High-standard farmland spore monitoring and disease prediction system based on AI identification

Abstract

The invention discloses a high-standard farmland spore monitoring and disease prediction system based on AI identification, and relates to the technical field of farmland disease prediction, and the system comprises a data acquisition module which is used for obtaining farmland multi-source data; the AI identification module is used for identifying the spore image data by adopting an AI spore identification model; the feature extraction and fusion module is used for performing feature extraction and feature fusion on the spore type data, the spore quantity data, the farmland meteorological data, the farmland soil data, the crop growth data and the historical disease data; and the disease prediction module is used for performing disease prediction according to the multi-source fusion feature vector by adopting a disease prediction model to obtain a disease prediction result. According to the method, the reliability and suitability of high-standard farmland spore monitoring and disease prediction are cooperatively improved by acquiring the multi-source data of the spore image, the farmland weather, the farmland soil, the crop growth and the historical diseases and combining the AI spore recognition model, the feature extraction fusion technology and the disease prediction model.

Description

Technical Field

[0001] This application relates to the field of farmland disease prediction technology, and in particular to a high-standard farmland spore monitoring and disease prediction system based on AI recognition. Background Technology

[0002] The contiguous planting of high-standard farmland significantly increases the risk of disease transmission. Spore transmission of fungal and bacterial diseases is a major contributing factor to outbreaks. For example, wheat powdery mildew spores can spread rapidly within 2-3 days under suitable temperature and humidity, leading to an increase in disease incidence of over 30%, and in severe cases, causing crop yield reductions of 15%-40%, posing a significant threat to food security. Therefore, research is needed on spore monitoring and disease prediction technologies for high-standard farmland.

[0003] Traditional spore monitoring and disease prediction in high-standard farmland mainly rely on manual operation. The specific process involves staff regularly (usually every 1-3 days) going into the fields to collect samples using a spore trapping device. These samples are then brought back to the laboratory for microscopic observation, counting, and identification of spore species. This is combined with past planting experience to determine the risk of disease occurrence. This method has significant limitations: Firstly, manual sampling has a low frequency and limited coverage, making it difficult to achieve continuous monitoring over a large area at all times. For high-standard farmland exceeding 1,000 acres, there are often multiple monitoring blind spots, making it impossible to capture spore dispersal dynamics in a timely manner. Secondly, manual identification and counting of spore species depend on the operator's professional level. The identification error between novice and experienced personnel can reach 20%-30%, and the counting efficiency is low, requiring more than 30 minutes to process a single sample, which is insufficient to meet the needs of rapid disease early warning.

[0004] In the prior art, Chinese patent CN120339853A discloses an agricultural pest and disease early warning method and system based on big data. First, it obtains a set of crop leaf images of the target farmland area from farmland image big data, and performs leaf region segmentation processing on the crop leaf image set to distinguish healthy areas from potentially diseased areas. Then, it extracts disease features from the images of potentially diseased areas to generate a set of key disease features. It uses a pre-trained pest and disease identification model to detect abnormal states of the key disease feature set, determines the pest and disease type, and predicts its spread trend. Finally, it generates a pest and disease early warning command containing geographic location information based on the pest and disease type identification and spread trend prediction data, and sends it to the farmland management system to trigger prevention and control response operations.

[0005] However, the aforementioned existing technologies do not take into account the relationship between spore dynamics, environmental factors and diseases, and the reliability and adaptability of spore monitoring and disease prediction in high-standard farmland need to be improved. Summary of the Invention

[0006] This application provides a high-standard farmland spore monitoring and disease prediction system based on AI recognition, which addresses the problem that existing farmland disease prediction technologies do not consider the correlation between spore dynamics, environmental factors and diseases, and that the reliability and adaptability of high-standard farmland spore monitoring and disease prediction need to be improved.

[0007] On the one hand, this application provides a high-standard farmland spore monitoring and disease prediction system based on AI recognition, including: a data acquisition module, an AI recognition module, a feature extraction and fusion module, and a disease prediction module.

[0008] The data acquisition module is configured to acquire spore image data, farmland meteorological data, farmland soil data, crop growth data, and historical disease data.

[0009] The AI ​​recognition module is configured to use an AI spore recognition model to identify the spore image data and obtain spore type data and spore quantity data.

[0010] The feature extraction and fusion module is configured to extract and fuse features from the spore species data, the spore quantity data, the farmland meteorological data, the farmland soil data, the crop growth data, and the historical disease data to obtain a multi-source fusion feature vector.

[0011] The disease prediction module is configured to use a disease prediction model to predict diseases based on the multi-source fusion feature vector, and obtain disease prediction results.

