Multi-element crop disease and insect pest monitoring and early warning method based on big data

Abstract

The invention belongs to the technical field of informatization prevention and control of diseases and insect pests, discloses a multi-element crop disease and insect pest monitoring and early warning method based on big data, and aims to solve the problems that traditional monitoring and early warning are one-sided and low in efficiency. The method comprises the following steps: firstly, constructing a sky-ground integrated network, and collecting satellite remote sensing, unmanned aerial vehicle images, internet-of-things perception, weather forecast and agricultural knowledge graph data; constructing a spatio-temporal feature vector through preprocessing and feature engineering; then, through a GCN-LSTM hybrid model, fusing spatio-temporal information to predict a future disease and pest occurrence probability; based on a prediction result, starting four-level early warning and generating a precise prevention and treatment prescription; and finally, model closed-loop optimization is realized through prevention and treatment effect evaluation. According to the invention, three-dimensional monitoring and accurate prediction are realized, the early warning time is advanced, the pesticide usage amount is reduced, the crop yield is guaranteed, and agriculture is promoted to be converted into data drive.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of information-based pest control, and particularly relates to a multi-element crop pest monitoring and early warning method based on big data. BACKGROUND

[0002] As a basic industry of the national economy, the stable development of agriculture is directly related to food security and livelihood security, and the outbreak of crop pests is one of the core bottlenecks restricting high yield and high quality of agriculture. East China is an important agricultural production area in China, and the planting scale of representative crops such as rice, corn and tomato is large and the degree of intensification is high. However, the climate in this region is humid and the seasons are distinct, and the environmental conditions of high temperature and humidity and alternating cold and warm are extremely easy to induce various pests, and the occurrence and spread of pests show significant spatiotemporal dynamics, which not only leads to the decline of crop growth and the sharp reduction of yield, but also may cause ecological problems such as excessive pesticide residues, soil and water pollution due to unreasonable pesticide application.

[0003] The traditional crop pest monitoring and early warning method mainly relies on manual inspection and experience judgment of agricultural technicians. This mode is not only low in efficiency and limited in coverage, and is difficult to meet the monitoring needs of large-scale intensive planting, and has weak ability to identify early hidden symptoms of pests, often missing the best prevention and control opportunity. Some areas have introduced single-dimensional monitoring methods, such as relying only on satellite remote sensing data for macro monitoring or only collecting environmental data through ground sensors, but the resolution of satellite remote sensing data is limited and it is difficult to capture the micro lesions of crop canopy, and ground single-point monitoring is easily affected by spatial heterogeneity, and cannot form a three-dimensional and comprehensive monitoring perspective.

[0004] At the same time, there is a lack of effective multi-source data fusion and feature mining method in the prior art, and the data formats of different sources are different and have weak correlation, which is difficult to be converted into effective information supporting accurate prediction, and the traditional prediction model is mainly based on statistical method, which cannot fully capture the evolution rule of pests in time and the diffusion mode in space, resulting in low prediction accuracy and poor timeliness, and ultimately causing insufficient pertinence of prevention and control measures and the phenomenon of "flooding" pesticide application, which not only increases the cost of agricultural production, but also aggravates the pressure on the ecological environment, so there is an urgent need for a pest monitoring and early warning technology system that can realize comprehensive monitoring, accurate prediction and intelligent decision-making. SUMMARY

[0005] In order to overcome the above technical problems, the application provides a multi-element crop pest monitoring and early warning method based on big data.

[0006] The application adopts the following technical scheme: A multi-element crop pest monitoring and early warning method based on big data, comprising the following steps: Step 1: Construct a sky-ground integrated multi-source heterogeneous data collection network to collect multi-dimensional data covering crop planting areas and converge to a cloud big data platform; Step 2: Preprocess the collected multi-source data, and construct a multi-dimensional spatio-temporal feature vector based on agronomic mechanisms and data characteristics; Step 3: Construct a hybrid deep learning model that combines graph convolution network and time series model, input the spatio-temporal feature vector to predict the occurrence probability of pests and diseases; Step 4: Start a multi-level early warning mechanism according to the prediction results, and generate a visual precise prevention and control prescription; Step 5: Monitor the implementation effect of prevention and control measures, feed back the evaluation data to the hybrid deep learning model to realize closed-loop self-optimization of the model.

