Plant disease image recognition method and system based on deep learning

Abstract

The application provides a plant pest and disease image recognition method and system based on deep learning, and belongs to the technical field of deep learning. First, a plurality of growth stage multi-source image sets of plants are acquired, including whole plant, leaf, stem image and growth stage time, shooting angle identification information; then a plant image time sequence correlation network is constructed for organ level time sequence alignment processing, obtaining an aligned image set and organ correlation coefficient; then cross-growth stage lesion feature correlation is performed through the correlation coefficient, a cross-stage lesion correlation set and time sequence evolution information are acquired; the correlation set is input into a dedicated deep learning model for feature aggregation and type matching, obtaining a preliminary result; finally, organ adaptation verification is performed on the preliminary result in combination with the time sequence evolution information, generating a final recognition result and lesion organ time sequence evolution atlas. The application improves the accuracy and reliability of plant pest and disease recognition.

Description

Technical Field

[0001] This invention relates to the field of deep learning technology, and more specifically, to a method and system for recognizing plant insect diseases based on deep learning images. Background Technology

[0002] In agriculture, accurate identification of plant pests and diseases is crucial for ensuring crop yield and quality. Traditional methods for identifying plant pests and diseases mainly rely on manual observation and experience. These methods are not only inefficient but also easily influenced by the observer's subjectivity, resulting in low accuracy. With the development of image recognition technology, image-based methods for identifying plant pests and diseases are gradually emerging.

[0003] However, most existing image-based identification methods only analyze plant images at a single growth stage, neglecting the crucial factor that plant appearance and pest / disease manifestations change at different growth stages. Plants at different growth stages exhibit variations in overall morphology, leaf and stem physiological characteristics, and the manifestations and severity of pests and diseases also differ at each stage. Relying solely on images from a single stage makes it difficult to comprehensively and accurately grasp the characteristics and development patterns of pests and diseases, easily leading to misjudgments and missed diagnoses.

[0004] Furthermore, existing methods typically do not adequately consider the relationships between different plant organs (such as leaves and stems) and the temporal evolution of pests and diseases on different organs. Plant pests and diseases often spread and proliferate between different organs, and their development is a dynamic, temporal process. The lack of analysis of these relationships and temporal evolution results in a lack of systematicity and comprehensiveness in the identification results. Summary of the Invention

[0005] In view of the aforementioned problems, and in conjunction with the first aspect of the present invention, embodiments of the present invention provide a method for plant pest and disease image recognition based on deep learning, the method comprising:

[0006] Acquire a multi-source image set of the plant at multiple growth stages. The multi-source image set includes an overall image of the plant, a partial image of the leaf, a partial image of the stem, and time markers and shooting angle markers for each growth stage.

[0007] Based on the growth stage time markers and shooting angle markers, a plant image temporal association network is constructed. The multi-source image set is then subjected to organ-level temporal alignment processing to obtain an organ-level temporal aligned image set and organ association coefficients between images.

[0008] By using the organ correlation coefficient, cross-growth stage lesion feature association processing is performed on the organ-level time-aligned image set to obtain a cross-stage lesion association set and the temporal evolution information of lesions on different organs.

[0009] The cross-stage lesion association set is input into a preset plant insect disease-specific deep learning model for feature aggregation and type matching to obtain lesion aggregation features and preliminary insect disease type results.

[0010] By combining the temporal evolution information of the lesions on different organs, the preliminary results of the insect and disease type are verified by organ adaptation, and the final plant insect and disease identification results and temporal evolution map of the lesion organs are generated.

[0011] Furthermore, embodiments of the present invention also provide a plant pest and disease image recognition system based on deep learning, comprising:

[0012] A processor; a machine-readable storage medium for storing machine-executable instructions of the processor; wherein the processor is configured to perform the above-described deep learning-based plant pest and disease image recognition method by executing the machine-executable instructions.

[0013] In another aspect, embodiments of the present invention also provide a computer program product, the computer program product including machine-executable instructions, the machine-executable instructions being stored in a computer-readable storage medium, a processor of a computer device reading the machine-executable instructions from the computer-readable storage medium, the processor executing the machine-executable instructions, causing the computer device to execute the above-described deep learning-based plant insect and disease image recognition method.

[0014] Based on the above, by acquiring multi-source image sets of plants at multiple growth stages, covering image information at different levels such as the whole plant, leaf parts, and stem parts, as well as growth stage time and shooting angle identifiers, a plant image temporal association network is constructed based on the growth stage time and shooting angle identifiers. Organ-level temporal alignment processing is performed on the multi-source image set, which can accurately obtain organ-level temporally aligned image sets and organ association coefficients between images, effectively solving the problem of image alignment under different growth stages and shooting angles. Next, cross-growth stage lesion feature association processing is performed on the organ-level temporally aligned image set through organ association coefficients, which can deeply explore the feature changes of lesions at different growth stages and the temporal evolution information on different organs. The cross-stage lesion association set is input into a preset plant pest and disease-specific deep learning model for feature aggregation and type matching processing, which can give full play to the powerful feature extraction and classification capabilities of the deep learning model and obtain accurate lesion aggregation features and preliminary pest and disease type results. Finally, by combining the temporal evolution information of lesions on different organs, the preliminary results of pest and disease types were verified by organ adaptation, which further improved the accuracy and reliability of the identification results. The final plant pest and disease identification results and the temporal evolution map of lesion organs were generated, providing a comprehensive, accurate and systematic decision-making basis for the prevention and control of plant pests and diseases in agricultural production, and effectively improving the level of intelligence in agricultural production and the effect of pest and disease control. Attached Figure Description

[0015] Figure 1 This is a schematic diagram of the execution flow of the deep learning-based plant insect and disease image recognition method provided in the embodiments of the present invention.

[0016] Figure 2 This is a schematic diagram of exemplary hardware and software components of the deep learning-based plant insect and disease image recognition system provided in an embodiment of the present invention. Detailed Implementation

[0017] The present invention will now be described in detail with reference to the accompanying drawings. Figure 1 This is a flowchart illustrating a deep learning-based image recognition method for plant insect diseases provided in one embodiment of the present invention. The following is a detailed description of this deep learning-based image recognition method for plant insect diseases.

[0018] Step S110: Obtain a multi-source image set of the plant at multiple growth stages. The multi-source image set includes an overall image of the plant, a partial image of the leaves, a partial image of the stem, and time markers and shooting angle markers for each growth stage.

[0019] In this embodiment, wheat is used as the target plant for pest and disease identification. Multi-source images are acquired from five growth stages of wheat: seed germination, tillering, jointing, heading, and grain filling. For each growth stage, a professional camera is used to capture images from different locations and angles in the field, obtaining an overall image of the plant. This overall image shows the overall morphology of the wheat plant at that growth stage, including plant height and the number of tillers. Simultaneously, macro lenses are used to capture detailed images of the leaves, focusing on the surface details such as the presence of spots or curling. Images of the stems are also captured to observe for signs of borers or discoloration. Each image is accompanied by a corresponding growth stage time marker, such as the seed germination stage being captured in the early part of a certain month of a certain year, the tillering stage in the middle of a certain month of the same year, and the shooting angle marker, such as 45 degrees from the front, 30 degrees from the side, or a top view. These images and markers together constitute a multi-source image set.

[0020] When collecting these image data, if privacy-sensitive data such as the geographic information of the shooting area is involved, data anonymization technology is used to convert the specific coordinates in the geographic location information into fuzzy area identifiers, and the image data is encrypted and stored. Access requires multiple authentications to prevent privacy leaks.

[0021] Step S120: Construct a plant image temporal association network based on the growth stage time markers and shooting angle markers, and perform organ-level temporal alignment processing on the multi-source image set to obtain an organ-level temporal aligned image set and organ association coefficients between images.

[0022] Step S121: Extract the growth stage time identifier and shooting angle identifier information of each image in the multi-source image set, perform primary sorting on all images according to the order of growth stages, and perform secondary sorting on the shooting angle identifier information within the same growth stage to obtain the initial time-series angle image sequence.

[0023] For the aforementioned multi-source image set of wheat, the growth stage time markers and shooting angle markers of each image are first extracted. All images are then sorted in the chronological order of seed germination, tillering, jointing, heading, and grain-filling stages. Within the same growth stage, such as tillering, secondary sorting is performed according to the shooting angle markers, such as first a 45-degree frontal view, then a 30-degree side view, and finally a top-down view. This sorting method organizes the originally chaotic multi-source images into an ordered initial temporal angle image sequence, establishing temporal and angular correlations between the images, thus preparing for subsequent organ-level temporal alignment processing.

[0024] Step S122: Extract plant organ contour features from each image in the initial time-series angle image sequence. The plant organ contour features include main stem contour features, leaf contour features, and stem branch contour features. The main stem contour features include the coordinates of the main stem contour points and the change value of the contour curvature. The leaf contour features include the coordinates of the leaf contour points, the direction of the leaf vein contour, and the density of the leaf edge serrations. The stem branch contour features include the coordinates of the branch contour points and the connection angle between the branch and the main stem.

[0025] Step S1221: Perform color space conversion processing on each image in the initial time-series angle image sequence, extract the hue channel image, and perform threshold segmentation processing on the hue channel image to divide the hue channel image into the plant foreground region and the background region, thus obtaining a binarized image.

[0026] For wheat images in the initial time-series angle image sequence, color space conversion was performed, converting the common RGB color space to the HSV color space, and then the hue channel image was extracted. Since the wheat plants differ in hue from the background (such as soil and weeds), threshold segmentation was performed on the hue channel image. By setting an appropriate threshold range, the area containing the wheat plants was divided into the foreground region, while the background region was excluded, thus obtaining a binarized image. In this binarized image, the foreground region of the plants is white, and the background region is black, clearly separating the approximate outline of the wheat plants.

[0027] Step S1222: After performing morphological closing operation on the binarized image, perform morphological opening operation to obtain an optimized binarized image. Then, perform edge detection on the optimized binarized image to form an initial set of contour segments.

[0028] The binarized image obtained above is first processed by morphological closing operation, which fills in small holes in the image, making the outline of the plant foreground region more complete. Next, morphological opening operation is performed, which removes small noise points in the image, further optimizing the image quality and obtaining an optimized binarized image. Then, an edge detection algorithm is used to process the optimized binarized image to detect the edges of the plant foreground region. These edges consist of a series of line segments, forming an initial set of contour line segments. These line segments initially outline the contour shapes of the various organs of the wheat plant.

[0029] Step S1223: Perform contour screening on the initial set of contour segments. Set screening rules according to the morphological characteristics of plant organs to screen out the main stem contour segments, leaf contour segments, and stem branch contour segments. Perform fitting processing on the screened main stem contour segments. Use a cubic polynomial fitting algorithm to fit the segments into a smooth curve that conforms to the upright shape of the plant main stem. Extract the coordinates of the sampling points on the smooth curve at fixed intervals as the coordinates of the main stem contour points. Calculate the curvature change value at each sampling point as the curvature change value of the main stem contour to form the main stem contour feature.

[0030] Screening rules were established based on the morphological characteristics of various organs of the wheat plant to filter the initial set of contour segments. The main stem of wheat is usually relatively upright, with relatively long and stable contour segments; these characteristics were used to select the main stem contour segments. Leaf contour segments have a wide distribution range and may have serrated edges. Stem branch contour segments are relatively thin and have connection points with the main stem contour segments. After selecting the main stem contour segments, a cubic polynomial fitting algorithm was used to fit them, resulting in a smooth curve that conforms to the upright morphology of the wheat main stem. Then, the coordinates of sampling points were extracted from this smooth curve at fixed intervals; these coordinates are the coordinates of the main stem contour points. For each sampling point, its curvature change value was calculated. The curvature change value reflects the degree of curvature of the main stem contour at that point. Combining the coordinates of the main stem contour points and the corresponding curvature change values ​​constitutes the main stem contour feature.

[0031] Step S1224: Close the selected leaf contour segments, extend the unclosed segments to intersect to form a closed contour, extract the coordinates of the sampling points on the closed contour as the coordinates of the leaf contour points, extract the contour direction of the main vein and lateral veins of the leaf through the leaf vein detection algorithm, count the number and distribution density of the serrated line segments on the leaf edge as the leaf edge serration density, and form the leaf contour features.