[0012] In one possible implementation, the data acquisition module is connected to a spore collection device, which is used to collect spores, generate spore image data, and send it to the data acquisition module.

[0013] The spore collection equipment is distributed in the farmland using a grid-plus-density deployment method.

[0014] In one possible implementation, the AI ​​recognition module is equipped with an image preprocessing unit.

[0015] The image preprocessing unit is used to perform noise reduction, enhancement, and segmentation processing on the spore image data before using it as input to the AI ​​spore recognition model.

[0016] In one possible implementation, the AI ​​spore recognition model is built on a convolutional neural network, and the structure of the AI ​​spore recognition model includes an input layer, a feature extraction layer, and a classification and counting layer.

[0017] The AI ​​spore recognition model incorporates ImageNet pre-trained weights and Dropout regularization.

[0018] In one possible implementation, the high-standard farmland spore monitoring and disease prediction system based on AI identification also includes a multi-source data preprocessing module.

[0019] The multi-source data preprocessing module is configured to perform outlier removal, missing value imputation, and data standardization on the spore type data, spore quantity data, farmland meteorological data, farmland soil data, crop growth data, and historical disease data, and then use the results as input to the feature extraction and fusion module.

[0020] In one possible implementation, the feature extraction includes spatial feature extraction based on graph convolutional networks and temporal feature extraction based on long short-term memory networks.

[0021] In one possible implementation, the feature fusion includes feature concatenation, attention weighting, and multi-source data supplementation fusion.

[0022] In one possible implementation, the disease prediction model is constructed based on a fully connected neural network, and the structure of the disease prediction model includes a hidden layer and an output layer.

[0023] In one possible implementation, the high-standard farmland spore monitoring and disease prediction system based on AI recognition also includes an early warning decision module.

[0024] The early warning decision module is configured to provide multi-level early warning and decision support based on the disease prediction results.

[0025] In one possible implementation, the AI-based high-standard farmland spore monitoring and disease prediction system also includes a result evaluation module.

[0026] The result evaluation module is configured to perform multi-dimensional evaluations of the performance of the data acquisition module, the AI ​​recognition module, the disease prediction module, and the early warning decision module, thereby quantifying and verifying the system performance.

[0027] The high-standard farmland spore monitoring and disease prediction system based on AI recognition disclosed in this application has the following advantages: By acquiring multi-source data such as spore images, farmland weather, farmland soil, crop growth, and historical diseases, and combining AI spore recognition models, feature extraction and fusion technologies, and disease prediction models, the reliability and adaptability of spore monitoring and disease prediction in high-standard farmland have been improved in a coordinated manner.

[0028] The proposed spore collection device is used to collect spores, generate spore image data and send it to the data acquisition module. The spore collection device is distributed in farmland using a grid + densification deployment method to achieve all-time, blind-spot-free dynamic monitoring of spores in high-standard farmland, breaking through the spatial limitations of traditional single-point monitoring.

[0029] The proposed AI spore recognition model is based on a convolutional neural network. The structure of the AI ​​spore recognition model includes an input layer, a feature extraction layer, and a classification and counting layer. The AI ​​spore recognition model introduces ImageNet pre-trained weights and Dropout regularization, which improves the accuracy and efficiency of spore recognition.

[0030] The proposed feature extraction methods include spatial feature extraction based on graph convolutional networks and temporal feature extraction based on long short-term memory networks. Feature fusion methods include feature splicing, attention weighting, and multi-source data supplementation and fusion, which fully explore the multi-dimensional relationship between spores, environment, and crops and improve data utilization.

[0031] The proposed disease prediction model is based on a fully connected neural network. The structure of the disease prediction model includes a hidden layer and an output layer, which improves the accuracy of disease prediction.

[0032] The proposed early warning and decision-making module provides multi-level early warning and decision support based on disease prediction results, realizing closed-loop management of prediction-early warning-decision. Attached Figure Description

[0033] To more clearly illustrate the technical solutions in the embodiments of this application 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 some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0034] Figure 1 This is a schematic diagram of a high-standard farmland spore monitoring and disease prediction system based on AI recognition, provided as an embodiment of this application. Detailed Implementation

[0035] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0036] like Figure 1 As shown in the figure, this application provides a high-standard farmland spore monitoring and disease prediction system based on AI recognition, including: a data acquisition module, an AI recognition module, a feature extraction and fusion module, and a disease prediction module.

[0037] The data acquisition module is configured to acquire spore image data, farmland meteorological data, farmland soil data, crop growth data, and historical disease data.