[0007] Preferably, the sky-ground integrated multi-source heterogeneous data collection network in step 1 includes a spatial macroscopic layer, a field mesoscopic layer, a crop near-observable layer, and a data integration layer, wherein: The spatial macroscopic layer accesses satellite remote sensing data, including calculating normalized vegetation index, leaf area index, and surface temperature; The field mesoscopic layer uses unmanned aerial vehicles equipped with visible light, multispectral, and thermal infrared cameras to conduct regular inspections along preset routes and obtain crop canopy ultra-high-definition images; The crop near-observable layer collects environmental microclimate data, crop leaf image, and pest monitoring data through grid-based Internet of Things monitoring sites; The data integration layer accesses precise weather forecast data and an agricultural knowledge graph that integrates expert knowledge.

[0008] Preferably, the data preprocessing in step 2 includes: performing cloud detection, radiation calibration, and atmospheric correction on satellite remote sensing images; performing image stitching, orthophoto generation, and crop and background segmentation on unmanned aerial vehicle images; all preprocessed data are assigned with accurate time stamps and geographic coordinates.

[0009] Preferably, the process of constructing a multi-dimensional spatio-temporal feature vector in step 2 includes: calculating effective accumulated temperature based on continuous temperature data, and calculating disease infection risk window length combining temperature and humidity data; using a pre-trained convolutional neural network to extract texture, color, and lesion shape features of crop images; spatio-temporally aligning image features, environmental features, and vegetation index features of the plot.

[0010] Preferably, the construction process of the hybrid deep learning model in step 3 is as follows: abstract the planting area plot as a dynamic graph structure, use the spatio-temporal feature vector as the graph node attribute, use the plot spatial relationship and propagation correlation as the graph edge and dynamically adjust the edge weight combining wind direction data; aggregate plot spatial information through GCN, input the graph representation sequence output by GCN into LSTM to capture time-dependent relationships, and realize future pest and disease occurrence probability prediction.

[0011] Preferably, the multi-level early warning mechanism in step 4 includes four levels: blue warning pushes attention notifications and inspection suggestions, yellow warning generates risk boundaries and biological control suggestions, orange warning outputs control prescription maps accurate to square meters, and red warning synchronizes disaster information to regional agricultural management departments.

[0012] Preferably, the prevention and control prescription map includes the application area boundary, the low-toxicity pesticide type matching the agricultural knowledge graph, the recommended spraying dosage and the optimal operation time window, and can be directly imported into the plant protection drone operation system to achieve targeted spraying.

[0013] Preferably, in step 5, the model closed-loop self-optimization adopts a reinforcement learning framework, taking the early warning decision as the agent action, and the prevention and control effect and yield change as reward / punishment signals, and updating the model parameters with the continuously collected post-effect evaluation data.

[0014] Preferably, the agricultural knowledge graph integrates data from annual pest and disease surveys, biological characteristics of pests and diseases, environmental correlation patterns, and active ingredients and usage guidelines for pesticides.

[0015] Preferably, the IoT monitoring station is equipped with a macro camera and an intelligent insect monitoring lamp with a lure to achieve automatic insect shooting and identification.

[0016] Compared with the prior art, the beneficial effects of the present invention are: This invention effectively solves the problems of one-sided monitoring, inaccurate prediction, and inefficient decision-making in traditional technologies by constructing a full-dimensional monitoring network, deep data processing, and innovative model design, bringing technical benefits and application value.

[0017] First, the integrated air-ground multi-source heterogeneous data acquisition network breaks the limitations of a single monitoring dimension. Satellite remote sensing data provides a macroscopic growth background, UAV centimeter-level imagery enables precise capture of microscopic lesions, and IoT sites acquire real-time environmental microclimate and biological information. Combined with weather forecasts and agricultural knowledge graphs, a three-dimensional data support system of "macro-meta-near-scale" is formed, ensuring the comprehensiveness and timeliness of monitoring information and laying a solid foundation for subsequent precise analysis.

[0018] Secondly, based on agronomic mechanisms, multidimensional feature engineering and spatiotemporal alignment technology transform raw data into deep features with clear biological significance, such as effective accumulated temperature and the duration of pathogen infection risk windows. At the same time, by constructing high-dimensional spatiotemporal vectors through feature fusion, the utilization value of the data is greatly improved, and the problem of "fragmentation" of multi-source data is solved.

