Grape aspergillosis early warning system and method based on data driving

Abstract

The invention discloses a grape aspergillosis early warning system and method based on data driving, and relates to the technical field of agricultural pest and disease monitoring. A data acquisition module acquires grape aspergillosis early data; the preprocessing module is used for preprocessing the environment data and the spore quantitative data; a coupling module extracts coupling characteristics of the preprocessed environment data and spore quantitative data; the feature association module constructs an association model based on a resonance vector, and an association result of the environmental data and the spore quantitative data is obtained through the association model; and the prediction module is used for inputting a correlation result of the environmental data and the spore quantification data into an airborne disease prediction model to obtain a grape aspergillosis prediction result. Through multi-source data acquisition and fusion, multi-model collaborative modeling and precise analysis of internal association between environmental factors and grape aspergillosis, early precise early warning of diseases is realized, technical support is provided for precise prevention and control of grape aspergillosis, the pesticide usage amount is reduced, and the yield and quality of grapes are improved.

Description

Technical Field

[0001] This invention relates to the field of agricultural pest and disease monitoring technology, and more specifically to a data-driven early warning system and method for grape aspergillosis. Background Technology

[0002] Currently, grape aspergillosis, a typical airborne fungal disease, occurs in major grape-producing regions worldwide. Its pathogenic spores spread through the air, infecting grape fruits, leaves, and branches, leading to fruit rot, reduced yield, and in severe cases, economic losses exceeding 30%. It also produces toxic metabolites that affect fruit quality and food safety. Current control measures for grape aspergillosis mainly rely on manual field inspections for identification and chemical control. Disease early warning systems often rely on growers' past experience or simple meteorological indicators, which has several drawbacks: first, the warnings are often delayed, as manual inspections can only detect the disease after symptoms appear, by which time the pathogen has already completed initial spread, missing the optimal control opportunity; second, the accuracy is low, as traditional methods ignore the combined effects of key environmental factors such as wind force, wind direction, and rainfall on spore germination and spread, and do not consider the spatial distribution differences of spores in the air, resulting in a high rate of false alarms; and third, data utilization is insufficient, failing to accurately depict the dynamic process of disease occurrence and development.

[0003] Therefore, given the shortcomings of existing technologies such as delayed early warning, low accuracy, and insufficient data fusion for grape aspergillosis, how to provide a data-driven early warning system and method for grape aspergillosis, to achieve accurate prediction before the disease occurs, and to provide scientific guidance for green prevention and control, is an urgent problem to be solved by those skilled in the art. Summary of the Invention

[0004] In view of this, the present invention provides a data-driven early warning system and method for grape aspergillosis. Through multi-source data collection and fusion and multi-model collaborative modeling, the system accurately analyzes the intrinsic relationship between environmental factors and the occurrence of grape aspergillosis, realizes early and accurate early warning of the disease, provides technical support for the precise control of grape aspergillosis, reduces pesticide use, and improves grape yield and quality.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: a data-driven early warning system for grape aspergillosis, comprising: The data acquisition module is used to collect environmental data and spore quantification data to obtain early data on grape aspergillosis. The preprocessing module is used to preprocess the environmental data and spore quantification data; The coupling module is used to extract the coupling features of preprocessed environmental data and spore quantization data through Hilbert transform, and generate resonance vectors. The feature association module is used to construct an association model based on resonance vectors. The association model includes a vineyard spore spatial distribution model based on grid association technology and a spore diffusion model based on chaotic dynamics analysis and phase transition prediction theory. The association model is used to obtain the association results between the environmental data and the spore quantification data. The prediction module is used to establish an airborne disease prediction model based on convolutional neural network and long short-term memory network algorithm. The correlation results of the environmental data and the spore quantification data are input into the airborne disease prediction model to obtain the prediction result of grape aspergillosis.

[0006] Preferably, the spore quantification data is preprocessed, including: Adaptive cropping is performed on the spore quantization data to obtain spore candidate regions, and the spore candidate regions are corrected and optimized to obtain corrected spore quantization images; The pixels in the spore quantization image are labeled to form spore recognition blocks; Calculate the area of ​​each spore identification block, calculate the number of spores in the spore quantization image, and generate a spore quantization feature vector.

