual feature unit, the connections between nodes represent the association relationship, and the weight of the connection is the association strength parameter. The chain also includes cross-stage feature evolution path information and cross-pest and disease type feature association reference information. After construction, the dynamic association chain of pest and disease features is stored in the feature association database of the AI ​​vision model. The feature association database adopts a distributed storage architecture to ensure data security and access efficiency. Simultaneously, an update mechanism is established for the database. When new pest and disease monitoring data is collected or new pest and disease types are discovered, the dynamic association chain of pest and disease features can be updated promptly to ensure the timeliness and accuracy of the association chain.

[0064] Step S130: Call the AI ​​vision model to perform multi-dimensional feature interaction processing on the plant visual data set based on the dynamic association chain of pest and disease features, and generate feature interaction results and preliminary pest and disease identification results corresponding to the visual data set of the plant to be identified. The feature interaction results include the interaction matching parameters of the visual features of each growth part and the core visual feature units in the association chain.

[0065] Step S131: Input the visual data set of the plant to be identified into the feature analysis module of the AI ​​visual model, and process the visual data collected by growth part into partitions to obtain the visual data of the leaf area, the visual data of the stem area and the visual data of the fruit area of ​​the plant to be identified.

[0066] In this embodiment, taking the visual data set of tomato plants as an example, the primary task of the feature parsing module is to partition the visually acquired data according to the growth location. Image segmentation algorithms, such as semantic segmentation algorithms based on deep learning, are used to segment the input visually acquired data. The semantic segmentation algorithm, through a trained neural network model, classifies each pixel in the image, identifying leaf regions, stem regions, fruit regions, and background regions. For leaf regions, it is necessary to further distinguish between different types of leaves, such as new leaves, old leaves, and tender leaves; for stem regions, it is necessary to distinguish between the main stem and lateral branches; for fruit regions, it is necessary to distinguish between green fruits, ripe fruits, and fruits that are changing color. After segmentation, the regions of each growth location are extracted from the original image, resulting in visual data for leaf regions, stem regions, and fruit regions. The visual data for each region includes the image information of that region and its corresponding location coordinates. The location coordinates are used for subsequent analysis of the distribution characteristics of diseased areas.

[0067] Step S132: Perform feature extraction on the visual data of each region to obtain the visual feature unit set of the leaf region, the visual feature unit set of the stem region, and the visual feature unit set of the fruit region. Each region's feature unit set includes color feature units, morphological feature units, and texture feature units.

[0068] In this embodiment, the feature extraction process is similar to the method described in step S122, but here it targets the visual data of each region of the plant to be identified, rather than historical monitoring data of pests and diseases. Taking the visual data of the leaf region as an example, preprocessing is first performed, including noise reduction, enhancement, and contrast adjustment, and then color feature units, morphological feature units, and texture feature units are extracted. The color feature units are obtained by converting the RGB color space to the HSV color space and extracting the hue mean, saturation variance, and brightness range; the morphological feature units are obtained by obtaining contour information through edge detection and calculating the perimeter, area, and roundness; the texture feature units are obtained by calculating energy, entropy, and contrast through the gray-level co-occurrence matrix. The feature extraction process for the stem region and fruit region is the same as that for the leaf region, obtaining their respective color feature units, morphological feature units, and texture feature units. All feature units extracted from each region are integrated to form a visual feature unit set for the leaf region, a visual feature unit set for the stem region, and a visual feature unit set for the fruit region. The feature unit set for each region contains comprehensive visual feature information for that region.

[0069] Step S133: Retrieve the dynamic association chain of pest and disease features from the feature association database of the AI ​​visual model, and perform preliminary matching between the visual feature unit set of the leaf area, the visual feature unit set of the stem area, and the visual feature unit set of the fruit area and the core visual feature unit of each stage in the association chain.

[0070] In this embodiment, the dynamic association chain of pest and disease features includes core visual feature units of various pests and diseases at different developmental stages, association strength parameters, and cross-stage feature evolution path information. Preliminary matching is performed between the visual feature unit sets of leaf regions, stem regions, and fruit regions and the core visual feature units at each stage in the association chain. Preliminary matching uses a feature vector similarity calculation method, comparing each feature unit in the visual feature unit set of the region to be identified with the feature vectors of all core visual feature units in the association chain to calculate a similarity value. Similarity values ​​can be calculated using methods such as cosine similarity and Euclidean distance. A preliminary matching threshold is set; when the similarity value between a feature unit in the region to be identified and a core visual feature unit in the association chain is higher than this threshold, the two are considered to have preliminarily matched successfully. Through preliminary matching, core visual feature units that may be related to the visual features of the plant to be identified are screened out.

[0071] Step S134: Based on the preliminary matching results, select the core visual feature unit groups whose matching degree with the visual feature unit set of each region exceeds the preset matching degree threshold, determine the pest and disease type and development stage corresponding to each group of core visual feature units, and form a candidate combination list.

[0072] In this embodiment, the matching degree is obtained by calculating the average similarity values ​​of all initially matched feature units. A preset matching degree threshold is set, and core visual feature unit groups with matching degrees exceeding this threshold are selected. A core visual feature unit group refers to a combination of multiple core visual feature units belonging to the same pest or disease and the same developmental stage. The pest or disease type and developmental stage corresponding to each group of core visual feature units are determined; for example, a certain core visual feature unit group belongs to the spread stage of early blight in tomatoes. The above core visual feature unit groups and their corresponding pest or disease types and developmental stages are integrated to form a candidate combination list. Each candidate combination in the candidate combination list represents a possible pest or disease and its developmental stage. When forming the candidate combination list, it is necessary to ensure that the candidate combinations in the list have a certain degree of diversity to avoid omitting potential pest or disease types.

[0073] Step S135: Call the multi-dimensional feature interaction module of the AI ​​vision model, and construct an interaction association model between visual feature units in each region based on the pest and disease information in the candidate combination list. The interaction association model is used to analyze the collaborative matching relationship between visual feature units in different regions.

[0074] Step S1351: Select a candidate combination from the candidate combination list and extract the pest and disease type information and development stage information corresponding to the candidate combination; based on the pest and disease type information and development stage information, retrieve the core visual feature units of the leaves, stems, and fruits of the pest and disease at the corresponding stage from the dynamic association chain of pest and disease features.

[0075] In this embodiment, a candidate combination is first selected from the candidate combination list, for example, a candidate combination for the spread stage of early blight in tomatoes. The pest / disease type information (early blight in tomatoes) and development stage information (spread stage) corresponding to this candidate combination are extracted. Based on this information, the core visual feature units of the leaves, stems, and fruits at the corresponding stages are retrieved from the dynamic association chain of pest / disease features. These core visual feature units are typical features of each growth part of the pest / disease during the spread stage, such as circular concentric lesions on leaves, brown sunken lesions on stems, and brown hard spots on fruits. During the retrieval process, the completeness and accuracy of the core visual feature units must be ensured, including various feature units such as color, shape, and texture.

[0076] Step S1352: Determine the association weight between the core visual feature unit of the leaf and the core visual feature unit of the stem, the association weight between the core visual feature unit of the leaf and the core visual feature unit of the fruit, and the association weight between the core visual feature unit of the stem and the core visual feature unit of the fruit. The association weight is set based on the cross-regional disease transmission pattern of the pest at the corresponding stage.

[0077] In this embodiment, for example, during the spread of early blight in tomatoes, the disease typically begins on the leaves and then gradually spreads to the stems and fruits. Therefore, the association weight between the core visual feature units of leaves and stems can be set to a higher value, the association weight between the core visual feature units of leaves and fruits to a lower value, and the association weight between the core visual feature units of stems and fruits to a lower value. The specific values ​​of the association weights are determined by analyzing the historical disease incidence data and cross-regional transmission paths of the pest, and can be optimized using expert experience or machine learning methods. The value of the association weights ranges from 0 to 1, and the sum of the weights can be 1 or not 1, depending on the importance of the transmission pattern.

