Soybean disease early-stage nondestructive testing method and system based on convolutional neural network

Abstract

The invention provides a soybean disease early-stage nondestructive testing method and system based on a convolutional neural network, and relates to the technical field of computers, and the method comprises the steps: obtaining an overall multispectral aerial image, a soybean leaf local image and space coordinate information thereof, and environment data; performing spatial partitioning according to the overall multispectral aerial image to obtain a growth potential difference partition map; carrying out weight calculation according to the growth potential difference partition map and the space coordinate information to obtain a point location growth potential weight; performing feature extraction according to the point location growth potential weight and the soybean leaf local image to obtain a leaf fusion feature; performing association analysis according to the leaf fusion features to obtain infection stress indexes; and constructing a risk field according to the infection stress index, the space coordinate information and the environment data, and generating an early risk distribution diagram. According to the method, systemic physiological abnormalities before occurrence of diseases are captured from the specific morphological development level of soybean plants.

Description

Technical Field

[0001] This invention relates to the field of computer technology, and more specifically, to a method and system for early non-destructive detection of soybean diseases based on convolutional neural networks. Background Technology

[0002] Early warning of crop diseases is a core challenge in agricultural disaster prevention and mitigation. Current technologies largely rely on identifying and diagnosing specific biomarkers or spectral characteristics after disease symptoms appear. These methods are generally based on detecting isolated symptom points, failing to systematically analyze the crop as a living organism with a specific morphological structure and developmental patterns. In particular, they lack the ability to capture the systemic and synergistic physiological imbalances caused by diseases in different organs and leaf positions during the early stages of infection. For crops like soybeans, with their typical trifoliate leaves and other specific morphological structures, current technologies struggle to analyze the physiological response correlations between compound leaves and between different canopy layers. This makes it impossible to diagnose the initial infection before visible symptoms appear by leveraging the inherent synergistic changes in morphological development, leading to delayed warnings and a high risk of misjudgment due to individual morphological differences.

[0003] Based on the shortcomings of the existing technologies, there is an urgent need for a non-destructive detection method and system for early soybean diseases based on convolutional neural networks. Summary of the Invention

[0004] The purpose of this invention is to provide a method and system for early non-destructive detection of soybean diseases based on convolutional neural networks, in order to improve the aforementioned problems. To achieve the above objective, the technical solution adopted by this invention is as follows: In a first aspect, this application provides a method and system for early non-destructive detection of soybean diseases based on convolutional neural networks, including: Acquire overall multispectral aerial images of the target farmland area, local images of soybean leaves collected at preset grid points and their spatial coordinate information, as well as environmental data of each point within the corresponding time period; Spatial partitioning is performed on the overall multispectral aerial image to obtain a growth potential difference partitioning map; The weights of the growth potential at each point are calculated based on the growth potential difference zoning map and the spatial coordinate information. Feature extraction is performed based on the growth potential weight of the points and the local image of the soybean leaves. By simultaneously extracting the collaborative response features of the three leaflets, the leaf fusion features are obtained. Based on the leaf fusion characteristics, correlation analysis was performed. By quantifying the characteristic response attenuation gradient between the lower and upper leaves and between different leaflets of the same compound leaf, an infection stress index characterizing the degree of plant physiological disorder was obtained. Based on the infection stress index, the spatial coordinate information, and the environmental data, a risk field is constructed to generate an early risk distribution map.

[0005] Secondly, this application also provides a method and system for early non-destructive detection of soybean diseases based on convolutional neural networks, including: The acquisition module is used to acquire overall multispectral aerial images of the target farmland area, local images of soybean leaves collected at preset grid points and their spatial coordinate information, as well as environmental data of each point within the corresponding time period. The partitioning module is used to perform spatial partitioning based on the overall multispectral aerial image to obtain a growth potential difference partitioning map. The calculation module is used to perform weight calculation based on the growth potential difference zoning map and the spatial coordinate information to obtain the growth potential weight of the point. The extraction module is used to extract features based on the growth potential weight of the point and the local image of the soybean leaf. By simultaneously extracting the collaborative response features of the three leaflets, the leaf fusion features are obtained. The analysis module is used to perform correlation analysis based on the leaf fusion characteristics. By quantifying the characteristic response attenuation gradient between the lower and upper leaves and between different leaflets of the same compound leaf, an infection stress index characterizing the degree of plant physiological disorder is obtained. The output module is used to construct a risk field based on the infection stress index, the spatial coordinate information, and the environmental data, and generate an early risk distribution map.

[0006] The beneficial effects of this invention are as follows: This invention integrates multi-scale field images and spatial environment data, and sequentially performs spatial partitioning based on differences in canopy growth potential, point weight optimization, collaborative extraction of compound leaf structural features, systemic stress correlation analysis of plants, and risk field construction. This enables the capture of systemic physiological abnormalities before disease occurrence from the specific morphological development level of soybean plants, thereby achieving true early non-destructive detection and warning. Attached Figure Description

[0007] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0008] Figure 1 This is a flowchart illustrating a non-destructive detection method for early soybean diseases based on convolutional neural networks, as described in an embodiment of the present invention. Figure 2This is a schematic diagram of the structure of a non-destructive detection system for early soybean diseases based on a convolutional neural network, as described in an embodiment of the present invention. Figure 3 This is a schematic diagram of the structure of a non-destructive testing device for early soybean diseases based on a convolutional neural network, as described in an embodiment of the present invention.

[0009] The diagram is labeled as follows: 800, a non-destructive testing device for early soybean diseases based on convolutional neural networks; 801, processor; 802, memory; 803, multimedia component; 804, I / O interface; 805, communication component; 901, acquisition module; 902, partitioning module; 903, calculation module; 904, extraction module; 905, analysis module; 906, output module. Detailed Implementation

[0010] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0011] It should be noted that similar reference numerals and letters in the following figures indicate similar items; therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures. Furthermore, in the description of this invention, terms such as "first," "second," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0012] Example 1

[0013] This embodiment provides a method and system for early non-destructive detection of soybean diseases based on convolutional neural networks.

