Agricultural problem prediction and decision support system based on large model

Abstract

The invention provides an agricultural problem prediction and decision support system based on a large model, and the system comprises a data collection and preprocessing module which is used for collecting data resources of a plurality of channels, and carrying out the data cleaning, data conversion and data storage through a distributed calculation framework; the intelligent agent large model module comprises a large model for agricultural problem prediction, and the large model is trained by adopting historical data; the policy response module is used for extracting information in the policy file, converting the information into structured data and providing a calling interface; and the decision support module is used for generating specific farming activity suggestions based on the structured data in the policy response module and the prediction result of the agent large model module. The system provided by the invention can dynamically provide suggestions in response to farming activities according to crop growth conditions.

Description

Technical Field

[0001] The present invention relates to the field of agricultural technology, and particularly relates to an agricultural problem prediction and decision support system based on a large model. Background Art

[0002] Traditional agricultural production relies on farmers' empirical judgments for activities such as irrigation, fertilization, and pest control. This method is inefficient and prone to resource waste. Although there are currently various technologies applied to agricultural risk management, such as satellite remote sensing monitoring and ground sensor networks, most of these methods focus on static analysis of specific problems and lack the ability of dynamic adaptation and cross-domain information fusion. In addition, existing technologies often ignore the direct impact of policy factors on agricultural production, making it difficult for farmers to adjust their strategies in a timely manner to cope with market and regulatory changes. Existing agricultural AI solutions often require a large amount of labeled data and are sensitive to environmental changes, with poor adaptability. Therefore, there is an urgent need for a more intelligent and comprehensive solution that can provide accurate risk warnings and scientific decision support for agricultural production in complex environments. Summary of the Invention

[0003] Aiming at the problems existing in the prior art, an agricultural problem prediction and decision support system based on a large model is provided, which uses an intelligent agent large model to predict agricultural problems and provides corresponding suggested measures.

[0004] The technical solution adopted by the present invention is as follows: An agricultural problem prediction and decision support system based on a large model, comprising:

[0005] A data collection and preprocessing module, which is used to collect data resources from multiple channels and perform data cleaning, data transformation, and data storage through a distributed computing framework;

[0006] An intelligent agent large model module, including a large model for predicting agricultural problems, and the large model is trained using historical data;

[0007] A policy response module, which is used to extract information from policy documents and convert it into structured data, and provide a call interface;

[0008] A decision support module, which generates specific farming activity suggestions based on the structured data in the policy response module and the prediction results of the intelligent agent large model module.

[0009] As a preferred solution, the data resources include satellite images, meteorological data, soil data, and crop growth cycle records.

[0010] As a preferred solution, in the data collection and preprocessing module, data cleaning includes meteorological data and soil data cleaning, including:

[0011] Meteorological data: Use interpolation methods to fill in missing meteorological data, or use meteorological data from neighboring stations as a substitute; Use statistical methods or machine learning algorithms to identify and correct extreme outliers; Adjust all meteorological data to the unified local standard time;

[0012] Soil data: Reduce random errors by calculating the average value through multiple repetitions; Normalize soil data at different depths to the same reference depth.

[0013] As a preferred solution, in the data acquisition and preprocessing module, data conversion includes meteorological data conversion, soil data conversion, and comprehensive conversion, including:

[0014] Meteorological data: Construct indicators that have a direct impact on crop production based on the cleaned meteorological data; Interpolate and expand meteorological data in the spatial dimension and perform sliding window averaging on meteorological data in the time dimension;

[0015] Soil data: Convert the chemical property information in the soil into numerical features and set threshold ranges according to crop requirements; Convert the soil physical property information into parameters that affect water retention capacity and aeration;

[0016] Comprehensive conversion: Create interaction features under the common conditions of meteorological data and soil data.

[0017] As a preferred solution, in the intelligent agent large model module, the large model includes:

[0018] Data integration unit, used to obtain the data output by the data acquisition and preprocessing module and convert it into a form for model processing;

[0019] Feature enhancement unit, using a pre-trained model to extract richer feature representations; Adopting a custom position encoding mechanism to capture spatio-temporal dynamic changes;

[0020] Multi-modal attention mechanism unit, fusing multi-modal data through a custom attention mechanism;

[0021] Deep learning model unit, including each layer composed of a multi-head self-attention mechanism and a feed-forward neural network, and equipped with residual connections and layer normalization; At the same time, according to the task requirements, it also includes a classifier, a regressor, or other specific task heads;

[0022] Long short-term memory unit, used to assist the model in learning information with a long time span.

[0023] As a preferred solution, the policy response module includes:

[0024] Data collection and preprocessing unit, used to collect policy documents issued by government departments and perform text cleaning;

[0025] A natural language understanding unit for tokenizing and annotating text, identifying key entities in the text, and extracting relationships between different entities;

[0026] A text classification and clustering unit for classifying policy document texts into different categories according to content and automatically discovering the hidden topic distribution in the documents using algorithms;

[0027] A structured transformation unit for extracting corresponding information based on a predefined template, organizing the extracted information into knowledge graph information, and completing the structured processing.

[0028] As a preferred solution, the decision support module includes:

[0029] A risk assessment module for assessing the risk of the prediction results according to the set criteria;

[0030] A recommendation generation module for generating corresponding recommended measures according to the risk assessment results.