[0078] Step S1353: Construct the input layer of the interactive association model, taking the visual feature unit sets of the leaf region, stem region, and fruit region of the plant to be identified as input vectors respectively; construct the hidden layer of the interactive association model, setting up three feature processing sub-modules, corresponding to the leaf region, stem region, and fruit region respectively. Each feature processing sub-module performs feature mapping processing on the input region feature unit set to obtain the region feature mapping vector.

[0079] In this embodiment, each feature processing submodule consists of multiple fully connected layers and activation functions. The number of neurons in the fully connected layers is determined based on the dimension of the input vector and the feature mapping requirements. The feature processing submodule performs feature mapping on the input set of regional feature units, mapping the original feature vector to a higher-dimensional feature space through nonlinear transformation, resulting in a regional feature mapping vector. This regional feature mapping vector can more effectively represent the essential features of regional visual feature units, improving the accuracy of subsequent interactive matching. The activation function uses the ReLU function or other nonlinear activation functions to increase the model's nonlinear expressive power.

[0080] Step S1354: Add a cross-regional interaction module in the hidden layer. Based on the preset association weights, perform weighted fusion processing on the regional feature mapping vectors of the three regions to calculate the feature synergy values ​​of the leaf region and the stem region, the feature synergy values ​​of the leaf region and the fruit region, and the feature synergy values ​​of the stem region and the fruit region.

[0081] In this embodiment, the cross-regional interaction module performs weighted fusion processing on the regional feature mapping vectors of the leaf region, stem region, and fruit region based on preset association weights. For example, when calculating the feature synergy value between the leaf region and the stem region, the feature mapping vectors of the leaf region and the stem region are weighted and summed according to their association weights to obtain a fusion vector, which is then output as the feature synergy value through a fully connected layer and an activation function. Similarly, the feature synergy values ​​between the leaf region and the fruit region, and between the stem region and the fruit region, are calculated. The feature synergy value reflects the degree of synergistic matching between the visual features of the two regions; the higher the value, the more the features of the two regions match the cross-regional lesion characteristics of the pest at the corresponding stage.

[0082] Step S1355: Construct the output layer of the interaction association model, using the regional feature matching degree and cross-regional feature collaboration value as output parameters. The regional feature matching degree refers to the matching degree between the regional feature mapping vector of each region and the corresponding core visual feature unit, and the cross-regional feature collaboration value refers to the average of the three cross-regional collaboration values. Based on historical interaction case data, adjust the hidden layer parameters and cross-regional association weights of the interaction association model so that the regional feature matching degree and cross-regional feature collaboration value output by the model can accurately reflect the collaborative matching relationship between visual feature units in different regions and core visual feature units.

[0083] In this embodiment, the intra-regional feature matching degree is obtained by calculating the similarity between the regional feature mapping vector of each region and the feature vector of the corresponding core visual feature unit. The similarity calculation method is the same as that used in the initial matching. The cross-regional feature collaboration value is the average of the three feature collaboration values ​​of the leaf region and the stem region, the leaf region and the fruit region, and the stem region and the fruit region. Based on historical interaction case data, the interaction association model is trained and adjusted. The historical interaction case data contains a large amount of visual data of the plants to be identified with known pest and disease types and development stages, as well as the true values ​​of the corresponding intra-regional feature matching degree and cross-regional feature collaboration value. Through the backpropagation algorithm, the weights and biases of the model's hidden layer, as well as the cross-regional association weights, are adjusted to minimize the error between the intra-regional feature matching degree and the cross-regional feature collaboration value output by the model and the true value. Cross-validation is used during training to ensure the model's generalization ability.

[0084] Step S1356: Construct a corresponding interaction association model for each candidate combination in the candidate combination list according to the above steps, forming a multi-model parallel processing framework.

[0085] In this embodiment, each candidate combination corresponds to a type and stage of pest or disease development, thus requiring the construction of an independent interaction and association model. All models form a multi-model parallel processing framework, which can simultaneously process the features of multiple candidate combinations, improving processing efficiency. The multi-model parallel processing framework employs a distributed computing architecture, distributing different models to different computing nodes for execution. Nodes communicate and share data via a network. The framework also includes model scheduling and load balancing mechanisms, rationally allocating computational tasks based on the computational complexity of each model and the computing resources of each node, ensuring load balancing across nodes and preventing any node from becoming overloaded.

[0086] Step S136: Calculate the interaction matching parameters between the visual feature units of each region and the corresponding core visual feature unit group through the interaction association model. The interaction matching parameters include the feature matching degree within the region, the feature synergy degree across regions, and the feature fit degree at each stage. Integrate the interaction matching parameters of each region to form the feature interaction result corresponding to the visual data set of the plant to be identified.

[0087] In this embodiment, the interaction matching parameters include intra-regional feature matching degree, cross-regional feature synergy degree, and stage feature fit degree. The intra-regional feature matching degree is the average of the intra-regional feature matching degrees for the leaf region, stem region, and fruit region. The cross-regional feature synergy degree is the average of the three cross-regional feature synergy values. The stage feature fit degree is calculated by comparing the set of visual feature units of the region to be identified with the cross-stage feature evolution path information of the corresponding development stage of the pest or disease. The calculation of the stage feature fit degree considers evolutionary information such as the order of appearance of feature units and morphological transition methods. The interaction matching parameters corresponding to each candidate combination are integrated, including intra-regional feature matching degree, cross-regional feature synergy degree, and stage feature fit degree, to form the feature interaction result corresponding to the visual data set of the plant to be identified. The feature interaction result is stored in the form of a data matrix, with each row representing a candidate combination and each column representing an interaction matching parameter.

[0088] Step S137: Based on the distribution of interaction matching parameters in the feature interaction results, normalize the feature matching degree within the region, the cross-regional feature synergy degree, and the stage feature fit degree to the same numerical range; select the candidate combination with the highest normalized feature matching degree within the region, the best normalized cross-regional feature synergy degree, and the best normalized stage feature fit degree from the candidate combination list, and determine it as the disease and pest type and development stage of the plant to be identified; generate preliminary disease and pest identification results containing the determined disease and pest type identifier, development stage identifier, and corresponding interaction matching parameters.

[0089] Step S1371: Extract the interaction matching parameters from the feature interaction results. For each candidate combination in the candidate combination list, organize the corresponding regional feature matching degree, cross-regional feature synergy degree, and stage feature fit degree. The regional feature matching degree refers to the average matching degree of the leaf region, stem region, and fruit region. The cross-regional feature synergy degree refers to the average of the three cross-regional synergy values. The stage feature fit degree refers to the overall fit degree with the core features of the corresponding stage in the association chain.

[0090] In this embodiment, for each candidate combination, the regional feature matching degree is calculated, which is the arithmetic mean of the regional feature matching degrees of the leaf region, stem region, and fruit region; the cross-regional feature synergy degree is the arithmetic mean of the three cross-regional feature synergy values; the stage feature fit degree refers to the overall fit between the set of visual feature units of the plant to be identified and the core features of the corresponding stage in the association chain, which is calculated by comprehensively considering factors such as the similarity of feature units, the order of appearance, and morphological transition. The above parameter values ​​are associated and stored with the pest and disease type and development stage information of the candidate combination.

[0091] Step S1372: Normalize the intra-regional feature matching degree, cross-regional feature synergy degree, and stage feature fit degree of each candidate combination to the same numerical range; set weight coefficients for the normalized intra-regional feature matching degree, normalized cross-regional feature synergy degree, and normalized stage feature fit degree, respectively. The weight coefficients are set according to the priority of pest and disease identification, with the weight coefficient of the normalized intra-regional feature matching degree being the highest; calculate the weighted average of the three normalized parameters of each candidate combination based on the weight coefficients to obtain the comprehensive matching score of each candidate combination.