[0014] See Figure 1 The figure shows that the method includes steps S100 to S600.

[0015] Step S100: Acquire the overall multispectral aerial image of the target farmland area, local images of soybean leaves collected at preset grid points and their spatial coordinate information, as well as environmental data of each point within the corresponding time period; Understandably, the core of this step lies in constructing a multi-scale, multi-modal farmland data acquisition system. Overall multispectral aerial images are typically acquired by drones equipped with multispectral cameras, covering the entire target field and including visible light and multiple near-infrared bands, to macroscopically reflect differences in canopy structure and overall growth status. The pre-set grid points ensure spatial representativeness and uniform distribution of sampling points. At these points, high-resolution handheld or fixed imaging devices are used to acquire local images of soybean leaves, ensuring clear distinction of leaf vein texture and subtle leaf surface features. Their spatial coordinates can be synchronously recorded using GPS or RTK positioning technology. Environmental data involves microclimate factors such as temperature, relative humidity, light intensity, and canopy wind speed at various points within the corresponding time period, which can be continuously monitored through small weather stations or sensor node networks deployed in the field. All these data must be synchronized temporally and precisely correlated spatially to provide a unified data foundation for subsequent analysis.

[0016] The layout of the pre-set grid points must be closely integrated with the actual shape and area of ​​the target farmland. Essentially, it establishes a systematic spatial sampling framework within the irregular boundaries of the farmland. The grid starts at a corner of the farmland, and the grid density (i.e., the number of sampling points) is determined based on the total area of ​​the farmland, the complexity of the terrain, and the expected monitoring accuracy requirements, using spatial statistical methods (such as determining a reasonable sampling interval based on variogram analysis) or following agronomic sampling criteria. Preferably, for approximately rectangular standard fields, a regular square grid can be used; for fields with irregular boundaries, it is necessary to first determine the circumscribed rectangle or establish a baseline grid based on the main operation direction, and then eliminate invalid points falling outside the field boundaries to ensure that effective sampling points are evenly distributed throughout the field. This design aims to obtain samples that best represent the spatial variability of the entire field at the optimal cost-effectiveness ratio, providing a reliable spatial data foundation for subsequent analysis.

[0017] Step S200: Spatial partitioning is performed based on the overall multispectral aerial image to obtain a growth potential difference partitioning map; It should be noted that the purpose of spatial zoning is to transform macroscopic farmland images into spatial patterns of growth potential with clear agronomical significance. This step quantifies and regionalizes canopy closure and physiological activity based on vegetation indices such as leaf area index and chlorophyll content derived from multispectral information. The idea is to acknowledge the inherent spatial heterogeneity of soybean growth in the field and classify the continuously changing canopy state into several representative growth potential level regions.

[0018] Step S300: Calculate the weights based on the growth potential difference zoning map and spatial coordinate information to obtain the growth potential weights of the points. Understandably, the essence of step S300, weight calculation, is to establish a bridge between the macroscopic growth potential distribution and the importance of microscopic point sampling. It is based on a core assumption: different growth potential regions have different indicative significance and sensitivity to early diseases. This step spatially matches discrete point information with continuous growth potential zoning maps, assigning each sampling point a weight coefficient that reflects the growth status of its region. This processing ensures that subsequent feature analysis of leaf images is no longer isolated but considers the representativeness of the point within the overall field growth context, optimizing the utilization efficiency of the information carried by limited sampling points and reflecting the idea of ​​improving the targeting of monitoring strategies under resource constraints.

[0019] Step S400: Based on the growth potential weight of the point and the local image of soybean leaves, feature extraction is performed. By simultaneously extracting the collaborative response features of the three leaflets, the leaf fusion features are obtained. It should be noted that this step incorporates the unique trifoliate compound leaf morphology of soybean as prior knowledge into the feature learning process. Unlike existing techniques that use general feature recognition on leaf images, this approach designs a feature extraction mechanism specifically capable of capturing the synergistic response relationships among the three leaflets. This treatment recognizes that early disease stress does not uniformly affect each leaflet of the compound leaf, but rather disrupts the inherent physiological synchronicity and symmetry among the three. Therefore, by simultaneously extracting and fusing features from the three leaflets, the aim is to discover this synergistic change pattern, which is crucial for diagnosing physiological imbalances invisible to the naked eye.

[0020] Step S500: Based on the leaf fusion characteristics, perform correlation analysis. By quantitatively analyzing the characteristic response attenuation gradient between the lower and upper leaves and between different leaflets of the same compound leaf, obtain the infection stress index that characterizes the degree of physiological disorder of the plant. Understandably, the correlation analysis in this step aims to diagnose disease stress from the perspective of the plant as a whole. Its approach is based on the pathological pattern of systemic infection in soybean diseases, where infection often begins in the lower, older leaves and tends to spread upwards to the newer leaves. This step captures this systemic physiological imbalance signal by quantitatively analyzing the characteristic response gradients (such as decay rate and asymmetry) between different vertical leaf positions (lower and upper leaves) and between different leaflets within the same compound leaf. This correlation analysis based on plant structure and infection pathways can elevate point-like leaf feature anomalies to a judgment of the overall health status of the plant, significantly improving the reliability of early identification.

[0021] Step S600: Construct a risk field based on infection stress indicators, spatial coordinate information and environmental data to generate an early risk distribution map.

[0022] It is important to note that risk field construction is a crucial step in realizing spatial extrapolation from "single-plant diagnosis" to "field early warning." This step does not treat the infection stress indicators at each location as independent events, but rather places them within specific spatial locations and environmental contexts to simulate the potential spread dynamics of the disease in the field. This treatment comprehensively considers factors such as the spatial distance between plants and canopy microclimate (e.g., the impact of wind speed and direction on spore dispersal). By constructing a physically meaningful dispersal model, point-like stress indicators are "diffused" into a continuous risk distribution surface. The resulting early risk distribution map not only identifies current high-risk points but also predicts the potential range and intensity of disease spread in the future, providing direct spatial evidence for precise control decisions.