[0031] As a preferred solution, the specific working process of the risk assessment module is as follows:

[0032] According to the future weather conditions predicted by the agent large model module, if it is found that there is a high probability of severe weather in certain areas, it is marked as a meteorological disaster risk;

[0033] According to the results of possible pests and diseases predicted by the agent large model module within a specific time period, if conditions suitable for the reproduction of pests and diseases are detected, the area is marked as a pests and diseases risk area;

[0034] According to the changing trend of the key soil indicators predicted by the agent large model module, if it is detected that the nutrients of a certain piece of land are unbalanced or acidification problems occur, the corresponding soil health risk is prompted;

[0035] According to the predicted future trend of the agricultural product market by the agent large model module, and prompt the market economy risk;

[0036] According to the impact predicted by the agent large model module after the application of new agricultural technologies, if there are negative impacts, the technology application risk is prompted.

[0037] As a preferred solution, the specific working process of the recommendation generation module is as follows:

[0038] For meteorological disaster risks: It is recommended to take preventive measures, including early irrigation to relieve drought, improving the drainage system to prevent floods, and planting drought-tolerant or flood-resistant crops;

[0039] For pest and disease risks: It is recommended to adopt biological control methods or reasonably use chemical pesticides for prevention and control; and it is recommended to implement a crop rotation system, intercropping and interplanting techniques, and eco-friendly farming methods;

[0040] For soil health risks: For different soil health problems, specific improvement plans are provided, including adding lime to adjust the pH value, applying organic fertilizers to improve soil fertility, or adopting precision fertilization techniques to ensure the effective supply of nutrients;

[0041] For market economy risks: According to market demand forecasts, product types with high demand and low supply are recommended; at the same time, participating in agricultural insurance programs is encouraged to disperse economic losses caused by price fluctuations;

[0042] For technology application risks: Evaluate the feasibility of new technologies and provide users with detailed implementation guidelines and strategies to address possible problems;

[0043] Combine the prediction results of the large language model of the intelligent agent with the data output by the policy response module to determine whether there are risks affecting production or generate farming suggestions.

[0044] As an optimal solution, it also includes a feature automatic selection module for dynamically adjusting the input feature set according to different crop types and their growth stages. The specific prediction steps include:

[0045] Obtain the initial feature set;

[0046] For each crop and its growth stage, initially screen out a set of candidate features based on historical data and expert knowledge;

[0047] And further narrow down the range of candidate features using feature importance evaluation methods;

[0048] Construct multiple feature combinations and use the model for testing, and select the feature combination with the best performance;

[0049] Continuously adjust and optimize the feature selection process according to the actual application effect to generate an optimized feature set for specific crops and their growth stages; this feature set is provided to the large language model of the intelligent agent for risk prediction.

[0050] Compared with the prior art, the beneficial effects of adopting the above technical solutions are:

[0051] (1) Improve the efficiency and accuracy of agricultural production: By integrating data resources from multiple channels such as satellite images, meteorological data, soil data, and crop growth cycle records, and using the large language model of the intelligent agent for in-depth analysis and prediction, the system can provide more accurate farming activity suggestions, thereby improving agricultural production efficiency and crop yields.

[0052] (2)Enhance the adaptability to environmental changes: The system proposed in the present invention not only considers the impacts of natural factors (such as weather and soil conditions), but also incorporates a policy response module, which can dynamically adjust strategies according to the latest agricultural policies, enhancing farmers' ability to cope with market and regulatory changes.

[0053] (3)Optimize resource allocation and reduce waste: Through comprehensive assessments of various aspects such as meteorological disasters, pest and disease risks, and soil health conditions, the system can guide farmers to take the most appropriate preventive and improvement measures, avoid unnecessary resource inputs, and reduce production costs.

[0054] (4)Support the development of sustainable agriculture: The system encourages the implementation of ecological-friendly farming methods such as the rotation system and intercropping techniques, which helps to protect the ecological environment and promote the sustainable development of agriculture.

[0055] (5)Improve the intelligence level of the decision-making support system: The feature automatic selection module is used to dynamically adjust the input feature set according to different crop types and their growth stages, ensuring that the model focuses on the most influential variables, further improving the intelligence level and adaptability of the system.

[0056] (6)Strengthen risk management and economic security: By analyzing factors such as fluctuations in agricultural product market prices and changes in supply and demand relationships, the system estimates future market trends, helping farmers to adjust planting plans in a timely manner to avoid market economic risks and ensure the economic benefits of agricultural production.

[0057] (7)Simplify the information acquisition process: The policy response module converts complex policy documents into structured data that is easy to understand and operate, providing farmers with a convenient information access path and reducing the understanding difficulty and time cost. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] Figure 1 It is a schematic diagram of the agricultural problem prediction and decision-making support system based on a large model proposed by the present invention.

[0059] Figure 2 It is a schematic diagram of the operation of the intelligent agent large model in the embodiment of the present invention.

[0060] Figure 3 It is a schematic diagram of the operation of the policy response module in the embodiment of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0061] Embodiments of the present application are described in detail below, and examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar modules or modules with the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present application, and cannot be construed as limitations on the present application. On the contrary, the embodiments of the present application include all changes, modifications and equivalents that fall within the spirit and connotation of the appended claims.

[0062] In order to provide accurate risk warning and scientific decision support for agricultural production in a complex environment, the embodiment of the present invention proposes an agricultural problem prediction and decision support system based on a large model. By predicting future agricultural problems and combining policy documents to implement agricultural activity suggestions, it can effectively improve agricultural production efficiency and reduce resource waste.