[0007] Preferably, the coupling features of the preprocessed environmental data and spore quantization data are extracted using Hilbert transform to generate a resonance vector, including: The preprocessed environmental data and spore quantization data are mapped to the complex domain to generate an analytical signal; Perform Hilbert-Huang transform on the analytic signal to decompose it into multiple intrinsic mode functions, and extract the instantaneous phase of each intrinsic mode function; Phase slip rate is calculated based on instantaneous phase, and the correlation degree of information is calculated through the coefficient of variation, standard deviation and mean of slip rate, so as to obtain the key coupling characteristics of environmental data and spore quantification data; Amplifying key coupling features based on resonant frequency modes generates mutual information resonant vectors.

[0008] Preferred spatial distribution models of vineyard spores based on grid association technology include: The system consists of an input layer, a multi-crop feature decoupling layer, a data import layer, a computation layer, and a post-processing optimization layer. Among them, the vineyard is divided into grid units, and the latitude and longitude of each grid unit, planting density, altitude, and soil parameters are fused into a spatial feature vector. The multi-crop feature decoupling layer incorporates a channel weight filter; the spatial feature vector is purified through the channel weight filter. The data import layer is used to import crop distribution databases, extract information on grape varieties and planting areas, and generate spatial constraint rules; The purified spatial feature vector is input into the computation layer, and the spore concentration distribution output by the computation layer is post-processed and optimized in combination with the spatial constraint rules to generate the spatial features of spore diffusion.

[0009] Preferred spore diffusion models based on chaotic dynamics analysis and phase transition prediction theory include: Using spore growth rate, wind force, wind direction, and relative humidity as input parameters, the data is mapped to the Lorenz attractor phase space, the phase space trajectory equation is solved, and dynamic trajectory data of spores is generated. Calculate the trajectory deviation, and set a dynamic threshold based on the trajectory deviation to filter out singular trajectories; Based on the renormalization group flow theory, the fixed point of phase space trajectory is traced to determine the critical point of phase transition for spore diffusion. By integrating wind direction vector field and topographic data, the temporal characteristics of spore dispersal are output.

[0010] Preferably, an airborne disease prediction model based on convolutional neural network and long short-term memory network algorithms is established. The correlation results of the environmental data and the spore quantification data are input into the airborne disease prediction model to obtain the prediction results for grape aspergillosis, including: Receive the spatial features, temporal features, and coupling features; The spatial features, temporal features, and coupling features are processed through a gating mechanism to control the flow of spatiotemporal information; A spatial attention weight graph is generated using the spatial attention submodule; The expected probability of disease onset, expected onset time, and expected severity of disease are output through the fully connected layer.

[0011] Preferred data-driven early warning methods for grape aspergillosis include: Environmental data and spore quantification data were collected to obtain early data on grape aspergillosis. The environmental data and spore quantification data are preprocessed; The coupling features of preprocessed environmental data and spore quantization data are extracted using Hilbert transform to generate resonance vectors. A correlation model based on resonance vectors is constructed, which includes a vineyard spore spatial distribution model based on grid correlation technology and a spore diffusion model based on chaotic dynamics analysis and phase transition prediction theory; the correlation results of the environmental data and the spore quantification data are obtained through the correlation model; An airborne disease prediction model based on convolutional neural network and long short-term memory network algorithm is established. The correlation results of the environmental data and the spore quantification data are input into the airborne disease prediction model to obtain the prediction results of grape aspergillosis.

[0012] As can be seen from the above technical solutions, compared with the prior art, the present invention discloses a data-driven early warning system and method for grape aspergillosis, which has the following beneficial effects: (1) It achieves early and accurate early warning by capturing early signals of spore germination and spread, and combining them with future weather forecasts to predict the risk of disease occurrence, thus solving the problem of delayed early warning in traditional methods; (2) It improves prediction accuracy by fully exploring key features in the spatial and temporal dimensions through multi-source data fusion and CNN-LSTM model collaboration, effectively reducing the misjudgment rate of single factor prediction; (3) It guides growers to apply pesticides as needed through graded early warning and precise prevention and control recommendations, reducing the abuse of chemical pesticides and ensuring the safety of grape yield and quality. Attached Figure Description

[0013] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0014] Figure 1 This is a schematic diagram of the structure of the data-driven early warning system for grape aspergillosis provided by the present invention.