[0032] For the selected leaf contour segments, since some segments may not be closed during actual shooting and image processing, closure processing is required. The two ends of the unclosed segments are extended according to certain rules until they intersect, forming a closed leaf contour. The coordinates of the sampling points on this closed contour are extracted as the leaf contour point coordinates. Next, a leaf vein detection algorithm is used to process the leaf region, extracting the contour direction of the main vein and lateral veins. The main vein is usually the thicker vein in the center of the leaf, while the lateral veins extend from the main vein towards the leaf edge. Simultaneously, the number of serrated segments on the leaf edge is counted, and their distribution density is calculated, i.e., the number of serrated segments per unit length of leaf edge. The leaf contour point coordinates, leaf vein contour direction, and leaf edge serration density are combined to form the leaf contour features.

[0033] Step S1225: Fit the selected stem branch outline segments to obtain the coordinates of the branch outline points. Calculate the angle between the branch outline segments and the main stem outline curve at the connection point using an angle calculation algorithm. This angle is used as the connection angle between the branch and the main stem, forming the stem branch outline feature.

[0034] The selected stem branch contour segments are fitted using a similar method to that used for the main stem contour segments to obtain smooth curves of the branch contours. The coordinates of sampling points on the curve are then extracted as the branch contour point coordinates. Next, an angle calculation algorithm is used to find the connection points between the branch contour segments and the main stem contour curves. The angle between the tangent directions of the branch contour curve and the main stem contour curve at this connection point is calculated; this angle is the connection angle between the branch and the main stem. Combining the branch contour point coordinates and the connection angle constitutes the stem branch contour feature.

[0035] Step S1226: Combine the main stem outline features, leaf outline features, and stem branch outline features into the plant organ outline features of each image, and record the coordinates of the connection points between the branches and the main stem, as well as the coordinates of the connection points between the leaves and the branches.

[0036] The obtained main stem outline features, leaf outline features, and stem branch outline features are integrated together to form complete plant organ outline features for each image. Simultaneously, the coordinates of the connection points between branches and the main stem, as well as the coordinates of the connection points between leaves and branches, are recorded. These connection point coordinates reflect the positional relationships between various organs of the wheat plant.

[0037] Step S123: Using each image as a network node, and using the similarity of plant organ contour features in images taken from the same shooting angle at adjacent growth stages as initial edge weights, an initial temporal association network is constructed. The initial edge weights are calculated by weighting the standardized values ​​of the main stem contour overlap ratio, the leaf contour overlap ratio, and the branch connection angle deviation value. The branch connection angle deviation value is converted into a deviation score that is comparable to the contour overlap ratio. The higher the main stem contour overlap ratio, the higher the leaf contour overlap ratio, and the higher the branch connection angle deviation score, the larger the initial edge weight.

[0038] Each wheat image in the initial temporal angle image sequence is considered a network node. For two images taken from the same angle at adjacent growth stages, the similarity of their plant organ contour features is calculated, and this similarity is used as the initial edge weights to construct the initial temporal association network. When calculating the initial edge weights, the overlap ratio of the main stem contour, the overlap ratio of the leaf contour, and the branch connection angle deviation value are calculated first. The main stem contour overlap ratio refers to the proportion of the number of main stem contour points that overlap in the two images out of the total number of contour points; the leaf contour overlap ratio is calculated similarly. The branch connection angle deviation value is the difference between the connection angles of corresponding branches and the main stem in the two images. This difference is converted into a deviation score by mapping the angle difference to a scoring interval that is the same as the contour overlap ratio range; the smaller the angle difference, the higher the deviation score. Then, these three indicators are standardized to make them of the same order of magnitude, and then weighted according to certain weights to obtain the initial edge weights. The higher the overlap ratio of the main stem outline, the higher the overlap ratio of the leaf outline, and the higher the deviation score of the branch connection angle, the more similar the plant organ outline features of the two images are, and the greater the initial edge weight will be.

[0039] Step S124: Calculate the indirect association weights of images with the same shooting angle in non-adjacent growth stages in the initial temporal association network. The indirect association weights are determined by the sum of the initial edge weights of all paths between two nodes. The path selection rule is that the interval between the growth stages corresponding to the nodes in the path does not exceed N, and the shooting angle identification information is consistent, where N is a set positive integer.

[0040] In the initial temporal association network, besides the direct initial edge weight connection between image nodes in adjacent growth stages, there are also indirect associations between image nodes in non-adjacent growth stages at the same shooting angle. When calculating the indirect association weights of these non-adjacent nodes, all possible paths between the two nodes are considered. Path selection must satisfy the following conditions: the interval between the growth stages corresponding to the nodes in the path does not exceed a set positive integer N, and the shooting angle identification information is consistent, to ensure the validity and relevance of the path. For each path that meets the rules, the initial edge weights between all nodes on the path are multiplied to obtain the weight product of that path. Then, the weight products of all paths between the two nodes are added together, and the sum is the indirect association weight between the two non-adjacent nodes. This indirect association weight reflects the degree of association established between images in non-adjacent growth stages through intermediate growth stage images.

[0041] Step S125: Integrate the initial edge weights and indirect association weights into organ association coefficients between nodes to obtain the plant image temporal association network. The larger the organ association coefficient, the stronger the association between the corresponding organs in the two images.

[0042] The initial edge weights and indirect association weights calculated above are integrated. For any two nodes in the network, the initial edge weights (if adjacent) and indirect association weights are added together to obtain the organ association coefficients between the nodes. This completes the construction of the plant image temporal association network, which can comprehensively reflect the degree of association between various organs in wheat images at different growth stages and from the same shooting angle. The larger the organ association coefficient, the stronger the correlation of the morphology, position, and other features of the corresponding organs in the two images, and the more continuous the temporal changes.

[0043] Step S126: Based on the organ correlation coefficient in the plant image temporal correlation network, select a reference image whose organ correlation coefficient with the current image is higher than the correlation threshold. Use the main stem axis and leaf vein line in the reference image as calibration benchmarks to perform coordinate calibration on the corresponding organ in the current image, so that the overlap of the main benchmark lines of the same organ in different images reaches the calibration standard, and obtain an organ-level temporal aligned image set.

[0044] Based on a set association threshold, images with organ association coefficients higher than this threshold in the plant image temporal association network are selected as reference images. These reference images have strong organ association with the current image and are suitable as calibration bases. The central axis of the main stem and the midrib of the leaf in the reference image are used as calibration benchmarks. The central axis of the main stem is the center line obtained by fitting the coordinates of the main stem contour points, and the midrib of the leaf is the midrib contour line extracted by the leaf vein detection algorithm. Then, coordinate calibration processing is performed on the corresponding organs in the current image. By adjusting the coordinate system of the current image, the central axis of the main stem in the current image coincides with the central axis of the main stem in the reference image, and the midrib of the leaf coincides with the midrib of the leaf in the reference image, ensuring that the overlap of the main benchmark lines of the same organ in different images reaches the set calibration standard. After the above calibration processing, an organ-level temporally aligned image set is obtained. The images in this organ-level temporally aligned image set have a high degree of consistency in the temporal position of the organs.

[0045] Step S130: Perform cross-growth stage lesion feature association processing on the organ-level temporally aligned image set through the organ association coefficient to obtain a cross-stage lesion association set and temporal evolution information of lesions on different organs.

[0046] Step S131: Perform plant organ segmentation processing on each image in the organ-level time-aligned image set. Use a segmentation method based on plant organ morphology templates to segment the image into leaf region images, main stem region images, and branch region images. Record the coordinate range information and organ type identifier of each organ region image.

[0047] For each wheat image in the organ-level time-aligned image set, a segmentation method based on plant organ morphology templates was used. Morphological templates for wheat leaves, main stems, and branches were pre-constructed, containing typical morphological features of each organ at different growth stages. The image was matched against these templates, and based on the matching results, the image was segmented into leaf region images, main stem region images, and branch region images. For example, the region with the highest matching degree to the leaf morphology template was segmented into a leaf region image, and similarly, main stem region images and branch region images were obtained. Simultaneously, the coordinate range information of each organ region image in the original image, such as the coordinates of the upper left and lower right corners, and the corresponding organ type identifier, such as "leaf," "main stem," and "branch," were recorded to facilitate subsequent separate processing of lesion features in different organs.

[0048] Step S132: Perform preliminary lesion detection processing on the image of each organ region. Use a sliding window detection method adapted to the morphology of plant organs. The window shape is determined according to the organ type. Use an elliptical window for the leaf region, a rectangular window for the main stem region, and a long and thin window for the branch region. Move the window to traverse the organ region image according to the set step size and extract the gray-scale change features and color distribution features within the window.

[0049] Step S1321: Determine the parameters of the sliding window based on the organ type identifier. The leaf region uses an elliptical window, with the major axis length set to a fixed proportion of the longest side of the leaf region image and the minor axis length set to a fixed proportion of the major axis length. The main stem region uses a rectangular window, with the window length set to a fixed proportion of the main stem region image length and the width set to a fixed proportion of the main stem region image width. The branch region uses a long and narrow window, with the window length set to a fixed proportion of the branch region image length and the width set to a fixed proportion of the branch region image width. The sliding step size is uniformly set to a fixed proportion of the shortest side of the corresponding window.

[0050] Appropriate sliding window parameters are selected for different organ region images based on organ type identification. For leaf region images, an elliptical window is used, with the major axis length set to a fixed proportion of the longest side of the leaf region image, and the minor axis length set to a fixed proportion of the major axis length. This window shape closely matches the natural shape of the leaf. A rectangular window is used for the main stem region image, with the window length and width both set to a fixed proportion of the main stem region image length and width, to accommodate the elongated shape of the main stem. A slender window is used for branch region images, with the window length and width set to fixed proportions of the branch region image length and width, respectively, reflecting the thinness of the branches. The sliding step size is uniformly set to a fixed proportion of the shortest side of the corresponding window, ensuring that the window can evenly traverse the entire organ region image.

[0051] Step S1322: Move the window with the set sliding step size, traverse the entire organ region image in the order from left to right and from top to bottom, and record the position coordinates of each window and the image pixel data within the window.

[0052] Following a set sliding step size, the sliding window moves from left to right and from top to bottom to traverse the entire organ region image. During the traversal, the position coordinates of each window in the organ region image are recorded, such as the coordinates of the upper left corner of the window, as well as the image pixel data contained within the window. This pixel data is the basis for subsequent extraction of grayscale variation features and color distribution features.

[0053] Step S1323: Extract grayscale change features from the pixel data within each window, convert the pixel data into a grayscale image, calculate the standard deviation of grayscale values, the mean of grayscale gradients, and the grayscale entropy of the grayscale image within the window, and combine the standard deviation of grayscale values, the mean of grayscale gradients, and the grayscale entropy of the grayscale image in a fixed order to form grayscale change features; the standard deviation of grayscale values ​​reflects the dispersion of grayscale distribution within the window, the mean of grayscale gradients reflects the intensity of grayscale changes within the window, and the grayscale entropy reflects the uniformity of grayscale distribution within the window.

[0054] The pixel data within each window is converted into a grayscale image. The grayscale value of each pixel in the grayscale image is obtained by weighted averaging of the original pixel's RGB values. Then, the standard deviation of the grayscale values ​​is calculated. A larger standard deviation indicates greater dispersion in the grayscale distribution within the window, meaning more pronounced differences in brightness. Next, the mean gradient of the grayscale values ​​is calculated. The mean gradient reflects the drasticness of grayscale changes within the window; a larger mean gradient indicates more drastic grayscale changes, potentially indicating the presence of edges or textures. Finally, the grayscale entropy is calculated. A larger entropy indicates a more uneven grayscale distribution within the window, containing more complex information. Combining these three metrics in a fixed order constitutes the grayscale variation characteristics of the image within the window.

[0055] Step S1324: Extract color distribution features from the pixel data in each window, convert the pixel data into a hue, saturation, and lightness color space, and calculate the histograms for the hue, saturation, and lightness channels respectively; divide each histogram into a fixed number of intervals, calculate the proportion of pixels in each interval to the total number of pixels in the window, and combine the interval pixel proportions of the three channel histograms in channel order to form color distribution features.

[0056] The pixel data within the window is converted from the RGB color space to a hue / saturation / lightness color space, which better matches how the human eye perceives color. Then, histograms are calculated for each of the hue, saturation, and lightness channels. Histograms reflect the pixel distribution of different color values ​​within each channel. Each histogram is divided into a fixed number of intervals; for example, the hue channel histogram is divided into multiple intervals, each representing a range of hue values. The proportion of pixels in each interval relative to the total number of pixels in the window is calculated; these proportions reflect the color distribution within that interval. Combining the interval pixel proportions from the three channel histograms in the order of hue, saturation, and lightness yields the color distribution characteristics.

