Agricultural meteorological intelligent decision support method and system combined with multi-source heterogeneous data

Abstract

The invention relates to the technical field of agricultural datamation, in particular to an agricultural meteorology intelligent decision support method and system combined with multi-source heterogeneous data, and the method comprises the following steps: S1, obtaining and processing the multi-source heterogeneous data: obtaining agricultural multi-source heterogeneous data through official and monitoring equipment, and preprocessing the multi-source heterogeneous data; s2, data extraction and fusion: extracting multi-source heterogeneous data according to relevance; the method integrates meteorological, soil and crop multi-dimensional data through the space-time diagram neural network, combines the dynamic adjacency matrix and the attention mechanism, effectively captures the space-time correlation characteristics of the farmland area, solves the problem of data island of a traditional method, improves the disaster recognition precision, and improves the disaster recognition accuracy through introducing a transfer learning mechanism and a federated learning framework. Rapid migration of a main production area model to a new planting area is achieved, meanwhile, farm node data are encrypted and aggregated through a parameter server, collaborative modeling is completed on the premise of guaranteeing privacy of agricultural enterprises, and the risk of sensitive data leakage is avoided.

Description

Technical Field

[0001] The present invention relates to the field of agricultural data technology, and in particular to an agricultural meteorological intelligent decision support method and system combining multi-source heterogeneous data. Background Art

[0002] With the growing global population, limited land resources, and the challenges posed by climate change to agricultural production, improving agricultural production efficiency, ensuring food security, and achieving sustainable agricultural development have become major issues of concern to the international community. Simultaneously, the rapid development of science and technology, particularly innovations in information technology, the Internet of Things, big data analytics, and artificial intelligence, has provided strong technical support for addressing these issues.

[0003] For example, a crop disease remote sensing monitoring method based on image processing and deep learning with application number CN202110522015.4 and publication date 20210806 includes the following steps: S1) obtaining remote sensing images of crops; S2) inputting the remote sensing images of crops into the trained YOLOv3 deep learning network based on Keras and TensorFlow framework to obtain recognition results, and realizing remote sensing monitoring of crop diseases. The present invention realizes remote sensing monitoring of crop diseases; it has the advantages of fast recognition speed, high accuracy, high efficiency, and low dependence on the supervisor of farmers. It cooperates with the hierarchical detection based on the feature extraction algorithm, which not only gives full play to the stability and flexibility of the deep learning monitoring effect, but also reflects the advantages of high precision, simple and efficient data labeling of traditional image processing algorithms, avoiding the time-consuming and labor-intensive process of collecting thousands of physical pictures in the early stage of the neural network and manually labeling them before training the neural network for application.

[0004] For example, an irrigation decision-making system and method based on an agricultural system model, whose application number is CN201611127292.0 and whose authorization announcement date is 20191203, belongs to the field of agricultural irrigation decision-making technology. The system includes a meteorological data interface, a meteorological database, a calibrated agricultural system model, a calibrated agricultural system model library, an irrigation decision algorithm, an irrigation decision algorithm library, and an irrigation control system. The meteorological data interface provides the calibrated agricultural system model with the meteorological data required for calculation. The calibrated agricultural system model performs calculations based on meteorological conditions to obtain crop growth information and the moisture conditions of the soil and crops. The irrigation decision algorithm determines whether to irrigate and calculates the irrigation amount based on the model calculation results. The irrigation control system irrigates according to the irrigation amount calculated by the irrigation decision algorithm. The present invention overcomes the shortcomings of existing irrigation decision-making systems and methods, formulates an accurate irrigation plan without installing field sensors, and can predict future irrigation plans and irrigation effects.

[0005] Traditional agrometeorological decision-making relies primarily on a single data source, lacking dynamic assessment and real-time feedback optimization mechanisms for multi-hazard coupled risk. Furthermore, existing methods still face technical bottlenecks in cross-regional migration adaptability, data privacy protection, and decision interpretability. Therefore, a method and system for intelligent agrometeorological decision-making support that integrates multi-source heterogeneous data is urgently needed to address these issues. Summary of the Invention

[0006] The purpose of the present invention is to provide an agricultural meteorological intelligent decision support method and system combining multi-source heterogeneous data to solve the above-mentioned deficiencies in the prior art.

[0007] In order to achieve the above object, the present invention provides the following technical solutions:

[0008] The invention relates to an agricultural meteorological intelligent decision support method combining multi-source heterogeneous data, and the agricultural meteorological intelligent decision support method combining multi-source heterogeneous data is characterized in that it comprises the following steps:

[0009] Step S1. Acquisition and processing of multi-source heterogeneous data: Acquisition of agricultural multi-source heterogeneous data through official and monitoring equipment, and pre-processing of the multi-source heterogeneous data;

[0010] In the step S1:

[0011] Official data include historical temperature / precipitation / wind speed data released by the Meteorological Bureau, crop growth cycle data and soil moisture reports provided by the agricultural department;

[0012] Monitoring equipment data is obtained through drone remote sensing images, ground meteorological sensor arrays, and IoT soil probes;

[0013] Preprocessing includes temporal and spatial interpolation compensation for missing values, wavelet transform filtering of sensor noise data, and normalization of data of different dimensions through min-max standardization.