[0023] Further, step S200 includes steps S210 to S230.

[0024] Step S210: Calculate the canopy closure based on the overall multispectral aerial image. By analyzing the difference in reflectance between the visible and near-infrared bands, calculate the leaf area index pixel by pixel to obtain the soybean canopy closure distribution map. Step S220: Divide the growth potential regions according to the soybean canopy closure distribution map, and merge the continuous pixel regions with similar canopy closure values ​​through a clustering algorithm to obtain a preliminary partition map. Step S230: Based on the preliminary zoning map, the growth potential level is labeled. By combining the prior knowledge of the canopy closure of soybeans at different growth stages, each zone is assigned a level label representing the strength of growth potential, and a growth potential difference zoning map is generated.

[0025] Specifically, step S210 first calculates the vegetation index pixel by pixel on the acquired overall multispectral aerial image, especially by using the reflectance combination of red light and near-infrared bands that are sensitive to vegetation chlorophyll content and biomass to obtain the initial distribution of normalized vegetation index (NDVI); then, based on the established empirical conversion model of soybean canopy leaf area index and NDVI value, the vegetation index value of each pixel is converted into the corresponding leaf area index (LAI). Since the leaf area index directly reflects the total area of ​​leaves per unit ground, its value can effectively characterize the canopy closure degree, thereby obtaining a soybean canopy closure degree distribution map with a quantified value of canopy closure for each pixel. Step S220 divides the growth potential areas based on the above distribution map. Its core is not a simple numerical grouping, but a clustering method that considers spatial continuity. This process not only groups pixels with similar canopy closure values ​​into one category, but also ensures that the merged areas are spatially continuous through morphological constraints (such as considering the spatial connectivity of pixel areas). It also pays special attention to preserving the strip-shaped distribution characteristics determined by the soybean planting row spacing. This makes the final preliminary partition map reflect the differences in canopy closure, and its partition shape also fits the actual planting pattern of soybeans in the field, avoiding the unreasonable division of plants with continuous growth in the same row. Step S230 then performs the final labeling of the growth vigor level of the preliminary zoning. The key to this process is the introduction of prior knowledge of specific soybean growth stages. That is, based on the ideal canopy closure range that the soybean population should theoretically reach at different growth stages (such as branching stage, flowering stage, and pod-setting stage), the average canopy closure value of each zoning is compared with the ideal range under the current growth stage. This transforms the zoning from a simple canopy closure level into a growth vigor level with clear agronomical significance (such as vigorous, normal, and weak). The resulting growth vigor difference zoning map is not only a spatial variation map, but also a spatialized knowledge map containing the evaluation of soybean growth and development status.

[0026] In practice, the above steps can be implemented through the following computer processing: First, input multispectral image data containing red band reflectance and near-infrared band reflectance, where the reflectance values ​​have been normalized to the range [0,1]. Next, calculate the canopy closure: for each pixel, calculate the Normalized Vegetation Index (NDVI) using the following formula: ; In the formula, NIR represents near-infrared reflectance, R represents red light reflectance, and a higher NDVI value indicates better vegetation cover. Next, the vegetation index is converted to leaf area index (LAI) to quantify canopy closure; a linear empirical model specific to soybeans is used. ; In the formula, a and b are empirical coefficients based on soybean canopy research calibration, for example, a=0.5, b=0.5. These coefficients ensure that the LAI value can reasonably reflect the leaf density of the soybean canopy, and the output is the LAI value of each pixel, forming a canopy closure distribution map. Next, growth potential regions are divided: the K-means clustering algorithm is used to aggregate pixels into K clusters based on their LAI values. The value of K is determined by the elbow rule, preferably K=3, to distinguish different growth potential levels; the clustering process minimizes the intra-cluster squared error of the objective function: ; In the formula, SSE is the squared error within the cluster of objective functions, x j μ represents the LAI value of the j-th pixel. i Indicates cluster C i The central LAI value, C i This represents the pixel set of the i-th cluster. After clustering, morphological opening and closing operations (such as using a 3×3 structuring element) are applied to smooth the region boundaries, ensuring spatial continuity and preserving soybean row sowing characteristics, outputting a preliminary partition map where each region consists of consecutive pixels with similar LAI values. Finally, growth potential is calibrated: the average LAI value of each partition is calculated and compared with the ideal LAI range for the current soybean growth stage; for example, at the flowering stage, the ideal LAI range is set to [T]. low T high If ]=[3.0,5.0], then the grade determination formula is: if the average LAI of the partition is... <T low If the growth potential is "weak", the level is "normal"; if the average LAI value of the partition is within the ideal LAI range, the level is "normal"; if the average LAI of the partition is greater than T... high The level is "vigorous"; among them, T low and T high The threshold is set based on prior knowledge of soybean growth period, and the output is a growth potential difference zoning map, which includes the growth potential level of each zoning.

[0027] Further, step S300 includes steps S310 to S330.

[0028] Step S310: Reconstruct spatial continuity based on the growth potential difference zoning map. By establishing a spatial autocorrelation characteristic model of growth potential values, the discrete zoning boundary is transformed into a continuously changing growth potential intensity distribution, resulting in a continuous growth potential surface. Step S320: Extract the growth potential intensity of each point based on the continuous growth potential surface and spatial coordinate information. By mapping the spatial coordinates of each point to the continuous growth potential surface, obtain the growth potential intensity value corresponding to the point, and obtain the basic growth potential intensity of the point. Step S330: Optimize the disease sensitivity weight based on the basic growth potential intensity of the point. By analyzing the response characteristics of different growth potential areas of soybean canopy to early diseases, establish the mapping relationship between growth potential intensity and disease occurrence probability, and obtain the growth potential weight of the point.