[0038] The AI ​​recognition module is configured to use an AI spore recognition model to identify the spore image data and obtain spore type data and spore quantity data.

[0039] The feature extraction and fusion module is configured to extract and fuse features from the spore species data, the spore quantity data, the farmland meteorological data, the farmland soil data, the crop growth data, and the historical disease data to obtain a multi-source fusion feature vector.

[0040] The disease prediction module is configured to use a disease prediction model to predict diseases based on the multi-source fusion feature vector, and obtain disease prediction results.

[0041] For example, the data acquisition module is connected to a spore collection device, which is used to collect spores, generate spore image data, and send it to the data acquisition module.

[0042] The spore collection equipment is distributed in the farmland using a grid-plus-density deployment method.

[0043] Specifically, in this embodiment, the spore collection device is deployed as follows: Equipment Composition: The standardized spore collection equipment includes a quantitative inhalation spore trap (sampling flow rate 5-10L / min, with a sampling interval of 1-2 hours that can be set remotely), a high-definition microscopic imaging unit (resolution ≥2048×1536 pixels, equipped with an LED supplementary lighting system and autofocus function to ensure clear spore images), an IoT data transmission module (supporting 4G / 5G / NB-IoT protocols, capable of uploading collected spore images and equipment status data in real time), and a solar power supply unit (suitable for farmland without power grids, ensuring continuous operation of the equipment). Deployment principle: Basic collection equipment is deployed in a grid pattern of 500 meters × 500 meters in high-standard farmland. At the same time, the equipment is deployed more densely in key areas such as farmland edges (susceptible to invasion by foreign spores) and irrigation inlets (high humidity makes it easy for spores to grow), with a spacing of 200-300 meters to ensure that there are no blind spots in monitoring.

[0044] In this embodiment, the sources of spore image data, farmland meteorological data, farmland soil data, crop growth data, and historical disease data are as follows: Spore image data: Acquired in real time by the microscopic imaging unit of the acquisition device, generating one set of spore images every 1-2 hours (each set contains 10-20 images of different fields of view).

[0045] Farmland meteorological data: collected through small weather stations deployed in farmland (accuracy: temperature ±0.5℃, humidity ±3%RH, precipitation ±0.1mm, wind speed ±0.2m / s), with a sampling frequency of 1 hour / time, including parameters such as temperature, humidity, precipitation, wind speed, and sunshine duration.

[0046] Farmland soil data: collected by soil sensors buried at a depth of 20cm, with a sampling frequency of once a day, including soil moisture content (accuracy ±2%), pH value (accuracy ±0.1%), and organic matter content (accuracy ±0.1%).

[0047] Crop growth data: obtained through field high-definition cameras (1920×1080 pixels, 3 times a week) and manual-assisted measurements, including crop growth stage (such as wheat greening stage and jointing stage), plant height (accuracy ±1cm), and leaf health (graded from 1 to 5).

[0048] Historical disease data: Collect disease records of the target farmland over the past 3-5 years, including the onset time, disease type, affected area, control agents, application methods and control effects, and construct a historical disease database.

[0049] For example, the AI ​​recognition module is equipped with an image preprocessing unit.

[0050] The image preprocessing unit is used to perform noise reduction, enhancement, and segmentation processing on the spore image data before using it as input to the AI ​​spore recognition model.

[0051] Specifically, in this embodiment, the denoising process is as follows: a 3×3 Gaussian filtering algorithm is used to denoise the original spore image data, and the formula is: .

[0052] Where σ represents the standard deviation of the Gaussian function, set to 0.8, to remove electronic noise and environmental interference in the image, and G(x,y) is a two-dimensional Gaussian function, where (x,y) represents the coordinate offset of a pixel relative to the center point.

[0053] The enhancement process involves improving the contrast between the spores and the background through histogram equalization, specifically by stretching the image grayscale histogram to the full grayscale range (0-255) to enhance the spore outline and texture features.

[0054] The segmentation process involves using the Otsu adaptive threshold segmentation algorithm to automatically calculate the optimal threshold (usually between 120 and 180), separating the spore target from the background region, and obtaining the independent image region of a single spore, with a segmentation accuracy of ≥92%.

[0055] For example, the AI ​​spore recognition model is built on a convolutional neural network, and the structure of the AI ​​spore recognition model includes an input layer, a feature extraction layer, and a classification and counting layer.

[0056] The AI ​​spore recognition model incorporates ImageNet pre-trained weights and Dropout regularization.

[0057] Specifically, in this embodiment, the specific structure of the AI ​​spore recognition model is as follows: Input layer: Receives preprocessed spore images, normalized to 224×224 pixels, with 3 channels (RGB format).