[0014] Step S2. Data extraction and fusion: Extract multi-source heterogeneous data based on correlation, then construct a spatiotemporal graph neural network to fuse the extracted multi-source heterogeneous data based on spatiotemporal correlation, and introduce an attention mechanism to dynamically weight the contribution of different data sources to the decision-making target;

[0015] In step S2, the spatiotemporal graph neural network includes a bidirectional spatiotemporal convolution module, wherein the spatial dimension uses a graph convolutional network to model the topological relationship of the field plots, and the temporal dimension uses a gated recurrent unit to capture the evolution of meteorological factors. The spatiotemporal graph neural network introduces a transfer learning mechanism to migrate the parameters of the meteorological warning model established in the old main production area to the new planting area, and adjusts the feature offset caused by soil type differences through a domain adaptive algorithm;

[0016] The attention mechanism simultaneously calculates the overlap between the meteorological probability distribution and the crop harvest time window to generate a disaster avoidance priority ranking. The attention mechanism includes the following:

[0017] (1) Source domain attention layer: Calculate the feature similarity matrix between satellite remote sensing spectral data and ground meteorological station data,

[0018] (2) Spatiotemporal attention layer: Dynamically capture the spatiotemporal correlation pattern between typhoon paths and crop planting distribution through a multi-head self-attention mechanism.

[0019] (3) An interpretable weight visualization module is set up at the output end to display the cross-influence factors of soil pH and precipitation;

[0020] The dynamic calculation of edge weights between vertices adopts the following formula, which is as follows:

[0021]

[0022] Among them, W ij is the edge weight between vertices i and j, The square of the Euclidean space distance between vertices i and j, α is the weight coefficient of the spatial distance term α = 0.6, σ is the bandwidth of the Gaussian kernel function σ = 150m, β is the weight coefficient of the feature correlation term β = 0.4, Corr(X i ,X j ) is the feature vector X of nodes i and j i ,X j Pearson correlation coefficient;

[0023] The data fusion process adopts a federated learning framework to achieve multi-agent data collaboration, and aggregates the local models of each farm node through a parameter server under the premise of protecting the privacy data of agricultural enterprises; the fused multi-dimensional feature vector is stored in the blockchain traceability system, and each feature dimension is accompanied by a data source hash value and timestamp.

[0024] The specific steps of step S2 are as follows:

[0025] Step S1-1. Multi-source data extraction and preprocessing: The spatiotemporal correlation features of meteorological, soil, and crop data are screened based on the Pearson correlation coefficient. A convolutional autoencoder is then used to reduce the dimensionality of remote sensing imagery and sensor data to extract key features.

[0026] Step S1-2. Spatiotemporal Graph Neural Network Construction: The farmland area is divided into N×N grid nodes, and a dynamic adjacency matrix is ​​generated based on spatial distance and data correlation. A graph attention network is then used to aggregate features of adjacent three-layer nodes to capture local spatial correlations, and a bidirectional LSTM is used to extract time series features.

[0027] Step S1-3. Multi-source contribution weighting: The semantic association weights of meteorological, soil, and crop data are calculated through a multi-head attention mechanism, and the allocation rules are determined based on the dynamic weights. The rules are as follows:

[0028] When the meteorological error rate is <5%, a weight of 0.4-0.6 is assigned;

[0029] When the crop data correlation is > 0.8, a weight of 0.3-0.5 is assigned;

[0030] When the soil data weight update frequency is > 1 time / h, a weight of 0.2-0.4 is assigned;

[0031] Step S1-4. Spatiotemporal attention optimization: Strengthen the feature weights of key time steps, attenuate the impact of historical redundant data, identify disaster-sensitive areas, and increase local weights.

[0032] Step S3. Data calculation and analysis: Calculate the risk of agricultural meteorology based on the contribution, and then perform a severity analysis of the risk;

[0033] The risk calculation in step S3 includes the following steps:

[0034] Step S2-1. Establish a multi-hazard coupled assessment model and perform weighted fusion of drought index, accumulated temperature deviation, and rainstorm disaster factor;

[0035] Step S2-2. Generate 1000 sets of meteorological disturbance scenarios using Monte Carlo simulation;

[0036] Step S2-3. Output a five-level risk warning through fuzzy comprehensive evaluation method.

[0037] In another embodiment provided by the present invention, the specific steps of risk severity analysis in step S3 are as follows:

[0038] Step S3-1. Disaster probability classification and spatial location: Based on the risk level distribution map, identify high-risk areas (such as drought-prone areas and low-lying, flood-prone areas). Overlay farmland boundary data to locate affected fields. Combined with historical disaster databases (10-year, 50-year, and 100-year disaster records), assess the historical recurrence period and extremeness of the current risk.

[0039] Step S3-2. Crop loss prediction: Input a crop growth model (such as the WOFOST model) to simulate yield loss rates under different risk scenarios:

[0040] Drought scenario: soil moisture content < 0.15m 3 / m 3 When the crop yield loss rate can reach 25%-40%,

[0041] Flood scenario: When rainfall exceeds 100 mm for three consecutive days, the crop yield loss rate increases by 50%;

[0042] Step S3-2. Economic Loss and Ecological Impact Assessment: Calculate direct economic losses (e.g., reduced production, increased irrigation costs) and indirect losses (e.g., market fluctuations, insurance claims), assess the risks of secondary disasters such as soil salinization and the spread of pests and diseases, and generate an ecological impact index (ranging from 0 to 1);

[0043] Step S3-3. Expert consultation and verification: Conduct multidisciplinary consultations with meteorological and agricultural departments, use digital twin systems to simulate the evolution of disasters, verify the accuracy of risk models, and then revise warning thresholds and optimize severity classification rules.