[0057] Step S133: Collect image samples of healthy organs of the same plant variety, extract the grayscale change features and color distribution features of the corresponding organ regions in the image samples, calculate the mean range and distribution pattern of the sample features, and determine the normal feature range of each organ.

[0058] A large number of organ image samples from healthy wheat of the same variety were collected. These samples should cover healthy leaves, main stems, and branches at different growth stages and from different shooting angles. The same organ region segmentation and sliding window processing as described above were performed on these healthy sample images, extracting the grayscale variation features and color distribution features within each window. Then, the mean range of these sample features was calculated, i.e., the average value and fluctuation range of each feature indicator, and their distribution patterns were analyzed, such as whether they conform to a normal distribution. Based on the mean range and distribution patterns of the sample features, the normal feature ranges for each organ at different growth stages and locations were determined. When subsequently detected window features exceed this normal range, a lesion may exist in that window region.

[0059] Step S134: Select window regions whose grayscale variation features or color distribution features deviate from the normal feature range of the corresponding organ as suspected lesion regions, record the location coordinates, size information and corresponding feature data of each suspected lesion region, and form an organ suspected lesion information table for each image. The feature data includes color distribution histogram data, texture distribution feature data and shape parameter data.

[0060] The grayscale variation features and color distribution features of each window extracted above are compared with the normal feature range of the corresponding organ. If any indicator in the grayscale variation features of a window (such as the standard deviation of grayscale values, the mean of gradient, or entropy) deviates from the normal range, or the percentage of pixels in the color distribution features deviates from the normal range, then the window region is marked as a suspected lesion region. The location coordinates, size information (such as the length and width of the window), and corresponding feature data of each suspected lesion region in the organ region image are recorded. The feature data includes color distribution histogram data (the percentage of pixels in each channel interval), texture distribution feature data (the three indicator values ​​of grayscale variation features), and shape parameter data (such as the shape, perimeter, and area of ​​the window). The above information is organized into an organ suspected lesion information table for each image to facilitate subsequent correlation analysis of suspected lesion regions.

[0061] Step S135: Based on the organ correlation coefficient, filter out the associated images whose organ correlation coefficient with the current image is higher than the correlation threshold, and extract the suspected lesion information table of the organ corresponding to the same organ type identifier in the associated images as the associated lesion information.

[0062] Based on the organ association coefficients calculated in the plant image temporal association network, images with organ association coefficients higher than the association threshold are selected as associated images. These associated images have a strong organ correlation with the current image and may contain information related to suspected lesion areas in the current image. Then, a table of suspected lesion information for organs with the same organ type identifier as the current image (e.g., both are leaves, main stems, or branches) is extracted from these associated images and used as associated lesion information. By comparing the suspected lesion information of the current image and the associated images, changes in lesions at different growth stages can be observed.

[0063] Step S136: Compare the positional correspondence and feature data similarity between the suspected lesion areas in the current image and the suspected lesion areas in the associated images. The positional correspondence is determined by the overlap ratio of the coordinates of the two suspected lesion areas in the organ coordinate system. The feature data similarity is calculated by comprehensively considering the overlap of the color distribution histogram, the matching ratio of the texture distribution features, and the standardized value of the deviation of the shape parameters. The deviation of the shape parameters is converted into a deviation score that is comparable to the overlap and matching ratio. The higher the overlap of the color distribution histogram, the higher the matching ratio of the texture distribution features, and the higher the deviation score of the shape parameters, the higher the feature data similarity.

[0064] An organ coordinate system is established, with the top-left corner of the organ region image as the origin, the horizontal direction as the X-axis, and the vertical direction as the Y-axis. For each suspected lesion region in the current image, the suspected lesion region is searched in the suspected lesion information table corresponding to the same organ type identifier in the associated image. The overlap ratio of their coordinates in the organ coordinate system is calculated, that is, the proportion of the overlapping area of ​​two suspected lesion regions to the total area of ​​the two regions. This ratio reflects the positional correspondence; the higher the ratio, the closer the positional correspondence. Simultaneously, feature data similarity is calculated, including the overlap of color distribution histograms, that is, comparing the similarity of the pixel proportions of corresponding intervals in the color distribution histograms of two suspected lesion regions; the higher the overlap, the more similar the color distribution. The matching ratio of texture distribution features, that is, the proportion of the number of matches of each indicator in the grayscale variation features to the total number of indicators. The deviation of shape parameters is converted into a deviation score, similar to the branch connection angle deviation value, making the deviation score comparable to the overlap and matching ratio. After standardizing these three indicators, a comprehensive calculation is performed to obtain the feature data similarity. The higher the overlap of color distribution histograms, the higher the matching ratio of texture distribution features, and the higher the shape parameter deviation score, the higher the similarity of feature data.

[0065] Step S137: Divide the suspected lesion areas that meet the association conditions in terms of both location correspondence and feature data similarity into the same lesion association group. Each lesion association group contains associated suspected lesion areas on the same organ at different growth stages.

[0066] The association conditions for location correspondence are set as follows: the coordinate overlap ratio is higher than a certain threshold. The association condition for feature data similarity is that the similarity value is higher than a certain threshold. For suspected lesion regions in the current image and associated images that meet these two association conditions, they are classified into the same lesion association group. Each lesion association group contains suspected lesion regions on the same organ at different growth stages that are related to each other. These regions are likely manifestations of the same lesion at different growth stages.

[0067] Step S138: Extract the location coordinate change trajectory, size change trend and feature data change pattern of each suspected lesion area in each lesion association group, and generate the temporal evolution information of the lesion on different organs. The temporal evolution information includes the position offset trajectory of the lesion in the organ coordinate system, the size expansion or contraction curve and the feature data change matrix.

[0068] For each lesion association group, suspected lesion regions are arranged in chronological order of growth stages. The positional coordinates of each suspected lesion region in the organ coordinate system are extracted, and these coordinates are connected chronologically to form a positional offset trajectory, reflecting the movement of the lesion on the organ. The size (e.g., area) of each suspected lesion region is recorded, and its expansion or contraction trend with different growth stages is analyzed to form a size expansion or contraction curve. Simultaneously, characteristic data of each suspected lesion region, such as color distribution histogram data and texture distribution characteristic data, are collected and analyzed to understand their changes over time. A characteristic data change matrix is ​​constructed, where rows represent different growth stages, columns represent different characteristic indicators, and matrix elements are the values ​​of the corresponding characteristic indicators. The positional offset trajectory, size expansion or contraction curve, and characteristic data change matrix are integrated to generate temporal evolution information of the lesion on different organs.

[0069] Step S139: Combine all lesion association groups and the suspected lesion regions contained in each group into a cross-stage lesion association set, and record the temporal evolution information of the lesions in different organs and the corresponding organ type identifiers for each group.

[0070] All the lesion association groups obtained above are summarized. Each lesion association group contains related suspected lesion regions on the same organ at different growth stages. These lesion association groups are combined to form a cross-stage lesion association set. At the same time, the temporal evolution information of the corresponding lesions in different organs and the organ type identifier of each group are associated and recorded with the lesion association groups in this set, so that when inputting into the deep learning model for processing, the origin and evolution of each lesion association group can be clearly identified.

[0071] Step S140: Input the cross-stage lesion association set into a preset plant insect disease-specific deep learning model for feature aggregation and type matching to obtain lesion aggregation features and preliminary insect disease type results.

[0072] Step S141: Preprocess each lesion association group in the cross-stage lesion association set, extract image data of each suspected lesion area in each group, adjust the image size according to the organ type where the lesion is located, adjust the leaf lesion image to a uniform rectangular size, and adjust the main stem and branch lesion images to a uniform elongated size.

[0073] For each lesion association group in the cross-stage lesion association set, the original image data of each suspected lesion region is extracted. The image data is then resized according to the organ type where the lesion is located. Leaf lesion images are resized to a uniform rectangular shape, ensuring consistency in length and width; main stem and branch lesion images, due to their morphological characteristics, are resized to a uniform elongated shape, ensuring that the image data of different lesion association groups have a consistent specification, facilitating subsequent model processing and feature extraction.

[0074] Step S142: Perform pixel value normalization processing on the adjusted image data to convert the pixel values ​​into a fixed numerical range, thereby obtaining a standardized lesion image sequence.

[0075] Resized image data may have pixel values ​​in varying ranges. To eliminate the impact of pixel value differences on model training and recognition, pixel value normalization is performed on the image data. This transforms the pixel value of each pixel to a fixed numerical range, for example, from 0-255 to 0-1. Normalization is achieved by dividing the pixel value by the maximum value of that channel (e.g., 255), resulting in a standardized lesion image sequence. The pixel values ​​in this standardized lesion image sequence have a uniform range, which helps the model learn lesion features better.

[0076] Step S143: Extract the temporal arrangement information of the standardized lesion image sequence, combine the growth stage interval data in the temporal evolution information of the lesion on different organs, and perform temporal completion processing on the standardized lesion image sequence. When the image interval between adjacent growth stages exceeds the interval threshold, an interpolation method based on the evolution law of organ lesions is used to generate an intermediate image to obtain a complete temporal lesion image sequence.

[0077] Temporal arrangement information, i.e., the order of growth stages corresponding to each image, is extracted from a standardized sequence of lesion images. Combining this with the growth stage interval data from the temporal evolution information of lesions in different organs, the interval between adjacent growth stages is checked to see if it exceeds a set threshold. If the interval exceeds the threshold, it indicates a temporal gap, requiring completion processing. The completion method employs interpolation based on the evolutionary patterns of organ lesions. Based on known lesion image features and temporal evolution information, the features of the missing growth stage are predicted, generating intermediate images. For example, based on the lesion's positional offset trajectory, size change trend, and feature data variation patterns, the possible morphology and features of the lesion in the intermediate growth stage are inferred, generating corresponding images. Through this temporal completion process, a complete temporal sequence of lesion images is obtained, ensuring the continuity and integrity of the temporal information.

[0078] Step S144: Input the complete temporal sequence of lesion images into the feature input layer of the plant insect disease-specific deep learning model. Through the dimension transformation operation of the input layer, the image sequence is converted into the input format adapted to each branch of the model to obtain the branch input features.

[0079] A complete temporal sequence of lesion images is input into the feature input layer of a pre-defined deep learning model specifically for plant diseases and pests. The feature input layer first performs a dimensionality transformation on the input image sequence, adjusting its dimensions to an input format acceptable to subsequent parallel branches of the model. For example, the number of channels, height, and width of the image sequence are adjusted to match the input requirements of the lesion texture detail extraction branch, the lesion morphology feature extraction branch, and the lesion color feature extraction branch. After dimensionality transformation, the branch input features are obtained, which can be correctly received and processed by each branch.

[0080] Step S145: Input the branch input feature into the lesion texture detail extraction branch, and process it sequentially through the preset convolutional layer, pooling layer and batch normalization layer to extract the surface texture details of the lesion in the image to obtain the lesion texture feature map.

[0081] The input features are first fed into the lesion texture detail extraction branch. This branch contains multiple convolutional layers, multiple pooling layers, and two batch normalization layers. The convolutional layers perform convolution operations on the input features using kernels of different sizes, extracting local features from the image. As the convolutional layers deepen, the extracted features gradually transition from low-level edges and textures to high-level semantic features. The pooling layers use max pooling to downsample the feature maps output by the convolutional layers, reducing the size and number of parameters while retaining important feature information. Batch normalization layers are inserted between adjacent convolutional layers to standardize the input of each layer, accelerating network training convergence and improving the model's generalization ability. Through the sequential processing of these layers, the texture details of the lesion surface in the image, such as the texture of spots and the texture of wormholes, can be effectively extracted, resulting in a lesion texture feature map.

[0082] Step S146: The branch input features are synchronously input into the lesion morphology feature extraction branch. The overall morphological features of the lesion are extracted through the expansion of the receptive field of the dilated convolutional layer and the global information aggregation of the global average pooling layer to obtain the lesion morphology feature map.

[0083] The input features of the branch are simultaneously fed into the lesion morphology feature extraction branch. This lesion morphology feature extraction branch contains multiple dilated convolutional layers, one global average pooling layer, and one batch normalization layer. The dilated convolutional layers, by setting the dilation rate in the convolutional kernel, can expand the receptive field without increasing the number of parameters, capturing a wider range of contextual information, which helps extract the overall morphological features of the lesion, such as its shape, size, and location. The batch normalization layer is inserted after the intermediate dilated convolutional layers to normalize the output of the dilated convolutions. The global average pooling layer performs global average pooling on the feature maps output by the dilated convolutional layers, converting each feature map into a single feature value, thus aggregating global information. Through these processes, the overall morphological features of the lesion can be extracted, resulting in a lesion morphology feature map.