[0029] Specifically, the spatial continuity reconstruction process in step S310 aims to overcome the boundary abrupt change problem in growth potential maps based on discrete partitions. Its core is the recognition that the growth state of crops in farmland is a gradual rather than abrupt change in space. This process quantifies the mutual influence of growth potential values ​​between any two points by establishing a spatial autocorrelation model, and uses algorithms such as Kriging interpolation to interpolate the discrete partition center values ​​into a continuous and smooth surface, thereby generating a continuous growth potential surface that better reflects the actual field conditions. Step S320 then performs precise extraction of point-specific growth potential intensity. By spatially overlaying the precise geographic coordinates of each sampling point with the continuous surface, bilinear interpolation is used to extract the growth potential intensity value at the corresponding coordinate position from the surface raster data, thus assigning a quantitative growth potential background value that reflects the characteristics of its microenvironment to each independent sampling point. The weight optimization process in step S330 incorporates the epidemiological laws of soybean diseases, based on the fact that there are significant differences in the sensitivity of different growth vigor regions to disease infection. This process establishes a nonlinear mapping function between growth vigor intensity and historical disease occurrence probability (usually manifested as weak growth areas having a higher risk of disease occurrence), converting the basic growth vigor intensity value into a weight coefficient that reflects the disease sensitivity of the site. The final weight value includes both the growth status information of the site itself and agronomical knowledge of disease susceptibility.

[0030] In practice, the above weight calculation steps are implemented through the following computer processing: the system input is a growth potential difference zoning map containing the growth potential levels of each zone and their corresponding average LAI values, as well as the latitude and longitude coordinates of each sampling point. For step S310, spatial continuity reconstruction is performed using ordinary kriging interpolation. First, the variogram is calculated to quantify spatial autocorrelation: ; In the formula, n is the nugget value (taken as 0.1), s is the sill value (taken as 1.2), and r is the range (taken as 20 meters). This parameter combination can accurately characterize the spatial variation of soybean field growth potential. The predicted values ​​at each location are obtained by solving the Kriging equations, generating a continuous growth potential surface with a resolution of 10 cm. For step S320, the bilinear interpolation method is used to extract the point growth potential intensity from the grid surface, with the following formula: ; In the formula, a', b', c', and d' are interpolation coefficients, calculated from the growth potential intensity values ​​of the four grid points surrounding the point to be interpolated, and x and y are the point coordinates. This formula can smoothly estimate the growth potential intensity at any point. For step S330, a disease sensitivity weighting function is established: ; In the formula, w represents the weighting function, g is the growth potential intensity of the point, g0 is the sensitivity threshold (set to 3.0), and α is the adjustment coefficient (taken as -0.8). This parameter setting makes the weight increase rapidly when the growth potential is lower than 3.0, which is consistent with the rule that weak areas are more susceptible to disease.

[0031] Further, step S400 includes steps S410 to S430.

[0032] Step S410: Based on the local image of soybean leaves, the compound leaf structure is analyzed. By constructing a feature extraction network with three-way symmetry, the three leaflets are treated as a whole and feature encoded under geometric constraints to obtain preliminary compound leaf structure features. Step S420: Based on the preliminary compound leaf structure characteristics, conduct a collaborative response analysis. By establishing a feature transmission model between leaflets, capture the synchronous change pattern of the three leaflets in physiological response, and obtain the compound leaf collaborative feature vector. Step S430: Based on the synergistic feature vector of the compound leaf and the growth potential weight of the point, feature fusion is performed. By using the growth potential weight as the feature importance coefficient, the synergistic features of the three leaflets are adaptively weighted and integrated to obtain the leaf fusion features.

[0033] Specifically, step S410, the compound leaf structure analysis, targets the unique trifoliate compound leaf morphology of soybean and employs a convolutional neural network structure with three-way symmetry constraints for feature extraction. This network structure uses a three-branch design with shared weights to learn the geometric features of the three leaflets respectively, and introduces geometric constraints on the spatial relative positional relationship of the three leaflets during the feature encoding process. This allows the network to understand the compound leaf as a whole structure, thereby extracting preliminary compound leaf structural features that reflect the overall morphology and spatial distribution relationship of the three leaflets. Step S420, the collaborative response analysis, further explores the intrinsic correlation of the three leaflets in physiological response. This process constructs a feature transfer model between leaflets, such as using graph attention mechanisms or specific feature interaction layers, to model and quantify how the feature changes of one leaflet affect the feature responses of the other two leaflets. The aim is to capture subtle collaborative change patterns such as asynchronous and asymmetric responses of the three leaflets that may occur under early disease stress, thereby transforming individual structural features into feature vectors that can characterize the physiological synergy within the compound leaf. The feature fusion process in step S430 aims to combine microscopic leaf features with macroscopic field growth potential. Its core operation is to use the growth potential weight, which represents the disease sensitivity of a location, as an importance coefficient to adaptively weight the compound leaf synergistic feature vector. For example, for locations with weaker growth potential and higher sensitivity, their synergistic feature vector will be given higher weight, thus occupying a more important position in subsequent analysis. This fusion method ensures that the final leaf fusion features include not only the physiological details of the leaf itself, but also its diagnostic value in the context of the overall growth potential in the field.

[0034] In practice, the above feature extraction and fusion steps are implemented through the following computer processing: The system input consists of three local images of soybean leaves (224×224 pixels, RGB channels) and point growth potential weights (scalar values). For step S410, a three-branch convolutional neural network with shared weights is used to parse the compound leaf structure. Each branch processes one leaf image. The network structure includes convolutional layers (3×3 kernel size, stride 1, padding 1), a ReLU activation function, and a max pooling layer (pooling size 2×2). Finally, a 256-dimensional feature vector V for each leaflet is obtained through global average pooling. k (k=1,2,3). To further encode spatial relationships, a positional encoding matrix (size 3×256) is introduced, whose elements are generated by a sine function: ; In the formula, P(pos,2d) is the position encoding matrix, pos is the leaflet position index (0,1,2), and d is the dimension index. The geometric constraint feature vector V is obtained by subtracting the position encoding matrix from the feature vector. kThis yields preliminary compound leaf structure characteristics. For step S420, a graph attention mechanism is used for collaborative response analysis, and V... k 'Construct a fully connected graph as nodes. The attention coefficient between nodes m and n is calculated as follows:' ; In the formula, e mn Let represent the attention coefficients between nodes m and n, W be the learnable weight matrix (size 128×256), a be the learnable attention vector (dimension 256), || denotes the concatenation operation, T denotes the matrix transpose, LeakyReLU is the activation function with a negative slope of 0.2, and v m 'and v n Let ' represent the geometric constraint feature vectors of leaflets m and n, respectively. Then, the attention weights between nodes are calculated and normalized. Based on the weights, the features of adjacent nodes are weighted and aggregated, and a 128-dimensional compound leaf collaborative feature vector is obtained through pooling. Finally, the growth potential weight is used as the feature importance coefficient, and a scalar multiplication operation is performed on the collaborative feature vector to achieve the fusion of leaf micro-features and field growth potential background, ultimately outputting a 128-dimensional leaf fusion feature vector, i.e., the leaf fusion feature.