[0044] Step S4. Decision generation and verification: Based on the severity analysis results, the AI ​​generates multiple decision plans, inputs the multiple decision plans into the virtual verification model for verification, and selects the best decision plan based on the verification results;

[0045] The specific operation steps of the virtual verification model are as follows:

[0046] (1) First, the yield and soil salinity changes after the decision-making plan is implemented are simulated by coupling the WOFOST crop model with the SWAT hydrological model;

[0047] (2) When the simulated yield reduction rate is >8% or the soil conductivity increase is >0.5dS / m, the marking scheme is high risk;

[0048] In step S4, the Pareto front analysis is used to select the best solution.

[0049] Step S5. Scheme verification and optimization: Apply the decision-making scheme to adjust agricultural planting, collect actual data during the adjustment process, and optimize the decision-making scheme based on the actual data.

[0050] The scheme optimization in step S5 adopts a closed-loop feedback mechanism. When collecting data in step S5, field sensors are deployed to monitor the soil moisture content and canopy temperature data after the decision is executed in real time. When the deviation between the monitoring data and the predicted value is greater than 15%, the Bayesian network is triggered to dynamically adjust the model parameters. When the decision scheme is applied in step S5 for planting, an incremental learning data set is generated every quarter, and the disaster identification feature space is optimized through comparative learning.

[0051] An agricultural meteorological intelligent decision support system that combines multi-source heterogeneous data includes a data acquisition and processing module, a data fusion module, a risk calculation and analysis module, a decision generation and verification module, and a feedback optimization module; the data acquisition and processing module collects agricultural multi-source heterogeneous data through official data interfaces and monitoring equipment, and the data acquisition and processing module also pre-processes the data; the data fusion module fuses multi-source heterogeneous data based on a spatiotemporal graph neural network, and the data fusion module also dynamically weights the contribution of different data sources through an attention mechanism; the risk calculation and analysis module calculates agricultural meteorological risks based on the data fusion results and assesses the severity of the risks; the decision generation and verification module generates multiple decision plans and simulates and verifies the best plan through a virtual verification model; the feedback optimization module executes the decision plan and monitors the execution effect in real time, and dynamically optimizes model parameters through incremental learning;

[0052] The data acquisition and processing module includes an official data interface unit, an equipment acquisition unit, and a preprocessing unit. The official data interface unit is connected to the meteorological bureau's temperature / precipitation / wind speed historical data, the agricultural department's crop growth cycle data, and soil moisture reports. The equipment acquisition unit integrates an unmanned aerial vehicle remote sensing image acquisition module, a ground meteorological sensor array, and an Internet of Things soil probe. The preprocessing unit is configured with a spatiotemporal interpolation algorithm to compensate for missing values, a wavelet transform filter to eliminate sensor noise, and a min-max normalization module to unify data dimensions.

[0053] The decision generation and verification module couples the WOFOST crop model and the SWAT hydrological model to simulate the changes in yield and soil salinity after the decision plan is executed, and also selects the optimal plan based on a multi-objective optimization algorithm.

[0054] The data fusion module includes a spatiotemporal graph neural network construction unit, a transfer learning unit, an attention mechanism unit, a dynamic edge weight calculation unit, a federated learning unit, and a blockchain storage unit. The spatiotemporal graph neural network construction unit is used to deploy a bidirectional spatiotemporal convolution module. The transfer learning unit is used to migrate the meteorological warning model parameters established in the old main production area to the new planting area. The attention mechanism unit is used to synchronously calculate the overlap between the meteorological probability distribution and the crop harvest time window and generate a disaster avoidance priority ranking. The dynamic edge weight calculation unit dynamically adjusts the edge weights between vertices according to a formula. The federated learning unit aggregates the local model gradients of each farm node through a parameter server and performs privacy data localization training. The blockchain storage unit is used to collect the fused data.

[0055] The risk calculation and analysis module includes a multi-hazard coupling assessment unit, a Monte Carlo simulation unit, and a fuzzy comprehensive evaluation unit. The multi-hazard coupling assessment unit is used to weightedly integrate the drought index, accumulated temperature deviation, and rainstorm disaster factor. The Monte Carlo simulation unit is used to generate 1,000 sets of meteorological disturbance scenarios. The fuzzy comprehensive evaluation unit outputs a five-level risk warning based on the results generated by the Monte Carlo simulation unit.

[0056] The feedback optimization module includes a closed-loop feedback unit, an incremental learning unit and a storage unit. The closed-loop feedback unit monitors soil moisture and canopy temperature data in real time by deploying field sensors; the incremental learning unit generates an incremental data set every quarter and optimizes the disaster identification feature space through comparative learning; the incremental learning unit generates an incremental data set every quarter and optimizes the disaster identification feature space through comparative learning.

[0057] In the above technical solution, the agricultural meteorological intelligent decision support method and system combining multi-source heterogeneous data provided by the present invention have the following beneficial effects:

[0058] (1) The present invention integrates multi-dimensional data of meteorology, soil, and crops through a spatiotemporal graph neural network, combines a dynamic adjacency matrix with an attention mechanism, and effectively captures the spatiotemporal correlation characteristics of farmland areas, thereby solving the data island problem of traditional methods and improving the accuracy of disaster identification. In addition, by introducing a transfer learning mechanism and a federated learning framework, the present invention realizes the rapid migration of the main production area model to the new planting area. At the same time, the parameter server encrypts and aggregates the farm node data, completing collaborative modeling while protecting the privacy of agricultural enterprises and avoiding the risk of sensitive data leakage.