[0019] Furthermore, the GCN-LSTM hybrid deep learning model innovatively abstracts land parcels into a dynamic graph structure. By learning the spatial diffusion patterns of pests and diseases through GCN and capturing the temporal evolution patterns through LSTM, it achieves accurate prediction of the probability of future pest and disease occurrences, advancing the warning time by 5 to 7 days and gaining crucial time for prevention and control. The four-level early warning mechanism and visualized precision prevention and control prescriptions realize "tiered response and targeted application." Blue and yellow warnings reduce the risk of outbreaks through early intervention, while orange and red warnings provide application plans accurate to the square meter. Combined with agricultural knowledge graphs to match low-toxicity and high-efficiency pesticides, it can be directly imported into plant protection drones to achieve fully automated targeted spraying, reducing pesticide usage, lowering agricultural production costs, and reducing pesticide pollution to the ecological environment, which meets the needs of green agricultural development.

[0020] Finally, the implementation effect evaluation and model closed-loop self-optimization mechanism realizes the full-process iteration of "monitoring-early warning-decision-prevention-evaluation" through the reinforcement learning framework. This enables the model to continuously learn the optimal decision-making strategy and continuously improve the accuracy of early warning and decision-making as application data accumulates. It forms an adaptive and self-evolving technical system, which, while ensuring the healthy growth and stable yield of crops, provides efficient and intelligent pest and disease control solutions for large-scale agricultural production, and promotes the transformation of agricultural production from "experience-driven" to "data-driven". Attached Figure Description

[0021] Figure 1 This is a schematic diagram of the method flow of the present invention. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of this invention clearer, a specific embodiment will be used to describe the invention in detail below. It should be noted that this embodiment uses three representative crops—rice, corn, and tomato—grown in an agricultural demonstration area in East China as examples to illustrate the specific implementation process of this invention, but the scope of protection of this invention is not limited thereto.

[0023] This embodiment details the complete process of a multi-dimensional crop pest and disease monitoring and early warning method based on big data, from data collection to the generation of accurate prevention and control decisions.

[0024] Step 1: Construct an integrated sky-ground multi-source heterogeneous data acquisition network To achieve comprehensive and multi-dimensional monitoring, we have deployed a multi-level data acquisition system within the demonstration zone.

[0025] First, at the macroscopic spatial level, we access remote sensing data services from the Chinese Gaofen series satellites and the European Sentinel series satellites. We regularly acquire multispectral images covering the entire demonstration area, from which we calculate key vegetation physiological parameters such as the Normalized Difference Vegetation Index (NDVI), Leaf Area Index (LAI), and Land Surface Temperature (LST). This data provides us with a macroscopic background of crop growth and a basis for detecting early abnormal stresses.

[0026] Secondly, at the field observation level, we utilize DJI agricultural drones equipped with visible light, multispectral, and thermal infrared cameras to conduct routine inspections twice a week along preset routes. The drones acquire ultra-high-definition images of the crop canopy at centimeter-level resolution, clearly capturing early lesions, color changes, and minute signs of insect infestation on the leaves.

[0027] Secondly, at the close-up crop level, we installed IoT monitoring stations integrating multiple sensors in a grid-like layout across each plot. These stations can collect and upload real-time environmental microclimate data such as air temperature and humidity, soil temperature and humidity, soil pH, carbon dioxide concentration, and light intensity. Additionally, some stations are equipped with macro cameras pointing towards crop leaves and smart insect monitoring lights with decoys for automatically photographing and identifying specific pests, enabling direct observation of the pests themselves.

[0028] Finally, we also integrated important data from other dimensions. We accessed accurate 72-hour weather forecast data provided by the National Meteorological Center, including rainfall probability, wind speed and direction, and temperature changes. Simultaneously, we constructed an agricultural knowledge graph that integrates historical pest and disease survey data, the biological characteristics and occurrence patterns of various pests and diseases, their correlation with environmental factors, and expert knowledge on the active ingredients and usage guidelines of various pesticides.

[0029] All collected data, whether remote sensing images, drone photos, sensor readings, or structured text, are given precise timestamps and geographic coordinates, and are uniformly aggregated into a cloud-based big data platform for storage and management.