[0063] Please refer to Figure 1 ,The system mainly includes data collection and preprocessing module, intelligent ,agent model module, policy response module and decision support module. ,The following will explain each module one by one.

[0064] The data collection and preprocessing module is mainly used to collect data resources from multiple channels, and perform data cleaning, data conversion and data storage through a distributed computing framework.

[0065] Specifically, this embodiment integrates data resources from multiple channels, including but not limited to satellite images, weather forecasts, soil moisture sensor readings, crop growth cycle records, etc.

[0066] After the data is collected, data cleaning operations need to be performed. In this embodiment, data cleaning is divided into two parts: meteorological data cleaning and sudden data cleaning.

[0067] (1) Meteorological data cleaning

[0068] ① Missing value processing: Weather stations occasionally lose data due to equipment failure or maintenance. For missing meteorological data, interpolation methods (such as linear interpolation, spline interpolation) can be used to fill in the missing data, or data from nearby stations can be used as a substitute.

[0069] ② Outlier detection and correction: Use statistical methods (such as the 3σ principle) or machine learning algorithms to identify and correct extreme outliers. For example, if the temperature recorded on a certain day deviates significantly from the historical mean for the same period, the validity of the data point needs to be further checked.

[0070] ③ Time consistency correction: Ensure that all meteorological data are recorded according to a unified time standard (such as UTC time) and adjusted to local standard time for analysis.

[0071] (2) Soil data cleaning

[0072] ① Sampling error correction: Sampling errors may occur during the soil sample collection process. The impact of random errors can be reduced by taking the average of multiple repeated samplings.

[0073] ② Depth standardization: Soil data at different depths may affect the performance of the model. Therefore, it is necessary to normalize the data at different depths to the same reference depth or establish a hierarchical model to process the data at each level separately.

[0074] After completing data cleaning, data transformation is also required. In this embodiment, data transformation mainly includes the following aspects:

[0075] (1) Meteorological data transformation

[0076] ① Feature engineering: New features are constructed based on the original meteorological data, such as calculating indicators directly affecting crop growth, like Growing Degree Days (GDD) and effective precipitation.

[0077] ② Spatiotemporal aggregation: Due to the obvious spatiotemporal characteristics of meteorological data, spatial interpolation expansion (such as Kriging interpolation) and sliding window averaging in the time dimension can be performed to generate more stable and reliable input features.

[0078] (2) Soil data transformation

[0079] ① Chemical property quantification: Chemical properties such as organic matter content, pH value, and conductivity in the soil are converted into numerical features, and threshold ranges are set according to crop requirements.

[0080] ② Physical property transformation: Physical properties such as soil texture (e.g., the proportion of sand, silt, and clay particles) and porosity are converted into parameters affecting water holding capacity and aeration through appropriate mathematical models.

[0081] (3) Comprehensive transformation

[0082] ① Cross - feature construction: Interaction features are created by combining meteorological and soil data. For example, the change in nutrient availability in the soil under specific temperature and humidity conditions is crucial for understanding the crop growth environment.

[0083] ② Multi - source data fusion: Data from multiple sources such as satellite remote sensing and ground observation stations are integrated to form a comprehensive agricultural environment monitoring network, improving the accuracy of the prediction model.

[0084] Through the data collection and pre - processing module, meteorological and soil data unique to agriculture can be effectively cleaned and transformed, laying a solid foundation for subsequent modeling work. In addition, in actual operation, strategies should be flexibly adjusted according to specific application scenarios to ensure that the data processing process is both scientific and practical.

[0085] In one embodiment, to reduce R & D costs, the distributed computing framework Apache Spark is adopted to implement data cleaning, transformation, and storage. Utilizing its mature technical advantages can effectively accelerate the system development process. This framework includes: Cluster Manager: Responsible for the allocation and management of the entire cluster's resources, ensuring load balancing among various nodes. Common cluster managers include the Standalone mode provided by Spark, YARN, Mesos, etc. Executor: Each application will be allocated multiple executors in the cluster, and they are responsible for actual data processing tasks. The executors run on each worker node and interact with the driver program through network communication. Driver Program: The Spark application program written by the user is the driver program, which defines the data processing logic and schedules tasks by collaborating with the cluster manager. Resilient Distributed Dataset (RDD): This is the core abstract concept of Spark, representing an immutable and partitionable collection of elements that can be operated on in parallel across different nodes in the cluster. Data Source & Sink: Used to connect to external data storage systems (such as HDFS, databases, etc.) to read raw data or write back the processing results.

[0086] In actual application, its working process includes: Initialization phase: The application program submitted by the user will start the driver program and request resources from the cluster manager. Task allocation: According to the data distribution, the driver program divides the computing tasks into several small tasks and distributes them to the corresponding executors for execution. Data processing: The executors load data from the specified data source, perform operations such as transformation and aggregation on the data according to the predetermined logic, and then cache the intermediate results or directly pass them to the next phase. Result output: The finally processed data will be written to the target location for subsequent analysis or display.