[0084] Step S147: Synchronously input the branch input features into the lesion color feature extraction branch. Through color feature extraction of the convolutional layer and feature transformation processing of the fully connected layer, extract the color distribution features and color gradient features of the lesion to obtain the lesion color feature vector.

[0085] The input features of the branch are also simultaneously fed into the lesion color feature extraction branch. This lesion color feature extraction branch contains multiple convolutional layers, multiple fully connected layers, and one randomly dropped layer. The convolutional layers are used to extract color features from the image, capturing the distribution and variation information of colors through different convolutional kernels. The fully connected layers transform and integrate the color features extracted by the convolutional layers, mapping high-dimensional features to a low-dimensional space. The number of neurons in the fully connected layers decreases sequentially, gradually compressing the feature dimension. The randomly dropped layer is inserted between two adjacent fully connected layers, randomly dropping some neurons with a certain probability to prevent the model from overfitting. Through the above processing, the color distribution features of the lesion can be extracted, such as the composition ratio of colors, the variation in color intensity, and color gradation features, such as the color transition from the center to the edge of the lesion, ultimately obtaining the lesion color feature vector.

[0086] Step S148: Input the lesion texture feature map, lesion morphology feature map and lesion color feature vector into the lesion feature interaction layer. First, the lesion texture feature map and lesion morphology feature map are concatenated according to the channel dimension. Then, a spatial weight map is generated through the spatial attention mechanism. The concatenated feature map is spatially weighted through the spatial weight map. Then, a channel weight vector is generated through the channel attention mechanism. The spatially weighted feature map is channel-weighted through the channel weight vector to obtain the interaction feature map.

[0087] The lesion texture feature map, lesion morphology feature map, and lesion color feature vector output from the three branches are input into the lesion feature interaction layer. First, the lesion texture feature map and lesion morphology feature map are concatenated along the channel dimension, that is, the two feature maps are combined in the depth direction to form a higher-dimensional concatenated feature map. Then, a spatial attention mechanism is used to process the concatenated feature map. This mechanism focuses on the spatial regions in the feature map that are more important for insect and disease identification, generating a spatial weight map. Each element in the spatial weight map represents a weight value at its corresponding location; the higher the weight value, the more important the feature at that location. The spatial weight map is used to perform spatial weighting on the concatenated feature map, that is, element-wise multiplication, enhancing the features of important spatial regions and suppressing the features of unimportant regions. Next, a channel attention mechanism is used to process the spatially weighted feature map, which focuses on the importance of features in different channels, generating a channel weight vector. Each element in the channel weight vector represents the weight value of the corresponding channel. This vector is used to perform channel weighting on the spatially weighted feature map, that is, the feature map of each channel is multiplied by its corresponding weight value, highlighting the features of important channels. After dual weighting of spatial and channel features, an interactive feature map is obtained, which integrates texture, shape, and color features and highlights key information.

[0088] Step S149: Input the interactive feature map into the lesion feature aggregation layer, and obtain a fixed-length lesion aggregation feature vector through the in-depth extraction of the plant lesion-specific convolutional neural network block and the dimension unification processing of the adaptive pooling layer.

[0089] The interactive feature map is input into the lesion feature aggregation layer. This layer first inputs the interactive feature map into a plant lesion-specific convolutional neural network block, which contains multiple convolutional operations. These operations further refine the extraction of comprehensive lesion features, fusing features from different dimensions, such as more tightly integrating attention-weighted texture, morphology, and color features, and optimizing feature representation to make the features more discriminative. Then, an adaptive pooling layer processes the feature map output from the convolutional neural network block. This layer automatically adjusts the pooling region size according to a preset output size, converting feature maps of different sizes into fixed-length feature vectors. This results in a fixed-length aggregated lesion feature vector, which integrates multiple feature information from the lesion and is suitable for input into the subsequent pest and disease type matching layer for classification.

[0090] Step S1410: Input the lesion aggregation features into the insect and disease type matching layer. Through feature mapping of two fully connected layers and probability transformation of the probability activation function, obtain the probability value corresponding to each insect and disease type. Select the insect and disease type with the highest probability value as the preliminary insect and disease type result of the lesion association group.

[0091] The aggregated feature vector of lesions is input into the pest and disease type matching layer. This layer consists of two fully connected layers and a probabilistic activation function. The first fully connected layer maps the aggregated feature vector of lesions to a feature vector of an intermediate dimension, further integrating and transforming the features. The second fully connected layer maps the feature vector of the intermediate dimension to a dimension consistent with the number of known plant pest and disease types, with each dimension corresponding to one pest or disease type. Then, a probabilistic activation function (such as the softmax function) converts the output of the second fully connected layer into a probability value corresponding to each pest or disease type, with the sum of the probability values ​​being 1. The pest or disease type with the highest probability value is selected as the preliminary pest or disease type result for the lesion association group, making a preliminary judgment on which pest or disease type the lesion may belong to.

[0092] Step S150: Combine the temporal evolution information of the lesions on different organs to perform organ adaptation verification processing on the preliminary insect and disease type results, and generate the final plant insect and disease identification results and the temporal evolution map of the lesion organs.

[0093] Step S151: Extract the temporal evolution information of lesions in different organs for each lesion association group in the cross-stage lesion association set, and analyze the position offset trajectory, size expansion or contraction curve and feature data change matrix of the suspected lesion area in the organ coordinate system at each growth stage.

[0094] The temporal evolution information of lesions in different organs corresponding to each lesion association group was extracted from the cross-stage lesion association set. This temporal evolution information was analyzed to separate the positional offset trajectory of the suspected lesion region at each growth stage in the organ coordinate system. This positional offset trajectory consists of a series of coordinate points, reflecting the movement path of the lesion on the organ; the size expansion or contraction curve, which uses the growth stage as the horizontal axis and the lesion size (e.g., area) as the vertical axis, showing the trend of lesion size change over time; and the feature data change matrix, which records the numerical changes of various feature data (e.g., grayscale change features, color distribution features, etc.) of lesions at different growth stages.

[0095] Step S152: Obtain a preset temporal evolution feature library of plant insect diseases and organs. The temporal evolution feature library of plant insect diseases and organs contains standard temporal evolution rules of various known plant insect diseases and pests on different organs. Each insect disease and pest type corresponds to a set of standard evolution rules, including standard position offset trajectory, standard size change curve and standard feature data change rules on leaves. The standard evolution rules on the main stem and branches are similar.

[0096] A pre-defined temporal evolution feature database of plant pests and diseases in different organs was established through the analysis and summarization of a large number of samples of known plant pests and diseases. This database contains standard temporal evolution patterns of various known plant pests and diseases in different organs (leaves, main stem, branches). For each pest or disease type, there is a standard positional offset trajectory on the leaf, indicating where the pest or disease typically begins to appear and how it spreads; a standard size change curve, showing the typical expansion or contraction pattern of lesion size with different growth stages; and standard feature data change patterns, such as grayscale variations and color distribution characteristics, over time. Similar standard evolution patterns exist on the main stem and branches.

[0097] Step S153: Based on the organ type identifier corresponding to the lesion association group, match the standard temporal evolution law of each insect and disease type under the organ type from the plant insect and disease organ temporal evolution feature library, compare the positional offset trajectory of the lesion association group with the matched standard positional offset trajectory, and calculate the trajectory matching degree.

[0098] Based on the organ type identifier corresponding to the lesion association group, such as leaf, main stem, or branch, the standard temporal evolution patterns of all known insect and disease types under that organ type are matched from the plant insect and disease organ temporal evolution feature database. Then, the positional offset trajectory of the current lesion association group is compared with the standard positional offset trajectory of each matched insect and disease type. During the comparison, the distance between the coordinate points of corresponding growth stages on the two trajectories is calculated, and the average of these distances is used as a measure of trajectory difference. The trajectory difference is then converted into a trajectory matching degree; the smaller the difference, the higher the matching degree. The trajectory matching degree reflects the similarity between the positional change of the current lesion and the positional change of the standard insect and disease.

[0099] Step S154: Compare the size expansion or contraction curve with the standard size change curve to calculate the size matching degree; compare the feature data change matrix with the standard feature data change pattern to calculate the feature matching degree.

[0100] Similarly, the size expansion or contraction curves of the lesion-associated group are compared with the standard size change curves. The difference between the size values ​​of the two curves at the corresponding growth stages is calculated, and the average difference is converted into a size matching degree; the smaller the difference, the higher the size matching degree. For the feature data change matrix, it is compared with the standard feature data change pattern. For each feature index in the matrix, the similarity between its change trend with the growth stage and the change trend of the corresponding index in the standard pattern is calculated, such as by calculating the correlation coefficient to measure the similarity. Then, the similarity of all feature indices is weighted and averaged to obtain the feature matching degree. The feature matching degree reflects the degree of similarity between the feature changes of the current lesion and the feature changes of the standard insect disease.

[0101] Step S155: Assign organ adaptation weights to trajectory matching degree, size matching degree, and feature matching degree. Calculate the comprehensive temporal matching degree by weighted summation. The trajectory matching degree of leaf lesions has the highest weight, the size matching degree of main stem lesions has the highest weight, and the feature matching degree of branch lesions has the highest weight.

[0102] Based on the characteristics of different organs, different organ-matching weights were assigned to trajectory matching, size matching, and feature matching. For leaf lesions, trajectory matching is of high importance for identification due to the relatively obvious movement of pests and diseases on the leaves; therefore, it has the highest weight. For main stem lesions, the size changes are usually more significant; for example, the stem rot caused by stem borers gradually expands, so size matching has the highest weight. For branch lesions, the changes in their feature data (such as color and texture) may be more unique, so feature matching has the highest weight. Then, trajectory matching, size matching, and feature matching are multiplied by their respective weights, and the products are summed to obtain the comprehensive temporal matching score. The comprehensive temporal matching score comprehensively considers the degree of matching between the lesion and the standard pests and diseases in terms of location, size, and features.

[0103] Step S156: If the overall temporal matching degree is higher than the organ adaptation verification threshold, then the preliminary insect and disease type result of the lesion association group is confirmed as a valid insect and disease type result.

[0104] An organ adaptation verification threshold is set, which is determined based on a large amount of experimental data and experience. The calculated comprehensive temporal matching degree is compared with this threshold. If the comprehensive temporal matching degree is higher than the threshold, it indicates that the temporal evolution characteristics of the current lesion association group are highly matched with the standard temporal evolution law corresponding to the preliminary insect-disease type result, and the preliminary result is reliable. Therefore, the preliminary insect-disease type result is confirmed as a valid insect-disease type result.

[0105] Step S157: If the overall temporal matching degree is not higher than the organ adaptation verification threshold, the lesion aggregation features of the lesion association group, the corresponding organ type identifier, and the growth stage information are input into the dynamic optimization module. After the model parameters are fine-tuned, the type matching is performed again to obtain the secondary identification type result. The organ adaptation verification step is repeated until the effective pest and disease type result is obtained.

[0106] If the overall temporal matching degree is not higher than the organ adaptation verification threshold, it indicates that the preliminary insect-disease type result may be inaccurate and further processing is required. The lesion aggregation features of the associated lesion group, along with the corresponding organ type identifier and growth stage information, are input into the model's dynamic optimization module. Upon receiving this information, the dynamic optimization module triggers model parameter fine-tuning. During fine-tuning, the model is retrained in small batches using samples with uncertainties, adjusting the weight parameters of each layer to improve the model's ability to identify this type of lesion. After fine-tuning, the complete temporal lesion image sequence of the associated lesion group is re-input into the optimized model for type matching, obtaining a secondary identification type result. Then, the organ adaptation verification process from S151 to S156 is repeated, calculating the overall temporal matching degree of the secondary identification type result and comparing it with the threshold. If the result is still not satisfactory, model parameter fine-tuning is triggered again, and the matching and verification steps are repeated until a valid insect-disease type result with an overall temporal matching degree higher than the threshold is obtained.

[0107] Step S158: Collect the effective insect and disease type results of all lesion association groups, associate the corresponding growth stage time identifier, organ type identifier, organ region coordinate range information and characteristic data of suspected lesion areas, and construct a prediction association table, wherein the prediction association table contains a unique identifier for each record.