[0030] Step 2: Data Preprocessing and Multidimensional Feature Engineering Raw multi-source data cannot be directly used for model analysis and requires in-depth processing and feature construction.

[0031] For satellite remote sensing imagery, we employ a cloud detection algorithm to remove invalid pixels obscured by clouds, and then perform radiometric calibration and atmospheric correction to ensure consistency across different time phases. For UAV imagery, we first perform image stitching to generate complete orthophoto maps of the plots, and then use image segmentation algorithms to separate crops from the soil background.

[0032] The next step is the core feature engineering. We don't just use raw sensor readings; instead, we construct deep features based on agronomic mechanisms. For example, we calculate the effective accumulated temperature from continuous temperature data, and combine humidity and temperature data to calculate the "pathogen infection risk window duration," which is the cumulative number of hours under suitable environmental conditions for pathogen infection. For image data, we utilize pre-trained convolutional neural network models, such as ResNet, to extract leaf texture features, color histogram features, and the shape and edge features of lesions.

[0033] We align all features in time and space. For example, we correlate the drone image features of a plot of land at a specific time point, the average environmental features of all sensors within that plot, and the vegetation index features of the corresponding satellite image pixels to form a high-dimensional spatiotemporal feature vector.

[0034] Raw, multi-source data cannot be directly used for model analysis and requires in-depth processing and feature construction. This process is not merely simple cleaning, but rather, based on agronomic mechanisms, transforms raw data into feature variables with clear biological significance.

[0035] For satellite remote sensing imagery, after radiometric calibration and atmospheric correction, we calculate key vegetation indices. For example, the Normalized Difference Vegetation Index (NDVI) is calculated as follows: NDVI = (NIR - RED) / (NIR + RED) Here, NIR represents the surface reflectance in the near-infrared band, and RED represents the surface reflectance in the red band. NDVI values ​​can effectively indicate the health and cover of crops, and an abnormal decrease in these values ​​is often an early signal of pest and disease stress.

[0036] When constructing environmental characteristics, we transform discrete meteorological data into continuous variables that directly impact pest and disease occurrence. For example, for pests that rely on temperature accumulation for development (such as the corn borer), we calculate their effective accumulated temperature (GDD). The formula for calculating the effective accumulated temperature over time period T is: GDD=Σ[(T_max_i+T_min_i) / 2-T_base] Accumulate only when (T_max_i+T_min_i) / 2>T_base Where T_max_i and T_min_i are the highest and lowest temperatures on day i, respectively, and T_base is the biological lower limit temperature for the development of this pest. The effective accumulated temperature accurately quantifies the thermal conditions required to drive the pest to complete its life cycle.

[0037] For diseases, especially fungal diseases (such as rice blast), infection is closely related to high humidity. We define the feature "disease infection risk window duration" H_risk, which is calculated as follows: H_risk=ΣI(H_i,T_i) Here, I(H_i, T_i) is a conditional indicator function. In the i-th hour, if the relative humidity H_i is above the threshold required for pathogen infection (e.g., 90%) and the temperature T_i is within its favorable range (e.g., 20-28°C), the function I is 1; otherwise, it is 0. By summing the hours over a day or several days, H_risk quantifies the total favorable length of time during which the pathogen may successfully infect the crop.

[0038] Through this type of feature engineering, we transform multi-source heterogeneous raw data into a unified, high-dimensional spatiotemporal feature matrix, providing high-quality input for subsequent deep learning models.

[0039] Step 3: Construct a spatiotemporal evolution prediction model for pests and diseases based on graph neural networks and time series models. The occurrence and spread of pests and diseases exhibit significant spatiotemporal dynamics, so we designed an innovative hybrid deep learning model to capture this pattern.

[0040] We abstract the farmland plots of the entire demonstration area into a dynamic graph structure. Each node in the graph represents a plot, and the attributes of the node are the high-dimensional feature vectors we constructed in the previous step. The edges between nodes represent the spatial proximity and potential propagation relationships between plots. For example, we dynamically adjust the weights of the edges between adjacent plots based on wind direction data to simulate the possibility of pests migrating with the wind.