[0035] Further, step S500 includes steps S510 to S530.

[0036] Step S510: Perform vertical stress gradient analysis based on leaf fusion characteristics. By constructing the characteristic response curve from the lower leaves to the upper leaves of the plant, calculate the decay rate of characteristic values ​​between different leaf positions, and obtain the vertical stress gradient vector. Step S520: Analyze the coordinated changes within the compound leaf based on the vertical stress gradient vector. By establishing the characteristic response correlation matrix of the three leaflets on the same petiole, quantify the asymmetric change patterns among the leaflets in the three-leaf structure and obtain the coordinated variation index of the compound leaf. Step S530: Infection assessment is conducted based on the vertical stress gradient vector and the compound leaf synergistic variation index. By integrating the stress transmission characteristics in the vertical direction of the plant with the abnormal distribution pattern at the compound leaf level, a comprehensive evaluation index characterizing the degree of systemic spread of the disease is constructed, thus obtaining the infection stress index.

[0037] Specifically, the vertical stress gradient analysis in step S510 aims to capture the typical pattern of systemic disease spread within soybean plants. This treatment is based on the pathological pattern that soybean diseases often start from the lower, older leaves and spread to the upper, newer leaves. Specifically, it is achieved by comparing the leaf fusion feature values ​​from leaves at different vertical heights (i.e., lower, middle, and upper parts) of the same plant to construct a response curve that reflects the change of feature values ​​as the leaf position increases. Then, the slope or decay rate of this curve is calculated, thereby transforming discrete leaf features into a gradient vector that can quantify the spread of stress in the vertical direction of the plant. Step S520, the analysis of synergistic changes within compound leaves, focuses on abnormal responses within a single reproductive unit. This processing targets the trifoliate compound leaf structure of soybean, constructing a pattern describing the relationship between the three leaflets by calculating the differences or similarities in their characteristic values ​​pairwise. Its core lies in quantifying whether the three leaflets, which should maintain a relatively balanced synergistic relationship under healthy conditions, have experienced imbalances in response amplitude or temporal asynchrony due to early infection, thus transforming an abstract synergy into a specific compound leaf synergistic variation index. Step S530, the infection assessment, is a comprehensive decision based on information from the first two steps. This processing integrates the stress diffusion directionality represented by the vertical gradient vector with the local physiological disorder pattern represented by the compound leaf synergistic variation index. For example, through weighted or rule-based inference models, it determines whether local compound leaf anomalies conform to the overall stress diffusion path of the plant, thereby generating a comprehensive infection stress index that reflects both stress intensity and whether its spatial distribution pattern conforms to the systemic infection pattern of the disease, providing reliable single-plant-level input for the final field-scale risk projection.

[0038] Further, step S600 includes steps S610 to S630.

[0039] Step S610: Model the propagation path based on spatial location and environmental data. Construct a spatial correlation diagram representing the probability of disease spore propagation by integrating data on plant spacing and canopy wind speed and direction, and obtain the disease propagation network. Step S620: Calculate the risk transmission based on the disease transmission network and infection stress indicators. Iterate and update the infection probability value of each node on the graph structure to simulate the dynamic process of disease transmission in the field and obtain the preliminary risk distribution. Step S630: Conduct a spatiotemporal risk assessment based on the preliminary risk distribution. By combining the impact of temperature and humidity data on the disease development rate, predict the spatial distribution pattern of the disease in a specific future period and generate an early risk distribution map.

[0040] Step S610, the propagation path modeling, transforms the abstract field space into a biologically meaningful disease propagation network. Its core lies in expanding the spatial adjacency relationship between plants from simple Euclidean distance to a propagation probability influenced by canopy microclimate. This process first determines the network nodes based on the actual geographical location of the sampling points. Then, it considers not only the plant spacing but also, more importantly, integrates current wind speed and direction data, assigning higher connection weights to adjacent nodes on the downwind side. This constructs a weighted directed graph, which quantifies the potential probability of pathogen spores spreading with airflow between different locations, laying the foundation for risk spatial diffusion simulation. Step S620, the risk propagation calculation, is performed on this network structure. The single-plant infection stress index is used as the initial infection probability of each node in the graph. Through the information propagation mechanism on the graph, the dynamic process of pathogen diffusion along the network edges over several propagation cycles is simulated. In each iteration, the infection probability of a node is updated by the probabilities and connection weights of its neighboring nodes. This simulation ultimately outputs a preliminary risk distribution that not only includes initial stress information but also reflects the cumulative effect of mutual infection risks between locations. Step S630, the spatiotemporal risk assessment, introduces a time dimension to achieve the goal of early warning. This process is based on the biological characteristics of the pathogen and uses key environmental data such as temperature and humidity as control parameters, which are input into the disease development rate model to correct and predict the risk evolution from the current "preliminary risk distribution" to a specific period in the future (such as the next 3-7 days). The resulting early risk distribution map is a dynamic risk assessment product that integrates the current spatial risk pattern and the development trend prediction driven by the environment.