[0092] In this embodiment, the normalization method uses min-max normalization. For a given parameter, the minimum and maximum values ​​of that parameter across all candidate combinations are used as a benchmark, and the parameter value for each candidate combination is transformed to between 0 and 1. After normalization, weight coefficients are assigned to the three parameters. These weight coefficients are set according to the priority of pest and disease identification. Generally, the regional feature matching degree has the greatest impact on the identification result, therefore its weight coefficient is the highest, followed by the cross-regional feature synergy degree, and the stage feature fit degree has the lowest weight coefficient. The specific values ​​of the weight coefficients can be determined through expert experience or machine learning methods, and the sum of the weights is 1. Based on the weight coefficients, a weighted calculation is performed on the three normalized parameters of each candidate combination. The normalized regional feature matching degree is multiplied by its weight coefficient, plus the normalized cross-regional feature synergy degree multiplied by its weight coefficient, and plus the normalized stage feature fit degree multiplied by its weight coefficient, to obtain the comprehensive matching score for each candidate combination. The comprehensive matching score reflects the overall degree of matching between the candidate combination and the visual features of the plant to be identified.

[0093] Step S1373: Sort the candidate combinations in the candidate combination list in descending order of comprehensive matching score to obtain a sorted candidate combination list; select the top K candidate combinations with the highest comprehensive matching score in the sorted candidate combination list as target candidate combinations.

[0094] In this embodiment, the candidate combination with the highest comprehensive matching score is ranked first in the sorted list, and so on. The top K candidate combinations with the highest comprehensive matching scores in the sorted candidate combination list are selected as target candidate combinations. The value of K can be set according to the actual situation, for example, K=3 or K=5, selecting the top 3 or top 5 candidate combinations. The target candidate combinations are the objects of further analysis and screening, which can improve the accuracy and reliability of the identification results.

[0095] Step S1374: Analyze the common disease incidence of the disease and pest types corresponding to the target candidate combinations on the plant varieties to be identified. If the difference between the disease and pest types corresponding to any target candidate combination on the plant variety and the disease and pest types corresponding to the other two target candidate combinations on the plant variety exceeds the preset frequency difference range, then the target candidate combination with the higher disease incidence rate will be given priority.

[0096] For example, if the plant to be identified is tomato, and the target candidate combinations correspond to tomato early blight, tomato late blight, and tomato gray mold, respectively, the incidence rates of these three diseases on tomato varieties are obtained by querying a tomato disease and pest database. Incidence frequency refers to the ratio of the historical incidence of a disease or pest on a particular plant variety to the total incidence of diseases and pests. The incidence frequencies of each target candidate combination are compared. If the difference between the incidence frequency of any target candidate combination and the incidence frequencies of the other two target candidate combinations exceeds a preset frequency difference range (e.g., the difference is greater than a preset percentage), then the target candidate combination with the higher incidence frequency is given priority. Diseases and pests with higher incidence frequencies are more likely to occur on the plant variety and therefore have higher priority.

[0097] Step S1375: If the difference in the incidence frequency of the disease and pest types corresponding to the target candidate combination on the plant variety is within the preset frequency difference range, then further analyze the stage feature fit of each target candidate combination and select the target candidate combination with the highest stage feature fit.

[0098] If the frequency differences of the disease and pest types corresponding to the target candidate combinations on the plant variety are all within the preset frequency difference range, meaning the frequency of each disease and pest type is relatively close, then the stage feature fit of each target candidate combination is further analyzed. Stage feature fit reflects the overall degree of fit between the visual characteristics of the plant to be identified and the core characteristics of the corresponding developmental stage of the disease and pest. The target candidate combination with the highest stage feature fit is selected, as the developmental stage of the disease and pest in this combination best matches the actual disease stage of the plant to be identified.

[0099] Step S1376: If the difference in stage feature fit of each target candidate combination is within the preset fit difference range, analyze the cross-regional feature synergy of each target candidate combination and select the target candidate combination with the highest cross-regional feature synergy; determine the pest and disease type and development stage corresponding to the finally selected target candidate combination as the pest and disease type and development stage of the plant to be identified.

[0100] If the difference in stage feature fit among the target candidate combinations is within a preset range, meaning the stage feature fit is also relatively close, then the cross-regional feature synergy of each target candidate combination is analyzed. Cross-regional feature synergy reflects the degree of synergistic matching between visual features of different growth sites. Higher synergy indicates that the disease characteristics of each region are more consistent with the cross-regional transmission pattern of the pest. The target candidate combination with the highest cross-regional feature synergy is selected. Through the above-mentioned layer-by-layer screening, the pest type and development stage corresponding to the finally selected target candidate combination are the pest type and development stage of the plant to be identified. Preliminary pest identification results are generated, including the identified pest type identifier, development stage identifier, and corresponding interactive matching parameters, such as regional feature matching degree, cross-regional feature synergy, and stage feature fit.

[0101] Step S140: Based on the parameter adjustment patterns in the historical pest and disease identification feedback data, perform multiple rounds of iterative optimization on the interaction matching parameters in the feature interaction results to obtain the optimized interaction matching parameters and the corrected pest and disease identification results.

[0102] Step S141: Obtain historical pest and disease identification feedback data, which includes historical visual data set of plants to be identified, historical feature interaction results, historical preliminary pest and disease identification results, and historical corrected pest and disease identification results.

[0103] In this embodiment, historical pest and disease identification feedback data is first acquired, derived from past pest and disease identification cases. This historical feedback data includes: a set of historical visual data of the plants to be identified, i.e., visual data collected from various growth parts of the plants in the past; historical feature interaction results, i.e., the interaction matching parameters in the feature interaction results at that time; historical preliminary pest and disease identification results, i.e., the preliminary identification results obtained through the AI ​​visual model at that time; and historical corrected pest and disease identification results, i.e., the correct identification results after manual review or actual verification. The collection of historical data needs to cover multiple pest and disease types, different developmental stages, and different plant varieties to ensure data diversity and representativeness.

[0104] Step S142: Extract the historical interaction matching parameters corresponding to the historical feature interaction results and the target interaction matching parameters corresponding to the historical corrected pest and disease identification results from the historical pest and disease identification feedback data; calculate the parameter difference between the historical interaction matching parameters and the target interaction matching parameters in each historical case, and classify the parameter difference according to the feature attributes of the historical visual data set of plants to be identified, wherein the feature attributes include plant variety attributes, collection season attributes and lesion area attributes.

[0105] In this embodiment, historical interaction matching parameters are parameters such as regional feature matching degree, cross-regional feature synergy, and stage feature fit in historical feature interaction results; target interaction matching parameters are the correct interaction matching parameters corresponding to the historically corrected pest and disease identification results. These parameters are accurate parameter values ​​verified in historical cases. The parameter difference between the historical interaction matching parameters and the target interaction matching parameters in each historical case is calculated, i.e., the target interaction matching parameters minus the historical interaction matching parameters. The parameter differences are classified according to the feature attributes of the historical visual data set of plants to be identified. Feature attributes include plant variety attributes, such as tomato, apple, rice, etc.; collection season attributes, such as spring, summer, autumn, winter, etc.; and lesion area attributes, such as leaf area, stem area, fruit area, etc. The parameter differences are classified according to different combinations of plant variety, collection season, and lesion area. For example, "tomato-summer-leaf area" is a feature attribute combination category, and the parameter differences of all historical cases under this category are collected.

[0106] Step S143: Normalize the parameter difference under each type of feature attribute, analyze the changing trend of the parameter difference under each type of feature attribute, determine the parameter adjustment direction and parameter adjustment range corresponding to different feature attribute combinations, and form a mapping relationship table.