[0108] Collect all valid insect and disease type results for lesion association groups after organ adaptation verification. Associate each valid insect and disease type result with its corresponding growth stage time marker (e.g., a specific year and month of heading), organ type marker (leaf, main stem, or branch), organ region coordinate range information (the organ's coordinates in the original image), and characteristic data of the suspected lesion region (color distribution histogram data, texture distribution feature data, shape parameter data, etc.). Assign a unique identifier, such as a unique string or number, to each association record to distinguish different lesion association groups. Compile the above association information into a predictive association table, which clearly displays the detailed information and identification results for each lesion association group.

[0109] Step S159: Construct a map coordinate system with growth stage as the horizontal axis and plant organ type as the vertical axis. The horizontal axis is labeled in the order of growth stage, and the vertical axis is labeled in the order of leaf, main stem, and branch. In the map coordinate system, draw the outline of each organ according to the coordinate range information of the organ region. The shape of the outline is consistent with the actual shape of the corresponding organ.

[0110] A coordinate system for the temporal evolution of diseased organs is constructed. The growth stage is plotted on the horizontal axis, labeled sequentially as seed germination, tillering, jointing, heading, and grain-filling stages. The plant organ type is plotted on the vertical axis, labeled sequentially as leaf, main stem, and branch. Based on the coordinate range information of each organ region recorded in the prediction correlation table, the outline of each organ is drawn in the coordinate system. For example, the outline of the leaf is drawn based on the coordinate range of the leaf region at the corresponding vertical axis position and the corresponding horizontal axis position of the growth stage. The shape of the outline corresponds to the actual morphology of a wheat leaf at that growth stage. The outlines of the main stem and branches are drawn similarly.

[0111] Step S1510: Within the outline of each organ, draw lesion markers according to the location coordinates of the suspected lesion area. The color of the lesion marker corresponds one-to-one with the effective pest and disease type results. The size of the lesion marker is proportional to the size of the suspected lesion area. The lesion marker is marked with a unique identifier of the lesion association group.

[0112] Within each organ outline marker, the specific location of the lesion is determined based on the coordinates of the suspected lesion area in the organ coordinate system, and a lesion marker is drawn. The color of the lesion marker corresponds one-to-one with the effective pest / disease type result; for example, the lesion marker for wheat rust is set to red, and powdery mildew is set to white, etc., to visually distinguish different pest / disease types through color. The size of the lesion marker is proportional to the size of the suspected lesion area; the larger the area, the larger the marker, reflecting the severity of the lesion. A unique identifier for the associated group of the lesion is marked within each lesion marker, facilitating the query of detailed information about the lesion through the prediction association table.

[0113] Step S1511: Connect the lesion identifiers of the same lesion association group at different growth stages in the atlas with a broken line. The thickness of the broken line is proportional to the organ association coefficient of the lesion association group. Mark the key nodes of feature data changes next to the broken line.

[0114] For the same lesion association group, lesion markers at different growth stages are connected by broken lines in the atlas according to their chronological order of growth. The thickness of the broken line is proportional to the organ correlation coefficient of the lesion association group; the larger the organ correlation coefficient, the stronger the correlation between lesions at different growth stages, and the thicker the broken line. Next to the broken line, key nodes of feature data changes are marked, such as growth stages where lesion size begins to expand rapidly or growth stages where color characteristics change significantly. These key nodes reflect important changes in the lesion evolution process.

[0115] Step S1512: Add a legend to one side of the map, marking the colors corresponding to different effective pest and disease types, the size range corresponding to the size of the lesion identifier, the range of organ correlation coefficients corresponding to the thickness of the broken line, and the meaning of the unique identifier. Combine the prediction correlation table with the completed map to form the final plant pest and disease identification result.

[0116] A legend is added to one side of the atlas, detailing the colors corresponding to different effective pest and disease types, such as red for wheat rust and white for powdery mildew; the size range corresponding to the lesion identifier size, such as the range for large identifiers, medium identifiers, etc.; the organ correlation coefficient range corresponding to the line thickness, such as the range for thick lines, medium lines, etc.; and the meaning of the unique identifier, such as its components and the information they represent. Combining the predicted correlation table with the completed lesion organ time-series evolution atlas forms the final plant pest and disease identification result. This result not only identifies the pest and disease type but also demonstrates the evolution of lesions in different organs and at different growth stages.

[0117] Step S210: The process of constructing and optimizing the pre-defined deep learning model for plant pests and diseases.

[0118] Step S211: Construct the basic model architecture, which includes a feature input layer, a parallel plant lesion feature extraction branch, a lesion feature interaction layer, a lesion feature aggregation layer, an insect and disease type matching layer, and a dynamic optimization module. The parallel plant lesion feature extraction branches are a lesion texture detail extraction branch, a lesion morphology feature extraction branch, and a lesion color feature extraction branch.

[0119] The basic architecture for constructing a deep learning model specifically for plant pests and diseases consists of multiple functional layers and modules. First, there's the feature input layer, responsible for receiving input image data and performing initial dimensionality transformation. Next is the parallel plant lesion feature extraction branch, containing three independent branches: lesion texture detail extraction, lesion morphology feature extraction, and lesion color feature extraction, each extracting lesion features from different perspectives. Following this is the lesion feature interaction layer, used to fuse the features extracted from the three branches and highlight key information. The lesion feature aggregation layer further extracts and unifies the dimensions of the interacting features. The pest and disease type matching layer classifies pest and disease types and outputs probabilities based on the aggregated features. A dynamic optimization module triggers fine-tuning of model parameters when the model's recognition results are uncertain, improving recognition accuracy. These layers and modules collaborate to form the complete model infrastructure.

[0120] Step S212: Configure the network parameters of the lesion texture detail extraction branch. The lesion texture detail extraction branch includes multiple convolutional layers, multiple pooling layers, and two batch normalization layers. The kernel size of the multiple convolutional layers increases sequentially, and the stride remains fixed. The pooling layers use max pooling, and the pooling kernel size is set to a fixed value. The batch normalization layers are inserted between two adjacent convolutional layers.

[0121] Configure the network parameters for the lesion texture detail extraction branch. This branch contains multiple convolutional layers, with the kernel size starting small and increasing progressively with each layer to gradually extract more complex texture features. The stride of all convolutional layers remains constant to ensure the feature map size shrinks systematically. Max pooling is used, which preserves salient texture features, and the kernel size is set to a fixed value, such as 2x2. Two batch normalization layers are inserted between adjacent convolutional layers to accelerate network convergence and prevent overfitting. With these parameter configurations, this lesion texture detail extraction branch can effectively extract the surface texture details of lesions.

[0122] Step S213: Configure the network parameters of the lesion morphology feature extraction branch. The lesion morphology feature extraction branch includes multiple dilated convolutional layers, one global average pooling layer, and one batch normalization layer. The porosity of the dilated convolutional layers increases sequentially at a fixed ratio. The pooling kernel size of the global average pooling layer is set to a fixed value. The batch normalization layer is inserted after the intermediate dilated convolutional layers.

[0123] Configure the network parameters for the lesion morphology feature extraction branch. This branch uses multiple dilated convolutional layers, with the dilation rate of each layer increasing sequentially at a fixed ratio. This increased dilation rate expands the receptive field, enabling the capture of lesion morphology information over a wider area. A global average pooling layer with a fixed kernel size (e.g., the same as the feature map size output by the last dilated convolutional layer) is used to aggregate global information from the entire feature map, obtaining global morphological features. A batch normalization layer is inserted after the intermediate dilated convolutional layers to standardize the output and stabilize the training process. These parameter configurations enable this branch to effectively extract the overall morphological features of the lesions.

[0124] Step S214: Configure the network parameters of the lesion color feature extraction branch. The lesion color feature extraction branch includes multiple convolutional layers, multiple fully connected layers and one random dropout layer. The number of neurons in the fully connected layers decreases sequentially. The dropout probability of the random dropout layer is set to a fixed value and is inserted between two adjacent fully connected layers.

[0125] Configure the network parameters for the lesion color feature extraction branch. This branch contains multiple convolutional layers for initial color feature extraction. Subsequent layers are fully connected, with the number of neurons decreasing sequentially to gradually compress the feature dimension. A random dropout layer is inserted between adjacent fully connected layers, with a fixed dropout probability (e.g., 0.5) to prevent overfitting. With these parameter configurations, this lesion color feature extraction branch can extract and integrate the color distribution and gradation features of the lesions.

[0126] Step S215: Configure the feature interaction strategy of the lesion feature interaction layer. The feature interaction strategy is to stitch the feature maps output by the parallel plant lesion feature extraction branch according to the channel dimension, generate a spatial weight map through a spatial attention mechanism, perform spatial weighting processing on the stitched feature map through the spatial weight map, generate a channel weight vector through a channel attention mechanism, and perform channel weighting processing on the spatially weighted feature map through the channel weight vector to obtain the interaction feature map.

[0127] A feature interaction strategy for configuring the lesion feature interaction layer is implemented. First, the feature maps output from the three parallel branches are concatenated along the channel dimension to form a concatenated feature map. Then, a spatial attention mechanism is applied, generating a spatial weight map by performing global average pooling and a fully connected layer on the concatenated feature map. This spatial weight map is used to spatially weight the concatenated feature map, enhancing the features of important spatial regions. Next, a channel attention mechanism is applied, generating a channel weight vector by performing global average pooling and a fully connected layer on the spatially weighted feature map. This channel weight vector is used to channel-weight the feature map, highlighting the features of important channels. Through this dual spatial and channel weighting interaction strategy, an interactive feature map incorporating key information is obtained.

[0128] Step S216: Configure the aggregation method of the lesion feature aggregation layer. The aggregation method is to input the interactive feature map output by the feature interaction layer into the convolutional neural network block. The convolutional neural network block contains multiple convolution operations to deepen the extraction of comprehensive lesion features, fuse lesion features of different dimensions, and optimize feature expression. Then, the feature map is converted into a fixed-length lesion aggregation feature vector through an adaptive pooling layer.

[0129] Configure the aggregation method of the lesion feature aggregation layer. The interactive feature map is input into a plant lesion-specific convolutional neural network block. This block contains multiple convolutional operations, with the kernel size and number set according to feature extraction requirements. These convolutional operations further deepen the extraction of comprehensive lesion features, fusing lesion features from different dimensions (texture, morphology, color) and optimizing feature representation to make the features more discriminative. Then, an adaptive pooling layer processes the feature map output from the convolutional neural network block. This layer automatically adjusts the pooling operation based on the preset output length, converting feature maps of different sizes into fixed-length lesion aggregation feature vectors for subsequent type matching.

[0130] Step S217: Configure the output strategy of the pest and disease type matching layer. The pest and disease type matching layer includes two fully connected layers and a probabilistic activation function. The first fully connected layer maps the fixed-length lesion aggregate feature vector to the feature vector of the intermediate dimension. The second fully connected layer maps it to the dimension consistent with the number of known plant pest and disease types. The probabilistic activation function converts the output into the probability value corresponding to each pest and disease type.

[0131] Configure the output strategy of the pest and disease type matching layer. This layer consists of two fully connected layers. The first fully connected layer receives a fixed-length aggregated feature vector of lesions and maps it to a feature vector of an intermediate dimension. The size of the intermediate dimension is determined based on the number of pest and disease types and the feature complexity. The second fully connected layer maps the feature vector of the intermediate dimension to a dimension consistent with the number of known plant pest and disease types, with each dimension corresponding to one pest or disease type. Finally, a probabilistic activation function (such as the softmax function) is used to process the output of the second fully connected layer, converting it into a probability value corresponding to each pest or disease type. The probability value represents the model's confidence in that pest or disease type.

[0132] Step S218: Configure the optimization rules of the dynamic optimization module. The optimization rules are to collect the deviation data between the probability value of the insect and disease type output by the model and the insect and disease type labeled in the sample in real time. When the deviation value of a fixed number of consecutive samples is higher than the deviation threshold, the model parameters are fine-tuned.

[0133] Configure the optimization rules for the dynamic optimization module. The dynamic optimization module collects real-time deviation data between the probability values ​​of insect and disease types output by the model and the actual insect and disease types labeled on the samples. This deviation data can be measured by calculating the cross-entropy loss between the predicted probability distribution and the true label distribution. When the deviation value of a fixed number of consecutive samples exceeds a set deviation threshold, it indicates that the model's performance on the current recognition task has deteriorated or that there are insufficiently learned sample patterns. At this point, model parameter fine-tuning is triggered. During fine-tuning, these high-deviation samples are used to form a mini-batch dataset, and the model is trained iteratively for several rounds, adjusting the weight parameters of each layer to reduce deviation and improve model performance.