[0041] Example 2

[0042] like Figure 2 As shown, this embodiment provides a non-destructive detection system for early soybean diseases based on convolutional neural networks. The system includes: The acquisition module 901 is used to acquire overall multispectral aerial images of the target farmland area, local images of soybean leaves collected at preset grid points and their spatial coordinate information, as well as environmental data of each point within the corresponding time period. The partitioning module 902 is used to perform spatial partitioning based on the overall multispectral aerial image to obtain a growth potential difference partitioning map. Calculation module 903 is used to calculate the weights based on the growth potential difference zoning map and spatial coordinate information to obtain the growth potential weights of the points. The extraction module 904 is used to extract features based on the growth potential weight of the point and the local image of soybean leaves. By simultaneously extracting the collaborative response features of the three leaflets, the leaf fusion features are obtained. Analysis module 905 is used to perform correlation analysis based on leaf fusion characteristics. By quantifying the characteristic response attenuation gradient between the lower and upper leaves and between different leaflets of the same compound leaf, infection stress indicators characterizing the degree of physiological disorder of the plant are obtained. Output module 906 is used to construct a risk field based on infection stress indicators, spatial coordinate information and environmental data, and generate an early risk distribution map.

[0043] In one specific embodiment of this application, the partitioning module 902 includes: The first partition unit is used to calculate the canopy closure based on the overall multispectral aerial image. By analyzing the difference in reflectance between the visible and near-infrared bands, the leaf area index is calculated pixel by pixel to obtain the soybean canopy closure distribution map. The second partitioning unit is used to divide the growth potential region according to the soybean canopy closure distribution map. It uses a clustering algorithm to merge consecutive pixel regions with similar canopy closure values ​​to obtain a preliminary partitioning map. The third partition unit is used to label the growth potential level based on the preliminary partition map. By combining the prior knowledge of the canopy closure of soybeans at different growth stages, each partition is assigned a level label representing the strength of growth potential, thus generating a growth potential difference partition map.

[0044] In one specific embodiment of this application, the computing module 903 includes: The first calculation unit is used to reconstruct spatial continuity based on the growth potential difference partition map. By establishing a spatial autocorrelation characteristic model of growth potential values, the discrete partition boundaries are transformed into a continuously changing growth potential intensity distribution, resulting in a continuous growth potential surface. The second calculation unit is used to extract the growth potential intensity of a point based on the continuous growth potential surface and spatial coordinate information. By mapping the spatial coordinates of each point to the continuous growth potential surface, the growth potential intensity value corresponding to the point is obtained, and the basic growth potential intensity of the point is obtained. The third calculation unit is used to optimize the disease sensitivity weight based on the basic growth potential intensity of the point. By analyzing the response characteristics of different growth potential areas of soybean canopy to early diseases, the mapping relationship between growth potential intensity and disease occurrence probability is established to obtain the growth potential weight of the point.

[0045] In one specific embodiment of this application, the extraction module 904 includes: The first extraction unit is used to analyze the compound leaf structure based on local images of soybean leaves. By constructing a feature extraction network with three-way symmetry, the three leaflets are treated as a whole and feature encoded under geometric constraints to obtain preliminary compound leaf structure features. The second extraction unit is used to perform synergistic response analysis based on the preliminary compound leaf structure characteristics. By establishing a feature transfer model between leaflets, it captures the synchronous change pattern of the three leaflets in physiological response and obtains the compound leaf synergistic feature vector. The third extraction unit is used to perform feature fusion based on the synergistic feature vector of the compound leaf and the growth potential weight of the point. By using the growth potential weight as the feature importance coefficient, the synergistic features of the three leaflets are adaptively weighted and integrated to obtain the leaf fusion features.

[0046] In one specific embodiment of this application, the analysis module 905 includes: The first analysis unit is used to perform vertical stress gradient analysis based on leaf fusion characteristics. By constructing the characteristic response curve from the lower leaves to the upper leaves of the plant, the decay rate of characteristic values ​​between different leaf positions is calculated to obtain the vertical stress gradient vector. The second analysis unit is used to analyze the coordinated changes within the compound leaf based on the vertical stress gradient vector. By establishing the characteristic response correlation matrix of the three leaflets on the same petiole, the asymmetric change pattern among the leaflets in the three-leaf structure is quantified to obtain the coordinated variation index of the compound leaf. The third analysis unit is used to assess infection based on the vertical stress gradient vector and the compound leaf synergistic variation index. By integrating the stress transmission characteristics in the vertical direction of the plant with the abnormal distribution pattern at the compound leaf level, a comprehensive evaluation index characterizing the degree of systemic spread of the disease is constructed, and the infection stress index is obtained.

[0047] In one specific embodiment of this application, the output module 906 includes: The first output unit is used to model the propagation path based on spatial location and environmental data. By integrating data on plant spacing and canopy wind speed and direction, a spatial correlation diagram representing the probability of disease spore propagation is constructed to obtain the disease propagation network. The second output unit is used to calculate the risk transmission based on the disease transmission network and infection stress indicators. By iteratively updating the infection probability value of each node on the graph structure, the dynamic process of disease transmission in the field is simulated to obtain the preliminary risk distribution. The third output unit is used to conduct spatiotemporal risk assessment based on the preliminary risk distribution. By combining the impact of temperature and humidity data on the disease development rate, it predicts the spatial distribution pattern of the disease in a specific future period and generates an early risk distribution map.

[0048] Example 3

[0049] Corresponding to the above method embodiments, this embodiment also provides a soybean disease early non-destructive testing device based on convolutional neural networks. The soybean disease early non-destructive testing device based on convolutional neural networks described below and the soybean disease early non-destructive testing method based on convolutional neural networks described above can be referred to and correspond to each other.

[0050] Figure 3 This is a block diagram illustrating an early non-destructive testing device 800 for soybean diseases based on a convolutional neural network, according to an exemplary embodiment. Figure 3 As shown, the soybean disease early non-destructive testing device 800 based on convolutional neural networks may include: a processor 801 and a memory 802. The soybean disease early non-destructive testing device 800 may also include one or more of a multimedia component 803, an I / O interface 804, and a communication component 805.