[0107] Step S1431: Divide the feature attributes of the historical visual data set of plants to be identified into multiple variety categories according to plant variety attributes, multiple season categories according to collection season attributes, and multiple lesion area categories according to lesion area attributes; for each combination of variety category, season category, and lesion area category, filter out all historical cases under that combination to form a feature attribute combination case set.

[0108] In this embodiment, plant variety attributes can be divided into multiple variety categories, such as tomato, apple, and rice; collection season attributes can be divided into multiple season categories, such as spring, summer, autumn, and winter; and lesion area attributes can be divided into multiple lesion area categories, such as leaf area, stem area, and fruit area. For each combination of variety category, season category, and lesion area category, all historical cases under that combination are selected to form a feature attribute combination case set. For example, a combination of variety category (tomato), season category (summer), and lesion area category (leaf area) will have a case set containing all historical cases of tomato leaf lesions collected in summer.

[0109] Step S1432: Calculate the mean difference between historical interaction matching parameters and target interaction matching parameters in the case set of each feature attribute combination. The mean difference includes the mean difference of feature matching degree within the region, the mean difference of feature synergy degree across regions, and the mean difference of feature fit at a given stage. Analyze the positive or negative nature of the mean difference. If the mean difference of feature matching degree within the region is positive, the parameter adjustment direction for feature matching degree within the region under this feature attribute combination is determined to be decreasing; if it is negative, the parameter adjustment direction for feature matching degree within the region under this feature attribute combination is determined to be increasing. Determine the parameter adjustment direction for feature synergy degree across regions and feature fit at a given stage using the same logic.

[0110] In this embodiment, for the mean difference in feature matching degree within a region, if it is positive, it indicates that the historical interaction matching parameters are generally lower than the target interaction matching parameters, so the parameter adjustment direction is to increase them; if it is negative, it indicates that the historical interaction matching parameters are generally higher than the target interaction matching parameters, so the adjustment direction is to decrease them. The logic for determining the parameter adjustment direction for cross-regional feature synergy and stage feature fit is the same as that for feature matching degree within a region. By analyzing the sign of the mean difference in parameters, the adjustment direction is determined for each interaction matching parameter under each feature attribute combination.

[0111] Step S1433: Calculate the standard deviation of the parameter difference in the case set for each feature attribute combination. Determine the parameter adjustment level based on the standard deviation. The larger the standard deviation, the higher the parameter adjustment level, and the higher the parameter adjustment level, the larger the corresponding parameter adjustment magnitude. Assign each feature attribute combination the corresponding regional feature matching parameter adjustment direction, cross-regional feature synergy parameter adjustment direction, stage feature fit parameter adjustment direction, and the corresponding parameter adjustment magnitude.

[0112] In this embodiment, the standard deviation reflects the dispersion of parameter differences. Based on the size of the standard deviation, the parameter adjustment range is divided into multiple levels, such as low, medium, and high. A larger standard deviation indicates greater fluctuation in the parameter differences, requiring a larger adjustment range for correction; therefore, a higher parameter adjustment range level corresponds to a larger parameter adjustment range. The specific value of the parameter adjustment range is determined based on the adjustment range level and the mean of the parameter differences. For example, a low-level adjustment range can be set to a small fixed value, a medium-level to a medium fixed value, and a high-level to a large fixed value, or it can be calculated by multiplying the standard deviation by the mean of the differences. For each feature attribute combination, corresponding parameter adjustment directions and ranges are assigned for regional feature matching degree, cross-regional feature synergy degree, and stage feature fit degree, forming a mapping table entry containing feature attribute combinations, adjustment directions, and adjustment ranges.

[0113] Step S1434: Organize the characteristic attribute combinations and their corresponding parameter adjustment directions and magnitudes into a table to form a mapping relationship table. The characteristic attribute combinations are composed of variety category, season category, and lesion area category. Supplement and improve the data in the mapping relationship table. If the number of historical cases corresponding to any characteristic attribute combination is lower than the preset case number threshold, refer to the parameter adjustment direction and magnitude of similar characteristic attribute combinations and make fine adjustments based on the particularity of the characteristic attribute combination to make the mapping relationship table cover all possible characteristic attribute combinations.

[0114] In this embodiment, the mapping table uses variety category, season category, and lesion area category as composite primary keys. Each primary key corresponds to one record, including fields such as the adjustment direction of feature matching degree within the region, the adjustment magnitude of feature matching degree within the region, the adjustment direction of feature synergy across regions, the adjustment magnitude of feature synergy across regions, the adjustment direction of feature fit at a stage, and the adjustment magnitude of feature fit at a stage. The data in the mapping table is supplemented and improved. For feature attribute combinations with a historical case count lower than a preset case count threshold, the mean and standard deviation of parameter differences cannot be accurately calculated due to insufficient data. In this case, the parameter adjustment direction and magnitude are referenced to similar feature attribute combinations. Similar feature attribute combinations refer to combinations with similar variety categories, the same season category, or the same lesion area category. Considering the specific characteristics of this feature attribute combination, such as the growth characteristics of the variety in a specific season or the special structure of the lesion area, the reference adjustment direction and magnitude are fine-tuned to ensure that the mapping table can cover all possible feature attribute combinations and avoid data loss.

[0115] Step S144: Extract the plant variety attributes, collection season attributes, and lesion area attributes from the visual data set of the plants to be identified, and find the corresponding parameter adjustment direction and parameter adjustment range in the mapping relationship table; based on the found parameter adjustment direction and parameter adjustment range, make the first adjustment to the feature matching degree within the region, the feature synergy across regions, and the feature fit of the stage in the feature interaction results, and obtain the interaction matching parameters after the first adjustment.

[0116] The feature attributes of the visual data set of the plants to be identified are extracted, including plant variety attributes (e.g., tomato), collection season attributes (e.g., summer), and lesion area attributes (e.g., leaf area). In the mapping table, the corresponding parameter adjustment direction and magnitude are found based on the combination of these three attributes. For example, if the feature matching degree within the region corresponding to the "tomato-summer-leaf area" combination is found to be increased, the adjustment magnitude is a certain value; the cross-regional feature synergy adjustment direction is increased, the adjustment magnitude is another value; and the stage feature fit adjustment direction is decreased, the adjustment magnitude is yet another value. Based on the found adjustment direction and magnitude, the interaction matching parameters in the feature interaction results are initially adjusted. The adjustment method is as follows: if the adjustment direction is increased, the current interaction matching parameter is added to the adjustment magnitude; if the adjustment direction is decreased, the current interaction matching parameter is subtracted from the adjustment magnitude. The value of the adjustment magnitude is determined according to the mapping table to ensure that the adjusted parameters are closer to the target interaction matching parameters. After the initial adjustment, the adjusted interaction matching parameters are obtained.

[0117] Step S145: Based on the first adjusted interaction matching parameters, re-screen candidate combinations, determine the pest and disease type and development stage of the plant to be identified, and obtain the pest and disease identification results after the first adjustment; extract the historical corrected pest and disease identification results of the same type and stage as the first adjusted pest and disease identification results from the historical feedback data, and obtain the corresponding target interaction matching parameter range.

[0118] In this embodiment, the screening process is similar to step S137, including steps such as calculating the comprehensive matching score, sorting, and selecting target candidate combinations to obtain the first adjusted pest and disease identification result, i.e., the new pest and disease type and development stage. Historically corrected pest and disease identification results of the same type and stage as the first adjusted result are extracted from historical feedback data; these are historical cases with the same pest and disease type and development stage. Corresponding target interaction matching parameters are obtained from these historical cases, and the value range of these parameters is statistically analyzed. For example, the minimum and maximum values ​​of the target interaction matching parameters are calculated to determine the target interaction matching parameter range. The target interaction matching parameter range reflects the reasonable value range of the interaction matching parameters for this pest and disease type at this development stage.