[0134] Step S219: Collect a labeled sample set containing different types of plant pests and diseases. The labeled sample set contains lesion images of each type of pest and disease in different plant organs and at different growth stages. Each sample is labeled with the name of the pest or disease type, the type of organ where the lesion is located, the growth stage of the lesion, and the severity level of the lesion.

[0135] A large set of labeled samples containing various types of plant pests and diseases was collected. The sample set should cover a variety of common plant pests and diseases, such as wheat rust, powdery mildew, and aphid damage. For each pest or disease type, images of lesions in different plant organs (leaves, main stem, branches) and at different growth stages (seed germination to grain filling) were collected. Each sample was detailed and labeled, including the pest or disease type name, the type of organ containing the lesion, the growth stage of the lesion, and the severity level of the lesion (e.g., mild, moderate, severe).

[0136] Step S2110: Divide the labeled sample set into a training set, a validation set, and a test set according to a fixed ratio. The training set contains samples of different insect and disease types, different organ types, and growth stages. The sample distribution of the validation set and the test set is consistent with that of the training set.

[0137] The labeled sample set is divided into training, validation, and test sets in a fixed ratio, such as 7:2:1. During this division, it is ensured that the training set includes samples from all different insect / disease types, organ types, and growth stages, and that the proportions of each type are as balanced as possible. The sample distribution in the validation and test sets is consistent with that of the training set, meaning the proportions of insect / disease types, organ types, and growth stages are the same to ensure the reliability of the validation and testing results. The training set is used for model parameter learning, the validation set is used for model performance evaluation and parameter tuning during training, and the test set is used for the final evaluation of the model's generalization ability.

[0138] Step S2111: Iteratively train the model using the training set. After each iteration, verify the model performance using the validation set. When the validation metrics of the validation set remain stable for a fixed number of consecutive iterations, stop training. The validation metrics include the accuracy of identifying insect and disease types, the accuracy of locating organ lesions, and the accuracy of matching lesions at different growth stages.

[0139] Step S21111: Perform data augmentation processing on the training set, including random horizontal flipping, random vertical flipping, random rotation within a fixed angle range, random cropping of a fixed proportion area, and random adjustment of brightness within a fixed range, so that the number of samples in the expanded training set is a fixed multiple of the original number.

[0140] Data augmentation is performed on the training set to increase sample diversity and improve the model's generalization ability. Data augmentation operations include random horizontal flipping (flipping the image horizontally with a 50% probability); random vertical flipping (also with a certain probability); random rotation within a fixed angle range, such as between -30 degrees and 30 degrees; random cropping of a fixed proportion of the image as new samples; and random brightness adjustment within a fixed range. Through these data augmentation operations, the number of samples in the expanded training set is a fixed multiple of the original number, such as three times, thus enriching the training data.

[0141] Step S21112: Divide the expanded training set into training batches according to a fixed batch size. Each training batch contains a fixed number of samples, which are evenly distributed according to insect / disease type, organ type, and growth stage.

[0142] The expanded training set is divided into multiple training batches with a fixed batch size, each containing a fixed number of samples, such as 32 samples. When dividing the batches, it is ensured that the samples in each batch are evenly distributed according to insect / disease type, organ type, and growth stage, avoiding a batch with too many sample types. This allows the model to learn diverse features in each iteration, improving training effectiveness.

[0143] Step S21113: Initialize the weight parameters of each layer of the model, and initialize the weights of the convolutional layer and the fully connected layer. In this process, the initial weight distribution is adjusted according to the statistical law of plant lesion characteristics, and the bias parameter is initialized to a fixed small value.

[0144] Initialize the weight parameters of each layer in the model. For the weights of convolutional and fully connected layers, initialize them according to the statistical regularities of plant lesion characteristics. For example, by analyzing the feature distribution of a large number of plant lesion images, adjust the initial distribution of the weights to make the initial weights closer to the direction of the optimal solution. Initialize the bias parameters of all layers to fixed small values, such as 0.01, to avoid the neuron output being zero, which helps the network start up stably.

[0145] Step S21114: Set training parameters, set the total number of iterations to a fixed value, set the initial value of the learning rate to a fixed value, and adopt an adaptive learning rate adjustment strategy. When the validation metric of the validation set does not improve for a fixed number of consecutive times, the learning rate is reduced by a fixed proportion, and the minimum learning rate is set to a fixed proportion of the initial value.

[0146] Configure the model training parameters. Set the total number of iterations to a fixed, large value, such as 100. Set the initial learning rate to a fixed value, such as 0.001. Employ an adaptive learning rate adjustment strategy, such as the Adam optimizer's adaptive learning rate strategy. When the validation metric (such as accuracy) on the validation set does not improve for a fixed number of consecutive iterations (such as 5), reduce the learning rate by a fixed percentage (such as 0.1 times) to fine-tune the parameters. Set the minimum learning rate to a fixed percentage of the initial value (such as 0.0001 times) to prevent training from stalling due to an excessively low learning rate.

[0147] Step S21115: Start iterative training. In each iteration, each training batch is input sequentially. The batch samples are input into the model. Through parallel processing of the lesion texture detail extraction branch, the lesion morphology feature extraction branch, and the lesion color feature extraction branch, the feature output of each branch is obtained. After processing by the lesion feature interaction layer and the lesion feature aggregation layer, the lesion aggregated features are obtained. Then, through the insect and disease type matching layer, the probability values ​​corresponding to each insect and disease type are output.

[0148] The iterative training process of the model begins. In each iteration, each training batch is input into the model sequentially. The batch samples are first processed by the feature input layer, converting them into branch input features. Then, the branch input features are input in parallel to three feature extraction branches. After convolution, pooling, batch normalization, and other processing in each branch, lesion texture feature maps, lesion morphology feature maps, and lesion color feature vectors are obtained. The outputs of these features are input to the lesion feature interaction layer, where spatial and channel attention weighting is applied to obtain the interaction feature map. The interaction feature map is input to the lesion feature aggregation layer, where it undergoes in-depth extraction by convolutional neural network blocks and processing by adaptive pooling layers to obtain a fixed-length aggregated lesion feature vector. Finally, the aggregated lesion feature vector is input to the pest and disease type matching layer, where it undergoes mapping by two fully connected layers and processing by probabilistic activation functions to output the probability values ​​corresponding to each pest and disease type.

[0149] Step S21116: Calculate the loss value between the output probability value and the insect / disease type labeled in the sample using the cross-entropy loss function. The cross-entropy loss function is calculated based on the correspondence between the probability value of each sample and the actual insect / disease type. The position label corresponding to the actual insect / disease type is set to 1, and other types are set to 0. The loss value is backpropagated to each layer of the model using the backpropagation algorithm. The gradient descent optimization algorithm is used to adjust the weight parameters and bias parameters of each layer. The step size of the gradient descent is determined by the current learning rate.

[0150] For each training batch, the cross-entropy loss function is used to calculate the loss between the probability value output by the model and the actual insect / disease type labeled in the sample. The cross-entropy loss function is calculated based on the difference between the predicted probability distribution and the true label distribution for each sample. The position label corresponding to the actual insect / disease type is set to 1, and other types are set to 0, forming a one-hot vector. The larger the loss value, the greater the difference between the prediction and the reality. The loss value is backpropagated from the output layer to each layer of the model using the backpropagation algorithm, calculating the gradient of the weight and bias parameters of each layer with respect to the loss value. The gradient descent optimization algorithm is used to adjust the weight and bias parameters of each layer based on the calculated gradient and the current learning rate, gradually reducing the loss value. The step size of gradient descent is determined by the current learning rate; a larger learning rate allows for a larger step size and faster parameter updates.

[0151] Step S21117: After each iteration cycle, input the validation set into the trained model to obtain the insect and disease type identification results, lesion location results, and growth stage matching results of the validation set; calculate the insect and disease type identification accuracy, which is the proportion of the number of correctly identified samples to the total number of samples in the validation set; calculate the organ lesion location accuracy, which is the proportion of the number of samples whose deviation between the lesion location coordinates and the labeled coordinates is within the allowable range to the total number of samples in the validation set; calculate the lesion matching accuracy at different growth stages, which is the proportion of the number of correctly matched lesion association groups at different growth stages to the total number of lesion association groups in the validation set.

[0152] After each iteration (i.e., training all training batches once), the validation set is input into the model in the current training state to obtain the pest / disease type identification results, lesion localization results (coordinates of the lesion on the organ), and growth stage matching results of the validation set. Validation metrics are calculated: pest / disease type identification accuracy is the proportion of correctly identified samples out of the total number of samples in the validation set; organ lesion localization accuracy is the proportion of samples whose lesion localization coordinates deviate from the labeled coordinates within an acceptable range (e.g., pixel distance less than a certain value) out of the total number of samples in the validation set; lesion matching accuracy at different growth stages is the proportion of the number of the same lesion association group correctly matched as the same lesion at different growth stages out of the total number of lesion association groups in the validation set. These metrics evaluate the model's performance from different perspectives.

[0153] Step S21118: Record the verification index for each iteration cycle, and determine whether the fluctuation range of the verification index for a fixed number of consecutive iterations is lower than its respective fluctuation threshold. If it is lower than the fluctuation threshold, the verification index is determined to be stable and training is stopped.

[0154] Record the validation metric values ​​for each iteration cycle, such as the accuracy rate of insect and disease type identification, the accuracy rate of organ lesion localization, and the accuracy rate of lesion matching at different growth stages. Set a fluctuation threshold for each validation metric. When, in a fixed number of consecutive iteration cycles (e.g., 3), the fluctuation range of all validation metrics (i.e., the absolute value of the difference between two adjacent iterations) is lower than their respective fluctuation thresholds, the validation metrics are considered stable, and the model has converged. At this point, stop model training and save the current model parameters as the final training result.

[0155] Step S2112: Use the test set to test the performance of the trained model. If the accuracy of the identification of pest and disease types, the accuracy of organ lesion localization, and the accuracy of lesion matching at different growth stages in the test set all meet the set standards, then the construction of the plant pest and disease-specific deep learning model is completed; if not, adjust the network parameters of each branch and retrain until the test is passed.

[0156] The trained model is tested using an independent test set. Test set samples are input into the model to obtain test results. The accuracy rates for identifying pest / disease types, lesion localization, and lesion matching at different growth stages are calculated. These test metrics are compared with set standards. If the accuracy reaches 90% or higher, and all metrics meet the standards, the construction of a deep learning model specifically for plant pests and diseases is complete. If the standards are not met, the test results are analyzed to identify the reasons for the model's performance deficiencies, such as low accuracy in identifying certain pest / disease types or inaccurate localization of specific organs. Based on the analysis results, the network parameters of each branch are adjusted, such as increasing the number of convolutional layers, adjusting the hole ratio, and changing the number of neurons in fully connected layers. Then, the model is retrained and validated using the training and validation sets until all metrics on the test set meet the set standards.

[0157] Based on the same inventive concept, please refer to Figure 2 The diagram shows a schematic block diagram of a deep learning-based plant pest and disease image recognition system 100 provided in this application embodiment for performing the above-described deep learning-based plant pest and disease image recognition method. The deep learning-based plant pest and disease image recognition system 100 may include a communication unit 110, a machine-readable storage medium 120, and a processor 130.

[0158] In this embodiment, both the machine-readable storage medium 120 and the processor 130 are located in the deep learning-based plant and insect disease image recognition system 100 and are separately configured. Alternatively, the machine-readable storage medium 120 can also be integrated into the processor 130 and can communicate and interact with external systems through the communication unit 110. 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 deep learning-based plant and insect disease image recognition method provided in the aforementioned method embodiments.