[0087] The agent large model module includes a large model for agricultural problem prediction. This agent large model is trained using historical data and can be called when in use. The agent large model uses the Transformer architecture as a basis and is optimized and adjusted for agricultural application scenarios. When processing crop images, a specially designed spatial attention mechanism is added to capture leaf features; for time series data analysis, a long short-term memory network (LSTM) is introduced to enhance the memory ability for historical data. The agent large model adopts a design based on an improved Transformer architecture, which not only enhances the time series modeling ability but also can effectively capture the characteristics of climate pattern changes. To better analyze time series data in the agricultural field, such as changes in meteorological conditions, LSTM is used to enhance the memory ability for historical data. This means that for data with time dependence, such as temperature, precipitation, etc., LSTM can more effectively capture and utilize this information for prediction. When dealing with data such as crop growth cycles that require considering the influence of environmental factors over a period of time in the past, the addition of LSTM allows the model to remember and learn this information over a long time span, thereby improving the prediction accuracy.

[0088] Please refer to Figure 2 , in this embodiment, the agent large model includes the following parts:

[0089] (1) Data integration unit: Integrate data from different sources, such as meteorological data, soil information, crop growth cycles, etc., and convert them into a form suitable for model processing through an embedding layer.

[0090] (2) Feature enhancement unit: Use a pre-trained model to extract richer feature representations, especially optimized for time series data (such as changes in climate conditions). Implement a custom position encoding mechanism to better capture spatio-temporal dynamic changes.

[0091] (3) Multimodal attention mechanism unit: Introduce multimodal inputs (text, images, numerical data), and effectively fuse these different types of information through a specially designed attention mechanism. Adopt a sparse or linearized attention mechanism to improve computational efficiency while maintaining the effective modeling ability for long-range dependencies.

[0092] (4) Deep learning model unit: Use a deep Transformer structure, each layer consisting of a multi-head self-attention mechanism and a feed-forward neural network, and equipped with residual connections and layer normalization. In one embodiment, according to specific task requirements, it may also include a classifier, a regressor, or other specific task heads.

[0093] (5) Long Short-Term Memory (LSTM) units, which are used to assist the model in learning information over long time spans. In this embodiment, the LSTM is not just an independent component; it is closely integrated with other parts of the intelligent agent large model. For example, in the data integration unit, data from different sources such as meteorological data and soil information are first integrated and transformed into a form suitable for model processing through an embedding layer; then, in the feature enhancement unit, a pre-trained model is used to extract richer feature representations and is optimized for time series data. Here, the LSTM plays a key role, especially in providing support for dealing with long-term dependence problems.

[0094] (6) Output interpretation unit: For agricultural decision support, in addition to the prediction results, interpretability is also required. Therefore, an interpretation module is added to help understand the model's decision-making process.

[0095] In this embodiment, the pre-trained models are models that have been trained on large-scale datasets, and these models can extract high-level and abstract features from the input data. In specific application scenarios (such as agricultural problem prediction), these pre-trained models can be directly used for feature extraction, or fine-tuned on this basis to meet the requirements of specific tasks. For applications in the agricultural field, the pre-trained models can include: ① Convolutional Neural Networks (CNNs) for image analysis: such as ResNet, Inception, EfficientNet, etc., which have been trained to recognize and classify a large amount of image data. In this case, these models can be used to process satellite images or farmland images taken by drones to extract information about crop health status, growth stages, etc. ② Long Short-Term Memory Networks (LSTM) or Transformer variants optimized for time series data: These models are good at capturing patterns in time series data, such as the changing trends of meteorological data. By using pre-trained time series models, it is possible to learn more effectively from historical meteorological data and apply it to future weather predictions or other related analyses. ③ Natural Language Processing (NLP) models: such as BERT, RoBERTa, etc. If the system needs to understand and process text information such as policy documents and market reports, these pre-trained language models can be used. After appropriate adjustment, they can help parse complex document content and extract valuable information.

[0096] In practical applications, it is necessary to train and deploy the large agent model, which mainly includes: (1) Data collection and preprocessing: Collect data from various sensors, public databases, and historical records, and perform cleaning and formatting. (2) Feature engineering: Select appropriate features according to the specific nature of agricultural problems and transform them into a form that the model can understand. (3) Model training: Use historical data to train the model and adjust hyperparameters to optimize performance. (4) Prediction and decision support: Use the trained model to predict future situations. (5) Feedback loop: Update the model based on actual results to continuously improve prediction accuracy and decision-making quality. (6) Deployment and monitoring: Deploy the system to the production environment, monitor its performance in real time, and perform regular maintenance.

[0097] In this embodiment, LSTM plays a crucial role in enhancing the time series processing ability of the large agent model, especially in analyzing and predicting agriculture-related time-dependent data. This not only improves the overall performance of the model but also provides more scientific and accurate support for agricultural production. The large agent model combined with LSTM can generate more accurate risk prediction results and provide decision-making suggestions to farmers accordingly. For example, in the case of an increasing risk of pests and diseases, the system can make more accurate predictions based on historical data and current environmental conditions, and then give effective prevention and control measures.

[0098] The policy response module is used to extract information from policy documents and convert it into structured data, and provide a call interface. In this embodiment, the policy response module is mainly implemented using a natural language processing engine specifically for the agricultural field. Please refer to Figure 3 , specifically including:

[0099] (1) Data collection and preprocessing unit

[0100] ① Data source: First, it is necessary to determine the data sources, including but not limited to announcements, regulatory documents, etc. issued by government departments.

[0101] ② Text cleaning: Clean the original text, remove irrelevant characters, adjust formats, etc., to ensure the basic quality of subsequent analysis.