[0015] Figure 2 This is a schematic diagram of the data-driven early warning method for grape aspergillosis provided by the present invention. Detailed Implementation

[0016] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0017] This invention discloses a data-driven early warning system for grape aspergillosis, such as... Figure 1 As shown, it includes: The data acquisition module is used to collect environmental data and spore quantification data to obtain early data on grape aspergillosis. The preprocessing module is used to preprocess the environmental data and spore quantification data; The coupling module is used to extract the coupling features of preprocessed environmental data and spore quantization data through Hilbert transform, and generate resonance vectors. The feature association module is used to construct an association model based on resonance vectors. The association model includes a vineyard spore spatial distribution model based on grid association technology and a spore diffusion model based on chaotic dynamics analysis and phase transition prediction theory. The association model is used to obtain the association results between the environmental data and the spore quantification data. The prediction module is used to establish an airborne disease prediction model based on convolutional neural network and long short-term memory network algorithm. The correlation results of the environmental data and the spore quantification data are input into the airborne disease prediction model to obtain the prediction result of grape aspergillosis.

[0018] Specifically, environmental data acquisition employs a dual-source acquisition model combining local grid monitoring and meteorological services. The vineyard is divided into 50m x 50m grid monitoring units, each deploying a monitoring node integrating multi-parameter sensors to simultaneously collect environmental parameters such as temperature, relative humidity, wind speed, wind direction, and rainfall, ensuring the capture of dynamic changes in the microenvironment. Simultaneously, it connects to the meteorological service center to obtain hourly weather forecast data for the next 72 hours (including temperature, humidity, precipitation probability, wind speed and direction, and sunshine duration).

[0019] Spore quantification data acquisition: An airborne spore capture device is deployed at the center of each grid cell to capture real-time images of airborne Aspergillus spores. The device has a built-in laser particle counter that simultaneously records the number of spores per unit volume and the spore growth rate. ¹) Equalized data are used to generate a time-series spore monitoring dataset. For scenarios with low spore concentrations, infrared-assisted illumination technology is used to enhance image contrast, ensuring effective capture of low-concentration spores.

[0020] It also includes: integrating geographic information systems to obtain spatial attribute data of vineyards such as latitude and longitude, altitude, planting density, variety type, and soil parameters (pH value, organic matter content, and water content); and obtaining data on the onset time, severity, control measures, and effects of grape aspergillosis in the past 5 years to provide basic samples for model training.

[0021] Specifically, the spore quantification data undergoes preprocessing, including: Adaptive cropping is performed on the spore quantization data to obtain spore candidate regions, and the spore candidate regions are corrected and optimized to obtain corrected spore quantization images; Specifically, a spatial attention mechanism and dynamic cropping technique are employed to construct a three-layer cascaded depthwise separable convolutional network with kernel sizes of 3×3, 5×5, and 3×3, and output channels of 16, 32, and 1, respectively. Each convolutional layer is followed by a ReLU activation function and a max pooling operation with a 2×2 pooling kernel to generate a spore region saliency heatmap. Based on the spore region saliency heatmap, a minimum bounding box is determined, with thresholds set at a heatmap pixel saliency ≥ 0.7 and a region area ≥ 10% of the total image area. Adaptive cropping is then performed on the original spore quantization data image to remove background interference such as soil, leaf impurities, and equipment shadows, yielding candidate spore regions.

[0022] Specifically, the candidate spore regions are corrected and optimized to obtain corrected spore quantization images. This includes: using a piecewise linear transformation algorithm to eliminate the influence of overexposure or underexposure of field light on the spore images. The specific steps are as follows: First, calculate the brightness histogram of the cropped spore image, using the following formula: ; ; Adaptively determine the low brightness threshold I1 and the high brightness threshold I2, where, For brightness level index, For brightness The number of pixels, This represents the maximum brightness value of the image. This is a round-down operation. For the rounding up operation; the image brightness range is divided into three intervals: [0, I1], (I1, I2], and (I2, L]. The pixel brightness is adjusted using a piecewise linear mapping formula, resulting in the following mapping formula: ; The above adjustments preserve spore morphology details and improve image quality.

[0023] The pixels in the spore quantization image are labeled to form spore recognition blocks; Calculate the area of ​​each spore identification block, calculate the number of spores in the spore quantization image, and generate a spore quantization feature vector.

[0024] Specifically, the `bwlabel` function (8-neighborhood calculation) in MATLAB image processing is used to label adjacent pixels in the image, forming spore identification blocks; the `regionprops` function is used to calculate the area of ​​each identification block, and interference blocks with an area less than 50 pixels are deleted; overlapping spores are separated using the watershed segmentation method, as shown in the formula below. ; ; ; ; Finally, the pure spore image is extracted using the imsubtract function, and the number of spores per unit volume is calculated by dividing the total area by the average spore area, thus generating a spore quantification feature vector.