[0059] (2) The present invention uses Monte Carlo simulation to generate 1,000 sets of meteorological disturbance scenarios, and combines the fuzzy comprehensive evaluation method to output a five-level risk warning, supporting the quantitative evaluation of the coupling effects of multiple disasters such as drought and flooding. The digital twin system simulates the execution effect of the decision-making plan, and uses Pareto frontier analysis to screen the optimal solution that balances economic benefits and ecological losses, so that the yield loss rate is reduced after the implementation of the irrigation plan. In addition, field sensors are deployed to monitor the execution effect of the decision in real time. When the deviation between the monitoring data and the predicted value is greater than 15%, the Bayesian network is triggered to dynamically adjust the model parameters. The data set is updated every quarter through incremental learning, and the feature space is optimized using comparative learning, so that the warning accuracy of the model in extreme climate scenarios continues to improve.

[0060] (3) The present invention writes the fused multi-dimensional feature vector (with data source hash value and timestamp) into the blockchain to ensure that the decision-making process is traceable and cannot be tampered with. Through the interpretable weight visualization module, the transparency of the decision-making logic is enhanced to help agricultural managers quickly locate the main causes of risks.

[0061] (4) This invention constructs an efficient, reliable and explainable agricultural meteorological intelligent decision support system through multi-source data fusion, spatiotemporal dynamic modeling, privacy-preserving collaborative learning and closed-loop feedback optimization, providing innovative technical solutions for responding to climate change challenges and achieving sustainable agricultural development. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments described in the present invention. For ordinary technicians in this field, other drawings can also be obtained based on these drawings.

[0063] Figure 1 A method flow chart is provided for an embodiment of the agricultural meteorological intelligent decision support method and system combining multi-source heterogeneous data according to the present invention.

[0064] Figure 2 The specific flow chart of step S2 is provided for the agricultural meteorological intelligent decision support method and system embodiment combining multi-source heterogeneous data of the present invention.

[0065] Figure 3 A schematic diagram of the risk calculation process provided by an embodiment of the agricultural meteorological intelligent decision support method and system combining multi-source heterogeneous data of the present invention.

[0066] Figure 4 A schematic diagram of the risk severity analysis process provided by an embodiment of the agricultural meteorological intelligent decision support method and system combining multi-source heterogeneous data of the present invention.

[0067] Figure 5 Schematic diagram of an agricultural meteorological intelligent decision support system provided by an embodiment of the agricultural meteorological intelligent decision support method and system combining multi-source heterogeneous data of the present invention.

[0068] Figure 6 Schematic diagram of data acquisition and processing modules provided by an embodiment of the agricultural meteorological intelligent decision support method and system combining multi-source heterogeneous data of the present invention.

[0069] Figure 7 Schematic diagram of a data fusion module provided by an embodiment of the agricultural meteorological intelligent decision support method and system combining multi-source heterogeneous data of the present invention.

[0070] Figure 8 Schematic diagram of the risk calculation and analysis module provided by the embodiment of the agricultural meteorological intelligent decision support method and system combining multi-source heterogeneous data of the present invention.

[0071] Figure 9A schematic diagram of a feedback optimization module provided by an embodiment of the agricultural meteorological intelligent decision support method and system combining multi-source heterogeneous data of the present invention. DETAILED DESCRIPTION

[0072] In order to enable those skilled in the art to better understand the technical solution of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings.

[0073] like Figure 1-4 As shown, the agricultural meteorological intelligent decision support method combining multi-source heterogeneous data provided by the embodiment of the present invention is characterized by comprising the following steps:

[0074] Step S1. Acquisition and processing of multi-source heterogeneous data: Acquisition of agricultural multi-source heterogeneous data through official and monitoring equipment, and pre-processing of the multi-source heterogeneous data;

[0075] In step S1:

[0076] Official data include historical temperature / precipitation / wind speed data released by the Meteorological Bureau, crop growth cycle data and soil moisture reports provided by the agricultural department;

[0077] Monitoring equipment data is obtained through drone remote sensing images, ground meteorological sensor arrays, and IoT soil probes;

[0078] Preprocessing includes temporal and spatial interpolation compensation for missing values, wavelet transform filtering of sensor noise data, and normalization of data of different dimensions through min-max standardization.

[0079] Step S2. Data extraction and fusion: Extract multi-source heterogeneous data based on correlation, then construct a spatiotemporal graph neural network to fuse the extracted multi-source heterogeneous data based on spatiotemporal correlation, and introduce an attention mechanism to dynamically weight the contribution of different data sources to the decision-making target;

[0080] In step S2, the spatiotemporal graph neural network includes a bidirectional spatiotemporal convolution module. In the spatial dimension, a graph convolutional network is used to model the topological relationship of fields. In the temporal dimension, a gated recurrent unit is used to capture the evolution of meteorological factors. A transfer learning mechanism is introduced into the spatiotemporal graph neural network to transfer the parameters of the meteorological warning model established in the old main production area to the new planting area. The feature offset caused by soil type differences is adjusted through a domain adaptation algorithm.

[0081] The attention mechanism simultaneously calculates the overlap between the meteorological probability distribution and the crop harvest time window to generate a disaster avoidance priority ranking. The attention mechanism includes the following:

[0082] (1) Source domain attention layer: Calculate the feature similarity matrix between satellite remote sensing spectral data and ground meteorological station data,

[0083] (2) Spatiotemporal attention layer: Dynamically capture the spatiotemporal correlation pattern between typhoon paths and crop planting distribution through a multi-head self-attention mechanism.