[0041] We employ Graph Convolutional Networks (GCNs) to process the spatially structured data at each time step. GCNs can aggregate information from neighboring plots, effectively learning the spatial spread patterns of pests and diseases. For example, if rice planthoppers appear in a neighboring plot, the risk of disease infection in that plot increases accordingly, and the GCN model can quantify this impact.

[0042] In the time dimension, we input the graph representation sequence of each plot at consecutive time points into a Long Short-Term Memory (LSTM) network. LSTM excels at capturing long-term dependencies in time series and can predict the probability of pest and disease outbreaks at a future point in time based on the evolution trends of crop growth, environmental changes, and pest and disease status over a past period.

[0043] By combining GCN and LSTM, our constructed GCN-LSTM model can simultaneously learn the spatial spread patterns and temporal evolution of pests and diseases. We use historical pest and disease occurrence records from the past three years as labels to supervise the training of the model, enabling it to accurately predict the probability of specific pests and diseases (such as rice blast, corn borer, and tomato late blight) occurring in each plot within the next 1 to 7 days.

[0044] The occurrence and spread of pests and diseases exhibit significant spatiotemporal dynamics. To capture this complex dependency, we abstracted the entire agricultural demonstration area into a dynamic graph and constructed a GCN-LSTM hybrid model.

[0045] First, in the spatial dimension, we use Geographic Networks (GCNs) to learn the propagation patterns of pests and diseases across different plots. The core idea of ​​GCNs is to aggregate information from neighboring nodes to update the representation of the central node. For any plot node in the graph, its feature H^(l+1) at layer l+1 is calculated using the following propagation rules: H^(l+1)=σ(D^(-1 / 2)ÃDmber^(-1 / 2)H^(l)W^(l)) In this formula, H^(l) is the feature matrix of all nodes in layer l. Ã=A+I is the adjacency matrix with self-loops, representing that each plot is influenced not only by its neighbors but also by its own state. D̃ is the degree matrix of Ã. The term D̃^(-1 / 2)ÃD̃^(-1 / 2) constitutes a symmetric normalized adjacency matrix, which ensures the stability of information transmission between nodes and can be understood as a weighted averaging process. W^(l) is the weight parameter matrix that the layer l network needs to learn, while σ is a non-linear activation function such as ReLU, used to increase the expressive power of the model. Through this formula, the feature representation of each plot incorporates the state information of its geographical neighbors, thus effectively simulating the spatial diffusion process of pests and diseases.

[0046] Secondly, in the time dimension, we input the plot feature sequence output by GCN at each time step t into LSTM to capture the evolution of pest and disease status over time. LSTM controls the flow of information through its unique gating mechanism. The core update of the cell state c_t can be represented as: c_t=f_t⊙c_(t-1)+i_t⊙g_t Here, c_(t-1) is the cell state at the previous time step, representing long-term memory. f_t is the forgetting gate, which determines how much information is forgotten from long-term memory. i_t is the input gate, which determines how much new information g_t can be added to long-term memory at the current time step. The symbol ⊙ represents element-wise multiplication. This mechanism allows LSTM to selectively remember and forget historical information, thereby learning the long-term temporal dependencies of pest and disease occurrences, such as how consecutive days of rainy weather gradually increase the risk of late blight outbreaks.

[0047] Ultimately, the model outputs the probability value of a specific pest or disease occurring in each plot of land within a future period, providing a direct basis for accurate early warning.

[0048] Step 4: Generate multi-level intelligent early warning and visualized precision prevention and treatment prescriptions The model's predictions are not a simple "yes" or "no," but rather a refined risk map and a complete decision support solution.

[0049] We have established a four-level early warning mechanism: Blue Alert (Attention Level): The model predicts that environmental conditions in a certain area (such as continuous rain, high temperature and humidity) will be very favorable for the occurrence of a certain disease in the coming week, but no actual symptoms have been detected. The system will send a notification to farmers, suggesting that they increase the frequency of drone patrols in the area.

[0050] Yellow Alert (Risk Level): Drones or ground cameras have detected scattered, suspected lesions or a small number of pests, and the model predicts a high risk of spread within the next 72 hours. The system will automatically generate a risk area boundary and recommend small-scale physical or biological control measures, such as releasing natural enemies or using plant-derived pesticides.