[0159] 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 plant pest image recognition method based on deep learning, characterized in that, The method comprises: acquiring a multi-source image set of plants at multiple growth stages, the multi-source image set containing whole-plant images, leaf local images, stem local images, and growth stage time identification and shooting angle identification information of each growth stage; constructing a plant image time sequence association network based on the growth stage time identification and shooting angle identification information, performing organ-level time sequence alignment processing on the multi-source image set to obtain an organ-level time sequence alignment image set and an organ association coefficient between images; performing cross-growth-stage lesion feature association processing on the organ-level time sequence alignment image set through the organ association coefficient to obtain a cross-stage lesion association set and time sequence evolution information of lesions on different organs; inputting the cross-stage lesion association set into a preset plant pest and disease specific deep learning model to perform feature aggregation and type matching processing, obtaining lesion aggregation features and preliminary pest and disease type results, the model base architecture of the plant pest and disease specific deep learning model containing a feature input layer, parallel plant lesion feature extraction branches, a lesion feature interaction layer, a lesion feature aggregation layer, a pest and disease type matching layer, and a dynamic optimization module, the parallel plant lesion feature extraction branches being a lesion texture detail extraction branch, a lesion morphological feature extraction branch, and a lesion color feature extraction branch, the lesion texture detail extraction branch containing multiple convolution layers, multiple pooling layers, and two layers of batch normalization layers, the lesion morphological feature extraction branch containing multiple hollow convolution layers, a global average pooling layer, and a batch normalization layer, and the lesion color feature extraction branch containing multiple convolution layers, multiple fully connected layers, and a random dropout layer; performing organ adaptation verification processing on the preliminary pest and disease type results in combination with the time sequence evolution information of lesions on different organs to generate a final plant pest and disease identification result and a lesion organ time sequence evolution map. 2.The deep learning-based plant pest image recognition method of claim 1, wherein, The method comprises: extracting the growth stage time identification and shooting angle identification information of each image in the multi-source image set, performing primary sorting on all images in the order of growth stages, and performing secondary sorting on the same growth stage according to the shooting angle identification information to obtain an initial time sequence angle image sequence; extracting plant organ contour features of each image in the initial time sequence angle image sequence, the plant organ contour features containing main stem contour features, leaf contour features, and stem branch contour features, the main stem contour features containing main stem contour point coordinates and contour curvature change values, the leaf contour features containing leaf contour point coordinates, leaf vein contour trends, and leaf edge serration densities, and the stem branch contour features containing branch contour point coordinates and branch connection angles with the main stem; The image is taken as a network node, and the similarity of the plant organ contour features of the images of the same shooting angle in adjacent growth stages is taken as an initial edge weight to construct an initial time sequence correlation network. The initial edge weight is calculated by weighting the standardized values of the stem contour overlap ratio, the leaf contour overlap ratio, and the branch connection angle deviation value. The branch connection angle deviation value is converted into a deviation score comparable to the contour overlap ratio. The higher the stem contour overlap ratio and the leaf contour overlap ratio, and the higher the branch connection angle deviation score, the greater the initial edge weight. The indirect correlation weight of the images of the same shooting angle in non-adjacent growth stages in the initial time sequence correlation network is calculated. The indirect correlation weight is determined by the sum of the products of the initial edge weights of all paths between two nodes. The path selection rule is that the interval between the growth stages corresponding to the nodes in the path does not exceed N, and the shooting angle identification information is consistent. N is a positive integer set. The initial edge weight and the indirect correlation weight are integrated into an organ correlation coefficient between nodes to obtain a plant image time sequence correlation network. The greater the organ correlation coefficient, the stronger the correlation of the corresponding organs in the two images. Based on the organ correlation coefficient in the plant image time sequence correlation network, reference images with organ correlation coefficients higher than the correlation threshold are selected. The central axis of the main stem and the main vein line of the leaf in the reference image are taken as the calibration reference to perform coordinate calibration processing on the corresponding organs in the current image. The main reference line of the same organ in different images is made to reach the calibration standard, and an organ-level time sequence alignment image set is obtained. 3.The deep learning-based plant pest image recognition method of claim 1, wherein, The organ-level time sequence alignment image set is processed based on the organ correlation coefficient to obtain a cross-stage lesion correlation set and time sequence evolution information of the lesion on different organs, including: Plant organ segmentation processing is performed on each image in the organ-level time sequence alignment image set. A segmentation method based on plant organ morphology templates is used to segment the image into leaf region images, main stem region images, and branch region images. The coordinate range information and organ type identification of each organ region image are recorded. Lesion preliminary detection processing is performed on each organ region image. A sliding window detection method adapted to the morphology of the plant organ is used. The window shape is determined according to the organ type. An oval window is used for the leaf region, a rectangular window is used for the main stem region, and an elongated window is used for the branch region. The window is moved by a set step to traverse the organ region image, and the gray scale change features and color distribution features in the window are extracted. Image samples of healthy organs of the same plant variety are collected. The gray scale change features and color distribution features of the corresponding organ regions in the image samples are extracted. The mean value range and distribution rule of the sample features are calculated to determine the normal feature range of each organ. Window regions with gray scale change features or color distribution features deviating from the normal feature range of the corresponding organ are selected as suspected lesion regions. The position coordinates, size information, and corresponding feature data of each suspected lesion region are recorded to form an organ suspected lesion information table for each image. The feature data includes color distribution histogram data, texture distribution feature data, and shape parameter data. According to the organ correlation coefficient, an associated image with an organ correlation coefficient higher than a correlation threshold is screened out from the current image, and organ suspected lesion information corresponding to the same organ type identifier in the associated image is extracted as associated lesion information; The position correspondence and feature data similarity of the suspected lesion regions of the current image and the associated image are compared, the position correspondence is determined by the coordinate overlap ratio of the two suspected lesion regions in the organ coordinate system, and the feature data similarity is comprehensively calculated by the color distribution histogram overlap degree, the texture distribution feature matching ratio, and the standardized value of the shape parameter deviation degree, wherein the shape parameter deviation degree is converted into a deviation score comparable to the overlap degree and the matching ratio, the higher the color distribution histogram overlap degree, the higher the texture distribution feature matching ratio, and the higher the shape parameter deviation score, the higher the feature data similarity; The suspected lesion regions that satisfy the correlation conditions in terms of the position correspondence and the feature data similarity are divided into the same lesion correlation group, and each lesion correlation group contains associated suspected lesion regions on the same organ in different growth stages; The position coordinate change trajectory, size change trend, and feature data change rule of each suspected lesion region in each lesion correlation group are extracted, and time sequence evolution information of the lesion on different organs is generated, the time sequence evolution information includes the position offset trajectory, the size expansion or contraction curve, and the feature data change matrix of the lesion in the organ coordinate system; All lesion correlation groups and the suspected lesion regions contained in each group are combined into a cross-stage lesion correlation set, and the time sequence evolution information of the lesion in different organs and the corresponding organ type identifier of each group are recorded. 4.The deep learning-based plant pest image recognition method of claim 1, wherein, The construction and optimization process of the preset plant pest disease specific deep learning model includes: A model infrastructure is constructed, the model infrastructure includes a feature input layer, parallel plant lesion feature extraction branches, a lesion feature interaction layer, a lesion feature aggregation layer, a pest disease type matching layer, and a dynamic optimization module, the parallel plant lesion feature extraction branches are a lesion texture detail extraction branch, a lesion morphology feature extraction branch, and a lesion color feature extraction branch; The network parameters of the lesion texture detail extraction branch are configured, the lesion texture detail extraction branch includes multiple convolution layers, multiple pooling layers, and two batch normalization layers, the convolution kernel size of the multiple convolution layers increases in turn, the step size remains a fixed value, the pooling layer adopts a maximum pooling method, and the pooling kernel size is set to a fixed value, and the batch normalization layer is inserted between adjacent two convolution layers; The network parameters of the lesion morphology feature extraction branch are configured, the lesion morphology feature extraction branch includes multiple hollow convolution layers, a global average pooling layer, and a batch normalization layer, the hollow rate of the hollow convolution layer increases in turn by a fixed proportion, the pooling kernel size of the global average pooling layer is set to a fixed value, and the batch normalization layer is inserted after the middle hollow convolution layer; Network parameters of the lesion color feature extraction branch are configured, the lesion color feature extraction branch includes multiple layers of convolutional layers, multiple layers of fully connected layers, and a layer of random dropout layer, the number of neurons of the fully connected layers decreases in turn, the dropout probability of the random dropout layer is set to a fixed value, and the random dropout layer is inserted between adjacent two fully connected layers; A feature interaction strategy of the lesion feature interaction layer is configured, the feature interaction strategy is that feature maps output by the parallel plant lesion feature extraction branch are spliced in the channel dimension, a spatial weight map is generated through a spatial attention mechanism, the spliced feature maps are spatially weighted through the spatial weight map, a channel weight vector is generated through a channel attention mechanism, and the spatially weighted feature maps are channel weighted through the channel weight vector to obtain an interaction feature map; An aggregation mode of the lesion feature aggregation layer is configured, the aggregation mode is that the interaction feature map output by the feature interaction layer is input into a convolutional neural network block, the convolutional neural network block includes multiple layers of convolutional operations, is used for deepening extraction of lesion comprehensive features, fusing lesion features in different dimensions, and optimizing feature expression, and then converts the feature map into a lesion aggregated feature vector with a fixed length through an adaptive pooling layer; An output strategy of the pest and disease type matching layer is configured, the pest and disease type matching layer includes two layers of fully connected layers and a probability activation function, the first layer of fully connected layers maps the lesion aggregated feature vector with a fixed length into an intermediate dimension feature vector, the second layer of fully connected layers maps the intermediate dimension feature vector into a dimension consistent with the number of known plant pest and disease types, and the probability activation function converts the output into probability values corresponding to each pest and disease type; An optimization rule of the dynamic optimization module is configured, the optimization rule is that deviation data between pest and disease type probability values output by the model and pest and disease types labeled by samples is collected in real time, and when deviation values of a continuous fixed number of samples are higher than a deviation threshold, model parameter fine tuning is triggered; A labeled sample set including different plant pest and disease types is collected, the labeled sample set includes lesion images of each pest and disease type on different plant organs and in different growth stages, and each sample is labeled with pest and disease type name, organ type where the lesion is located, growth stage of the lesion, and lesion severity level information; The labeled sample set is divided into a training set, a validation set, and a test set at a fixed ratio, the training set includes samples of different pest and disease types, different organs, and different growth stages, and sample distributions of the validation set and the test set remain consistent with the training set; The model is iteratively trained by using the training set, the model performance is verified by using the validation set after each iteration, and when verification indexes of the validation set all remain stable for a continuous fixed number of times, the training is stopped, the verification indexes include pest and disease type recognition accuracy, organ lesion positioning accuracy, and lesion matching accuracy in different growth stages; The trained model is tested for performance by using the test set, if pest and disease type recognition accuracy, organ lesion positioning accuracy, and lesion matching accuracy in different growth stages of the test set all reach a set standard, the construction of the plant pest and disease exclusive deep learning model is completed, and if not, the network parameters of each branch are adjusted and then retrained until the test is passed. 5.The deep learning-based plant pest image recognition method of claim 1, wherein, The cross-stage lesion association set is input into a preset plant pest and disease specific deep learning model for feature aggregation and type matching processing to obtain lesion aggregation features and preliminary pest and disease type results, including: Each lesion association group in the cross-stage lesion association set is preprocessed, image data of each suspected lesion area in each group is extracted, the image size is adjusted according to the organ type where the lesion is located, the leaf lesion image is adjusted to a uniform rectangular size, and the main stem and branch lesion image is adjusted to a uniform elongated size; The adjusted image data is subjected to pixel value normalization processing to convert the pixel value to a fixed numerical range to obtain a standardized lesion image sequence; The time sequence arrangement information of the standardized lesion image sequence is extracted, the time sequence completion processing is performed on the standardized lesion image sequence in combination with the growth stage interval data in the time sequence evolution information of the lesion on different organs, when the image interval between adjacent growth stages exceeds the interval threshold, the interpolation method based on the organ lesion evolution rule is used to generate the intermediate image to obtain a complete time sequence lesion image sequence; The complete time sequence lesion image sequence is input into the feature input layer of the plant pest and disease specific deep learning model, the image sequence is converted into an input format suitable for each branch of the model through the dimension conversion operation of the input layer to obtain branch input features; The branch input features are input into the lesion texture detail extraction branch, the lesion surface texture details in the image are extracted through the sequential processing of the preset convolution layer, pooling layer and batch normalization layer to obtain a lesion texture feature map; The branch input features are simultaneously input into the lesion morphological feature extraction branch, the overall morphological features of the lesion are extracted through the expanded receptive field processing of the atrous convolution layer and the global information aggregation processing of the global average pooling layer to obtain a lesion morphological feature map; The branch input features are simultaneously input into the lesion color feature extraction branch, the color distribution features and color gradient features of the lesion are extracted through the color feature extraction of the convolution layer and the feature conversion processing of the fully connected layer to obtain a lesion color feature vector; The lesion texture feature map, lesion morphological feature map and lesion color feature vector are input into the lesion feature interaction layer, the lesion texture feature map and the lesion morphological feature map are first spliced according to the channel, the spatial weight map is generated through the spatial attention mechanism for spatial weighting, and then the weighted feature map and the lesion color feature vector are spliced according to the channel, the channel weight vector is generated through the channel attention mechanism for channel weighting to obtain an interaction feature map; The interaction feature map is input into the lesion feature aggregation layer, the fixed-length lesion aggregation features are obtained through the deepening extraction of the plant lesion specific convolutional neural network block and the dimension unification processing of the adaptive pooling layer; The lesion aggregation features are input into the pest and disease type matching layer, the probability values corresponding to each pest and disease type are obtained through the feature mapping of the two fully connected layers and the probability conversion of the probability activation function, and the pest and disease type with the highest probability value is selected as the preliminary pest and disease type result of the lesion association group. 