[0084] (3) An interpretable weight visualization module is set up at the output end to display the cross-influence factors of soil pH and precipitation;

[0085] The following formula is used to dynamically calculate the edge weights between vertices, as shown below:

[0086]

[0087] Among them, W ij is the edge weight between vertices i and j, The square of the Euclidean space distance between vertices i and j, α is the weight coefficient of the spatial distance term α = 0.6, σ is the bandwidth of the Gaussian kernel function σ = 150m, β is the weight coefficient of the feature correlation term β = 0.4, Corr(X i ,X j ) is the feature vector X of nodes i and j i ,X j Pearson correlation coefficient;

[0088] The data fusion process uses a federated learning framework to achieve multi-agent data collaboration, and aggregates the local models of each farm node through a parameter server while protecting the privacy data of agricultural enterprises; the fused multi-dimensional feature vector is stored in the blockchain traceability system, and each feature dimension is accompanied by a data source hash value and timestamp.

[0089] The specific steps of step S2 are as follows:

[0090] Step S1-1. Multi-source data extraction and preprocessing: The spatiotemporal correlation features of meteorological, soil, and crop data are screened based on the Pearson correlation coefficient. A convolutional autoencoder is then used to reduce the dimensionality of remote sensing imagery and sensor data to extract key features.

[0091] It should be noted that meteorological data include temperature (24-hour average), precipitation (6-hour cumulative), wind speed (10-meter altitude), relative humidity (daytime minimum); soil data include electrical conductivity (20 cm depth), volumetric water content (0-40 cm average), soil pH (surface), available nitrogen content (0-20 cm); crop data include NDVI (normalized difference vegetation index), chlorophyll content (SPAD value), canopy cover, and crop growth stage code.

[0092] Step S1-2. Spatiotemporal Graph Neural Network Construction: The farmland area is divided into N×N grid nodes, and a dynamic adjacency matrix is ​​generated based on spatial distance and data correlation. A graph attention network is then used to aggregate features of adjacent three-layer nodes to capture local spatial correlations, and a bidirectional LSTM is used to extract time series features.

[0093] Step S1-3. Multi-source contribution weighting: The semantic association weights of meteorological, soil, and crop data are calculated through a multi-head attention mechanism, and the allocation rules are determined based on the dynamic weights. The rules are as follows:

[0094] When the meteorological error rate is <5%, a weight of 0.4-0.6 is assigned;

[0095] When the crop data correlation is > 0.8, a weight of 0.3-0.5 is assigned;

[0096] When the soil data weight update frequency is > 1 time / h, a weight of 0.2-0.4 is assigned;

[0097] Step S1-4. Spatiotemporal attention optimization: Strengthen the feature weights of key time steps, attenuate the impact of historical redundant data, identify disaster-sensitive areas, and increase local weights.

[0098] Step S3. Data calculation and analysis: Calculate the risk of agricultural meteorology based on the contribution, and then perform a severity analysis of the risk;

[0099] The risk calculation in step S3 includes the following steps:

[0100] Step S2-1. Establish a multi-hazard coupling assessment model and perform weighted fusion of drought index, accumulated temperature deviation, and rainstorm disaster factor;

[0101] Step S2-2. Generate 1000 sets of meteorological disturbance scenarios using Monte Carlo simulation;

[0102] Step S2-3. Output a five-level risk warning through fuzzy comprehensive evaluation method.

[0103] In another embodiment provided by the present invention, the specific steps of risk severity analysis in step S3 are as follows:

[0104] Step S3-1. Disaster probability classification and spatial location: Based on the risk level distribution map, identify high-risk areas (such as drought-prone areas and low-lying, flood-prone areas). Overlay farmland boundary data to locate affected fields. Combined with historical disaster databases (10-year, 50-year, and 100-year disaster records), assess the historical recurrence period and extremeness of the current risk.

[0105] Step S3-2. Crop loss prediction: Input the crop growth model and simulate the yield loss rate under different risk scenarios:

[0106] Drought scenario: soil moisture content < 0.15m 3 / m 3 When the crop yield loss rate can reach 25%-40%,

[0107] Flood scenario: When rainfall exceeds 100 mm for three consecutive days, the crop yield loss rate increases by 50%;

[0108] It should be noted that the crop growth model uses the logistic regression equation to calculate the probability of yield reduction, which is as follows:

[0109] P(y|x)=1 / [1+exp(-0.5·SPEI+1.2·ΔNDVI)]

[0110] Among them, SPEI is the standardized precipitation evapotranspiration index, ΔNDVI is the vegetation index change rate;

[0111] Step S3-2. Economic Loss and Ecological Impact Assessment: Calculate direct economic losses (e.g., reduced production, increased irrigation costs) and indirect losses (e.g., market fluctuations, insurance claims), assess the risks of secondary disasters such as soil salinization and the spread of pests and diseases, and generate an ecological impact index (ranging from 0 to 1);

[0112] Step S3-3. Expert consultation and verification: Conduct multidisciplinary consultations with meteorological and agricultural departments, use digital twin systems to simulate the evolution of disasters, verify the accuracy of risk models, and then revise warning thresholds and optimize severity classification rules.