[0119] Step S146: Determine whether the interaction matching parameters after the first adjustment are within the range of the target interaction matching parameters. If they are not within the range of the target interaction matching parameters, further adjust the parameter adjustment range according to the mapping relationship table. Based on the adjusted parameter adjustment range, perform a second adjustment on the interaction matching parameters after the first adjustment to obtain the second adjusted interaction matching parameters and the second adjusted pest and disease identification results.

[0120] In this embodiment, for each interaction matching parameter, such as the feature matching degree within the region, its value is checked to see if it falls between the minimum and maximum values ​​of the target interaction matching parameter range. If all parameters are within the range, the adjustment ends; if at least one parameter is outside the range, further adjustment is required. Based on the mapping table, the adjustment magnitude correction coefficient for the parameter outside the range under that feature attribute combination is found. The adjustment magnitude correction coefficient is determined based on the degree to which the parameter difference exceeds the range; the greater the exceedance, the larger the correction coefficient. The original parameter adjustment magnitude is adjusted based on the adjustment magnitude correction coefficient, for example, by multiplying the original adjustment magnitude by the correction coefficient to obtain a new adjustment magnitude. The new adjustment magnitude is used to perform a second adjustment on the interaction matching parameters after the first adjustment, with the adjustment direction remaining unchanged, resulting in the second adjusted interaction matching parameters. Based on the second adjusted interaction matching parameters, candidate combinations are re-screened to obtain the second adjusted pest and disease identification results.

[0121] Step S147: Repeat the above steps until the adjusted interaction matching parameters are within the range of the target interaction matching parameters. Use the interaction matching parameters at this time as the optimized interaction matching parameters, and use the corresponding pest and disease identification results as the corrected pest and disease identification results.

[0122] For example, step S1471: record the current number of adjustments and the corresponding interactive matching parameters, pest and disease type-stage combination; normalize the regional feature matching degree, cross-regional feature synergy degree, and stage feature fit degree in the currently adjusted interactive matching parameters; determine whether the normalized regional feature matching degree, normalized cross-regional feature synergy degree, and normalized stage feature fit degree are all within the range of the normalized target interactive matching parameters, wherein the range of the normalized target interactive matching parameters refers to the average value of the corresponding normalized parameters in the historically corrected pest and disease identification results plus or minus the preset fluctuation range.

[0123] In this embodiment, the adjusted interaction matching parameters are normalized using the same method as in step S1372, normalizing them to a range of 0 to 1. Simultaneously, the target interaction matching parameter range is also normalized to obtain a normalized target interaction matching parameter range. This normalized range is obtained by calculating the average value of the corresponding normalized parameters in the historically corrected pest and disease identification results, and then adding or subtracting a preset fluctuation range. The preset fluctuation range is determined based on the dispersion of the historical data. It is then determined whether the feature matching degree within the normalized region, the cross-regional feature synergy, and the stage feature fit are all within the normalized target interaction matching parameter range.

[0124] Step S1472: If the feature matching degree within the region, the cross-regional feature synergy degree, and the stage feature fit degree are all within the range of the target interaction matching parameters, then stop adjusting, mark the current interaction matching parameters as the optimized interaction matching parameters, and mark the corresponding pest and disease identification results as the corrected pest and disease identification results.

[0125] If all normalized interaction matching parameters fall within the range of the normalized target interaction matching parameters, it indicates that the current interaction matching parameters are relatively accurate and can reflect the true pest and disease situation of the plant to be identified. At this point, the adjustment process is stopped, and the current interaction matching parameters are marked as optimized interaction matching parameters, while the corresponding pest and disease identification results are marked as corrected pest and disease identification results. The corrected pest and disease identification results have higher accuracy and reliability compared to the initial identification results.

[0126] Step S1473: If at least one parameter among the regional feature matching degree, cross-regional feature synergy degree, and stage feature fit degree is not within the target interaction matching parameter range, then analyze the parameter type that is not within the target interaction matching parameter range; if the parameter type that is not within the target interaction matching parameter range is the regional feature matching degree, then find the adjustment magnitude correction coefficient of the regional feature matching degree under the feature attribute combination in the mapping relationship table; adjust the parameter adjustment magnitude of the regional feature matching degree based on the adjustment magnitude correction coefficient, and then readjust the current regional feature matching degree.

[0127] In this embodiment, for parameter types that are not within the range, the corresponding adjustment magnitude correction coefficient for that feature attribute combination is looked up in the mapping table. The adjustment magnitude correction coefficient is determined based on the adjustment experience of that parameter type in historical cases and is used to further optimize the adjustment magnitude. For example, if the feature matching degree within the region is not within the range and the current value is lower than the lower limit of the range, the adjustment magnitude needs to be increased by multiplying the original adjustment magnitude by a correction coefficient greater than 1; if the current value is higher than the upper limit of the range, the adjustment magnitude needs to be decreased by multiplying the original adjustment magnitude by a correction coefficient less than 1. Based on the adjusted adjustment magnitude, the feature matching degree within the current region is adjusted again.

[0128] Step S1474: If the parameter type not within the target interaction matching parameter range is cross-regional feature synergy, then find the adjustment magnitude correction coefficient of the cross-regional feature synergy under the feature attribute combination in the mapping relationship table, and adjust the parameter adjustment magnitude of the cross-regional feature synergy based on the adjustment magnitude correction coefficient; if the parameter type not within the target interaction matching parameter range is stage feature fit, then find the adjustment magnitude correction coefficient of the stage feature fit under the feature attribute combination in the mapping relationship table, and adjust the parameter adjustment magnitude of the stage feature fit based on the adjustment magnitude correction coefficient.

[0129] For parameters not within the target range, such as cross-regional feature synergy or stage feature fit, the processing method is similar to that for intra-regional feature matching. The corresponding adjustment magnitude correction coefficient is found in the mapping table. Based on the direction and magnitude of the deviation between the current parameter value and the target range, the adjustment magnitude of the cross-regional feature synergy or stage feature fit parameter is adjusted. For example, if the current value of the cross-regional feature synergy is lower than the lower limit of the target range, and the adjustment magnitude correction coefficient is greater than 1, the original adjustment magnitude is multiplied by this coefficient, and the cross-regional feature synergy is adjusted again to increase it.

[0130] Step S1475: Based on the readjusted interaction matching parameters, re-screen candidate combinations to determine new pest and disease identification results; record the new number of adjustments, new interaction matching parameters, and new pest and disease identification results, and return to the step to determine whether the new interaction matching parameters are within the target interaction matching parameter range.

[0131] In this embodiment, the new number of adjustments, new interaction matching parameters, and new pest and disease identification results can be recorded. Then, the process returns to step S1471 to determine whether the new interaction matching parameters are within the target interaction matching parameter range. The above adjustment, filtering, and judgment steps are repeated until all interaction matching parameters are within the target range or the preset maximum number of adjustments is reached.

[0132] Step S1476: Continue executing the above steps until all interactive matching parameters are within the target interactive matching parameter range, or the number of adjustments reaches the preset maximum number of times; if there are still parameters not within the target interactive matching parameter range after the preset maximum number of adjustments, then normalize the adjusted regional feature matching degree, cross-regional feature synergy degree, and stage feature fit degree respectively; calculate the absolute difference between each normalized parameter and the median value of the normalized target interactive matching parameter range; select the interactive matching parameter with the smallest sum of all absolute differences as the optimized interactive matching parameter, and use the corresponding pest and disease identification result as the corrected pest and disease identification result.