[0058] Feature extraction layer: It contains 5 convolutional layers with kernel sizes of 3×3, 3×3, 5×5, 3×3 and 3×3 respectively, and a stride of 1 for each layer. Each layer is followed by a ReLU activation function f(x)=max(0,x) and a 2×2 max pooling layer (stride 2) to gradually extract features such as the shape, texture and color of the spores.

[0059] The classification and counting layers consist of two fully connected layers (256 and 128 neurons respectively), employing Dropout (probability 0.5) to prevent overfitting. The output layer uses the Softmax activation function. It enables spore classification (supporting the identification of spores from 10 common wheat diseases, including powdery mildew and stripe rust). i Let z represent the probability of the i-th disease. i z represents the original score of the i-th disease. j denoted by j, where k represents the total number of disease categories, and the number of spores is counted using a contour detection algorithm (based on Canny edge detection).

[0060] The training process of the AI ​​spore recognition model is as follows: The AI ​​spore recognition model is initialized with ImageNet pre-trained weights. 50,000 labeled spore images (5,000 images of each type of spore) are used and divided into training set, validation set and test set in a 7:2:1 ratio. The training rounds are 50 epochs, the batch size is 32, and the optimizer is Adam (initial learning rate 0.001, decaying by 0.95 every 10 epochs). After training, the model's recognition accuracy on the test set is ≥95%, and the counting error rate is ≤8%.

[0061] For example, the high-standard farmland spore monitoring and disease prediction system based on AI recognition also includes a multi-source data preprocessing module.

[0062] The multi-source data preprocessing module is configured to perform outlier removal, missing value imputation, and data standardization on the spore type data, spore quantity data, farmland meteorological data, farmland soil data, crop growth data, and historical disease data, and then use the results as input to the feature extraction and fusion module.

[0063] Specifically, in this embodiment, outlier removal is performed by using the 3σ principle to remove outlier data, such as temperatures exceeding the normal range for the same period in the region (-10℃ to 40℃) and humidity > 100%RH. Missing value imputation: Short-term missing data (≤24 hours) are imputed using linear interpolation, while long-term missing data (>24 hours) are imputed using the historical mean of that period, with an imputation error ≤5%; Data standardization: Normalize all data to the interval [0,1], using the following formula: .

[0064] Where x norm Here are the standardized values ​​for this type of data, where x is the original data. min x max These are the minimum and maximum values ​​for this type of data, ensuring consistent data units and facilitating subsequent model training.

[0065] For example, the feature extraction includes spatial feature extraction based on graph convolutional networks and temporal feature extraction based on long short-term memory networks.

[0066] Specifically, in this embodiment, the spatial feature extraction process is as follows: Graph structure construction: Based on the high-standard farmland grid monitoring network, the coverage area of ​​each spore collection device is set as a graph node. The node characteristics include the proportion of spore species in the area (such as the proportion of powdery mildew spores) and spore density (spore count / collection area, unit: spores / m²). According to the actual distance between devices, nodes with a distance of less than 500 meters are set as adjacent nodes, and edges are constructed. The edge weight is the Pearson correlation coefficient of the spore density of adjacent nodes (reflecting the spatial correlation of spore dispersal).

[0067] GCN Model Training: A 2-layer Graph Convolutional Network (GCN) is used to extract spatial features. Each layer contains 64 neurons, with ReLU activation function and the convolution formula as follows: Z=A ^ ReLU(A ^ XW0+b0)W1+b1.

[0068] (where A) ^(where X is the normalized adjacency matrix, W is the node feature matrix, and b is the bias), outputting a 64-dimensional spatial feature vector for each node to capture the spatial patterns of spore diffusion direction and aggregation area in farmland.

[0069] The specific process of time feature extraction is as follows: Time series construction: The preprocessed multi-source data (12 features such as spore number, temperature, and humidity) are arranged in 1-hour time steps to form a continuous time series dataset (e.g., 30 days of data contains 720 time steps).

[0070] LSTM model training: A two-layer Long Short-Term Memory (LSTM) network is used to extract temporal features, with 128 neurons in each layer, through the input gate i. t =σ(W i [x t ,h t−1 ]+b i Forgotten Gate f t =σ(W f [x t ,h t−1 ]+b f ), output gate o t =σ(W o [x t ,h t−1 ]+b o It captures long-term dependencies in time series (such as the growth trend of spore number under continuous high humidity) and outputs a 64-dimensional time feature vector.

[0071] For example, the feature fusion includes feature splicing, attention weighting, and multi-source data supplementation fusion.

