A watermelon fusarium wilt recognition method fusing leaf image features and a storage medium

Abstract

The application discloses a watermelon fusarium wilt recognition method fusing leaf image features and a storage medium, relates to the technical field of intelligent plant disease recognition, and comprises the following steps: collecting leaf original digital images and constructing a multilevel analysis structure; different physiological state spectral components are extracted by performing multichannel spectral separation on the images, and a composite feature spectrum fusing apparent and deep physiological information is generated by using a feature growth model; the spectrum is subjected to regional deconstruction and feature matching according to the multilevel structure, a pathogen coincidence degree index of each region is calculated, and an overall infection probability mapping is fused and generated; the mapping is subjected to spatial clustering to recognize a disease spot core region, and a topological network of disease development is constructed based on the spatial morphological relationship of the disease spot core region; and a recognition conclusion is output in combination with network structure parameters and the probability mapping. The method can improve the detection sensitivity of early disease symptoms and can realize analysis of disease spatial distribution patterns and development dynamics.

Description

Technical Field

[0001] This invention belongs to the field of intelligent plant disease identification technology, specifically a method and storage medium for identifying watermelon wilt disease by integrating leaf image features. Background Technology

[0002] Currently, computer vision-based plant disease identification methods mainly rely on RGB images in the visible light band. These conventional techniques extract color, texture, or shape features from leaves and use classifiers to determine disease status. However, visible light images primarily reflect outward morphology and are not sensitive enough to changes in the plant's internal physiological state, such as early disease responses to chlorophyll degradation and water stress. This results in limited ability to identify diseases in their latent or initial stages, hindering early warning. Furthermore, the feature extraction process of conventional methods is often based on the entire image or fixed grid division, failing to fully consider the local and spatial dynamic characteristics of disease occurrence and development in leaf tissues.

[0003] Most existing disease identification schemes treat the problem as a holistic classification task, outputting a simple judgment of health or infection, or relying on manually set thresholds to segment suspected areas. These methods lack the ability to automatically detect and accurately locate the core area of ​​lesions, and cannot effectively distinguish similar lesions caused by different factors. They fail to model and analyze the spatial relationships and evolution trends between lesions, providing only static diagnoses and unable to assess the severity and development stage of the disease, thus limiting their guiding value for targeted intervention in precision agriculture. There is a need for an identification technology that can reveal the physiological state of leaves from multiple dimensions and dynamically analyze the spatial transmission patterns of diseases. Summary of the Invention

[0004] This invention aims to solve at least one of the technical problems existing in the prior art;

[0005] Therefore, this invention proposes a method for identifying watermelon wilt disease by fusing leaf image features, comprising:

[0006] Acquire raw digital images of the watermelon leaves to be tested, and construct a multi-level analysis structure for the images of the watermelon leaves to be tested;

[0007] Multi-channel spectral separation is performed on the original digital image to extract image components reflecting different physiological states. The separated image components are then imported into a feature growth model to generate a composite feature map of the leaf surface state.

[0008] Based on the multi-level analysis structure, the composite feature map is deconstructed into regions to obtain multiple partitioned feature data blocks. The partitioned feature data blocks are then matched and compared layer by layer with the pre-stored watermelon wilt pathogen feature database.

[0009] Based on the matching and comparison results, calculate the pathogen conformity index for each partition feature data block;

[0010] By integrating the pathogen conformity index of all partition feature data blocks, an overall infection probability mapping of the watermelon leaf to be tested is generated.

[0011] Spatial clustering analysis is performed on the overall infection probability mapping to identify potential lesion core regions. Based on the morphological evolution relationship of the lesion core regions, a topological network describing the disease development is constructed.

[0012] Based on the mapping between the structural parameters of the topological network and the overall infection probability, the identification conclusion of watermelon wilt disease is output.

[0013] Furthermore, the acquisition of raw digital images of the watermelon leaves to be tested, and the construction of a multi-level analysis structure for the images of the watermelon leaves to be tested, specifically includes:

[0014] Under standard lighting conditions, a high-resolution color image covering the entire visible surface of the watermelon leaf under test is acquired using an image sensor.

[0015] In the high-resolution color image, the leaf outline is automatically identified and segmented from the background area;

[0016] An initial analysis grid covering the entire blade region is established based on the blade profile as the basic boundary.

[0017] Based on the distribution information of the main vein and lateral veins of the leaf, the initial analysis grid is adaptively subdivided to form multiple sub-region grid levels with progressively smaller scales;

[0018] The multi-level analysis structure is a spatial analysis framework composed of the initial analysis grid and sub-region grids of each level. Each grid cell is associated with its level and spatial location coordinates.

[0019] Furthermore, the step of performing multi-channel spectral separation on the original digital image to extract image components reflecting different physiological states specifically includes:

[0020] Analyze the intensity values ​​of each pixel in the high-resolution color image under the red, green, and blue channels;

[0021] The red channel intensity value, green channel intensity value and blue channel intensity value of each pixel are combined into a three-dimensional spectral vector;

[0022] Based on the preset reference range of spectral vectors for various leaf physiological states, the three-dimensional spectral vector of each pixel is classified and judged.

[0023] Pixels identified as belonging to the same physiological state are extracted from the high-resolution color image, and their original spatial relationships are preserved to form an independent image layer.

[0024] Repeat this discrimination and extraction process to finally obtain multiple image components corresponding to healthy chlorophyll, chlorotic tissue, necrotic tissue, and water-soaked tissue, respectively.

[0025] Furthermore, the step of importing the separated image components into a feature growth model to generate a composite feature map of the leaf surface state specifically includes:

[0026] For each image component, calculate its pixel statistical distribution within each grid cell of the multi-level analysis structure;

[0027] The pixel statistical distribution includes pixel density, average spectral intensity, and standard deviation of spectral intensity;

[0028] The statistical distributions of pixels from different image components within the same grid cell are spliced ​​together to form the multi-source feature vector of the grid cell.