[0025] Consistency verification was performed on the quantified spore data, and interpolation was used to supplement missing data. The verified spore data (including spore quantity and spore growth rate) was mapped to the [0,1] interval to complete the standardization process, ensuring that the data format is consistent with the environmental data, and providing adaptive data for subsequent coupled feature extraction.

[0026] Specifically, the coupling features between preprocessed environmental data and spore quantization data are extracted using Hilbert transform to generate resonance vectors, including: The preprocessed environmental data and spore quantization data are mapped to the complex domain to generate an analytical signal; Specifically, the environmental data includes temperature, relative humidity, wind force, wind direction, and rainfall; the spore quantification data includes spore quantity and spore growth rate. The formula for calculating an analytic signal is: ; in, This is the result of the Hilbert transform.

[0027] Perform a Hilbert-Huang transform on the analytic signal to decompose it into multiple intrinsic mode functions (IMFs), and extract the instantaneous phase of each IMF. ; Phase slip rate is calculated based on instantaneous phase, and the correlation degree of information is calculated through the coefficient of variation, standard deviation and mean of slip rate, so as to obtain the key coupling characteristics of environmental data and spore quantification data; Amplifying key coupling features based on resonant frequency modes generates mutual information resonant vectors.

[0028] Specifically, the Savitzky-Golay filter is used to calculate the phase change rate and phase acceleration. The Savitzky-Golay filter has a window length of 11 and a polynomial order of 3. The formula for calculating the phase change rate is: ; The formula for calculating phase acceleration is: ; Phase slip events are identified by setting a threshold of 3 times the local standard deviation, and the slip period is statistically calculated based on the phase slip events. And calculate the slip rate .

[0029] Specifically, the formula for calculating information relevance is as follows: ; in, The coefficient of variation of the slip rate. The standard deviation of the slip rate, The mean of the slip velocity is given, and the information correlation degree is the information correlation degree between the environment and the spore data.

[0030] Based on the resonance frequency mode amplification of key coupling features, a 256-dimensional mutual information resonance vector is generated. The first 128 dimensions encode time-domain anomaly information, and the last 128 dimensions encode frequency-domain resonance features.

[0031] Specifically, the spatial distribution model of vineyard spores based on grid association technology includes: The system consists of an input layer, a multi-crop feature decoupling layer, a data import layer, a computation layer, and a post-processing optimization layer. The process involves dividing the vineyard into grid units, with each grid unit consisting of 50m × 50m grids. The latitude and longitude of each grid unit are embedded into a vector, and the planting density, altitude, and soil parameters are integrated into a spatial feature vector. The soil parameters include the soil pH value, organic matter content, and water content. A multi-crop feature decoupling layer is inserted before the fully connected layer of the feature extraction network. The multi-crop feature decoupling layer has a built-in geographical location-related channel weight filter to suppress interfering feature responses from non-grape-growing areas. The spatial feature vector is purified by the channel weight filter. In actual vineyards, there are non-grape-growing areas such as surrounding crops, roads, and open spaces. Soil parameters in these areas, such as those around roads, and the planting density of other crops, can interfere with the accuracy of spore distribution prediction. The role of the multi-crop feature decoupling layer is to screen out effective features related to grape cultivation during the feature extraction stage, preventing interfering features from entering the model calculation and improving modeling accuracy from the source.

[0032] The data import layer is used to import crop distribution databases, extract information on grape varieties and planting areas, and generate spatial constraint rules; The model calculation may produce unreasonable results such as predicting high concentrations of spores in non-grape-growing areas due to blurred features in the edge region. The spatial constraint rules, by anchoring to the actual planting boundary, provide a physical spatial verification standard for the subsequent model output, ensuring that the prediction results are only valid within the actual grape-growing area and avoiding erroneous outputs that are out of touch with reality.

[0033] The purified spatial feature vector is input into the computational layer, and the spore concentration distribution output by the computational layer is post-processed and optimized in conjunction with the spatial constraint rules to generate spatial features of spore diffusion. This ensures a high degree of consistency between the output results and the spatial matching of the actual planting area, ultimately generating a heatmap of the spatial distribution of spore concentration in the vineyard.