[0133] The adjustment process continues. If, after reaching the preset maximum number of adjustments, any parameter still falls outside the target interaction matching parameter range, the adjustment is stopped, and a compromise solution is adopted. The adjusted interaction matching parameters are normalized, and the absolute difference between each normalized parameter and the median of the normalized target interaction matching parameter range is calculated. The median of the target interaction matching parameter range is the arithmetic mean of the minimum and maximum values. The absolute differences of each parameter are summed to obtain the total absolute difference. The interaction matching parameter set with the smallest total absolute difference is selected as the optimized interaction matching parameter set, and the corresponding pest and disease identification result is used as the corrected pest and disease identification result. In cases where the target range cannot be fully met, this method selects the parameter combination with the smallest deviation from the median of the target range to maximize the accuracy of the identification results.

[0134] Step S150: Based on the corrected pest and disease identification results and optimized interactive matching parameters, generate a pest and disease identification report containing pest and disease type identifiers, development stage assessments and feature matching descriptions, and send the pest and disease identification report to the target terminal.

[0135] In this embodiment, the pest and disease identification report includes several parts: pest and disease type identification, clearly indicating the type of pest or disease affecting the plant to be identified, such as early blight of tomato; development stage assessment, determining the current development stage of the pest or disease, such as the spreading stage; feature matching description, detailing the optimized interactive matching parameters, including specific values ​​for regional feature matching degree, cross-regional feature synergy, and stage feature fit, as well as a comparison of these parameters with the target interactive matching parameter range, explaining the matching details of visual features of each growth part with core visual feature units, such as the matching of color feature units in the leaf area with the core color feature units of early blight of tomato during the spreading stage, etc. The report may also include pest and disease control recommendations, providing targeted control methods and measures based on the pest and disease type and development stage. The generated pest and disease identification report is stored in text or PDF format and sent to the target terminal, such as the grower's mobile phone, computer, or agricultural management platform, via a network communication module. Encrypted transmission is used during the transmission process to ensure the security of the report content. After receiving the report, the target terminal can display or print it for relevant personnel to view and refer to, so as to take timely and effective pest and disease control measures.

[0136] Based on the same inventive concept, please refer to Figure 2 The diagram shows a schematic block diagram of an AI vision model-based pest identification system 100 provided in this application embodiment for executing the above-described AI vision model-based pest identification method. The AI ​​vision model-based pest identification system 100 may include a communication unit 110, a machine-readable storage medium 120, and a processor 130.

[0137] In this embodiment, both the machine-readable storage medium 120 and the processor 130 are located within the AI ​​vision model-based pest and disease identification system 100 and are separately configured. However, it should be understood that the machine-readable storage medium 120 may also be independent of the AI ​​vision model-based pest and disease identification system 100 and may be accessed by the processor 130 via a bus interface. Alternatively, the machine-readable storage medium 120 may be integrated into the processor 130 and may communicate and interact with external systems through the communication unit 110.

[0138] The processor 130 is the control center of the AI ​​vision model-based pest and disease identification system 100. It connects to various parts of the system via various interfaces and lines. By running or executing software programs and / or modules stored in the machine-readable storage medium 120, and by calling data stored in the machine-readable storage medium 120, it performs various functions and processes data of the AI ​​vision model-based pest and disease identification system 100, thereby providing overall monitoring of the system. Optionally, the processor 130 may include one or more processing cores; for example, the processor 130 may integrate an application processor and a modem processor, wherein the application processor mainly handles the operating system, user interface, and applications, and the modem processor mainly handles wireless communication. It is understood that the modem processor may also not be integrated into the processor. The machine-readable storage medium 120 is used to store machine-executable instructions for executing the scheme of this application, and the processor 130 is used to execute the machine-executable instructions stored in the machine-readable storage medium 120 to implement the pest and disease identification method based on AI visual model provided in the aforementioned method embodiments.

[0139] It should be noted that, in order to simplify the description of the present invention and thus help to understand one or more embodiments of the invention, multiple features may sometimes be grouped into one embodiment, drawing or description thereof in the foregoing description of the embodiments of the present invention.

Claims

1. A method for identifying pests and diseases based on an AI visual model, characterized in that, The method includes: Obtain a set of visual data of the plant to be identified, wherein the set of visual data of the plant to be identified includes visual acquisition data of different growth parts of the plant at different acquisition times; A dynamic association chain of pest and disease features is constructed, which includes core visual feature units of various pests and diseases at different development stages, association strength parameters between core visual feature units, and cross-stage feature evolution path information. The AI ​​vision model is invoked to perform multi-dimensional feature interaction processing on the plant visual data set based on the dynamic association chain of pest and disease features, generating feature interaction results and preliminary pest and disease identification results corresponding to the visual data set of the plant to be identified. The feature interaction results include the interaction matching parameters between the visual features of each growth part and the core visual feature units in the association chain. Based on the parameter adjustment patterns in historical pest and disease identification feedback data, the interaction matching parameters in the feature interaction results are iteratively optimized in multiple rounds to obtain the optimized interaction matching parameters and the corrected pest and disease identification results. Based on the corrected pest and disease identification results and optimized interactive matching parameters, a pest and disease identification report is generated, which includes pest and disease type identifiers, development stage assessments, and feature matching descriptions. The pest and disease identification report is then sent to the target terminal.

2. The method for identifying pests and diseases based on an AI visual model according to claim 1, characterized in that, The construction of dynamic association chains of pest and disease characteristics includes: Acquire full-cycle growth monitoring data for various pests and diseases, including continuous visual acquisition data of each pest and disease from the initial occurrence stage, the spread stage to the severe damage stage, and corresponding growth environment parameter records. Visual feature units for each development stage are extracted from the full-cycle growth monitoring data of each pest and disease to obtain a set of multi-stage feature units corresponding to each pest and disease. The visual feature units include color feature units, morphological feature units, texture feature units and lesion area distribution feature units. Core feature screening is performed on the multi-stage feature unit set of each disease and pest. The visual feature unit that appears more frequently in any development stage of the disease and pest than in other stages and can uniquely identify that stage is retained as the core visual feature unit, forming a subset of core visual feature units for each stage of each disease and pest. The correlation between core visual feature units in adjacent development stages of each disease and pest was analyzed. The co-occurrence frequency, morphological similarity, and evolutionary dependence of core visual feature units in adjacent stages were calculated. The co-occurrence frequency, morphological similarity, and evolutionary dependence were normalized to the same numerical range. Based on the weighted calculation results of the normalized three, the correlation strength parameter between each core visual feature unit was determined. Track the changes in the core visual feature units of each disease and pest from the initial occurrence stage to the severe damage stage, record the order of appearance, morphological transition mode and duration of each core visual feature unit, and form cross-stage feature evolution path information for each disease and pest. The core visual feature unit subsets of each stage of each disease and pest, the corresponding association strength parameters, and the cross-stage feature evolution path information are integrated to form a feature association sub-chain for a single disease and pest. Collect all disease and pest feature association sub-chains for single diseases and pests, classify and sort them according to disease and pest type, associate and mark similar core visual feature units that exist between different disease and pest types, and supplement feature association reference information across disease and pest types; Based on feature association reference information across pest and disease types, the association strength parameters of core visual feature units in each pest and disease feature association sub-chain are adjusted to ensure that the feature association logic between different pest and disease feature association sub-chains remains consistent. All adjusted sub-chains of pest and disease feature associations are integrated to construct a dynamic association chain of pest and disease features containing various pest and disease feature association information, which is then stored in the feature association database of the AI ​​vision model.