[0029] Based on the hierarchical relationship of the multi-level analysis structure, the multi-source feature vectors of the lower-level sub-grid cells are summarized and fused into their parent grid cells of the upper level.

[0030] Traverse all grid cells at all levels to generate a data set that completely corresponds to the multi-level analysis structure, with each cell containing multi-source feature vectors. This data set is the composite feature map.

[0031] Furthermore, the step of performing regional deconstruction on the composite feature map based on the multi-level analysis structure to obtain multiple partitioned feature data blocks specifically includes:

[0032] Starting from the highest level grid of the composite feature map, analyze the multi-source feature vector of each grid cell;

[0033] Based on the similarity measure of multi-source feature vectors between grid cells, adjacent grid cells with similar features are merged to form primary feature blocks;

[0034] In the next level, the same similarity merging operation is performed on all sub-grid cells belonging to the same primary feature block to form a more refined secondary feature block;

[0035] This process proceeds iteratively from top to bottom, level by level, until the lowest mesh of the multi-level analysis structure is reached.

[0036] The stable feature blocks that are eventually formed at each level, together with the multi-source feature vectors of all the grid cells contained therein, are packaged into an independent partitioned feature data block.

[0037] Furthermore, the step of performing layer-by-layer matching and comparison between the partition feature data block and the pre-stored watermelon wilt pathogen feature database specifically includes:

[0038] The watermelon wilt pathogen feature database stores standard feature vector sets for different infection stages and different disease sites;

[0039] For each partition feature data block, the central trend features of the multi-source feature vectors of all grid cells contained therein are extracted as the representative feature vectors of the partition feature data block;

[0040] The similarity between the representative feature vector of the partitioned feature data block and each vector in the standard feature vector set is calculated.

[0041] Record all standard feature vectors whose similarity to the representative feature vectors of the partition feature data block exceeds a preset threshold, and label the infection stage and disease site information corresponding to all standard feature vectors;

[0042] The matching and comparison results are a set of similarity records and associated infection stage and disease site information corresponding to each partition feature data block.

[0043] Furthermore, the step of calculating the pathogen conformity index for each partition feature data block based on the matching and comparison results specifically includes:

[0044] For a partitioned feature data block, count the total number of standard feature vectors that exceed a preset threshold recorded in its matching and comparison results;

[0045] Analyze the infection stage corresponding to the standard feature vector recorded in the matching and comparison results, and calculate the concentration of the infection stage in the pathological development sequence;

[0046] Analyze the disease location corresponding to the standard feature vector recorded in the matching and comparison results, and calculate the degree of agreement between the disease location and the actual location of the current partition feature data block on the leaf.

[0047] By combining the total number, the concentration of the pathological development sequence, and the consistency of the actual location, a quantitative pathogen consistency index value is generated through weighted calculation.

[0048] The pathogen conformity index value is used to characterize the degree of conformity between the leaf region corresponding to the partition feature data block and the typical characteristics of watermelon wilt.

[0049] Furthermore, the process of fusing the pathogen consistency index of all partition feature data blocks to generate the overall infection probability mapping of the watermelon leaf under test specifically includes:

[0050] Obtain the leaf area covered by each partition feature data block and the calculated pathogen conformity index value;

[0051] Using the spatial location of the leaf as coordinates, the pathogen conformity index value of each partition feature data block is assigned to all pixels within its coverage area;

[0052] For a pixel covered by multiple partition feature data blocks, the average value of its corresponding multiple pathogen conformity index values ​​is used as the value assigned to the pixel.

[0053] Spatial interpolation smoothing is performed on the values ​​of all pixels within the leaf region to form a probability distribution surface that covers the entire leaf and whose values ​​change continuously.

[0054] The overall infection probability mapping is the probability distribution surface, where the value at each position represents the probability that the pixel belongs to the lesion region.

[0055] Furthermore, the spatial clustering analysis performed on the overall infection probability mapping to identify potential lesion core regions, and the construction of a topological network describing disease development based on the morphological evolution relationships of the lesion core regions, specifically includes:

[0056] A probability threshold is set for the overall infection probability mapping, and pixels with probability values ​​higher than the threshold are marked as candidate lesions;

[0057] Based on the spatial proximity of the candidate lesions, a density clustering algorithm is used to aggregate candidate lesions that are close to each other into multiple independent lesion clusters;

[0058] Calculate the geometric center of each lesion cluster, and expand outward from the geometric center until all pixels within the cluster are covered, forming the initial boundary of the lesion core region;

[0059] By analyzing the area, shape factor, and relative position to the leaf midrib of each lesion core region, regions that conform to the typical spatial distribution characteristics of watermelon wilt disease were selected as the final potential lesion core regions.

[0060] For each potential lesion core area, key points on its boundary contour are extracted, and a regional growth path is constructed based on the contour evolution trend.

[0061] Using the core area of ​​each potential lesion as a node, and based on the spatial distance between regions, the continuity of probability distribution, and the extension direction of the regional growth path, directed connection edges are established between nodes.

[0062] A topological network representing the spatial association and evolution direction of lesion regions is constructed using the nodes and the directed connecting edges.

[0063] Furthermore, the present invention also includes a storage medium storing a computer program thereon, which, when executed by a processor, implements the watermelon wilt disease identification method as described above by fusing leaf image features.