[0051] The processor 801 controls the overall operation of the soybean disease early non-destructive testing device 800 based on convolutional neural networks to complete all or part of the steps in the aforementioned soybean disease early non-destructive testing method based on convolutional neural networks. The memory 802 stores various types of data to support the operation of the soybean disease early non-destructive testing device 800 based on convolutional neural networks. This data may include, for example, instructions for any application or method operating on the soybean disease early non-destructive testing device 800 based on convolutional neural networks, as well as application-related data, such as contact data, sent and received messages, images, audio, video, etc. The memory 802 can be implemented using any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read-Only Memory (EPROM), Programmable Read-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. The multimedia component 803 may include a screen and an audio component. The screen may be, for example, a touchscreen, and the audio component is used to output and / or input audio signals. For example, the audio component may include a microphone for receiving external audio signals. The received audio signals may be further stored in the memory 802 or transmitted via the communication component 805. The audio component also includes at least one speaker for outputting audio signals. I / O interface 804 provides an interface between processor 801 and other interface modules, such as keyboards, mice, and buttons. These buttons can be virtual or physical. Communication component 805 is used for wired or wireless communication between the convolutional neural network-based soybean disease early non-destructive testing device 800 and other devices. Wireless communication methods include Wi-Fi, Bluetooth, Near Field Communication (NFC), 2G, 3G, or 4G, or a combination thereof. Therefore, the corresponding communication component 805 may include a Wi-Fi module, a Bluetooth module, and an NFC module.

[0052] In an exemplary embodiment, a soybean disease early non-destructive testing device 800 based on a convolutional neural network can be implemented by one or more application-specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field-programmable gate arrays (FPGAs), controllers, microcontrollers, microprocessors, or other electronic components to perform the aforementioned soybean disease early non-destructive testing method based on a convolutional neural network.

[0053] In another exemplary embodiment, a computer-readable storage medium including program instructions is also provided. When executed by a processor, these program instructions implement the steps of the aforementioned method for early non-destructive detection of soybean diseases based on a convolutional neural network. For example, the computer-readable storage medium may be the aforementioned memory 802 including program instructions, which may be executed by a processor 801 of a convolutional neural network-based early non-destructive detection device 800 for soybean diseases to complete the aforementioned method for early non-destructive detection of soybean diseases based on a convolutional neural network.

[0054] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A soybean disease early non-destructive detection method based on a convolutional neural network, characterized in that, The method comprises the following steps: acquiring an overall multispectral aerial image of a target farmland area, local images of soybean leaves collected at preset grid points and spatial coordinate information thereof, and environmental data of each point within a corresponding time period; spatially partitioning according to the overall multispectral aerial image to obtain a growth potential difference partition map; weight calculation according to the growth potential difference partition map and the spatial coordinate information to obtain point growth potential weight; feature extraction according to the point growth potential weight and the local images of soybean leaves to obtain leaf fusion features by synchronously extracting collaborative response features of three leaflets; correlation analysis according to the leaf fusion features to obtain an infection stress index representing the degree of physiological disorder of the plant by quantitatively analyzing the feature response attenuation gradient between lower and upper leaves and between different leaflets of the same compound leaf; risk field construction according to the infection stress index, the spatial coordinate information and the environmental data to generate an early risk distribution map.

2. The method for early-stage non-destructive detection of soybean diseases based on convolutional neural network according to claim 1, characterized in that, spatial partitioning according to the overall multispectral aerial image to obtain a growth potential difference partition map, comprising: canopy density calculation according to the overall multispectral aerial image to obtain a soybean canopy density distribution map by analyzing the reflectance difference between the visible and near-infrared bands and calculating the leaf area index pixel by pixel; growth potential region division according to the soybean canopy density distribution map to obtain a preliminary partition map by merging continuous pixel regions with similar canopy density values through a clustering algorithm; growth potential grade calibration according to the preliminary partition map to generate a growth potential difference partition map by assigning each partition a grade identifier representing the strength of the growth potential in combination with the prior knowledge of the canopy density of soybeans at different growth stages. 3.The soybean disease early stage non-destructive detection method based on convolutional neural network according to claim 1, characterized in that, weight calculation according to the growth potential difference partition map and the spatial coordinate information to obtain point growth potential weight, comprising: spatial continuity reconstruction according to the growth potential difference partition map to obtain a continuous growth potential surface by establishing a spatial autocorrelation model of the growth potential value and converting the discrete partition boundaries into a continuously changing growth potential intensity distribution; point growth potential intensity extraction according to the continuous growth potential surface and spatial coordinate information to obtain the point growth potential weight by mapping the spatial coordinates of each point to the continuous growth potential surface to obtain the corresponding growth potential intensity value of the point, and obtaining the point basis growth potential intensity; disease sensitivity weight optimization according to the point basis growth potential intensity to obtain the point growth potential weight by analyzing the response characteristics of different growth potential regions of the soybean canopy to early diseases and establishing a mapping relationship between the growth potential intensity and the disease occurrence probability.

4. The method for early-stage non-destructive detection of soybean diseases based on convolutional neural network according to claim 1, characterized in that, feature extraction according to the point growth potential weight and the local images of soybean leaves to obtain leaf fusion features by synchronously extracting collaborative response features of three leaflets, comprising: compound leaf structure analysis according to the local images of soybean leaves to obtain preliminary compound leaf structure features by constructing a feature extraction network with three-way symmetry and performing feature coding of three leaflets as a whole under geometric constraints; According to the preliminary compound leaf structure characteristics, a cooperative response analysis is performed, a characteristic transmission model between the three leaflets is established, synchronous change rules of the three leaflets in physiological response are captured, and a compound leaf cooperative characteristic vector is obtained; According to the compound leaf cooperative characteristic vector and the point growth potential weight, a feature fusion is performed, the growth potential weight is taken as a characteristic importance coefficient, the cooperative characteristics of the three leaflets are adaptively weighted and integrated, and a leaf fusion feature is obtained.