6.The deep learning-based plant pest image recognition method of claim 1, wherein, The time sequence evolution information of the lesions on different organs is used for organ adaptation verification processing of the preliminary insect disease type result, to generate a final plant insect disease recognition result and a lesion organ time sequence evolution graph, including: Extracting time sequence evolution information of the lesions of each lesion correlation group in the cross-stage lesion correlation set on different organs, and analyzing the position offset trajectory, size expansion or contraction curve, and feature data change matrix of the suspected lesion area in the organ coordinate system in each growth stage; Obtaining a preset plant insect disease organ time sequence evolution feature library, which contains standard time sequence evolution rules of various known plant insect disease types on different organs. Each insect disease type corresponds to a set of standard evolution rules, including standard position offset trajectory, standard size change curve, and standard feature data change rule on the leaf, and the standard evolution rule on the main stem and branch; According to the organ type identifier corresponding to the lesion correlation group, the standard time sequence evolution rule of each insect disease type under the organ type is matched from the plant insect disease organ time sequence evolution feature library. The position offset trajectory of the lesion correlation group is compared with the matched standard position offset trajectory, and the trajectory matching degree is calculated; The size expansion or contraction curve is compared with the standard size change curve, and the size matching degree is calculated. The feature data change matrix is compared with the standard feature data change rule, and the feature matching degree is calculated; The organ adaptation weight is assigned to the trajectory matching degree, the size matching degree, and the feature matching degree. The comprehensive time sequence matching degree is calculated by weighted summation. The trajectory matching degree weight of the leaf lesion is the highest, the size matching degree weight of the main stem lesion is the highest, and the feature matching degree weight of the branch lesion is the highest; If the comprehensive time sequence matching degree is higher than the organ adaptation verification threshold, it is confirmed that the preliminary insect disease type result of the lesion correlation group is the effective insect disease type result; If the comprehensive time sequence matching degree is not higher than the organ adaptation verification threshold, the lesion aggregation feature of the lesion correlation group and the corresponding organ type identifier and growth stage information are input into the dynamic optimization module, and the model parameter fine tuning is triggered to perform type matching again to obtain a secondary recognition type result. The organ adaptation verification step is repeated until the effective insect disease type result is obtained; Collecting the effective insect disease type result of all lesion correlation groups, correlating the corresponding growth stage time identifier, organ type identifier, organ region coordinate range information, and feature data of the suspected lesion area, and constructing a prediction correlation table, which contains the unique identifier of each record; A graph coordinate system is constructed with the growth stage as the horizontal axis and the plant organ type as the vertical axis. The horizontal axis is labeled in the order of growth stages, and the vertical axis is labeled in the order of leaves, main stems, and branches. The outlines of each organ are drawn in the graph coordinate system according to the organ region coordinate range information. The shape of the outline is consistent with the actual shape of the corresponding organ. In each organ contour mark, a lesion mark is drawn according to the position coordinates of the suspected lesion area, the color of the lesion mark corresponds to the effective insect disease type result one by one, the size of the lesion mark is proportional to the size of the suspected lesion area, and the unique identifier of the lesion associated group is marked in the lesion mark; In the atlas, the lesion marks of the same lesion associated group at different growth stages are connected by a polyline, the thickness of the polyline is proportional to the organ correlation coefficient of the lesion associated group, and the key nodes of the characteristic data change are marked beside the polyline; An atlas is added on one side of the atlas, the colors corresponding to different effective insect disease type results, the size range corresponding to the size of the lesion mark, the organ correlation coefficient range corresponding to the thickness of the polyline, and the meaning of the unique identifier are marked, and the prediction association table and the drawn atlas are combined into the final plant insect disease identification result. 7.The deep learning-based plant pest image recognition method of claim 2, wherein, The plant organ contour feature of each image in the initial time sequence angle image sequence is extracted, including: The color space conversion processing is performed on each image in the initial time sequence angle image sequence, the hue channel image is extracted, and the threshold segmentation processing is performed on the hue channel image, so that the hue channel image is divided into plant foreground area and background area, and the binary image is obtained; After the morphological closing operation processing is performed on the binary image, the morphological opening operation processing is performed, and the optimized binary image is obtained. The edge detection processing is performed on the optimized binary image, and the initial contour line segment set is formed; The contour screening processing is performed on the initial contour line segment set, the screening rules are set according to the morphological characteristics of the plant organ, the main stem contour line segment, the leaf contour line segment and the stem branch contour line segment are screened out, the screened main stem contour line segment is fitted, the line segment is fitted into a smooth curve conforming to the upright shape of the plant main stem by using a cubic polynomial fitting algorithm, the sampling point coordinates on the smooth curve are extracted as the main stem contour point coordinates at a fixed interval, the curvature change value at each sampling point is calculated as the main stem contour curvature change value, and the main stem contour feature is composed; The closed processing is performed on the screened leaf contour line segment, the line segment that is not closed is extended to intersect to form a closed contour, the sampling point coordinates on the closed contour are extracted as the leaf contour point coordinates, the contour direction of the main vein and the lateral vein of the leaf is extracted through the vein detection algorithm, the number and distribution density of the leaf edge jagged line segment are counted as the leaf edge jagged density, and the leaf contour feature is composed; The fitting processing is performed on the screened stem branch contour line segment, the branch contour point coordinates are obtained, the angle between the branch contour line segment and the main stem contour curve at the connection point is calculated as the connection angle between the branch and the main stem by using an angle calculation algorithm, and the stem branch contour feature is composed; The main stem contour feature, the leaf contour feature and the stem branch contour feature are combined into the plant organ contour feature of each image, and the connection point coordinates of the branch and the main stem and the connection point coordinates of the leaf and the branch are recorded. 8.The plant pest image recognition method based on deep learning according to claim 3, characterized in that, The lesion preliminary detection processing is performed on each organ region image, a sliding window detection mode suitable for the shape of the plant organ is adopted, and the gray scale change feature and the color distribution feature in the window are extracted, including: The parameters of the sliding window are determined according to the organ type identification, the blade region adopts an elliptical window, the length of the major axis of the elliptical window is set as a fixed proportion of the longest side of the blade region image, and the length of the minor axis is a fixed proportion of the length of the major axis; the main stem region adopts a rectangular window, the length of the window is set as a fixed proportion of the length of the main stem region image, and the width is set as a fixed proportion of the width of the main stem region image; the branch region adopts an elongated window, the length of the window is set as a fixed proportion of the length of the branch region image, and the width is set as a fixed proportion of the width of the branch region image; the sliding step is uniformly set as a fixed proportion of the shortest side of the corresponding window, and the proportion is determined through test of multiple normal images of the same organ type; The window is moved by the set sliding step, and the entire organ region image is traversed in the order from left to right and from top to bottom, and the position coordinates of each window and the image pixel data in the window are recorded; The pixel data in each window is subjected to gray value change feature extraction, the pixel data is converted into a gray image, the standard deviation of the gray value of the gray image in the window, the average of the gray value gradient and the gray value entropy are calculated, and the standard deviation of the gray value of the gray image, the average of the gray value gradient and the gray value entropy are combined in a fixed order as the gray value change feature; the standard deviation of the gray value reflects the dispersion degree of the gray distribution in the window, the average of the gray value gradient reflects the intensity of the gray change in the window, and the gray value entropy reflects the uniformity of the gray distribution in the window; The pixel data in each window is subjected to color distribution feature extraction, the pixel data is converted into a hue-saturation-lightness color space, and the histograms of the hue channel, the saturation channel and the lightness channel are calculated respectively; Each histogram is divided into a fixed number of intervals, the proportion of the number of pixels in each interval to the total number of pixels in the window is calculated, and the interval pixel proportions of the three channel histograms are combined in the channel order as the color distribution feature; Multiple healthy images of the same plant variety and the same organ type are collected as normal samples, each normal sample image is processed according to the same window parameters, and the gray value change feature and the color distribution feature of each window are extracted; The mean and the fluctuation range of each window feature of the same organ type in the normal samples are calculated, the mean plus or minus a fixed multiple of the fluctuation range is taken as the normal feature range of the organ type, and the gray value change feature and the color distribution feature correspond to their respective normal feature ranges. 9.The plant pest image recognition method based on deep learning according to claim 4, characterized in that, The model is iteratively trained using the training set, and the performance of the model is verified through the validation set after each iteration, and when the verification indexes of the validation set are all stable for a fixed number of times in succession, the training is stopped, including: The training set is subjected to data enhancement processing, including random horizontal flipping, random vertical flipping, random rotation within a fixed angle range, random cropping of a fixed proportion of the region and random adjustment of the brightness within a fixed range, so that the number of samples of the expanded training set is a fixed multiple of the original number; The expanded training set is divided into training batches according to a fixed batch size, each training batch contains a fixed number of samples, and the samples are uniformly distributed according to the type of insect and disease, the type of organ and the growth stage; The weight parameters of each layer of the initialized model are initialized, and the weights of the convolutional layer and the fully connected layer are initialized. The initial weight distribution is adjusted according to the statistical law of the plant lesion characteristics, and the bias parameter is initialized as a fixed decimal value. The training parameters are set. The total number of iterations is set as a fixed value, the initial value of the learning rate is set as a fixed value, and the adaptive learning rate adjustment strategy is adopted. When the validation index of the validation set does not improve for a fixed number of consecutive times, the learning rate is reduced by a fixed proportion, and the minimum learning rate is set as a fixed proportion of the initial value. Start the iterative training. Each iteration inputs each training batch in turn, inputs the batch samples into the model, and obtains the feature output of each branch through the parallel processing of the lesion texture detail extraction branch, the lesion morphological feature extraction branch and the lesion color feature extraction branch. After the processing of the lesion feature interaction layer and the lesion feature aggregation layer, the lesion aggregation features are obtained, and then the probability values corresponding to each pest and disease type are output through the pest and disease type matching layer. The cross-entropy loss function is used to calculate the loss value between the output probability value and the sample label of the pest and disease type. The calculation of the cross-entropy loss function is based on the corresponding relationship between the probability value of each sample and the true pest and disease type. The position label corresponding to the true pest and disease type is set to 1, and the labels of other types are set to 0. The loss value is propagated to each layer of the model through the back propagation algorithm, and the gradient descent optimization algorithm is used to adjust the weight parameters and bias parameters of each layer. The step size of the gradient descent is determined by the current learning rate. After each iteration period, the validation set is input into the trained model to obtain the pest and disease type recognition result, the lesion positioning result and the growth stage matching result of the validation set. The pest and disease type recognition accuracy is calculated, that is, the proportion of the number of correctly recognized samples to the total number of samples in the validation set. The organ lesion positioning accuracy is calculated, that is, the proportion of the number of samples whose lesion positioning coordinates and labeled coordinates are within the allowed range to the total number of samples in the validation set. The lesion matching accuracy of different growth stages is calculated, that is, the proportion of the number of correctly matched lesion association groups in different growth stages to the total number of lesion association groups in the validation set. The validation index of each iteration period is recorded, and it is judged whether the fluctuation range of the validation index for a fixed number of consecutive times is lower than the respective fluctuation threshold. If they are all lower than the fluctuation threshold, it is determined that the validation index remains stable, and the training is stopped. If not, the next iteration is continued until the stop training condition is met. 10.A plant pest and disease image recognition system based on deep learning, characterized in that, It includes: a processor; a machine-readable storage medium for storing machine-executable instructions of the processor; wherein the processor is configured to execute the machine-executable instructions to perform the deep learning-based plant pest and disease image recognition method of any one of claims 1 to 9.

Images