[0064] Compared with the prior art, the beneficial effects of the present invention are:

[0065] By performing multi-channel spectral separation on raw digital images, spectral image components reflecting different physiological and biochemical states of leaves can be obtained. These components contain information beyond the conventional visible light range. Introducing these multi-source components into a feature growth model for collaborative analysis and fusion results in a composite feature map that not only includes surface morphology information but also deeply integrates internal physiological state characteristics closely related to disease occurrence. This process enables the identification system to capture early, subtle physiological changes caused by Fusarium wilt infection, enhancing its ability to detect early symptoms and its diagnostic specificity in complex environments.

[0066] Spatial clustering analysis of the overall infection probability map adaptively identifies high-confidence lesion core regions based on the spatial distribution characteristics of probability values. A topological network is constructed based on the morphology and spatial relationships of these core regions, transforming discrete lesion information into a structured model containing nodes and connections. By calculating the structural parameters of this network, the spatial aggregation patterns, diffusion potential, and interrelationship strength of lesions can be quantitatively analyzed. This method not only achieves precise spatial localization of lesions but also provides quantitative evidence for assessing the local severity of the disease and inferring its spatial propagation dynamics, thus completing the transformation from static identification to dynamic process analysis. Attached Figure Description

[0067] Figure 1 This is a flowchart illustrating the steps of the watermelon wilt disease identification method based on the fusion of leaf image features described in this invention.

[0068] Figure 2 A flowchart for constructing a multi-level analysis structure;

[0069] Figure 3 A flowchart for multi-channel spectral separation and extraction of image components;

[0070] Figure 4 A heatmap showing the probability distribution of lesions on watermelon leaves;

[0071] Figure 5 This is a bar chart showing the multi-channel spectral characteristics of watermelon leaves. Detailed Implementation

[0072] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0073] See Figure 1 Raw digital images of watermelon leaves to be tested were acquired, and a multi-level analysis structure was constructed based on these images. Multi-channel spectral separation was then performed on the raw digital images to extract multiple image components reflecting different physiological states of the leaves. These separated image components were imported into a feature growth model to generate a composite feature map that comprehensively describes the leaf surface state. Next, based on the pre-constructed multi-level analysis structure, this composite feature map was deconstructed into a series of partitioned feature data blocks. Each partitioned feature data block was matched and compared layer by layer with a pre-stored standard feature vector set containing features of watermelon wilt pathogens at different stages and locations—the watermelon wilt pathogen feature library. Based on the comparison results, a quantified pathogen conformity index was calculated for each partitioned feature data block. Then, the pathogen conformity indices of all partitioned feature data blocks were fused in leaf space to generate an overall infection probability map covering the entire leaf area. Spatial clustering analysis was performed on this probability map to identify potential lesion core areas, and based on the morphological and spatial relationships between these areas, a topological network describing the possible development path of the disease was constructed. Finally, by combining the structural parameters of the topological network with the global information of the overall infection probability mapping, a conclusion is outputting the identification of whether watermelon wilt disease exists and its development status.

[0074] See Figure 2 In one embodiment of the present invention, an original digital image of the watermelon leaf to be tested is acquired, and a high-resolution color image covering the entire visible surface of the watermelon leaf is obtained using an image sensor under standard lighting conditions. An example scenario for the high-resolution color image is an RGB format digital file with a resolution of 4000 pixels by 3000 pixels. In a specific implementation, the leaf outline in the high-resolution color image is automatically identified and segmented from the background area. The segmentation operation is completed by image processing algorithms that distinguish the color difference and texture difference between green plant tissue and non-plant background. A data comparison example is that the original image before segmentation contains 12 million total pixels, while after segmentation, only about 5 million pixels belonging to the watermelon leaf are retained. In a specific implementation, an initial analysis grid covering the entire leaf area is established based on the identified leaf outline as the basic boundary. The initial analysis grid is a geometric structure that uniformly divides the internal area of ​​the leaf outline into regular cells, for example, dividing the leaf area into square grid cells of 100 rows by 100 columns.

[0075] In some embodiments, the initial analysis grid is adaptively subdivided based on the distribution information of the midrib and lateral veins of the leaf. This midrib and lateral vein distribution information is obtained by analyzing the intensity gradient and line features of the green spectral channel in a high-resolution color image. The adaptive subdivision process follows a preset rule: for grid cells identified as having a midrib orientation in the image, the grid cell is subdivided into smaller sub-cells along the midrib path. It can be understood that the intensity of the subdivision operation is related to the salience of the veins; the subdivision degree in the midrib region is higher than that in the lateral vein region, while the grid in non-vein regions may retain its initial size. A data comparison example shows that an initial grid cell located on the midrib is subdivided into 16 sub-grid cells (4x4), while an initial grid cell located on a lateral vein is subdivided into 4 sub-grid cells (2x2).

[0076] Optionally, the rule for adaptively subdividing the initial analysis grid can be expressed as a function based on the distance to the leaf veins, which determines the subdivision level at each location. After subdivision, multiple sub-region grid levels with progressively decreasing scales are formed. Progressive scale reduction means that the area of ​​a sub-grid cell is smaller than the area of ​​its parent grid cell. A specific example of a hierarchical relationship is: level 0 is the initial analysis grid, level 1 is the grid generated by the first subdivision of the leaf vein region based on level 0, and level 2 is the grid generated by further subdividing the core region of the main vein based on level 1. In practice, the multi-level analysis structure consists of the initial analysis grid and the sub-region grids of each level. This spatial analysis framework is represented in computer memory as a tree-like data structure, where the root node corresponds to the entire leaf outline, and branch nodes correspond to grid cells at different levels.