[0102] (2) Natural Language Understanding (NLU) unit

[0103] ① Word segmentation and annotation: Use word segmentation tools specific to the agricultural field or general ones to segment the text and mark the roles of each word or phrase (such as nouns, verbs, etc.).

[0104] ② Named Entity Recognition (NER): Identify and classify the key entities in the text, such as policy subjects, time, location, types of agricultural products involved, etc.

[0105] ③Relationship extraction: Identify the relationships between different entities in the text, such as "a certain policy stipulates restrictions on the planting area of a certain crop".

[0106] (3) Text classification and clustering unit

[0107] ①Document classification: Classify documents into different categories according to their content, such as subsidy policies, environmental regulations, etc., to facilitate users to quickly find the information they need.

[0108] ②Topic modeling: Automatically discover the hidden topic distribution in the document set through algorithms, which helps to deeply understand the text content.

[0109] (4) Structured conversion unit

[0110] ①Template matching: Extract specific types of information based on pre-defined templates. For example, for subsidy policies, a template can be set to capture subsidy recipients, amounts, application conditions, etc.

[0111] ②Knowledge graph construction: Organize the extracted information into the form of a knowledge graph, making the relationships between data more intuitive and understandable.

[0112] (5) Application layer integration

[0113] ①API service: Provide RESTful API interfaces for other systems to call, facilitating third-party applications to obtain the parsed structured data.

[0114] ②Front-end display: Develop a user-friendly interface that allows users to conveniently access the information they care about through keyword search and other methods.

[0115] (6) Continuous learning and optimization

[0116] ①Feedback mechanism: Establish a user feedback channel to collect the opinions of users to continuously improve the accuracy and practicality of the system.

[0117] ②Model update: Regularly train and update the model, and use the latest corpus and technological progress to improve the processing effect.

[0118] The core of this NLP engine lies in its ability to understand and parse complex natural language texts, especially those containing a large number of professional terms and legal provisions, and then present them to the end-users in an easy-to-understand and operable manner. This not only improves the information retrieval efficiency but also provides strong data support for decision-making.

[0119] For the models used in the policy response module, such as the named entity recognition (NER) model, it also needs to be pre-trained. The corresponding training process is given in this embodiment, including:

[0120] (1) Data Collection

[0121] ① Corpus Construction: First, a large amount of relevant text data needs to be collected as training materials. For the NLP engine in the agricultural field, these materials may include government announcements, laws and regulations, industry reports, etc.

[0122] ② Labeled Data Preparation: For supervised learning, it is usually necessary to manually label some or all of the data. For example, in the named entity recognition task, clearly indicate which words are policy subjects, time, locations, etc.

[0123] (2) Feature Engineering

[0124] ① Feature Selection: Extract useful features from the original text, such as word frequency, part-of-speech tags, syntactic structures, etc.

[0125] ② Vectorized Representation: Convert the text into a form that can be understood by machine learning algorithms. Common methods include the bag-of-words model, TF-IDF, word embeddings, etc.

[0126] (3) Model Selection and Training

[0127] ① Select an Appropriate Model Architecture: Select an appropriate model according to the task requirements. Common ones include logistic regression, support vector machines, deep learning models (such as LSTM, Transformer, etc.).

[0128] ② Hyperparameter Tuning: Adjust various parameters of the model to optimize performance, which is usually done through cross-validation.

[0129] ③ Training Process: Use the labeled dataset to train the selected model. In this process, the model tries to minimize the loss function to learn the mapping relationship between input features and output labels.

[0130] (4) Evaluation and Testing

[0131] ① Evaluation Metrics: Use metrics such as precision, recall, F1-score, etc. to measure the performance of the model.

[0132] ② Test Set Validation: Evaluate the model performance on a test set independent of the training set to ensure its generalization ability.

[0133] (5) Deployment and Iteration

[0134] ① Deployment to Production: Once the model reaches a satisfactory performance level, it can be deployed to the production environment for actual use.

[0135] ② Continuous Monitoring and Update: Regularly check the performance of the model and update and improve it based on new data or feedback information.

[0136] In the agricultural field, considering the professionalism of terms and the diversity of document formats, customized preprocessing steps and postprocessing rules may also be required to improve the model's ability to understand and parse specific types of text.

[0137] The decision support module mainly generates specific farming activity suggestions based on the structured data in the policy response module and the prediction results of the agent large model module. In this embodiment, the decision support module includes: a risk assessment module that conducts a risk assessment on the prediction results according to the set criteria; and a suggestion generation module that generates corresponding suggestion measures based on the risk assessment results. The working processes of these two modules are described one by one below.

[0138] (1) Meteorological disaster risk

[0139] The agent large model analyzes weather forecast data, historical meteorological data, etc. to predict possible natural disasters such as droughts, floods, frosts, etc.

[0140] At this time, the risk assessment module can judge based on the predicted future weather conditions. If it is found that there is a high probability of severe weather in certain areas, it is marked as a meteorological disaster risk.

[0141] (2) Pest and disease risk

[0142] The agent large model combines the crop growth cycle, environmental conditions, and historical pest and disease data to predict the main pests and diseases that may occur during a specific period.

[0143] At this time, the risk assessment module can judge the prediction results. If conditions suitable for the reproduction of pests and diseases are detected (for example, the temperature and humidity reach a certain threshold), the area is marked as a pest and disease risk area.