[0051] Orange Alert (High Risk): A small-scale outbreak of pests and diseases has been detected. The model predicts that without intervention, it will spread extensively within 48 hours. The system will immediately issue an alert and automatically generate a detailed and precise prevention and control prescription map. This prescription map is accurate to every square meter within the plot, indicating the areas requiring pesticide application, the recommended pesticide type (matched from a knowledge graph with low toxicity and high efficacy), the suggested spraying dosage, and the optimal operating time window.

[0052] Red Alert (Emergency Level): Pests and diseases have shown a large-scale outbreak trend, and the model predicts that they will pose a serious threat to yields. In addition to issuing the highest level alert to farmers, the system will automatically report the disaster situation, the scope of impact assessment, and prevention and control recommendations to the regional agricultural technology extension station or plant protection department in order to activate the regional joint prevention and control emergency plan.

[0053] This precise prevention and control prescription map can be directly imported into the operation system of agricultural drones. The drones can then carry out fully automatic targeted spraying based on this map, applying pesticides only to the affected crops, thereby significantly reducing pesticide usage, lowering costs, and reducing environmental pollution.

[0054] Step 5: Implementation effect evaluation and model closed-loop self-optimization After each early warning and control measure is implemented, this method continuously monitors the control effect using drones and sensors. For example, after pesticide spraying, the system continuously tracks whether the lesions in the target area stop expanding and whether the pest density decreases significantly.

[0055] These post-effect evaluation data will serve as new training samples, feeding back into the predictive model. We introduce a reinforcement learning framework, treating the entire "monitoring-early warning-decision-control-evaluation" process as a closed loop. The model's decisions (when to issue an early warning, which pesticide to recommend) are considered "actions," while the final control effect and yield changes serve as "rewards" or "penalties." Through continuous self-iteration, the model can learn to make optimal decisions under different conditions. For example, it might discover that for a certain resistant pest, alternating the use of two different pesticide components is more effective than using a single pesticide.

[0056] To continuously improve the accuracy of the model and the effectiveness of the decisions, we have established a closed-loop self-optimization mechanism.

[0057] After each early warning and prevention measure is implemented, we evaluate the predictive effectiveness of the model. We use precision and recall as key evaluation metrics.

[0058] Precision = TP / (TP + FP) Recall = TP / (TP + FN) In this model, TP (True Positive) represents plots where the model successfully issued a warning and pests and diseases actually occurred; FP (False Positive) represents plots where the model issued a warning but pests and diseases did not actually occur, which relates to whether unnecessary prevention and control costs will be incurred; and FN (False Negative) represents plots where the model failed to issue a warning but pests and diseases actually occurred, which relates to whether production losses due to missed reporting will occur. By optimizing the balance between these two indicators, the warning system is ensured to be both sensitive and reliable.

[0059] Furthermore, we introduce a reinforcement learning framework to optimize pest control decisions. We define the system state *s* as the characteristics of all current plots, action *a* as the pest control decision generated by the system (such as spraying a specific pesticide in a certain area), and reward *r* as a comprehensive calculation based on changes in yield, cost savings, and environmental impact after pest control. We use the Q-learning algorithm to update the value function Q(s,a) of the decision model, with the following update rules: Q(s,a)←Q(s,a)+α[r+γmax_a'Q(s',a')-Q(s,a)] This formula means that the value Q(s,a) of taking action a in state s is adjusted based on the immediate reward r and the maximum future value max_a'Q(s',a') that can be obtained in the next state s'. α is the learning rate, controlling the step size of updates; γ is the discount factor, balancing the importance of immediate and future rewards. Through continuous interaction and iterative updates with the environment, the model can autonomously learn which prevention and control measures (action a) will yield the greatest long-term comprehensive benefit (cumulative reward) under what pest and disease conditions, thus achieving an intelligent upgrade from "accurate prediction" to "optimal decision-making."

[0060] Through the close collaboration of the above five steps, this embodiment of the invention successfully constructs a multi-faceted crop pest and disease monitoring and early warning system capable of autonomous learning, accurate prediction, and intelligent decision-making. Application in demonstration areas shows that, compared to traditional control methods relying on manual experience, this method can advance the early warning time of pests and diseases by 5 to 7 days, reduce pesticide use by more than 40%, and effectively ensure the healthy growth and final yield of crops.