3. The method for identifying pests and diseases based on an AI visual model according to claim 2, characterized in that, The process involves extracting visual feature units for each developmental stage from the full-cycle growth monitoring data of each pest and disease, resulting in a multi-stage feature unit set corresponding to each pest and disease, including: The full-cycle growth monitoring data of each disease and pest are segmented in chronological order. The segmentation interval is determined according to the growth rate and morphological change range of the disease and pest. The continuous monitoring data is divided into multiple time periods, and each time period corresponds to a development stage. The visual acquisition data for each time period is preprocessed, and the color features of the preprocessed lesion area visual data are extracted. The visual data is converted from RGB color space to HSV color space using a color space conversion method. The hue mean, saturation variance and brightness range of the lesion area are extracted as color feature units. Morphological features are extracted from the preprocessed visual data of the lesion area. The contour information of the lesion area is obtained through the edge detection algorithm, and the perimeter, area and roundness of the contour are calculated as morphological feature units. Texture features were extracted from the preprocessed visual data of the lesion area, and the energy, entropy and contrast of the lesion area were calculated as texture feature units using the gray-level co-occurrence matrix method. The distribution features of the lesion area were extracted from the preprocessed visual data of the lesion area. The location coordinates, distribution density, and distance of the lesion area in the overall visual data of the plant were marked as the distribution feature units of the lesion area. The color feature units, morphological feature units, texture feature units and lesion area distribution feature units extracted in each time period are integrated to form the stage feature unit group corresponding to that time period. Arrange the phase feature units corresponding to all time periods of each disease and pest in chronological order to form a multi-stage feature unit set corresponding to each disease and pest.

4. The method for identifying pests and diseases based on an AI visual model according to claim 2, characterized in that, The process of tracking the changes in core visual feature units of each pest and disease from its initial occurrence stage to its severe damage stage, recording the order of appearance, morphological transition mode, and duration of each core visual feature unit, forms cross-stage feature evolution path information for each pest and disease, including: Assign time labels to each development stage in the full-cycle growth monitoring data of each pest and disease, marking the start and end times of each development stage; The occurrence time of each core visual feature unit is extracted from the subset of core visual feature units at each stage of each disease and pest. The occurrence time refers to the moment when the core visual feature unit first appears in the visual acquisition data. All core visual feature units of each disease and pest are sorted according to the order of their occurrence to form a temporal sequence of core visual feature units. The morphological differences between two adjacent core visual feature units in the time sequence of core visual feature units are analyzed to determine the morphological transition mode between them. The morphological transition mode includes gradual transition, abrupt transition and superposition transition. Gradual transition refers to the gradual change of morphological parameters, abrupt transition refers to the sudden change of morphological parameters, and superposition transition refers to the subsequent core visual feature unit adding morphological parameters on the basis of the previous core visual feature unit. Calculate the time span from the appearance time to the disappearance time of each core visual feature unit, and take the time span as the duration of the core visual feature unit. If the core visual feature unit continues until the end of the severe hazard stage, then the duration of the core visual feature unit is the time span from the appearance time to the end of the severe hazard stage. The temporal sequence of core visual feature units, the morphological transition mode of each adjacent core visual feature unit, and the duration of each core visual feature unit are integrated to form the basic information of cross-stage feature evolution path. Based on the basic information of cross-stage feature evolution paths, draw cross-stage feature evolution path diagrams for each disease and pest, and mark the position, occurrence time, duration and morphological transition mode of each core visual feature unit; In the cross-stage feature evolution path diagram, supplement the trend of the association strength parameter change corresponding to each core visual feature unit, and mark the curve of the association strength parameter change over time. By integrating the cross-stage feature evolution path diagrams and corresponding textual descriptions, cross-stage feature evolution path information for each disease and pest is formed.

5. The method for identifying pests and diseases based on an AI visual model according to claim 1, characterized in that, The AI ​​visual model is invoked to perform multi-dimensional feature interaction processing on the plant visual data set based on the dynamic association chain of pest and disease features, generating feature interaction results and preliminary pest and disease identification results corresponding to the visual data set of the plants to be identified, including: The visual data set of the plant to be identified is input into the feature analysis module of the AI ​​visual model. The visual data is divided into regions according to the growth parts to obtain the visual data of the leaf area, the visual data of the stem area, and the visual data of the fruit area of ​​the plant to be identified. Feature extraction was performed on the visual data of each region to obtain visual feature unit sets for leaf region, stem region, and fruit region. Each feature unit set of the region contains color feature units, morphological feature units, and texture feature units. The dynamic association chain of pest and disease features is retrieved from the feature association database of the AI ​​visual model. The visual feature unit sets of leaf area, stem area, and fruit area are initially matched with the core visual feature units of each stage in the association chain. Based on the preliminary matching results, core visual feature unit groups that match the visual feature unit sets of each region with a matching degree exceeding the preset matching degree threshold are selected. The pest and disease type and development stage corresponding to each core visual feature unit group are determined, and a candidate combination list is formed. The multi-dimensional feature interaction module of the AI ​​vision model is invoked to construct an interaction association model between visual feature units in different regions based on the pest and disease information in the candidate combination list. The interaction association model is used to analyze the collaborative matching relationship between visual feature units in different regions. The interaction matching parameters between each region's visual feature unit and the corresponding core visual feature unit group are calculated using an interaction association model. The interaction matching parameters include the feature matching degree within the region, the feature synergy degree across regions, and the feature fit degree at each stage. The interaction matching parameters of each region are integrated to form the feature interaction results corresponding to the visual data set of the plant to be identified. Based on the distribution of interaction matching parameters in the feature interaction results, the feature matching degree within the region, the cross-regional feature synergy degree, and the stage feature fit degree are normalized to the same numerical range. The candidate combination with the highest normalized feature matching degree within the region, the best normalized cross-regional feature synergy degree, and the best normalized stage feature fit degree is selected from the candidate combination list and determined as the disease and pest type and development stage of the plant to be identified. Generate preliminary pest and disease identification results, which include identified pest and disease type identifiers, development stage identifiers, and corresponding interactive matching parameters.

6. The method for identifying pests and diseases based on an AI visual model according to claim 5, characterized in that, The multi-dimensional feature interaction module that invokes the AI ​​visual model constructs an interaction association model between visual feature units in each region based on pest and disease information in the candidate combination list, including: Select a candidate combination from the candidate combination list and extract the pest and disease type information and development stage information corresponding to the candidate combination. Based on the information on the type and development stage of the pest, the core visual feature units of the leaf, stem and fruit at the corresponding stage are retrieved from the dynamic association chain of pest characteristics. The association weights between the core visual feature units of leaves and the core visual feature units of stems, the association weights between the core visual feature units of leaves and the core visual feature units of fruits, and the association weights between the core visual feature units of stems and the core visual feature units of fruits are determined. The association weights are set based on the cross-regional disease transmission pattern of the pest at the corresponding stage. The input layer of the interactive association model is constructed by taking the visual feature unit set of the leaf region, the visual feature unit set of the stem region, and the visual feature unit set of the fruit region of the plant to be identified as input vectors respectively. A hidden layer of the interaction association model is constructed, and three feature processing sub-modules are set up, corresponding to the leaf region, stem region and fruit region respectively. Each feature processing sub-module performs feature mapping processing on the input set of regional feature units to obtain the regional feature mapping vector. A cross-regional interaction module is added to the hidden layer. Based on the preset association weights, the regional feature mapping vectors of the three regions are weighted and fused to calculate the feature synergy values ​​of the leaf region and the stem region, the feature synergy values ​​of the leaf region and the fruit region, and the feature synergy values ​​of the stem region and the fruit region. The output layer of the interactive association model is constructed, and the feature matching degree within the region and the feature collaboration value across the region are used as output parameters. The feature matching degree within the region refers to the matching degree between the regional feature mapping vector of each region and the corresponding core visual feature unit, and the feature collaboration value across the region refers to the average of the three cross-region collaboration values. Based on historical interaction case data, the hidden layer parameters and cross-regional association weights of the interaction association model are adjusted so that the regional feature matching degree and cross-regional feature collaboration value output by the model can accurately reflect the collaborative matching relationship between visual feature units in different regions and core visual feature units. Following the steps described above, construct a corresponding interaction and association model for each candidate combination in the candidate combination list, forming a multi-model parallel processing framework.