[0113] Step S4. Decision generation and verification: Based on the severity analysis results, the AI ​​generates multiple decision plans, inputs the multiple decision plans into the virtual verification model for verification, and selects the best decision plan based on the verification results;

[0114] The specific operation steps of the virtual verification model are as follows:

[0115] (1) First, the yield and soil salinity changes after the decision-making plan is implemented are simulated by coupling the WOFOST crop model with the SWAT hydrological model;

[0116] (2) When the simulated yield reduction rate is >8% or the soil conductivity increase is >0.5dS / m, the marking scheme is high risk;

[0117] In step S4, the Pareto frontier analysis is used to select the best solution, and the objective function is:

[0118] max(economic benefit / (0.8·cost+0.2·ecological loss)).

[0119] Step S5. Scheme verification and optimization: Apply the decision-making scheme to adjust agricultural planting, collect actual data during the adjustment process, and optimize the decision-making scheme based on the actual data.

[0120] The scheme optimization in step S5 adopts a closed-loop feedback mechanism. When collecting data in step S5, field sensors are deployed to monitor the soil moisture and canopy temperature data after the decision is executed in real time. When the deviation between the monitored data and the predicted value is greater than 15%, the Bayesian network is triggered to dynamically adjust the model parameters. When the decision scheme is applied in step S5, an incremental learning data set is generated every quarter, and the disaster identification feature space is optimized through comparative learning.

[0121] An agricultural meteorological intelligent decision support system combining multi-source heterogeneous data, such as Figure 5-9 As shown, it includes a data acquisition and processing module, a data fusion module, a risk calculation and analysis module, a decision generation and verification module, and a feedback optimization module; the data acquisition and processing module collects agricultural multi-source heterogeneous data through official data interfaces and monitoring equipment, and the data acquisition and processing module also pre-processes the data; the data fusion module fuses multi-source heterogeneous data based on a spatiotemporal graph neural network, and the data fusion module also dynamically weights the contribution of different data sources through an attention mechanism; the risk calculation and analysis module calculates agricultural meteorological risks based on the data fusion results and assesses the severity of the risks; the decision generation and verification module generates multiple decision plans and simulates and verifies the best plan through a virtual verification model; the feedback optimization module executes the decision plan and monitors the execution effect in real time, and dynamically optimizes the model parameters through incremental learning;

[0122] The data acquisition and processing module includes an official data interface unit, an equipment acquisition unit, and a preprocessing unit. The official data interface unit accesses historical temperature, precipitation, and wind speed data from the Meteorological Bureau, crop growth cycle data from the agricultural sector, and soil moisture reports. The equipment acquisition unit integrates a drone remote sensing image acquisition module, a ground-based meteorological sensor array, and an IoT soil probe. The preprocessing unit is equipped with a spatiotemporal interpolation algorithm to compensate for missing values, a wavelet transform filter to eliminate sensor noise, and a min-max normalization module to unify data dimensions.

[0123] The decision generation and verification module couples the WOFOST crop model and the SWAT hydrological model to simulate the changes in yield and soil salinity after the execution of the decision plan, and also selects the optimal plan based on the multi-objective optimization algorithm.

[0124] The data fusion module includes a spatiotemporal graph neural network construction unit, a transfer learning unit, an attention mechanism unit, a dynamic edge weight calculation unit, a federated learning unit, and a blockchain storage unit. The spatiotemporal graph neural network construction unit is used to deploy a bidirectional spatiotemporal convolution module. The transfer learning unit is used to migrate the meteorological warning model parameters established in the old main production area to the new planting area. The attention mechanism unit is used to synchronously calculate the overlap between the meteorological probability distribution and the crop harvest time window and generate a disaster avoidance priority ranking. The dynamic edge weight calculation unit dynamically adjusts the edge weights between vertices according to the formula. The federated learning unit aggregates the local model gradients of each farm node through the parameter server and conducts private data local training. The blockchain storage unit is used to collect the fused data.

[0125] The risk calculation and analysis module includes a multi-hazard coupling assessment unit, a Monte Carlo simulation unit, and a fuzzy comprehensive evaluation unit. The multi-hazard coupling assessment unit is used to weightedly integrate the drought index, accumulated temperature deviation, and rainstorm disaster factors. The Monte Carlo simulation unit is used to generate 1,000 sets of meteorological disturbance scenarios. The fuzzy comprehensive evaluation unit outputs a five-level risk warning based on the results generated by the Monte Carlo simulation unit.

[0126] The feedback optimization module includes a closed-loop feedback unit, an incremental learning unit and a storage unit. The closed-loop feedback unit monitors soil moisture and canopy temperature data in real time by deploying field sensors; the incremental learning unit generates an incremental data set every quarter and optimizes the disaster identification feature space through comparative learning; the incremental learning unit generates an incremental data set every quarter and optimizes the disaster identification feature space through comparative learning.

[0127] The above description is merely illustrative of certain exemplary embodiments of the present invention. It goes without saying that those skilled in the art will be able to modify the described embodiments in various ways without departing from the spirit and scope of the present invention. Therefore, the above drawings and description are illustrative in nature and should not be construed as limiting the scope of protection of the claims.

Claims

1. An agricultural meteorological intelligent decision support method combining multi-source heterogeneous data is characterized by: The following steps are involved: Step S1. Acquisition and processing of multi-source heterogeneous data: Acquisition of agricultural multi-source heterogeneous data through official and monitoring equipment, and pre-processing of the multi-source heterogeneous data; Step S2. Data extraction and fusion: Extract multi-source heterogeneous data based on correlation, then construct a spatiotemporal graph neural network to fuse the extracted multi-source heterogeneous data based on spatiotemporal correlation, and introduce an attention mechanism to dynamically weight the contribution of different data sources to the decision-making target; Step S3. Data calculation and analysis: Calculate the risk of agricultural meteorology based on the contribution, and then perform a severity analysis of the risk; Step S4. Decision generation and verification: Based on the severity analysis results, the AI ​​generates multiple decision plans, inputs the multiple decision plans into the virtual verification model for verification, and selects the best decision plan based on the verification results; Step S5. Scheme verification and optimization: Apply the decision-making scheme to adjust agricultural planting, collect actual data during the adjustment process, and optimize the decision-making scheme based on the actual data.