[0034] The computational layer employs a combination of lightweight convolutional layers and fully connected layers, adapting to both local correlation capture and global computation of spatial features: The convolutional layer uses a single 3×3 convolutional kernel with 32 output channels to capture the spore diffusion effect of adjacent grid cells; The fully connected layer uses a 2-layer fully connected network with the number of hidden layer neurons increasing from 64 to 32. It fuses the local features output by the convolutional layer with the global spatial features and maps them to the spore concentration prediction value of a single grid cell. By introducing an interaction term of adjacent grid features, the concentration prediction value is dynamically adjusted by calculating the feature similarity between the current grid and the eight surrounding grids, including soil parameter matching degree and planting density difference, to simulate the spread and accumulation effect of spores in adjacent areas.

[0035] In the computation layer, the input is the purified spatial feature vector, ensuring that the computation is based on valid information; By using spatial constraint rules to eliminate unreasonable predictions from non-growing areas, and calibrating the concentration distribution deviation within the growing areas, the output results are ensured to match the spatial accuracy of the actual growing areas. Only grape growing areas have spore concentration data, and there is no invalid data in non-growing areas.

[0036] Specifically, spore diffusion models based on chaotic dynamics analysis and phase transition prediction theory include: Using spore growth rate, wind force, wind direction, and relative humidity as input parameters, the data are mapped to the Lorenz attractor phase space. The phase space trajectory equation is solved using the fourth-order Runge-Kutta method to generate dynamic trajectory data of spores. Calculate the trajectory deviation, set a dynamic threshold based on the trajectory deviation to filter out singular trajectories; identify the critical environmental conditions for rapid spore diffusion; The formula for calculating trajectory deviation is as follows: ; in, For real-time trajectory points, Center of the trajectory; Based on the renormalization group flow theory, the fixed point of the phase space trajectory is traced, the bifurcation value of the spore transitioning from a stable state to an unstable state is identified, and the critical point of phase transition for spore diffusion is determined. By integrating wind direction vector field and topographic data, the temporal characteristics of spore dispersal are output, including the spore dispersal range, concentration change curve and spatial distribution dynamic heat map for the next 24-72 hours.

[0037] Specifically, an airborne disease prediction model based on convolutional neural networks and long short-term memory network algorithms is established. The correlation results between the environmental data and the spore quantification data are input into the airborne disease prediction model to obtain the prediction results for grape aspergillosis, including: The spatial features, temporal features, and coupling features are received; specifically: a 64×64×16 spatial feature map - spore concentration distribution of grid cells; a 128×8 temporal feature sequence - environmental and spore data for nearly 128 time steps; and a 256-dimensional coupling feature vector - mutual information resonance vector. The spatial features, temporal features, and coupling features are processed through a gating mechanism to control the flow of spatiotemporal information; The spatial attention submodule generates a spatial attention weight map to enhance the characteristic response of high-concentration spore regions and key environmental parameters. It contains two layers of ConvLSTM units, each with 64 neurons. Each unit has a built-in forget gate, input gate, and output gate.

[0038] The expected probability of disease onset, expected onset time, and expected severity of disease in grapes are mapped to a three-dimensional output through a fully connected layer.

[0039] Specifically, the training of the airborne disease prediction model includes: Integrating historical data from the past 5 years, the dataset was divided into training, validation, and test sets in a 7:2:1 ratio. Data augmentation techniques, including time-axis stretching / compression, spatial translation, and Gaussian noise addition, were used to expand the sample size by 3 times. Employing a multi-task loss function ,in, The cross-entropy loss is the probability of disease onset. This represents the mean squared error loss due to the onset time. F1 loss is the percentage of disease severity. The training set was used to train the airborne disease prediction model, employing the Adam optimization algorithm (initial learning rate 0.001). , Combined with a cosine annealing strategy (the learning rate decays to 0.5 times the current value every 50 rounds), the batch size is set to 256, the total number of training rounds is 200, and an early stopping mechanism is triggered. Set up an incremental training interface to receive new spore feature description files, generate prototype vector patches, and implement minute-level parameter iteration through a differential update protocol to ensure that the model adapts to spore mutations.

[0040] After each training round, the loss value is calculated using the validation set. If the validation set loss does not decrease for 10 consecutive rounds, an early stopping mechanism is triggered to avoid model overfitting.

[0041] After training, the final performance of the model is evaluated using a test set without any parameter tuning. The test set is used to verify the model's generalization effect on unseen data.