7. The method for identifying pests and diseases based on an AI visual model according to claim 5, characterized in that, Based on the distribution of interaction matching parameters in the feature interaction results, candidate combinations with the highest feature matching degree within the region, the best cross-regional feature synergy, and the best stage feature fit are selected from the candidate combination list to determine the pest and disease type and development stage of the plant to be identified, including: Extract the interaction matching parameters from the feature interaction results. For each candidate combination in the candidate combination list, organize the corresponding regional feature matching degree, cross-regional feature synergy degree, and stage feature fit degree. The regional feature matching degree refers to the average matching degree of the leaf region, stem region, and fruit region. The cross-regional feature synergy degree refers to the average of the three cross-regional synergy values. The stage feature fit degree refers to the overall fit degree with the core features of the corresponding stage in the association chain. The intra-regional feature matching degree, cross-regional feature synergy degree, and stage feature fit degree of each candidate combination are normalized to the same numerical range. Weight coefficients are set for the normalized intra-regional feature matching degree, normalized cross-regional feature synergy degree, and normalized stage feature fit degree. The weight coefficients are set according to the priority of pest and disease identification, with the weight coefficient of the normalized intra-regional feature matching degree being the highest. The three normalized parameters of each candidate combination are weighted and calculated based on the weight coefficients to obtain the comprehensive matching score of each candidate combination. The candidate combinations in the candidate combination list are sorted in descending order of their comprehensive matching scores to obtain the sorted candidate combination list. Select the top K candidate combinations with the highest comprehensive matching scores from the sorted candidate combination list as the target candidate combinations; Analyze the common disease incidence of the disease and pest types corresponding to the target candidate combinations on the plant varieties to be identified. If the difference between the disease and pest types corresponding to any target candidate combination on the plant variety and the disease and pest types corresponding to the other two target candidate combinations on the plant variety exceeds the preset frequency difference range, the target candidate combination with the higher disease incidence rate will be given priority. If the difference in the incidence frequency of the disease and pest types corresponding to the target candidate combinations on the plant variety is within the preset frequency difference range, then the stage feature fit of each target candidate combination is further analyzed, and the target candidate combination with the highest stage feature fit is selected. If the difference in stage feature fit between each target candidate combination is within the preset fit difference range, then the cross-regional feature synergy of each target candidate combination is analyzed, and the target candidate combination with the highest cross-regional feature synergy is selected. The disease and pest types and development stages corresponding to the final selected target candidate combinations are determined as the disease and pest types and development stages of the plants to be identified.

8. The method for identifying pests and diseases based on an AI visual model according to claim 1, characterized in that, The parameter adjustment patterns based on historical pest and disease identification feedback data are used to perform multiple rounds of iterative optimization on the interaction matching parameters in the feature interaction results, resulting in optimized interaction matching parameters and corrected pest and disease identification results, including: Acquire historical pest and disease identification feedback data, which includes historical visual data set of plants to be identified, historical feature interaction results, historical preliminary pest and disease identification results, and historical corrected pest and disease identification results. Extract historical interaction matching parameters corresponding to historical feature interaction results and target interaction matching parameters corresponding to historical corrected pest and disease identification results from historical pest and disease identification feedback data. Calculate the parameter difference between the historical interaction matching parameters and the target interaction matching parameters in each historical case. Classify the parameter difference according to the feature attributes of the historical visual data set of plants to be identified. The feature attributes include plant variety attributes, collection season attributes, and lesion area attributes. Normalize the parameter difference under each type of feature attribute, analyze the changing trend of the parameter difference under each type of feature attribute, determine the parameter adjustment direction and parameter adjustment range corresponding to different feature attribute combinations, and form a mapping relationship table. Extract the plant variety attributes, collection season attributes, and lesion area attributes from the visual data set of the plants to be identified, and find the corresponding parameter adjustment direction and parameter adjustment range in the mapping relationship table; Based on the adjustment direction and magnitude of the parameters obtained from the search, the intra-regional feature matching degree, cross-regional feature synergy degree and stage feature fit degree in the feature interaction results are adjusted for the first time to obtain the interaction matching parameters after the first adjustment. Based on the first adjusted interaction matching parameters, candidate combinations are re-screened to determine the types and development stages of diseases and pests of the plants to be identified, and the first adjusted disease and pest identification results are obtained. Extract historical pest and disease identification results of the same type and stage as the first adjusted pest and disease identification results from historical feedback data, and obtain the corresponding target interaction matching parameter range; Determine whether the interaction matching parameters after the initial adjustment are within the target interaction matching parameter range. If they are not within the target interaction matching parameter range, further adjust the parameter adjustment range according to the mapping relationship table. Based on the adjusted parameter adjustment range, the interaction matching parameters after the first adjustment are adjusted a second time to obtain the interaction matching parameters after the second adjustment and the pest and disease identification results after the second adjustment. Repeat the above steps until the adjusted interaction matching parameters are within the range of the target interaction matching parameters. Use the interaction matching parameters at this point as the optimized interaction matching parameters, and use the corresponding pest and disease identification results as the corrected pest and disease identification results.

9. The method for identifying pests and diseases based on an AI visual model according to claim 8, characterized in that, The analysis examines the changing trends of parameter differences under each type of feature attribute, determines the parameter adjustment direction and magnitude corresponding to different feature attribute combinations, and forms a mapping relationship table, including: The characteristic attributes of the historical visual data set of plants to be identified are divided into multiple variety categories according to plant variety attributes, multiple season categories according to collection season attributes, and multiple lesion area categories according to lesion area attributes. For each combination of variety category, season category, and lesion area category, all historical cases under that combination are selected to form a case set of characteristic attribute combinations; Calculate the mean difference between historical interaction matching parameters and target interaction matching parameters in each feature attribute combination case set. The mean difference includes the mean difference of feature matching degree within the region, the mean difference of feature synergy degree across regions, and the mean difference of feature fit degree at each stage. Analyze the sign of the mean difference of parameters. If the mean difference of feature matching degree within a region is positive, then the parameter adjustment direction for feature matching degree within a region under that feature attribute combination is determined to be decreasing; if it is negative, then the parameter adjustment direction for feature matching degree within a region under that feature attribute combination is determined to be increasing. The same logic is used to determine the parameter adjustment direction for cross-regional feature synergy and stage feature fit. Calculate the standard deviation of the parameter difference in the case set for each feature attribute combination. Determine the parameter adjustment level based on the size of the standard deviation. The larger the standard deviation, the higher the parameter adjustment level, and the higher the parameter adjustment level, the greater the corresponding parameter adjustment. Assign each feature attribute combination the corresponding adjustment direction of the feature matching degree parameter within the region, the adjustment direction of the feature synergy degree parameter across the region, the adjustment direction of the feature fit degree parameter at each stage, and the corresponding parameter adjustment magnitude; The characteristic attribute combinations and their corresponding parameter adjustment directions and magnitudes are organized into a table to form a mapping relationship table, in which the characteristic attribute combinations are composed of variety category, season category and disease area category. The data in the mapping table is supplemented and improved. If the number of historical cases corresponding to any combination of feature attributes is lower than the preset case number threshold, the parameter adjustment direction and magnitude of similar feature attribute combinations are referenced, and fine-tuning is performed in combination with the particularity of the feature attribute combination so that the mapping table covers all possible feature attribute combinations.

10. A pest and disease identification system based on an AI visual model, characterized in that, include: processor; A machine-readable storage medium for storing machine-executable instructions of the processor; The processor is configured to execute the pest and disease identification method based on any one of claims 1 to 9 by executing the machine-executable instructions.

Images