[0077] In some embodiments, each grid cell in the spatial analysis framework is associated with its hierarchical level and spatial coordinates. The hierarchical information is stored using an integer index, and the spatial coordinates are stored using the grid cell's row and column numbers or bounding box coordinates in the image pixel coordinate system. It can be understood that after the multi-level analysis structure is constructed, each pixel in the leaf image can be mapped to one or more grid cells at different levels through its coordinates. The multi-level analysis structure provides a multi-layered, multi-resolution spatial reference foundation for subsequent feature extraction and region analysis, enabling the analysis to grasp the overall trend at a macroscopic level while focusing on detailed areas such as leaf veins and edges at a microscopic level.

[0078] See Figure 3In one embodiment of the present invention, the intensity values ​​of each pixel in the acquired high-resolution color image under the red, green, and blue channels are analyzed. A specific example of a high-resolution color image is an RGB image with a resolution of 4000 pixels by 3000 pixels. The pixel located at row 500 and column 500 has a red channel intensity value of 205, a green channel intensity value of 180, and a blue channel intensity value of 90. The red, green, and blue channel intensity values ​​of each pixel are combined into a three-dimensional spectral vector. For example, the three-dimensional spectral vector of this pixel is represented as (205, 180, 90). The three-dimensional spectral vector of each pixel is classified and judged according to the preset spectral vector reference range corresponding to various leaf physiological states. The preset spectral vector reference range is stored in the form of vector intervals. For example, the reference range of healthy chloroplasts is defined as a combination of intervals where the red channel intensity value is between [30, 100], the green channel intensity value is between [120, 220], and the blue channel intensity value is between [20, 80]. The classification and discrimination process is achieved by calculating the Euclidean distance between the pixel's spectral vector and the center vector of each reference range. The distance calculation formula is as follows:

[0079] ;

[0080] in: These represent the intensity values ​​of a pixel in the red, green, and blue channels, respectively. These represent the center values ​​of the reference range for the k-th physiological state in the red, green, and blue channels, respectively. This represents the Euclidean distance between the spectral vector of a pixel and the reference center vector of the k-th physiological state. Pixels classified as belonging to the same physiological state are extracted from the original image while preserving their original spatial relationships, thus forming an independent image layer. For example, the original high-resolution color image contains 12 million pixels, while the healthy chlorophyll tissue image layer generated after discrimination and extraction contains approximately 8 million pixels. By repeating this discrimination and extraction process, multiple independent image components corresponding to healthy chlorophyll tissue, chlorotic tissue, necrotic tissue, and water-soaked tissue are finally obtained. These image components have the same number and size as the original image but only contain pixels classified as specific physiological states.

[0081] In some embodiments, for each separated image component, the pixel statistical distribution in each grid cell within the multi-level analysis structure is calculated. The pixel statistical distribution includes pixel density, average spectral intensity, and the standard deviation of spectral intensity. Taking a specific grid cell in the multi-level analysis structure as an example, this grid cell covers a rectangular region in the image. For the healthy chlorotic tissue image component, the number of all pixels falling within this rectangular region is counted and divided by the total area of ​​the rectangular region to obtain the pixel density. Simultaneously, the average intensity of these pixels in the red, green, and blue channels is calculated to obtain the average spectral intensity, and the standard deviation of the intensity values ​​for each channel is calculated. It can be understood that for the chlorotic tissue, necrotic tissue, and water-soaked tissue image components, the exact same statistical calculations are performed within the same grid cell, thereby generating a set of data containing three statistical measures for each image component within that grid cell. The statistical distributions of pixels located within the same grid cell but originating from different image components are stitched together to form a multi-source feature vector for the grid cell. For example, if the statistical distribution of the healthy chloroplast image component in a grid cell is (0.65, 150, 20), the statistical distribution of the chlorotic tissue image component is (0.20, 120, 15), the statistical distribution of the necrotic tissue image component is (0.10, 80, 10), and the statistical distribution of the water-soaked tissue image component is (0.05, 100, 12), then the multi-source feature vector formed after stitching is (0.65, 150, 20, 0.20, 120, 15, 0.10, 80, 10, 0.05, 100, 12). In practice, based on the inherent hierarchical relationship of the multi-level analysis structure, the multi-source feature vectors of lower-level sub-grid units are summarized and fused into their parent grid units. The summarization and fusion operations are performed independently for each dimension of the multi-source feature vector. For a parent grid unit, each element value of its multi-source feature vector is obtained by taking the arithmetic mean of the corresponding element values ​​of all its directly subordinate sub-grid units. For example, if a parent grid unit has two sub-grid units, and the pixel density of healthy chlorophyll tissue in the multi-source feature vector of sub-grid unit one is 0.70, while the corresponding value for sub-grid unit two is 0.60, then the corresponding element value for the parent grid unit is 0.65. Optionally, the fusion process can use a weighted average method, with the weight determined by the proportion of the total number of pixels covered by the sub-grid units to the total number of pixels in the parent grid unit. By traversing all grid cells in all levels, a data set is generated that completely corresponds to the multi-level analysis structure, and each cell contains a multi-source feature vector formed by fusing information from itself and its subordinate cells. This complete data set is the composite feature map describing the surface state of the blade. The composite feature map is represented as a multi-dimensional array in computer memory. The first and second dimensions of the array correspond to the row and column positions of the grid cells, and the third dimension corresponds to each element of the multi-source feature vector.

[0082] It is understandable that the generation and fusion process of multi-source feature vectors is iterative. The initial multi-source feature vector is calculated starting from the lowest-level grid cell of the multi-level analysis structure, and then fused layer by layer upwards until the highest level. In some embodiments, the lowest-level grid cell of the multi-level analysis structure corresponds to the pixel-level resolution of the image, and its multi-source feature vector is directly constructed based on the pixel statistical distribution of each image component within that grid cell. In specific implementations, for a high-level grid cell in the multi-level analysis structure, its multi-source feature vector not only reflects the statistical characteristics of the cell itself but also incorporates the statistical characteristics of all its subordinate sub-cells, thus achieving multi-scale feature representation. Optionally, after generating the composite feature map, the multi-source feature vector can be normalized to ensure that the values ​​of different dimensions are within a similar range, for example, by dividing the pixel density by the total number of pixels in the image or dividing the spectral intensity value by 255.