5. The method for early-stage non-destructive detection of soybean diseases based on convolutional neural network according to claim 1, wherein, According to the leaf fusion feature, a correlation analysis is performed, a characteristic response attenuation gradient between lower leaves and upper leaves and between different leaflets of the same compound leaf is quantitatively analyzed, and an infection stress index representing a physiological disorder degree of the plant is obtained, including: According to the leaf fusion feature, a vertical direction stress gradient analysis is performed, a characteristic response curve from the lower leaves of the plant to the upper leaves is constructed, an attenuation rate of the characteristic values between different leaf positions is calculated, and a vertical stress gradient vector is obtained; According to the vertical stress gradient vector, a compound leaf internal cooperative change analysis is performed, a characteristic response correlation matrix of the three leaflets on the same petiole is established, an asymmetric change mode between the leaflets in the three-leaf compound leaf structure is quantified, and a compound leaf cooperative variation index is obtained; According to the vertical stress gradient vector and the compound leaf cooperative variation index, an infection evaluation is performed, a comprehensive evaluation index representing a degree of systematic diffusion of the disease is constructed by fusing the stress propagation characteristics in the vertical direction of the plant and the abnormal distribution mode in the horizontal direction of the compound leaf, and an infection stress index is obtained.

6. A soybean disease early non-destructive detection system based on convolutional neural network, characterized in that, including: An acquisition module is configured to acquire an overall multispectral aerial image of a target farmland region, local images of soybean leaves collected at preset grid points and spatial coordinate information thereof, and environmental data of each point within a corresponding time period; A partition module is configured to perform spatial partitioning based on the overall multispectral aerial image to obtain a growth potential difference partition map; A calculation module is configured to calculate a point growth potential weight based on the growth potential difference partition map and the spatial coordinate information; An extraction module is configured to extract features based on the point growth potential weight and the local images of the soybean leaves, and obtain a leaf fusion feature by synchronously extracting cooperative response features of the three leaflets; An analysis module is configured to perform correlation analysis based on the leaf fusion feature, quantitatively analyze a characteristic response attenuation gradient between lower leaves and upper leaves and between different leaflets of the same compound leaf, and obtain an infection stress index representing a physiological disorder degree of the plant; An output module is configured to construct a risk field based on the infection stress index, the spatial coordinate information, and the environmental data, and generate an early risk distribution map.

7. The system for early-stage non-destructive detection of soybean diseases based on convolutional neural network according to claim 6, wherein, The partition module includes: A first partition unit is configured to calculate a canopy density based on the overall multispectral aerial image, calculate a leaf area index pixel by pixel by analyzing reflectivity differences in visible and near-infrared bands, and obtain a soybean canopy density distribution map; A second partition unit is configured to divide a growth potential region based on the soybean canopy density distribution map, and obtain a preliminary partition map by merging continuous pixel regions with similar canopy density values through a clustering algorithm. The third partition unit is configured to perform growth potential grade labeling according to the preliminary partition map, and assign a grade identifier representing the strength of the growth potential to each partition by combining the prior knowledge of the canopy density at different growth stages of the soybean, to generate a growth potential difference partition map.

8. The system for early detection of soybean diseases based on CNN according to claim 6, wherein, The calculation module comprises: The first calculation unit is configured to perform spatial continuity reconstruction according to the growth potential difference partition map, and convert the discrete partition boundary into a continuously changing growth potential intensity distribution by establishing a spatial autocorrelation model of the growth potential value, to obtain a continuous growth potential surface; The second calculation unit is configured to perform point growth potential intensity extraction according to the continuous growth potential surface and spatial coordinate information, and obtain the growth potential intensity value corresponding to each point by mapping the spatial coordinates of each point to the continuous growth potential surface, to obtain a point basic growth potential intensity; The third calculation unit is configured to perform disease sensitivity weight optimization according to the point basic growth potential intensity, and establish a mapping relationship between the growth potential intensity and the disease occurrence probability by analyzing the response characteristics of the different growth potential regions of the soybean canopy to early diseases, to obtain a point growth potential weight.

9. The system for early detection of soybean diseases based on CNN according to claim 6, wherein, The extraction module comprises: The first extraction unit is configured to perform compound leaf structure analysis according to the local image of the soybean leaf, and obtain preliminary compound leaf structure features by constructing a feature extraction network with three-way symmetry and performing feature coding on three leaflets as a whole under geometric constraints; The second extraction unit is configured to perform collaborative response analysis according to the preliminary compound leaf structure features, and capture the synchronous change law of the three leaflets in physiological response by establishing a feature transfer model between the leaflets, to obtain a compound leaf collaborative feature vector; The third extraction unit is configured to perform feature fusion according to the compound leaf collaborative feature vector and the point growth potential weight, and perform adaptive weighted integration on the collaborative features of the three leaflets by taking the growth potential weight as a feature importance coefficient, to obtain a leaf fusion feature.

10. The system for early detection of soybean diseases based on CNN according to claim 6, wherein, The analysis module comprises: The first analysis unit is configured to perform vertical direction stress gradient analysis according to the leaf fusion feature, and calculate the attenuation rate of the feature values between different leaf positions by constructing a feature response curve from the lower leaves to the upper leaves of the plant, to obtain a vertical stress gradient vector; The second analysis unit is configured to perform compound leaf internal collaborative change analysis according to the vertical stress gradient vector, and quantify the asymmetric change mode between the leaflets in the trilobed compound leaf structure by establishing a feature response correlation matrix of the three leaflets on the same petiole, to obtain a compound leaf collaborative variation index; The third analysis unit is configured to perform infection evaluation according to the vertical stress gradient vector and the compound leaf collaborative variation index, and construct a comprehensive evaluation index representing the degree of systematic spread of the disease by fusing the stress propagation characteristics in the vertical direction of the plant and the abnormal distribution pattern in the horizontal direction of the compound leaf, to obtain an infection stress index.

Images