[0061] Although embodiments of the present invention have been shown and described, those skilled in the art will understand that various changes, modifications, substitutions and variations can be made to the above embodiments without departing from the principles and spirit of the present invention, the scope of which is defined by the claims and their equivalents.

Claims

1. A big data-based multi-element crop disease and pest monitoring and early warning method, characterized in that, The method comprises the following steps: Step 1: constructing a sky-ground integrated multi-source heterogeneous data collection network, collecting multi-dimensional data covering crop planting areas and converging to a cloud big data platform; Step 2: preprocessing the collected multi-source data, and constructing a multi-dimensional space-time feature vector based on the agronomic mechanism and data characteristics; Step 3: constructing a hybrid deep learning model combining graph convolution network and time series model, inputting the space-time feature vector to predict the occurrence probability of diseases and pests; Step 4: starting a multi-level early warning mechanism according to the prediction results, and generating a visual precise prevention and control prescription; Step 5: monitoring the implementation effect of the prevention and control measures, feeding back the evaluation data to the hybrid deep learning model to realize closed loop self-optimization of the model.

2. The method of claim 1, wherein, The sky-ground integrated multi-source heterogeneous data collection network in step 1 comprises a spatial macroscopic layer, a field mesoscopic layer, a crop near view layer and a data integration layer, wherein: The spatial macroscopic layer accesses satellite remote sensing data, including calculating normalized vegetation index, leaf area index and ground temperature; The field mesoscopic layer acquires crop canopy ultra-high definition images through a UAV equipped with visible light, multispectral and thermal infrared cameras, and performs normal patrol inspection according to a preset flight route; The crop near view layer collects environmental microclimate data, crop leaf image and pest monitoring data through a grid-based Internet of Things monitoring site; The data integration layer accesses accurate weather forecast data and an agricultural knowledge graph fused with expert knowledge.

3. The method of claim 1, wherein, The data preprocessing in step 2 includes: performing cloud detection, radiation calibration and atmospheric correction on satellite remote sensing images; performing image stitching, orthophoto generation and crop and background segmentation on UAV images; all preprocessed data are assigned with accurate time stamps and geographic coordinates.

4. The method of claim 1, wherein, The process of constructing a multi-dimensional space-time feature vector in step 2 includes: calculating effective accumulated temperature based on continuous temperature data, and calculating disease infection risk window length combining temperature and humidity data; extracting texture, color and lesion shape features of crop images using a pre-trained convolutional neural network; time and space aligning image features, environmental features and vegetation index features of the plot.

5. The method of claim 1, wherein, The construction process of the hybrid deep learning model in step 3 is: abstracting the planting area plot as a dynamic graph structure, taking the space-time feature vector as the graph node attribute, taking the plot spatial relationship and propagation correlation as the graph edge and dynamically adjusting the edge weight combining wind direction data; Through GCN, the spatial information of the plot is aggregated, the graph representation sequence output by GCN is input into LSTM to capture the time dependence, and the future occurrence probability of diseases and pests is predicted.

6. The method of claim 1, wherein, The multi-level early warning mechanism in step 4 includes four levels: blue warning pushes attention notification and patrol suggestion, yellow warning generates risk boundary and biological control suggestion, orange warning outputs precise to square meter prevention and control prescription map, and red warning synchronizes disaster information to regional agricultural management department.

7. The method of claim 6, wherein, The prevention and control prescription map contains pesticide application area boundary, low-toxicity pesticide type matched with agricultural knowledge graph, recommended spraying dose and optimal operation time window, and can be directly imported into the plant protection UAV operation system to realize targeted spraying.

8. The method of claim 1, wherein, The model closed loop self-optimization in step 5 adopts a reinforcement learning framework, takes early warning decisions as agent actions, takes prevention and treatment effects and yield changes as reward / punishment signals, and updates model parameters with continuously collected aftereffect evaluation data.

9. The method of claim 2, wherein, The agricultural knowledge graph fuses year-by-year pest survey data, biological characteristics of pests, environmental correlation rules, and effective components and use specifications of pesticides.

10. The method of claim 2, wherein, The Internet of Things monitoring station is equipped with a macro lens camera and a smart pest monitoring lamp with a lure core, realizing automatic shooting and identification of pests.

Images