[0083] In one embodiment of the present invention, analysis begins with the highest-level grid of the composite feature map. This highest-level grid is, for example, a structure that divides the leaf region into 25 coarse-grained grid cells (5 rows x 5 columns). The multi-source feature vector of each grid cell is analyzed; the multi-source feature vector is a numerical array containing 12 elements. Adjacent grid cells with similar features are merged based on a similarity measure of the multi-source feature vectors between grid cells. This similarity measure is achieved by calculating the cosine similarity between the multi-source feature vectors of two grid cells, using the following formula:

[0084] ;

[0085] in: and These represent the multi-source feature vectors of grid cell a and grid cell b, respectively. This represents the cosine similarity value between two vectors. A similarity threshold is set; when two adjacent grid cells show similarity... When the value is greater than 0.85, a merging operation is performed to form primary feature blocks. For example, before merging, there are 25 independent grid cells, and after similarity-based merging, 8 unconnected primary feature blocks are formed.

[0086] In some embodiments, the same feature similarity-based merging operation is performed on all sub-mesh cells belonging to the same primary feature block in the next level. The next level mesh is subdivided from the highest level mesh; for example, each highest level mesh cell is subdivided into 16 sub-mesh cells of 4x4. It can be understood that when merging sub-mesh cells belonging to the same primary feature block, the similarity metric and merging threshold are consistent with the rules used at the highest level, but the merging operation is only performed between sub-mesh cells within the boundary of the same primary feature block, thereby forming finer secondary feature blocks. The number and shape of secondary feature blocks better reflect the local detail changes on the blade surface than primary feature blocks. This process iterates from top to bottom, level by level, until the bottom layer mesh of the multi-level analysis structure is reached. The cell size of the bottom layer mesh may correspond to the size of several pixels. In practice, the feature blocks that are formed at each level and have consistent and stable internal characteristics are packaged into an independent partition feature data block. The packaging operation includes recording the hierarchical identifier, spatial coordinates, and multi-source feature vectors corresponding to all grid cells covered by the feature block. The partition feature data block can be represented in terms of data structure as an object containing block metadata and a list of feature vectors.

[0087] Optionally, the stability of a feature block can be assessed by determining whether the block boundary changes significantly during the merging of two adjacent levels. If the boundary change is less than a preset proportion, it is considered stable. The watermelon wilt pathogen feature library stores standard feature vector sets corresponding to different infection stages and different disease sites. Different infection stages may include early water-soaked appearance, mid-stage chlorosis, and late-stage necrosis, while different disease sites may include leaf margins, near leaf veins, and the middle of the leaf. Each standard feature vector has the same dimension and physical meaning as the multi-source feature vector. For each generated partition feature data block, the central trend feature of the multi-source feature vectors of all grid cells within it is extracted. The central trend feature is obtained by calculating the arithmetic mean of the values ​​of each dimension of the multi-source feature vectors of all grid cells within the partition feature data block. This is used as the representative feature vector of the partition feature data block. For example, if a partition feature data block contains 20 grid cells, then each element value of its representative feature vector is equal to the average of the corresponding element values ​​of the 20 grid cells.

[0088] It is understandable that the representative feature vector of the partition feature data block is compared with the standard feature vectors in the pathogen feature database to calculate similarity. The similarity calculation can use the reciprocal of the Euclidean distance or other correlation measures. All standard feature vectors whose similarity to the representative feature vector of the current partition feature data block exceeds a preset threshold are recorded. The preset threshold is set according to the actual application scenario, for example, an Euclidean distance less than 0.1 or a cosine similarity greater than 0.9. Simultaneously, the infection stage and disease site information corresponding to all standard feature vectors are labeled, and the labeling information is stored together with the standard feature vectors in the form of data labels. In some embodiments, the matching and comparison results are reflected in the similarity record corresponding to each partition feature data block and its associated set of infection stage and disease site information. The similarity record can be a list, and each entry in the list contains the identifier of the matched standard feature vector, the calculated similarity value, and the infection stage label and disease site label corresponding to that standard feature vector. In specific implementations, for a particular partition feature data block, its matching result may be associated with multiple infection stages, such as simultaneously matching the standard feature vectors of the early water-soaked and mid-stage chlorotic stages, which reflects the transitional or complex nature of the region's features. Optionally, the matching process can prioritize recording the top K standard feature vectors with the highest similarity, rather than recording all results that exceed the threshold.

[0089] In one embodiment of the present invention, for a specified partition feature data block, the total number of standard feature vectors with similarity exceeding a preset threshold recorded in the statistical matching and comparison results is calculated. The preset threshold is set to a cosine similarity of 0.9. The matching results of partition feature data block A contain 5 standard feature vectors exceeding this threshold. The infection stage corresponding to the standard feature vectors recorded in the matching and comparison results is analyzed, and the concentration of the infection stage in the pathological development sequence is calculated. The pathological development sequence is defined as the order from the early water-soaked stage to the middle chlorotic stage and then to the late necrosis stage. The concentration is measured by the dispersion of the infection stage to which the matched standard feature vector belongs in this time sequence. The disease site corresponding to the standard feature vector recorded in the matching and comparison results is analyzed, and the degree of agreement between the disease site and the actual spatial position of the current partition feature data block on the leaf is calculated. Disease site information includes leaf margin, near leaf veins, middle of leaf, etc. The agreement is calculated by comparing whether the actual position coordinates of the partition feature data block are consistent with the typical disease site area marked by the matched standard feature vector. If the positions are consistent, the agreement is high. By considering the total number, the concentration in the pathological development sequence, and the degree of agreement in the actual location, a quantitative pathogen consistency index value is generated through a predefined weighted calculation. An example of the weighted calculation is shown in the following formula:

[0090] ;

[0091] in: The calculated pathogen matching index value represents the total number of standard feature vectors that exceeded the threshold after normalization. This represents a measure of the concentration of infection stages in the timeline of pathological development. A measure of the degree of agreement between the site of the disease and its actual spatial location. These are preset weighting coefficients used to balance the impact of different factors on the final indicator. See Table 1:

[0092] Table 1: Matching Results and Intermediate Calculated Values ​​of Partition Feature Data Blocks

[0093]

[0094] In practice, the leaf region covered by each partition feature data block and its calculated pathogen compatibility index value are obtained. The leaf region covered by each partition feature data block is determined by the boundary coordinates of all grid cells it contains. Using the spatial position of the leaf as coordinates, the pathogen compatibility index value of each partition feature data block is assigned to all pixels within its coverage area. For example, if the pathogen compatibility index value of partition feature data block A is 0.872, and it covers a rectangular area from (row 100, column 200) to (row 150, column 250) in the image, then all pixels within this rectangular area are assigned the initial value of 0.872. For a pixel covered by multiple partition feature data blocks, the average value of the multiple pathogen conformity index values ​​corresponding to the pixel is used as the pixel's assigned value. For example, if pixel P1 is simultaneously located within the coverage of partition feature data block A and partition feature data block B, and the pathogen conformity index value of partition feature data block A is 0.872 and the pathogen conformity index value of partition feature data block B is 0.653, then the final assigned value of pixel P1 is (0.872+0.653) / 2=0.7625.

[0095] In some embodiments, spatial interpolation smoothing is performed on the values ​​assigned to all pixels within the blade region. This spatial interpolation smoothing can employ bilinear interpolation or cubic spline interpolation algorithms. The purpose of the interpolation is to fill in the values ​​assigned to pixels not directly covered by any partition feature data block and to ensure a continuous overall numerical transition. This forms a probability distribution surface covering the entire blade with continuously changing values. Mathematically, this probability distribution surface can be represented as a bivariate function. Where x and y represent the coordinates of the pixel. This represents the probability that the pixel belongs to the lesion area. The overall infection probability mapping is the probability distribution surface, which can be stored in a computer as a floating-point matrix of the same size as the original leaf image. The value of each element in the matrix corresponds to the value of the probability distribution surface at that pixel location. In other words, the value on the probability distribution surface directly represents the probability that the corresponding leaf location belongs to the watermelon wilt lesion area; the closer the value is to 1, the higher the probability, and the closer the value is to 0, the lower the probability.

[0096] Optionally, spatial interpolation smoothing can be performed by applying a Gaussian filter to the values ​​of all pixels after assignment, to further reduce numerical abrupt changes caused by the boundaries of the partition feature data blocks. In specific implementation, the generation of the overall infection probability map transforms the discrete partition feature data block evaluation results into a continuous spatial probability model.

[0097] See Figure 4 This is a heatmap showing the probability distribution of lesions on watermelon leaves. Heatmaps can quickly locate the core area and spread of lesions, providing a spatial basis for subsequent disease classification and control decisions. The continuous distribution of probability values ​​achieves a quantitative expression of lesions, avoiding the subjectivity of traditional manual identification. The color gradient from the core to the periphery corresponds to the pathological development sequence of watermelon wilt: "early water-soaked appearance → mid-stage chlorosis → late-stage necrosis." The color gradient from the core outwards (red → yellow → blue) visually reflects the direction and speed of disease spread; the wider the transition zone, the faster the disease spreads, providing a 2-3 day advance warning of "disease outbreak risk," giving farmers a window of opportunity for prevention and control.

[0098] In one embodiment of the present invention, a probability threshold is set for the overall infection probability mapping. The probability threshold is set to 0.7. Pixels with probability values ​​higher than the probability threshold are marked as candidate lesions. The overall infection probability mapping is a floating-point matrix of size 1000 pixels multiplied by 800 pixels. After threshold filtering, 1250 candidate lesions are obtained. Based on the spatial proximity of the candidate lesions, a density clustering algorithm is used to aggregate candidate lesions that are close to each other into multiple independent lesion clusters. The density clustering algorithm sets the neighborhood radius to 5 pixels and the minimum number of points threshold to 10. A data comparison example shows that there were 1250 discrete candidate lesions before clustering, and 18 independent lesion clusters were formed after clustering. The geometric center of each lesion cluster is calculated, and the area is expanded outward from the geometric center until the expanded area can cover all pixels in the cluster. The geometric center is obtained by calculating the average coordinates of all pixels in the lesion cluster. The expansion process adopts the algorithm of continuously increasing the radius of the circular area until the circular area contains the minimum outer circle of all pixels in the cluster, thereby forming the initial boundary of the core area of ​​the lesion. The initial boundary is an approximately circular area.

[0099] In some embodiments, the area, shape factor, and relative position to the leaf midrib of each lesion core region are analyzed. The area is obtained by calculating the total number of pixels within the initial boundary of the lesion core region, and the shape factor is calculated using the formula:

[0100] ;

[0101] in: Let A represent the shape factor, and P represent the area of ​​the lesion core region, and P represent the perimeter of the initial boundary of the lesion core region. Regions conforming to the typical spatial distribution characteristics of watermelon wilt are selected as the final potential lesion core regions. The selection criteria include the shape factor. A minimum distance of 0.6, an area greater than 50 pixels, and a shortest distance to the leaf midrib less than 30 pixels are required. For example, 12 out of 18 initial lesion core regions meet all screening criteria. Key points are extracted from the boundary contour of each identified potential lesion core region. Key points are obtained by detecting points with significant changes in the curvature of the boundary contour. A regional growth path is constructed based on the contour evolution trend, which is inferred by analyzing the movement direction of the key point sequence. The growth path is represented as a fitted curve from the start to the end of the key point sequence.