[0072] Specifically, in this embodiment, feature concatenation involves concatenating the 64-dimensional spatial feature vector output by GCN with the 64-dimensional temporal feature vector output by LSTM to form an initial 128-dimensional fused feature vector.

[0073] Attention weighting: A two-layer fully connected network (64 neurons in the first layer and 32 neurons in the second layer) is constructed to calculate the importance scores of each dimension of the initial fusion features. Attention weights are generated through the Softmax function (e.g., spore density feature weight 0.35, humidity feature weight 0.25). After weighted summation, a 128-dimensional final fusion feature vector is obtained, highlighting the impact of key features on disease prediction.

[0074] Multi-source data supplementation and fusion: The final fused feature vector and the preprocessed soil data and crop growth data are input into a 3-layer fully connected neural network (FCN, with 256, 128 and 64 neurons respectively). The complex correlation between the multi-source data is learned through nonlinear transformation (ReLU activation function), and a 128-dimensional final fused feature representation is output.

[0075] For example, the disease prediction model is built on a fully connected neural network, and the structure of the disease prediction model includes a hidden layer and an output layer.

[0076] Specifically, in this embodiment, a disease prediction model is constructed based on a fully connected neural network. The input is a 128-dimensional fusion feature representation, and the output is the probability (0-100%) of disease occurrence in various areas of farmland and the disease level ("no risk" [0, 20%), "low risk" [20%, 40%), "medium risk" [40%, 70%), "high risk" [70%, 100%]) within the next 7 days. The specific structure of the disease prediction model is as follows: Hidden layers: 3 layers, with 256, 128 and 64 neurons respectively. Each layer is followed by a ReLU activation function and Dropout (probability 0.3).

[0077] Output layer: 2 neurons, corresponding to the disease occurrence probability (Sigmoid activation function) and disease level (Softmax activation function), respectively.

[0078] The training and optimization of the disease prediction model are as follows: Dataset partitioning: Historical data (including spore monitoring, environmental, and disease records) were divided into training and validation sets in an 8:2 ratio; Loss function: A weighted sum of cross-entropy loss (used for disease severity prediction, weight 0.6) and mean squared error loss (used for probability prediction, weight 0.4), as shown in the formula: Loss = 0.6 × CE(y) true ,y pred )+0.4×MSE(p true ,p pred ).

[0079] Wherein, CE(y) true ,y pred ) represents the cross-entropy loss, MSE(p) true ,p pred ) represents the mean square error loss.

[0080] Optimization algorithm: The Adam optimizer is used with an initial learning rate of 0.001 and a decay of 0.95 every 10 cycles.

[0081] Regularization: L2 regularization (coefficient 0.001) is introduced to prevent overfitting.

[0082] Training termination condition: Stop training when the total loss value of the validation set no longer decreases for 5 consecutive epochs, save the optimal model parameters, and after training, the mean absolute percentage error of the model validation set is ≤10%.

[0083] For example, the high-standard farmland spore monitoring and disease prediction system based on AI recognition also includes an early warning decision module.

[0084] The early warning decision module is configured to provide multi-level early warning and decision support based on the disease prediction results.

[0085] Specifically, in this embodiment, the multi-level early warning mechanism of the early warning decision module is as follows: Low-risk warning: When the disease level is "low risk", the system sends a reminder to farmland managers via SMS, which includes the trend of spore quantity changes and the recommended inspection frequency (1 time / day).

[0086] Medium-risk warning: When the level is "medium risk", a warning notification will be pushed to the management platform, along with a risk area map (marking areas with high spore density) and a spore density heat map. It is recommended to increase the patrol frequency to 2 times / day and prepare control agents.

[0087] High-risk warning: When the risk level is "high", the farmland management station will be triggered by sound and light alarms, automatically generate a warning report (including predicted onset time, affected area, and recommended prevention and control window period), and push it to the regional agricultural authorities. At the same time, an emergency notification will be sent to the management personnel.

[0088] The intelligent decision support of the early warning decision module is as follows: Based on an expert knowledge base (containing prevention and control solutions for 10 common diseases), targeted suggestions are generated: Recommended pesticides: Based on the predicted disease type, we recommend 2-3 highly effective and low-toxicity pesticides (e.g., for powdery mildew, we recommend 25% triadimefon wettable powder, applied at a concentration of 1500 times, by spraying).

[0089] Application timing: Based on weather forecasts (no precipitation in the next 24 hours and wind speed <2m / s), the optimal application time is recommended (such as 9:00-11:00 am or 3:00-5:00 pm, avoiding high temperatures and dew periods that may affect efficacy).