[0144] (3) Soil health risk

[0145] The agent large model evaluates the changing trends of key indicators such as soil fertility, pH value, and salt content through data collected by soil sensors.

[0146] At this time, the risk assessment module can judge the prediction results. If problems such as nutrient imbalance or acidification are detected in a certain piece of land, the corresponding soil health risk is prompted, and improvement measures are recommended.

[0147] (4) Market economy risk

[0148] The agent large model analyzes factors such as fluctuations in agricultural product market prices and changes in supply and demand relationships to estimate the market trend in the next period of time.

[0149] At this time, the risk assessment module can judge the prediction results. Based on the price decline or increase trend predicted by the economic model, it can prompt market economic risks and help farmers adjust their planting plans in a timely manner to cope with market economic risks.

[0150] (5) Technical application risks

[0151] Examine the application effects of new technologies (such as new pesticides, irrigation technologies) and their possible side effects. Before introducing new agricultural technologies, the system simulates the impacts after their implementation. If there are negative impacts, it will warn of relevant risks.

[0152] For this risk, some available data are mainly used and the judgment is completed using the large agent model. In this embodiment, the available data mainly include:

[0153] Historical data: Use the data records of adopting new technologies in similar environments in the past, including but not limited to crop yields, changes in soil health conditions, incidence of pests and diseases, etc. These data provide basic learning materials for the model, enabling it to identify the impact patterns of different technologies on agricultural production.

[0154] Environmental factors: Considering that the application of new technologies is often affected by local climate conditions (such as temperature, humidity), soil types and their chemical properties, etc., the large agent model will combine multi-source information such as meteorological data and soil analysis results to evaluate the applicability of specific technologies in the local environment.

[0155] Economic model: In addition to direct agricultural outputs, it is also necessary to consider the economic benefits or increased costs brought about by the introduction of new technologies. For example, a new irrigation technology may significantly reduce water resource consumption but have a high initial investment; therefore, it is necessary to predict its long-term return rate and payback period through an economic model to comprehensively evaluate the technical application risks.

[0156] Simulation and experimental results: For some new technologies that have not been widely applied, they can be tested on a small scale first, and the data of the experimental fields are input into the large agent model for simulation. This helps to predict the possible problems and their impact degrees during large-scale promotion.

[0157] Policy responses: Government regulations on aspects such as environmental protection and food safety may also affect the application effects of new technologies. The structured policy information provided by the policy response module can help the system understand the restrictions or incentive measures of current regulations on the adoption of new technologies, so as to more accurately evaluate the technical application risks.

[0158] (6) Policy risks

[0159] A rule-based inference engine that can customize the interpretation of complex agricultural policy texts according to different crop types and regional characteristics. When major risks that may affect production are detected, an alarm is immediately triggered, and personalized guidance plans are pushed to relevant farmers.

[0160] For the risks evaluated by the risk assessment module, the recommendation generation module generates corresponding recommended measures according to the risk assessment results, specifically including:

[0161] (1) Meteorological disaster risks

[0162] Recommendation: For possible meteorological disasters (such as droughts, floods, etc.), the system can recommend that farmers take preventive measures, such as irrigating in advance to alleviate the impact of drought, or improving the drainage system to prevent floods. In addition, it can also recommend drought-tolerant or flood-resistant crop varieties.

[0163] (2) Pest and disease risks

[0164] Recommendation: Once an increase in pest and disease risks is identified, the system should recommend biological control methods or the rational use of chemical pesticides for prevention and control. At the same time, it is recommended to implement ecological-friendly farming methods such as crop rotation systems and intercropping techniques to reduce the occurrence probability of pests and diseases.

[0165] (3) Soil health risks

[0166] Recommendation: For soil health problems (such as acidification, nutrient imbalance), the system will provide specific improvement plans, such as adding lime to adjust the pH value, applying organic fertilizers to improve soil fertility, or using precision fertilization techniques to ensure the effective supply of nutrients.

[0167] (4) Market economy risks

[0168] Recommendation: Facing the risks brought by market economy fluctuations, the system can help farmers make more informed planting choices, such as recommending product types with high demand and low supply according to market demand forecasts; at the same time, encouraging participation in agricultural insurance plans to disperse the economic losses caused by price fluctuations.

[0169] (5) Technology application risks

[0170] Recommendation: When considering the adoption of new technologies, if there are potential risks, the system should evaluate its feasibility in detail and provide users with detailed implementation guidelines and strategies to address possible problems. This includes, but is not limited to, means such as technical training and experimental field testing to reduce risks.

[0171] It should be added that the output results of the policy response module will also be applied to decision-making, not only the early warning of policy risks, but also the corresponding farming suggestions based on encouragement measures, specifically including:

[0172] 1. Adjust the planting plan in combination with agricultural policies

[0173] When the policy response module analyzes the subsidy policies issued by the government to encourage the planting of specific crops, the system can, based on this information and combined with the predictions of the agent large model on factors such as future weather and market demand, recommend to farmers the types of crops that are suitable for planting in the region and eligible for subsidies. For example, if the local government in a certain area launches a support policy for organic vegetable planting, and market analysis shows that the demand for organic vegetables will increase in the next few months, then the system will suggest that local farmers increase the planting area of organic vegetables.