[0042] In one specific embodiment of the present invention, a data-driven early warning method for grape aspergillosis is provided, such as... Figure 2 As shown, it includes: Environmental data and spore quantification data were collected to obtain early data on grape aspergillosis. The environmental data and spore quantification data are preprocessed; The coupling features of preprocessed environmental data and spore quantization data are extracted using Hilbert transform to generate resonance vectors. A correlation model based on resonance vectors is constructed, which includes a vineyard spore spatial distribution model based on grid correlation technology and a spore diffusion model based on chaotic dynamics analysis and phase transition prediction theory; the correlation results of the environmental data and the spore quantification data are obtained through the correlation model; An airborne disease prediction model based on convolutional neural network and long short-term memory network algorithm is established. The correlation results of the environmental data and the spore quantification data are input into the airborne disease prediction model to obtain the prediction results of grape aspergillosis.

[0043] This invention analyzes the impact of environmental information on the germination and spread of Aspergillus cereus spores, establishes a spatial distribution and diffusion model for the germination and spread of Aspergillus cereus spores, and analyzes the coupling relationship between environmental data such as temperature, humidity, wind force, wind direction, and rainfall and the number of Aspergillus cereus spores, spore growth rate, and spatial distribution characteristics of spores in the air. Taking into account spore number, spore growth rate, spatial distribution characteristics of spores in the air, and environmental data, and using parameters such as meteorological forecast data, quantitative information on different types of spores of airborne diseases, and historical environmental data as inputs, and the expected incidence probability, expected onset time, and expected severity of Aspergillus cereus as outputs, an airborne disease prediction model based on convolutional neural networks and long short-term memory network algorithms is established to achieve early and accurate prediction and forecasting of Aspergillus cereus.

[0044] In one specific embodiment of the present invention, a wine grape plantation, planted with Cabernet Sauvignon, covers a total area of ​​100 mu (approximately 66,667 square meters), with a planting density of 300 vines per mu, and the soil type is sandy loam (pH 6.5-7.2, organic matter content 1.2%-1.5%, and water content 18%-22%).

[0045] I. System Deployment and Implementation

[0046] (a) Hardware deployment

[0047] Grid division and equipment layout: The park is divided into 50m×50m grids, forming a total of 13×13=169 grid monitoring units, covering the entire 100-acre park. Edge units are adjusted according to the actual boundaries. Each grid unit is equipped with one multi-parameter sensor, and one airborne spore capture device is deployed at the center, for a total of 169 sensors and 169 spore capture devices.

[0048] Three edge computing nodes are deployed in the center of the park, each covering 50-60 grid cells, responsible for local data preprocessing and transmission scheduling; a distributed computing cluster is deployed in the cloud, with the Hadoop distributed storage framework and TensorFlow deep learning framework installed, and a Redis hot data cache and HDFS cold data storage system is built.

[0049] Mobile terminals were provided to park management personnel, and a web-based monitoring screen was deployed in the duty room at the park entrance to enable multi-channel synchronization of early warning information.

[0050] (II) Data Acquisition Configuration

[0051] Environmental data acquisition: Sensors collect temperature, relative humidity, wind speed, wind direction, and rainfall; connect with the meteorological service center to obtain hourly weather forecasts for the next 72 hours (including temperature, humidity, probability of precipitation, wind speed and direction, and sunshine duration), forming an environmental dataset.

[0052] Spore quantification data acquisition: The spore capture device acquires spore images, and the built-in laser particle counter synchronously records the number of spores per unit volume and the spore growth rate.

[0053] Spatial attributes such as latitude, longitude, and altitude of the park are obtained through GIS; historical data such as disease records and prevention and control measures in the park are retrieved as basic samples for model training.

[0054] II. Technical Implementation Process

[0055] (a) Data preprocessing execution

[0056] Spore image preprocessing

[0057] Background interference removal: A spore region saliency heatmap is generated using a 3-layer depth separable convolutional network (3×3→5×5→3×3 convolutional kernels, 16→32→1 output channels). The heatmap is then cropped according to a saliency ≥0.7 and a region area ≥10% threshold to remove background elements such as soil particles and leaf fragments.

[0058] Illumination correction: Calculate the brightness histogram of the cropped image and adaptively determine the low brightness threshold. =35, High brightness threshold =200 (L=255), the pixel brightness is adjusted by piecewise linear mapping formula, the uniformity of image brightness is improved after correction, and the clarity of spore edge details is improved.