[0090] Prevention and control scope delineation: Based on the risk area map, areas with a disease occurrence probability ≥60% are designated as key prevention and control areas (requiring full-area application of pesticides at the recommended upper limit), areas with a probability of 30%-60% are designated as routine prevention and control areas (applying pesticides at the recommended dosage), and areas with a probability <30% are designated as monitoring areas (requiring only enhanced inspections, no pesticide application is needed). This reduces pesticide usage and prevention and control costs, and reduces agricultural non-point source pollution.

[0091] For example, the high-standard farmland spore monitoring and disease prediction system based on AI recognition also includes a result evaluation module.

[0092] The result evaluation module is configured to perform multi-dimensional evaluations of the performance of the data acquisition module, the AI ​​recognition module, the disease prediction module, and the early warning decision module, thereby quantifying and verifying the system performance.

[0093] Specifically, in this embodiment, the result evaluation module is equipped with the following multi-dimensional evaluation indicators: Spore monitoring efficiency indicators: monitoring coverage (actual monitoring area / total farmland area, target ≥98%), data collection frequency (actual collection interval / set interval, target deviation ≤5%), equipment failure rate (number of faulty devices / total number of devices, target ≤2% / month).

[0094] AI recognition accuracy metrics: Spore species recognition accuracy (number of correctly identified spores / total number of identified spores, target ≥ 95%), spore counting error rate (|AI count - manual count| / manual count × 100%, target ≤ 8%), single image processing time (target ≤ 3 seconds).

[0095] Disease prediction accuracy indicators: mean absolute error (target ≤ 5%), root mean square error (RMSE, target ≤ 7%), mean absolute percentage error (target ≤ 10%), and warning lead time (warning time - actual onset time, target ≥ 3 days).

[0096] Decision support practicality indicators: Adoption rate of prevention and control recommendations (farmland area with adopted recommendations / total application area, target ≥ 80%), disease control effect (reduction rate of diseased area after prevention and control, target ≥ 40%), and reduction rate of pesticide use (pesticide use after adopting system recommendations / traditional control use, target ≥ 20%).

[0097] The evaluation implementation method for the results evaluation module is as follows: Select typical high-standard farmland (≥1000 mu, planted with wheat or rice) to conduct field tests for one complete growing season (e.g., the entire growth period of wheat is 210 days). The specific steps are as follows: Baseline data collection: Before the test, baseline data (species and quantity) of spore monitoring were obtained through manual sampling for one month. Data such as disease incidence and pesticide dosage under traditional control methods were recorded as a comparison benchmark.

[0098] System deployment and data acquisition: Spore collection equipment and meteorological and soil sensors are deployed according to the "grid + encryption" principle to collect multi-source data in real time. The system automatically completes AI recognition, feature fusion, disease prediction and early warning decision-making.

[0099] Data comparison and indicator calculation: Manual sampling verification is carried out twice a week. The spore identification results and disease prediction results output by the system are compared with the manual measured data. Various evaluation indicators are calculated monthly, and a complete evaluation report is generated after the growing season ends.

[0100] Optimization and iteration: Based on the evaluation results, optimize the AI ​​recognition model (such as supplementing rare spore samples) and prediction model parameters (such as adjusting the regularization coefficient) to improve system performance and stability.

[0101] The following two examples illustrate the practical application of a high-standard farmland spore monitoring and disease prediction system based on AI recognition.

[0102] Example 1: Application of spore monitoring and powdery mildew prediction in a high-standard wheat farmland (1500 mu) in North China.

[0103] Implementation components: System deployment: 12 spore collection devices are deployed in a 500m x 500m grid, with 3 additional devices placed along the edge of farmland (near woodland) and at irrigation inlets; 5 small weather stations (monitoring temperature, humidity, and wind speed) and 20 soil sensors (buried 20cm deep to monitor moisture content and pH) are deployed simultaneously; and 3 high-definition cameras are installed in the field (to capture images of wheat growth 3 times a week).

[0104] Data Acquisition and Preprocessing: From March to June 2024 (wheat greening to grain filling stage), the spore collection device acquired one set of images (15 images per set) every 1.5 hours, the meteorological station uploaded data once every hour, and the soil sensor uploaded data once a day. The spore images were preprocessed using 3×3 Gaussian filtering (σ=0.8), histogram equalization, and Otsu segmentation (threshold 145). Abnormal meteorological data (such as a high temperature of 35℃ in April) were removed using the 3σ principle, and missing soil moisture content data for 2 days were filled using linear interpolation. All data were normalized to the [0,1] interval.