[0174] 2. Optimize the farming method in compliance with environmental protection regulations

[0175] Specific measures: The policy response module can identify new regulations regarding soil protection or water resource management and convert them into specific operation guidelines. For example, if the newly introduced regulations limit the irrigation volume of certain water-intensive crops, then the system will, based on the soil moisture prediction results provided by the agent large model, recommend water-saving technologies such as drip irrigation and also recommend drought-tolerant varieties to meet the requirements of the new regulations.

[0176] 3. Develop emergency response plans to cope with disaster risks

[0177] When the policy response module detects early warning policies for natural disasters that may affect agricultural production (such as floods, droughts, etc.), it will transmit the relevant information to the decision-making support module. Based on this, combined with meteorological disaster predictions, the system can propose more accurate risk assessment and emergency response strategies. For example, when predicting the upcoming rainy season, the system will not only warn of the potential flood risk but also, based on relevant policy guidance, recommend appropriate improvement measures for drainage facilities.

[0178] 4. Guide sales strategies using market trends

[0179] By analyzing the agricultural product price regulation policies issued by government departments and international market dynamics, the policy response module can help determine the best sales timing. Suppose that during a certain period, the government announces temporary export incentive measures for agricultural products. The system will then use the market price prediction model to recommend that farmers seize this opportunity and adjust the product listing time appropriately to maximize profits.

[0180] 5. Promote technological innovation and development

[0181] For application policies involving the promotion of new technologies, such as the use of new fertilizers and precision agriculture technologies, the policy response module can provide detailed implementation guidelines and, based on the results simulated by the agent large model, predict the possible yield changes and economic effects after the introduction of new technologies. This helps farmers make informed technology investment decisions and at the same time reduces the uncertainty risks brought by adopting new technologies.

[0182] In practical applications, an intuitive and easy-to-use mobile application interface can be designed, providing rich visualization tools to help users understand complex analysis results, as well as supporting two-way communication to allow users to provide feedback or request more detailed explanations.

[0183] In one embodiment, to improve the prediction accuracy of the agent large model, a feature automatic selection module is also proposed, which is used to dynamically adjust the input feature set according to different crop types and their growth stages. The working process of this feature automatic selection module is as follows:

[0184] (1) Initialization: Load the pre-trained model and set the initial feature set.

[0185] (2) Feature screening: For each crop and its growth stage, a set of candidate features is initially screened according to historical data and expert knowledge. The feature importance evaluation technology is used to further narrow down the range of candidate features. In this embodiment, the most relevant features are automatically selected according to the requirements of different crop types and their specific growth stages. For example, for some crops at a specific growth period (such as the flowering period or the fruiting period), the changes in factors such as temperature, humidity, and light may need to be particularly concerned. Dynamically adjusting the input feature set ensures that the model can focus on those variables that have the greatest impact on the current crop growth state. After selecting the features, statistical methods or machine learning techniques are also needed to evaluate the importance of each feature. For example, tree-based methods (such as random forests), LASSO regression, or other feature selection methods can be used to determine which features have the greatest impact on the target variables (such as the probability of pest and disease occurrence, yield, etc.). The effectiveness and stability of the selected features are verified through techniques such as cross-validation.

[0186] (3) Feature combination and optimization: Construct multiple feature combinations and use the model to test, and select the feature combination with the best performance.

[0187] Optimization strategies such as genetic algorithms and grid search may be involved to find the optimal feature set.

[0188] (4) Feedback and iteration: Continuously adjust and optimize the feature selection process according to the actual application effect. For example, if it is found that the performance of a certain crop at a specific growth stage is not as expected, improvements can be made by adding new features or adjusting the importance weights of existing features.

[0189] Through the feature automatic selection module proposed in this embodiment, an optimized feature set for a specific crop and its growth stage can be generated. This feature set will be provided as input to the intelligent agent large model for risk prediction and decision support. In this way, the automatic feature selection algorithm not only improves the prediction accuracy of the model, but also enhances the flexibility and adaptability of the system, enabling it to better cope with the complex environmental changes in agricultural production.

[0190] For those of ordinary skill in the art, the specific meanings of the above terms can be understood according to specific circumstances; the accompanying drawings in the embodiments are used to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are some, rather than all, of the embodiments of the present invention. Usually, the components of the embodiments of the present invention described and illustrated herein can be arranged and designed in various different configurations.

[0191] Although the embodiments of the present application have been shown and described above, it can be understood that the above embodiments are exemplary and should not be construed as limiting the present application. Those of ordinary skill in the art can make changes, modifications, substitutions, and variations to the above embodiments within the scope of the present application.

Claims

1. An agricultural problem prediction and decision support system based on large models, characterized in that, Including: A data collection and preprocessing module, which is used to collect data resources from multiple channels and perform data cleaning, data transformation, and data storage through a distributed computing framework; An agent large model module, including a large model for agricultural problem prediction, and the large model is trained using historical data; A policy response module, which is used to extract information from policy documents and convert it into structured data, and provide a call interface; A decision support module, which generates specific farming activity suggestions based on the structured data in the policy response module and the prediction results of the agent large model module.

2. The agricultural problem prediction and decision support system based on a large model according to claim 1, characterized in that, The data resources include satellite images, meteorological data, soil data, and crop growth cycle records.