2. The agricultural meteorological intelligent decision support method combining multi-source heterogeneous data according to claim 1 is characterized in that: In the step S1: Official data include historical temperature / precipitation / wind speed data released by the Meteorological Bureau, crop growth cycle data and soil moisture reports provided by the agricultural department; Monitoring equipment data is obtained through drone remote sensing images, ground meteorological sensor arrays, and IoT soil probes; Preprocessing includes temporal and spatial interpolation compensation for missing values, wavelet transform filtering of sensor noise data, and normalization of data of different dimensions through min-max standardization.

3. The agricultural meteorological intelligent decision support method combining multi-source heterogeneous data according to claim 1 is characterized in that: In step S2, the spatiotemporal graph neural network includes a bidirectional spatiotemporal convolution module, wherein the spatial dimension uses a graph convolutional network to model the topological relationship of the field plots, and the temporal dimension uses a gated recurrent unit to capture the evolution of meteorological factors. The spatiotemporal graph neural network introduces a transfer learning mechanism to migrate the parameters of the meteorological warning model established in the old main production area to the new planting area, and adjusts the feature offset caused by soil type differences through a domain adaptive algorithm; The attention mechanism simultaneously calculates the overlap between the meteorological probability distribution and the crop harvest time window to generate a disaster avoidance priority ranking. The attention mechanism includes the following: (1) Source domain attention layer: Calculate the feature similarity matrix between satellite remote sensing spectral data and ground meteorological station data, (2) Spatiotemporal attention layer: Dynamically capture the spatiotemporal correlation pattern between typhoon paths and crop planting distribution through a multi-head self-attention mechanism. (3) An interpretable weight visualization module is set up at the output end to display the cross-influence factors of soil pH and precipitation; The dynamic calculation of edge weights between vertices adopts the following formula, which is as follows: Among them, W ij is the edge weight between vertices i and j, The square of the Euclidean space distance between vertices i and j, α is the weight coefficient of the spatial distance term α = 0.6, σ is the bandwidth of the Gaussian kernel function σ = 150m, β is the weight coefficient of the feature correlation term β = 0.4, Corr(X i ,X j ) is the feature vector X of nodes i and j i ,X j Pearson correlation coefficient; The data fusion process adopts a federated learning framework to achieve multi-agent data collaboration, and aggregates the local models of each farm node through a parameter server under the premise of protecting the privacy data of agricultural enterprises; the fused multi-dimensional feature vector is stored in the blockchain traceability system, and each feature dimension is accompanied by a data source hash value and timestamp.

4. The agricultural meteorological intelligent decision support method combining multi-source heterogeneous data according to claim 1 is characterized in that: The specific steps of step S2 are as follows: Step S1-1. Multi-source data extraction and preprocessing: The spatiotemporal correlation features of meteorological, soil, and crop data are screened based on the Pearson correlation coefficient. A convolutional autoencoder is then used to reduce the dimensionality of remote sensing imagery and sensor data to extract key features. Step S1-2. Spatiotemporal Graph Neural Network Construction: The farmland area is divided into N×N grid nodes, and a dynamic adjacency matrix is ​​generated based on spatial distance and data correlation. A graph attention network is then used to aggregate features of adjacent three-layer nodes to capture local spatial correlations, and a bidirectional LSTM is used to extract time series features. Step S1-3. Multi-source contribution weighting: The semantic association weights of meteorological, soil, and crop data are calculated through a multi-head attention mechanism, and the allocation rules are determined based on the dynamic weights. The rules are as follows: When the meteorological error rate is <5%, a weight of 0.4-0.6 is assigned; When the crop data correlation is > 0.8, a weight of 0.3-0.5 is assigned; When the soil data weight update frequency is > 1 time / h, a weight of 0.2-0.4 is assigned; Step S1-4. Spatiotemporal attention optimization: Strengthen the feature weights of key time steps, attenuate the impact of historical redundant data, identify disaster-sensitive areas, and increase local weights.

5. The agricultural meteorological intelligent decision support method combining multi-source heterogeneous data according to claim 1 is characterized in that: The risk calculation in step S3 includes the following steps: Step S2-1. Establish a multi-hazard coupling assessment model and perform weighted fusion of drought index, accumulated temperature deviation, and rainstorm disaster factor; Step S2-2. Generate 1000 sets of meteorological disturbance scenarios using Monte Carlo simulation; Step S2-3. Output a five-level risk warning through fuzzy comprehensive evaluation method.