[0105] AI Recognition and Feature Fusion: A pre-trained 5-layer convolutional neural network model was used to identify spores. During the testing period, the accuracy rate of powdery mildew spore identification was 96.2%, and the counting error rate was 7.5%. A graph structure with 15 nodes was constructed based on a graph convolutional network (the edge weights are the Pearson correlation coefficients of spore density, with a mean of 0.73), and 64-dimensional spatial features were extracted. Time series data with a step size of 1 hour was processed through a 2-layer long short-term memory network (128 neurons / layer) to extract 64-dimensional temporal features. The features were then fused with an attention mechanism (spore density weight 0.38, humidity weight 0.26) and a 3-layer fully connected neural network (256 / 128 / 64 neurons) to obtain the final 128-dimensional features.

[0106] Predictive Model Training and Early Warning Decision-Making: A 3-layer fully connected neural network prediction model (256 / 128 / 64 neurons) was constructed based on the TensorFlow machine learning framework. The loss function was cross-entropy (0.6) + mean squared error (0.4), Adam optimization (initial learning rate 0.001), and L2 regularization (0.001). The historical data from 2021 to 2023 were divided into training and validation sets in an 8:2 ratio. After 85 training cycles, the mean absolute percentage error of the validation set was reduced to 9.2%. On April 20, 2024, the model predicted that the probability of powdery mildew occurrence would rise to 45%-55% (medium risk) from April 25 to 30, triggering an early warning from the platform. It was recommended to spray with 25% triadimefon wettable powder at a dilution of 1500 times on April 23-24 (no precipitation, wind speed 1.5-2m / s), focusing on the prevention and control of 300 acres in the northeastern region.

[0107] Results evaluation: Compared with manual monitoring data, the system's spore identification accuracy rate was 96.2% (78% for manual identification), and the counting error rate was 7.5%; the average absolute percentage error for powdery mildew prediction was 8.8%, and the early warning was 5 days in advance; after control measures were implemented according to the system's recommendations, the affected area decreased by 42% compared to 2023 (traditional control), the amount of pesticides used decreased by 23%, and the control cost decreased by 31%.

[0108] Effects: In this example, the system successfully achieved dynamic monitoring of spores and accurate prediction of powdery mildew in high-standard wheat farmland in North China, solving the problems of low efficiency and delayed prediction associated with traditional manual sampling. Through a 5-day advance warning and precise control measures, the spread of the disease was effectively controlled, significantly reducing control costs and pesticide usage. This provides a replicable and scalable technical solution for disease control in high-standard wheat farmland in North China, while also supporting regional food security and green agricultural development.

[0109] Example 2: Application of spore monitoring and rice blast prediction in a high-standard rice farmland (2000 mu) in East China.

[0110] Implementation components: System deployment: 16 spore collection devices were deployed in a 500m x 500m grid, with 4 additional devices placed near the paddy field inlet and along the edge of the field ridges; 6 weather stations were set up (focusing on monitoring precipitation and humidity), 25 soil sensors were installed (monitoring organic matter content and pH value), and 4 field cameras were installed.

[0111] Core adjustments: Based on the characteristics of rice blast spores, 5,000 additional labeled images of rice blast spores were added to train the convolutional neural network model. The edge weight calculation parameters of the graph convolutional network were adjusted (considering the influence of paddy field water system connectivity on spore diffusion). The time step of the long short-term memory network was optimized to 2 hours (to adapt to the nighttime diffusion characteristics of rice blast spores).

[0112] Results: From July to September 2024 (rice tillering to heading stage), the accuracy rate of rice blast spore identification was 95.8%, the average absolute percentage error of prediction was 9.5%, and the early warning was issued 4 days in advance; after applying 40% isoprothiolane EC at 800 times dilution as recommended by the system, the diseased area decreased by 38% and the amount of pesticide used decreased by 21%.

[0113] Effects: This example verifies the system's adaptability to different crops (rice) and different diseases (rice blast). By adjusting the model samples and parameters, it can meet the disease control needs of high-standard rice farmland in East China, further demonstrating the system's versatility and practicality.

[0114] This application embodiment improves the reliability and adaptability of high-standard farmland spore monitoring and disease prediction by acquiring multi-source data such as spore images, farmland weather, farmland soil, crop growth, and historical diseases, and combining AI spore recognition models, feature extraction and fusion technology, and disease prediction models.

[0115] The proposed spore collection device is used to collect spores, generate spore image data and send it to the data acquisition module. The spore collection device is distributed in farmland using a grid + densification deployment method to achieve all-time, blind-spot-free dynamic monitoring of spores in high-standard farmland, breaking through the spatial limitations of traditional single-point monitoring.