3. The agricultural problem prediction and decision support system based on a large model according to claim 1 or 2, characterized in that, In the data collection and preprocessing module, data cleaning includes meteorological data and soil data cleaning, including: Meteorological data: Use interpolation methods to fill in missing meteorological data, or use meteorological data from neighboring stations as a substitute; Use statistical methods or machine learning algorithms to identify and correct extreme outliers; Adjust all meteorological data to the unified local standard time; Soil data: Reduce random errors by calculating the average value through multiple repetitions; Normalize soil data at different depths to the same reference depth.

4. The agricultural problem prediction and decision support system based on a large model according to claim 3, characterized in that, In the data collection and preprocessing module, data transformation includes meteorological data transformation, soil data transformation, and comprehensive transformation, including: Meteorological data: Build indicators that have a direct impact on crop production based on the cleaned meteorological data; Perform interpolation expansion on meteorological data in the spatial dimension and perform sliding window averaging on meteorological data in the time dimension; Soil data: Convert the chemical property information in the soil into numerical features and set threshold ranges according to crop requirements; Convert the physical property information of the soil into parameters that affect water retention capacity and air permeability; Comprehensive transformation: Create interaction features under the common conditions of meteorological data and soil data.

5. The agricultural problem prediction and decision support system based on a large model according to claim 1, characterized in that, In the agent large model module, the large model includes: A data integration unit, which is used to obtain the data output by the data collection and preprocessing module and convert it into a form for model processing; A feature enhancement unit, which uses a pre-trained model to extract richer feature representations; Adopt a custom position encoding mechanism to capture spatio-temporal dynamic changes; A multi-modal attention mechanism unit, which fuses multi-modal data through a custom attention mechanism; A deep learning model unit, including each layer composed of a multi-head self-attention mechanism and a feed-forward neural network, and equipped with residual connections and layer normalization; At the same time, according to the task requirements, it also includes a classifier, a regressor, or other specific task heads; A long short-term memory unit, which is used to assist the model in learning information over a long time span.

6. The agricultural problem prediction and decision support system based on a large model according to claim 1, wherein The policy response module includes: A data collection and preprocessing unit, which is used to collect policy documents issued by government departments and perform text cleaning; A natural language understanding unit, which is used to tokenize and annotate the text, identify key entities in the text, and extract the relationships between different entities; A text classification and clustering unit, which is used to classify policy document texts into different categories according to the content and use algorithms to automatically discover the hidden topic distribution in the documents; A structured conversion unit is used to extract corresponding information based on a predefined template, organize the extracted information into knowledge graph information, and complete the structured processing.

7. The agricultural problem prediction and decision support system based on a large model according to claim 1, characterized in that, The decision support module includes: A risk assessment module that conducts a risk assessment on the prediction results according to the set criteria; A recommendation generation module that generates corresponding recommended measures based on the risk assessment results.

8. The agricultural problem prediction and decision support system based on the large model according to claim 7, characterized in that, The specific working process of the risk assessment module is as follows: Based on the future weather conditions predicted by the intelligent agent large model module, if it is found that there is a high probability of severe weather in certain areas, it is marked as a meteorological disaster risk; Based on the possible pest and disease results predicted by the intelligent agent large model module within a specific time period, if conditions suitable for the reproduction of pests and diseases are detected, the area is marked as a pest and disease risk area; Based on the changing trend of the key soil indicators predicted by the intelligent agent large model module, if nutrient imbalance or acidification problems are detected in a certain piece of land, corresponding soil health risks are prompted; Based on the predicted future trend of the agricultural product market by the intelligent agent large model module, and prompt the market economy risk; Based on the predicted impact after the application of new agricultural technologies by the intelligent agent large model module, if there are negative impacts, the technology application risk is prompted.

9. The agricultural problem prediction and decision support system based on a large model according to claim 7 or 8, characterized in that, The specific working process of the recommendation generation module is as follows: For meteorological disaster risks: It is recommended to take preventive measures, including early irrigation to relieve drought, improving the drainage system to prevent floods, and planting drought-tolerant or flood-resistant crops; For pest and disease risks: It is recommended to use biological control methods or reasonably use chemical pesticides for prevention and control; and it is recommended to implement a crop rotation system and intercropping and interplanting technologies, which are ecological and friendly farming methods; For soil health risks: For different soil health problems, specific improvement plans are provided, including adding lime to adjust the pH value, applying organic fertilizers to improve soil fertility, or using precision fertilization techniques to ensure the effective supply of nutrients; For market economy risks: According to the market demand forecast, recommend product types with high demand and low supply; at the same time, encourage participation in agricultural insurance plans to disperse the economic losses caused by price fluctuations; For technology application risks: Evaluate the feasibility of new technologies, and provide users with detailed implementation guidelines and strategies to address possible problems; Combining the prediction results of the intelligent agent large model with the data output by the policy response module, judge whether there are risks affecting production or generate farming suggestions.

10. The agricultural problem prediction and decision support system based on a large model according to claim 1, characterized in that, It also includes a feature automatic selection module, which is used to dynamically adjust the input feature set according to different crop types and their growth stages. The specific prediction process includes: Obtain the initial feature set; For each crop and its growth stage, initially screen out a group of candidate features based on historical data and expert knowledge; And use the feature importance assessment method to further narrow the range of candidate features; Construct multiple feature combinations and use the model for testing, and select the feature combination with the best performance; Continuously adjust and optimize the feature selection process according to the actual application effect, and generate an optimized feature set for a specific crop and its growth stage; this feature set is provided to the intelligent agent large model for risk prediction.

Images