[0102] It can be understood that each potential lesion core region is considered a network node, and directed connections between nodes are established based on the spatial distance between node regions, the continuity of probability distribution, and the extension direction of the region growth path. Spatial distance is calculated as the Euclidean distance between the geometric centers of two potential lesion core regions. The continuity of probability distribution assesses the smoothness of the change in the overall infection probability mapping value on the straight line connecting the two nodes. The extension direction of the region growth path determines whether the growth path of one node roughly points to another node. In specific implementation, the rules for establishing directed connections between nodes can be quantified as a comprehensive score. When the score exceeds a preset connection threshold, a directed edge is established between the two nodes, pointing from the node with the higher score to the node with the lower score. A topological network that can characterize the spatial association and possible evolution direction between lesion regions is constructed through nodes and directed connections. This topological network can be represented in a computer using an adjacency matrix or graph data structure, where node attributes include information such as region location, area, and shape factor, and the weights of the directed edges reflect the connection strength.

[0103] Optionally, the extension direction of the regional growth path can be obtained by calculating the principal direction of the keypoint sequence, which is the average azimuth angle of the overall movement direction of the keypoint sequence. In some embodiments, the continuity of the probability distribution can be evaluated by calculating the gradient change of the overall infection probability mapping value on the connection path; a gentle gradient change indicates high continuity. In specific implementations, the constructed topological network can visually demonstrate the spatial relationship between multiple potential lesion core regions and possible disease spread paths; for example, the network may exhibit a star-shaped radial structure or a chain-like connection structure.

[0104] See Figure 5 This is a bar chart of multi-channel spectral characteristics of watermelon leaves. This data directly corresponds to the pathological changes of watermelon wilt: the high average intensity of the red channel reflects the browning and necrosis of lesions, while the high standard deviation and pixel density of the green channel reflect the green heterogeneity during the chlorosis process, providing a quantitative basis for subsequent "image component extraction." The differences in characteristics across different channels can be used to distinguish between "healthy tissue," "chlorotic tissue," and "necrotic tissue," thereby guiding precise targeted pesticide application and reducing pesticide usage. By comparing the spectral characteristics of different batches of leaves, the weight coefficients of each channel in the "feature growth model" can be optimized, improving the model's accuracy in identifying lesions. The correspondence between spectral characteristics and pathological changes can, in turn, verify the pathological development sequence of watermelon wilt—"chlorosis followed by necrosis"—providing data support for research on the disease mechanism.

[0105] The above embodiments are only used to illustrate the technical methods of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical methods of the present invention without departing from the spirit and scope of the technical methods of the present invention.

Claims

1. A method for identifying watermelon wilt disease by fusing leaf image features, characterized in that, Includes the following steps: Acquire raw digital images of the watermelon leaves to be tested, and construct a multi-level analysis structure for the images of the watermelon leaves to be tested; Multi-channel spectral separation is performed on the original digital image to extract image components reflecting different physiological states. The separated image components are then imported into a feature growth model to generate a composite feature map of the leaf surface state. Based on the multi-level analysis structure, the composite feature map is deconstructed into regions to obtain multiple partitioned feature data blocks. The partitioned feature data blocks are then matched and compared layer by layer with the pre-stored watermelon wilt pathogen feature database. Based on the matching and comparison results, calculate the pathogen conformity index for each partition feature data block; By integrating the pathogen conformity index of all partition feature data blocks, an overall infection probability mapping of the watermelon leaf to be tested is generated. Spatial clustering analysis is performed on the overall infection probability mapping to identify potential lesion core regions. Based on the morphological evolution relationship of the lesion core regions, a topological network describing the disease development is constructed. Based on the mapping between the structural parameters of the topological network and the overall infection probability, the identification conclusion of watermelon wilt disease is output; The step of importing the separated image components into the feature growth model to generate a composite feature map of the leaf surface state specifically includes: For each image component, calculate its pixel statistical distribution within each grid cell of the multi-level analysis structure; The pixel statistical distribution includes pixel density, average spectral intensity, and standard deviation of spectral intensity; The statistical distributions of pixels from different image components within the same grid cell are spliced ​​together to form the multi-source feature vector of the grid cell. Based on the hierarchical relationship of the multi-level analysis structure, the multi-source feature vectors of the lower-level sub-grid cells are summarized and fused into their parent grid cells of the upper level. Traverse all grid cells at all levels to generate a data set that completely corresponds to the multi-level analysis structure, with each cell containing multi-source feature vectors. This data set is the composite feature map.

2. The watermelon wilt disease identification method based on fusion of leaf image features according to claim 1, characterized in that, The process of acquiring raw digital images of the watermelon leaves to be tested and constructing a multi-level analysis structure for the images of the watermelon leaves to be tested specifically includes: Under standard lighting conditions, a high-resolution color image covering the entire visible surface of the watermelon leaf under test is acquired using an image sensor. In the high-resolution color image, the leaf outline is automatically identified and segmented from the background area; An initial analysis grid covering the entire blade region is established based on the blade profile as the basic boundary. Based on the distribution information of the main vein and lateral veins of the leaf, the initial analysis grid is adaptively subdivided to form multiple sub-region grid levels with progressively smaller scales; The multi-level analysis structure is a spatial analysis framework composed of the initial analysis grid and sub-region grids of each level. Each grid cell is associated with its level and spatial location coordinates.