[0059] Spore quantization calculation: Using the MATLAB image processing workflow, the spore identification blocks are marked by the bwlabel function (8 neighborhoods), interference blocks <50 pixels are deleted, overlapping spores are separated by the watershed segmentation method, and finally the number of spores per unit volume is calculated to generate a 128-dimensional spore quantization feature vector.

[0060] (ii) Coupling Feature Extraction

[0061] Five types of environmental time-series data—temperature, relative humidity, wind force, wind direction, and rainfall—are selected, along with two types of spore time-series data—spore quantity and spore growth rate—and mapped to the complex domain to construct analytical signals.

[0062] Perform Hilbert-Huang transform on the analytic signal to decompose it into 8 intrinsic mode functions (IMFs), and extract the instantaneous phase of each IMF.

[0063] The phase change rate ω(t) and phase acceleration α(t) are calculated using a filter. Phase slip events are identified with three times the local standard deviation. The average slip period T and slip rate v are statistically calculated.

[0064] The information correlation degree is calculated, and the key coupling features are amplified based on the resonant frequency mode to generate a 256-dimensional mutual information resonance vector. The first 128 dimensions encode time-domain anomalies, and the last 128 dimensions encode frequency-domain resonances.

[0065] (III) Construction and Operation of the Association Model

[0066] Spatial distribution model: The 64-dimensional latitude and longitude embedding vector, 1-dimensional planting density, 1-dimensional altitude, and 3-dimensional soil parameters of each grid cell are fused into a 69-dimensional spatial feature vector; the interference features of non-grape planting areas (3 grid cells at the edge of the park) are filtered through a multi-crop feature decoupling layer; the park crop distribution database is imported to generate spatial constraint rules, which are then processed by the computation layer to output a heat map of the spatial distribution of spore concentration in the park.

[0067] The diffusion model uses spore growth rate, wind speed, wind direction, and relative humidity as inputs, maps them to the Lorenz attractor phase space, and generates trajectory data using the fourth-order Runge-Kutta method. The trajectory deviation is calculated to identify critical conditions (humidity ≥ 80% and wind speed ≥ 2.5 m / s). The phase transition critical point is determined based on the renormalization stream theory, and the spore diffusion range in the next 72 hours is predicted in combination with topographic data.

[0068] (iv) Training and prediction of airborne disease prediction models

[0069] Model training: Integrate historical data from the past 5 years, and divide it into a training set (1277 records), a validation set (365 records), and a test set (183 records) in a 7:2:1 ratio; expand the sample to 5475 records by using time-axis stretching, spatial translation, and Gaussian noise (σ=0.01); use the multi-task loss function as the optimization objective, and train for 200 rounds using the Adam algorithm (initial learning rate 0.001) and cosine annealing strategy. The validation set loss did not decrease for 10 consecutive rounds after 120 rounds (final loss = 0.12), triggering the early stopping mechanism.

[0070] Model prediction: Input the spatial feature map (64×64×16), the temporal feature sequence (128×8), and the 256-dimensional coupling vector into the Conv-LSTM model, and output the prediction results.

[0071] III. Early Warning Implementation

[0072] (I) Tiered early warning push

[0073] Warning information will be pushed out through three channels: Mobile app for management personnel: pop-up windows and voice reminders, along with heat maps of affected areas and animations of spread paths; The large screen in the duty room visually displays high-risk areas and precise application ranges; SMS notification: Send emergency prevention notices to farmers.

[0074] This invention, through multi-source data collection and fusion and multi-model collaborative modeling, accurately analyzes the intrinsic relationship between environmental factors and the occurrence of grape aspergillosis, enabling early and accurate disease warning, providing technical support for the precise control of grape aspergillosis, reducing pesticide use, and improving grape yield and quality.

[0075] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to the method section.

[0076] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A data-driven early warning system for grape aspergillosis, characterized in that, include: The data acquisition module is used to collect environmental data and spore quantification data to obtain early data on grape aspergillosis. The preprocessing module is used to preprocess the environmental data and spore quantification data; The coupling module is used to extract the coupling features of preprocessed environmental data and spore quantization data through Hilbert transform, and generate resonance vectors. The feature association module is used to construct an association model based on resonance vectors. The association model includes a vineyard spore spatial distribution model based on grid association technology and a spore diffusion model based on chaotic dynamics analysis and phase transition prediction theory. The association model is used to obtain the association results between the environmental data and the spore quantification data. The prediction module is used to establish an airborne disease prediction model based on convolutional neural network and long short-term memory network algorithm. The correlation results of the environmental data and the spore quantification data are input into the airborne disease prediction model to obtain the prediction result of grape aspergillosis.