6. The agricultural meteorological intelligent decision support method combining multi-source heterogeneous data according to claim 1 is characterized in that: The specific steps of risk severity analysis in step S3 are as follows: Step S3-1. Disaster probability classification and spatial positioning: Based on the risk level distribution map, identify high-risk areas, overlay farmland boundary data to locate affected fields, and combine with the historical disaster database to assess the historical recurrence period and extremeness of the current risk; Step S3-2. Crop loss prediction: Input the crop growth model and simulate the yield loss rate under different risk scenarios: Drought scenario: soil moisture content < 0.15m 3 / m 3 When the crop yield loss rate can reach 25%-40%, Flood scenario: When rainfall exceeds 100 mm for three consecutive days, the crop yield loss rate increases by 50%; Step S3-2. Economic Loss and Ecological Impact Assessment: Calculate direct economic losses (e.g., reduced production, increased irrigation costs) and indirect losses (e.g., market fluctuations, insurance claims), assess the risks of secondary disasters such as soil salinization and the spread of pests and diseases, and generate an ecological impact index (ranging from 0 to 1); Step S3-3. Expert consultation and verification: Conduct multidisciplinary consultations with meteorological and agricultural departments, use digital twin systems to simulate the evolution of disasters, verify the accuracy of risk models, and then revise warning thresholds and optimize severity classification rules.

7. The agricultural meteorological intelligent decision support method combining multi-source heterogeneous data according to claim 1 is characterized in that: The virtual verification model in step S4 is an agricultural digital twin system, and the specific operation steps of the virtual verification model are as follows: (1) First, the yield and soil salinity changes after the decision-making plan is implemented are simulated by coupling the WOFOST crop model with the SWAT hydrological model; (2) When the simulated yield reduction rate is >8% or the soil conductivity increase is >0.5dS / m, the marking scheme is high risk; In step S4, the Pareto front analysis is used to select the best solution.

8. The agricultural meteorological intelligent decision support method combining multi-source heterogeneous data according to claim 1 is characterized in that: The scheme optimization in step S5 adopts a closed-loop feedback mechanism. When collecting data in step S5, field sensors are deployed to monitor the soil moisture content and canopy temperature data after the decision is executed in real time. When the deviation between the monitoring data and the predicted value is greater than 15%, the Bayesian network is triggered to dynamically adjust the model parameters. When the decision scheme is applied in step S5 for planting, an incremental learning data set is generated every quarter, and the disaster identification feature space is optimized through comparative learning.

9. The agricultural meteorological intelligent decision support system combining multi-source heterogeneous data is characterized by: It includes a data acquisition and processing module, a data fusion module, a risk calculation and analysis module, a decision generation and verification module, and a feedback optimization module; the data acquisition and processing module collects agricultural multi-source heterogeneous data through official data interfaces and monitoring equipment, and the data acquisition and processing module also pre-processes the data; the data fusion module fuses multi-source heterogeneous data based on a spatiotemporal graph neural network, and the data fusion module also dynamically weights the contribution of different data sources through an attention mechanism; the risk calculation and analysis module calculates agricultural meteorological risks based on the data fusion results and assesses the severity of the risks; the decision generation and verification module generates multiple decision plans and simulates and verifies them through a virtual verification model to select the best plan; the feedback optimization module executes the decision plan and monitors the execution effect in real time, and dynamically optimizes the model parameters through incremental learning; The data acquisition and processing module includes an official data interface unit, an equipment acquisition unit, and a preprocessing unit. The official data interface unit is connected to the meteorological bureau's temperature / precipitation / wind speed historical data, the agricultural department's crop growth cycle data, and soil moisture reports. The equipment acquisition unit integrates an unmanned aerial vehicle remote sensing image acquisition module, a ground meteorological sensor array, and an Internet of Things soil probe. The preprocessing unit is configured with a spatiotemporal interpolation algorithm to compensate for missing values, a wavelet transform filter to eliminate sensor noise, and a min-max normalization module to unify data dimensions. The decision generation and verification module couples the WOFOST crop model and the SWAT hydrological model to simulate the changes in yield and soil salinity after the decision plan is executed, and also selects the optimal plan based on a multi-objective optimization algorithm.

10. The agricultural meteorological intelligent decision support system combining multi-source heterogeneous data according to claim 9, characterized in that: The data fusion module includes a spatiotemporal graph neural network construction unit, a transfer learning unit, an attention mechanism unit, a dynamic edge weight calculation unit, a federated learning unit, and a blockchain storage unit. The spatiotemporal graph neural network construction unit is used to deploy a bidirectional spatiotemporal convolution module. The transfer learning unit is used to migrate the meteorological warning model parameters established in the old main production area to the new planting area. The attention mechanism unit is used to synchronously calculate the overlap between the meteorological probability distribution and the crop harvest time window and generate a disaster avoidance priority ranking. The dynamic edge weight calculation unit dynamically adjusts the edge weights between vertices according to a formula. The federated learning unit aggregates the local model gradients of each farm node through a parameter server and performs privacy data localization training. The blockchain storage unit is used to collect the fused data. The risk calculation and analysis module includes a multi-hazard coupling assessment unit, a Monte Carlo simulation unit, and a fuzzy comprehensive evaluation unit. The multi-hazard coupling assessment unit is used to weightedly integrate the drought index, accumulated temperature deviation, and rainstorm disaster factor. The Monte Carlo simulation unit is used to generate 1,000 sets of meteorological disturbance scenarios. The fuzzy comprehensive evaluation unit outputs a five-level risk warning based on the results generated by the Monte Carlo simulation unit. The feedback optimization module includes a closed-loop feedback unit, an incremental learning unit and a storage unit. The closed-loop feedback unit monitors soil moisture and canopy temperature data in real time by deploying field sensors; the incremental learning unit generates an incremental data set every quarter and optimizes the disaster identification feature space through comparative learning; the incremental learning unit generates an incremental data set every quarter and optimizes the disaster identification feature space through comparative learning.

Images