3. The watermelon wilt disease identification method based on fusion of leaf image features according to claim 2, characterized in that, The step of performing multi-channel spectral separation on the original digital image to extract image components reflecting different physiological states specifically includes: Analyze the intensity values ​​of each pixel in the high-resolution color image under the red, green, and blue channels; The red channel intensity value, green channel intensity value and blue channel intensity value of each pixel are combined into a three-dimensional spectral vector; Based on the preset reference range of spectral vectors for various leaf physiological states, the three-dimensional spectral vector of each pixel is classified and judged. Pixels identified as belonging to the same physiological state are extracted from the high-resolution color image, and their original spatial relationships are preserved to form an independent image layer. Repeat this discrimination and extraction process to finally obtain multiple image components corresponding to healthy chlorophyll, chlorotic tissue, necrotic tissue, and water-soaked tissue, respectively.

4. The watermelon wilt disease identification method based on fusion of leaf image features according to claim 3, characterized in that, The process of performing regional deconstruction on the composite feature map based on the multi-level analysis structure to obtain multiple partitioned feature data blocks specifically includes: Starting from the highest level grid of the composite feature map, analyze the multi-source feature vector of each grid cell; Based on the similarity measure of multi-source feature vectors between grid cells, adjacent grid cells with similar features are merged to form primary feature blocks; In the next level, the same similarity merging operation is performed on all sub-grid cells belonging to the same primary feature block to form a more refined secondary feature block; This process proceeds iteratively from top to bottom, level by level, until the lowest mesh of the multi-level analysis structure is reached. The stable feature blocks that are eventually formed at each level, together with the multi-source feature vectors of all the grid cells contained therein, are packaged into an independent partitioned feature data block.

5. The watermelon wilt disease identification method based on fusion of leaf image features according to claim 4, characterized in that, The step of matching and comparing the partition feature data block with the pre-stored watermelon wilt pathogen feature database layer by layer specifically includes: The watermelon wilt pathogen feature database stores standard feature vector sets for different infection stages and different disease sites; For each partition feature data block, the central trend features of the multi-source feature vectors of all grid cells contained therein are extracted as the representative feature vectors of the partition feature data block; The similarity between the representative feature vector of the partitioned feature data block and each vector in the standard feature vector set is calculated. Record all standard feature vectors whose similarity to the representative feature vectors of the partition feature data block exceeds a preset threshold, and label the infection stage and disease site information corresponding to all standard feature vectors; The matching and comparison results are a set of similarity records and associated infection stage and disease site information corresponding to each partition feature data block.

6. The watermelon wilt disease identification method based on fusion of leaf image features according to claim 5, characterized in that, The step of calculating the pathogen conformity index for each partition feature data block based on the matching and comparison results specifically includes: For a partitioned feature data block, count the total number of standard feature vectors that exceed a preset threshold recorded in its matching and comparison results; Analyze the infection stage corresponding to the standard feature vector recorded in the matching and comparison results, and calculate the concentration of the infection stage in the pathological development sequence; Analyze the disease location corresponding to the standard feature vector recorded in the matching and comparison results, and calculate the degree of agreement between the disease location and the actual location of the current partition feature data block on the leaf. By combining the total number, the concentration of the pathological development sequence, and the consistency of the actual location, a quantitative pathogen consistency index value is generated through weighted calculation. The pathogen conformity index value is used to characterize the degree of conformity between the leaf region corresponding to the partition feature data block and the typical characteristics of watermelon wilt.

7. The watermelon wilt disease identification method based on fused leaf image features according to claim 6, characterized in that, The pathogen conformity index, which integrates all partition feature data blocks, is used to generate the overall infection probability mapping of the watermelon leaf under test. Specifically, this includes: Obtain the leaf area covered by each partition feature data block and the calculated pathogen conformity index value; Using the spatial location of the leaf as coordinates, the pathogen conformity index value of each partition feature data block is assigned to all pixels within its coverage area; For a pixel covered by multiple partition feature data blocks, the average value of its corresponding multiple pathogen conformity index values ​​is used as the value assigned to the pixel. Spatial interpolation smoothing is performed on the values ​​of all pixels within the leaf region to form a probability distribution surface that covers the entire leaf and whose values ​​change continuously. The overall infection probability mapping is the probability distribution surface, where the value at each position represents the probability that the pixel belongs to the lesion region.

8. The watermelon wilt disease identification method based on fusion of leaf image features according to claim 1, characterized in that, The step of performing spatial clustering analysis on the overall infection probability mapping to identify potential lesion core regions, and constructing a topological network describing disease development based on the morphological evolution relationships of the lesion core regions, specifically includes: A probability threshold is set for the overall infection probability mapping, and pixels with probability values ​​higher than the threshold are marked as candidate lesions; Based on the spatial proximity of the candidate lesions, a density clustering algorithm is used to aggregate candidate lesions that are close to each other into multiple independent lesion clusters; Calculate the geometric center of each lesion cluster, and expand outward from the geometric center until all pixels within the cluster are covered, forming the initial boundary of the lesion core region; By analyzing the area, shape factor, and relative position to the leaf midrib of each lesion core region, regions that conform to the typical spatial distribution characteristics of watermelon wilt disease were selected as the final potential lesion core regions. For each potential lesion core area, key points on its boundary contour are extracted, and a regional growth path is constructed based on the contour evolution trend. Using the core area of ​​each potential lesion as a node, and based on the spatial distance between regions, the continuity of probability distribution, and the extension direction of the regional growth path, directed connection edges are established between nodes. A topological network representing the spatial association and evolution direction of lesion regions is constructed using the nodes and the directed connecting edges.

9. A storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements a watermelon wilt disease identification method that integrates leaf image features as described in any one of claims 1 to 8.

Images