2. The data-driven early warning system for grape aspergillosis according to claim 1, characterized in that, Preprocessing of spore quantification data includes: Adaptive cropping is performed on the spore quantization data to obtain spore candidate regions, and the spore candidate regions are corrected and optimized to obtain corrected spore quantization images; The pixels in the spore quantization image are labeled to form spore recognition blocks; Calculate the area of ​​each spore identification block, calculate the number of spores in the spore quantization image, and generate a spore quantization feature vector.

3. The data-driven early warning system for grape aspergillosis according to claim 1, characterized in that, The coupling features between the preprocessed environmental data and spore quantization data are extracted using Hilbert transform to generate resonance vectors, including: The preprocessed environmental data and spore quantization data are mapped to the complex domain to generate an analytical signal; Perform Hilbert-Huang transform on the analytic signal to decompose it into multiple intrinsic mode functions, and extract the instantaneous phase of each intrinsic mode function; Phase slip rate is calculated based on instantaneous phase, and the correlation between information is calculated by the coefficient of variation, standard deviation and mean of slip rate, so as to obtain the key coupling characteristics of environmental data and spore quantification data; Amplifying key coupling features based on resonant frequency modes generates mutual information resonant vectors.

4. The data-driven early warning system for grape aspergillosis according to claim 1, characterized in that, A spatial distribution model of vineyard spores based on grid association technology includes: The system consists of an input layer, a multi-crop feature decoupling layer, a data import layer, a computation layer, and a post-processing optimization layer. Among them, the vineyard is divided into grid units, and the latitude and longitude of each grid unit, planting density, altitude, and soil parameters are fused into a spatial feature vector. The multi-crop feature decoupling layer incorporates a channel weight filter; the spatial feature vector is purified through the channel weight filter. The data import layer is used to import crop distribution databases, extract information on grape varieties and planting areas, and generate spatial constraint rules; The purified spatial feature vector is input into the computation layer, and the spore concentration distribution output by the computation layer is post-processed and optimized in combination with the spatial constraint rules to generate the spatial features of spore diffusion.

5. The data-driven early warning system for grape aspergillosis according to claim 4, characterized in that, Spore diffusion models based on chaotic dynamics analysis and phase transition prediction theory include: Using spore growth rate, wind force, wind direction, and relative humidity as input parameters, the data is mapped to the Lorenz attractor phase space, the phase space trajectory equation is solved, and dynamic trajectory data of spores is generated. Calculate the trajectory deviation, and set a dynamic threshold based on the trajectory deviation to filter out singular trajectories; Based on the renormalization group flow theory, the fixed point of phase space trajectory is traced to determine the critical point of phase transition for spore diffusion. By integrating wind direction vector field and topographic data, the temporal characteristics of spore dispersal are output.

6. The data-driven early warning system for grape aspergillosis according to claim 5, characterized in that, An airborne disease prediction model based on convolutional neural network and long short-term memory network algorithms is established. The correlation results of the environmental data and the spore quantification data are input into the airborne disease prediction model to obtain the prediction results for grape aspergillosis, including: Receive the spatial features, temporal features, and coupling features; The spatial features, temporal features, and coupling features are processed through a gating mechanism to control the flow of spatiotemporal information; A spatial attention weight graph is generated using the spatial attention submodule; The expected probability of disease onset, expected onset time, and expected severity of disease are output through the fully connected layer.

7. A data-driven early warning method for grape aspergillosis, applied to the data-driven early warning system for grape aspergillosis as described in any one of claims 1-6, characterized in that, include: Environmental data and spore quantification data were collected to obtain early data on grape aspergillosis. The environmental data and spore quantification data are preprocessed; The coupling features of preprocessed environmental data and spore quantization data are extracted using Hilbert transform to generate resonance vectors. A correlation model based on resonance vectors is constructed, which includes a vineyard spore spatial distribution model based on grid correlation technology and a spore diffusion model based on chaotic dynamics analysis and phase transition prediction theory; the correlation results of the environmental data and the spore quantification data are obtained through the correlation model; An airborne disease prediction model based on convolutional neural network and long short-term memory network algorithm is established. The correlation results of the environmental data and the spore quantification data are input into the airborne disease prediction model to obtain the prediction results of grape aspergillosis.

Images