[0116] The proposed AI spore recognition model is based on a convolutional neural network. The structure of the AI ​​spore recognition model includes an input layer, a feature extraction layer, and a classification and counting layer. The AI ​​spore recognition model introduces ImageNet pre-trained weights and Dropout regularization, which improves the accuracy and efficiency of spore recognition.

[0117] The proposed feature extraction methods include spatial feature extraction based on graph convolutional networks and temporal feature extraction based on long short-term memory networks. Feature fusion methods include feature splicing, attention weighting, and multi-source data supplementation and fusion, which fully explore the multi-dimensional relationship between spores, environment, and crops and improve data utilization.

[0118] The proposed disease prediction model is based on a fully connected neural network. The structure of the disease prediction model includes a hidden layer and an output layer, which improves the accuracy of disease prediction.

[0119] The proposed early warning and decision-making module provides multi-level early warning and decision support based on disease prediction results, realizing closed-loop management of prediction-early warning-decision.

[0120] Although preferred embodiments of this application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of this application.

[0121] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and variations.

Claims

1. A high-standard farmland spore monitoring and disease prediction system based on AI recognition, characterized in that, include: Data acquisition module, AI recognition module, feature extraction and fusion module, disease prediction module; The data acquisition module is configured to acquire spore image data, farmland meteorological data, farmland soil data, crop growth data, and historical disease data. The AI ​​recognition module is configured to: use an AI spore recognition model to recognize the spore image data to obtain spore type data and spore quantity data; The feature extraction and fusion module is configured to: extract and fuse features from the spore species data, the spore quantity data, the farmland meteorological data, the farmland soil data, the crop growth data, and the historical disease data to obtain a multi-source fusion feature vector; The disease prediction module is configured to use a disease prediction model to predict diseases based on the multi-source fusion feature vector, and obtain disease prediction results.

2. The high-standard farmland spore monitoring and disease prediction system based on AI recognition as described in claim 1, characterized in that, The data acquisition module is connected to a spore collection device, which is used to collect spores, generate spore image data, and send it to the data acquisition module. The spore collection equipment is distributed in the farmland using a grid-plus-density deployment method.

3. The high-standard farmland spore monitoring and disease prediction system based on AI recognition according to claim 1, characterized in that, The AI ​​recognition module is equipped with an image preprocessing unit; The image preprocessing unit is used to perform noise reduction, enhancement, and segmentation processing on the spore image data before using it as input to the AI ​​spore recognition model.

4. The high-standard farmland spore monitoring and disease prediction system based on AI recognition according to claim 1, characterized in that, The AI ​​spore recognition model is built on a convolutional neural network. The structure of the AI ​​spore recognition model includes an input layer, a feature extraction layer, and a classification and counting layer. The AI ​​spore recognition model incorporates ImageNet pre-trained weights and Dropout regularization.

5. The high-standard farmland spore monitoring and disease prediction system based on AI recognition according to claim 1, characterized in that, Also includes: Multi-source data preprocessing module; The multi-source data preprocessing module is configured to perform outlier removal, missing value imputation, and data standardization on the spore type data, spore quantity data, farmland meteorological data, farmland soil data, crop growth data, and historical disease data, and then use the results as input to the feature extraction and fusion module.

6. The high-standard farmland spore monitoring and disease prediction system based on AI recognition according to claim 1, characterized in that, The feature extraction includes spatial feature extraction based on graph convolutional networks and temporal feature extraction based on long short-term memory networks.

7. The high-standard farmland spore monitoring and disease prediction system based on AI recognition according to claim 1, characterized in that, The feature fusion This includes feature splicing, attention weighting, and multi-source data supplementation and fusion.

8. The high-standard farmland spore monitoring and disease prediction system based on AI recognition according to claim 1, characterized in that, The disease prediction model is built on a fully connected neural network, and its structure includes a hidden layer and an output layer.

9. A high-standard farmland spore monitoring and disease prediction system based on AI recognition as described in claim 1, characterized in that, Also includes: Early warning and decision-making module; The early warning decision module is configured to provide multi-level early warning and decision support based on the disease prediction results.

10. A high-standard farmland spore monitoring and disease prediction system based on AI recognition as described in claim 9, characterized in that, It also includes: a results evaluation module; The result evaluation module is configured to perform multi-dimensional evaluations of the performance of the data acquisition module, the AI ​​recognition module, the disease prediction module, and the early warning decision module, thereby quantifying and verifying the system performance.

Images