Crop yield prediction method and system based on large language and data-mechanism

Abstract

The application discloses a crop yield prediction method and system based on large language and data-mechanism, relates to the technical field of crop yield estimation, and comprises the following steps: S1, obtaining multi-source data of a target plot, performing semantic analysis, extracting environmental feature entities, and mapping to generate a localized parameter set; S2, simulating a crop growth process, outputting a prior state set and a predicted yield trend value; S3, updating a crop growth state variable through an asynchronous assimilation strategy; S4, executing a two-level fusion strategy, evaluating the climate anomaly degree of the current environment, dynamically adjusting the fusion weight of the mechanism model, and calculating a final crop yield prediction value; the crop yield prediction method and system based on large language and data-mechanism provided by the application solve the problems of crop sawtooth effect caused by traditional single-variable assimilation and poor generalization ability of a single model under extreme climate, and realize high-precision and high-robustness crop yield collaborative prediction.

Description

Technical Field

[0001] This invention relates to the field of crop yield estimation technology, and in particular to a crop yield prediction method and system based on big data and big language mechanisms. Background Technology

[0002] As two major global food crops, wheat and corn require accurate yield forecasting for crucial food trade decisions and refined farm management. With the development of remote sensing technology, the Internet of Things, and artificial intelligence, crop yield estimation techniques have evolved from traditional field sampling and statistics to model-based digital yield estimation.

[0003] Currently, mainstream crop yield prediction methods are mainly divided into two categories: process-based mechanistic model methods and data-based machine learning methods. However, existing technologies still have many limitations in practical applications. First, single-mechanistic models have poor spatial adaptability. Although mechanistic models, represented by WOFOST and DSSAT, have clear biological significance, their operation depends on a large number of precise input parameters. In large-scale practical applications, due to the spatial heterogeneity of soil properties, it is difficult to obtain high-precision localized parameters, resulting in large errors in model simulation results in non-validation areas. Second, purely data-driven models lack mechanistic constraints and generalization ability. Although deep learning-based yield estimation methods can uncover nonlinear patterns in historical data, they are essentially "black box" models, lacking biological interpretability. Furthermore, data-driven models are highly dependent on the distribution of historical training data. When encountering extreme climate events (such as extreme heat or sudden drought), the models often produce serious prediction biases due to insufficient generalization ability, failing to meet the robust prediction requirements under the background of climate change. Third, existing data assimilation techniques suffer from mechanistic disruption. While widely used to combine remote sensing data with mechanistic models, current techniques often only adjust leaf biomass when updating crop status using LAI, neglecting the allometric growth relationships between roots, stems, and leaves. This leads to simulated crops exhibiting "abnormal" states, such as "large leaves and small roots," which violate plant physiology, potentially causing subsequent simulation failures or oscillations. Fourth, the systems have high interaction barriers and lack intelligent decision support. Existing crop yield estimation systems are mostly professional software designed for researchers, complex to operate, and require users with strong agronomic and computer science backgrounds. Ordinary users struggle to interact with the system using natural language and cannot utilize unstructured agricultural descriptions (such as text-based weather warnings and disaster records) to assist model correction. Furthermore, the model outputs are often cold, hard numbers, lacking specific attribution analysis and decision-making based on agronomic knowledge. Summary of the Invention

[0004] The purpose of this invention is to provide a crop yield prediction method and system based on big language and data-mechanism, which solves the problems of crop sawtooth effect caused by traditional univariate assimilation and poor generalization ability of single models under extreme climate, and realizes high-precision and high-robust collaborative prediction of crop yield.

[0005] To achieve the above objectives, this invention provides a crop yield prediction method based on big language and data-mechanism, comprising the following steps: S1. Multi-source data semantic analysis and parameter initialization: Obtain multi-source data of the target plot, perform semantic parsing on the multi-source data through a large language model agent, extract environmental feature entities, and map to generate a local parameter set; S2. Dual-drive parallel simulation: The set of driving mechanism models is used to simulate the crop growth process and output a set of prior states; the data-driven model generates the predicted yield trend value of the crop. S3. Obtain leaf area index (LAI) and soil moisture (SM) observations for the target plot, and assimilate the LAI and SM observations into the mechanistic model set through an asynchronous assimilation strategy to update the crop growth state variables. S4. Implement a two-level fusion strategy, use a large language model agent to assess the climate anomaly of the current environment, dynamically adjust the fusion weights of the mechanism model set and the data-driven model, and calculate the final crop yield prediction.

[0006] This invention also provides a system for crop yield prediction based on big language and data-mechanism, including an external interface module, an LLM agent interaction module, a dual-drive computing engine, an assimilation correction module, and an intelligent fusion decision module.

[0007] The LLM intelligent agent interaction module is used to directly interact with the user and acts as the master node to schedule other modules. It receives unstructured agricultural description text and plot information from the user; performs semantic parsing using the built-in large language model intelligent agent to generate a localized parameter configuration file for the mechanistic model, and distributes the localized parameter configuration file to the dual-drive computing engine; the LLM intelligent agent interaction module sends data acquisition instructions to external interface modules based on the parsed spatiotemporal coordinates; at the end of the yield estimation process, the LLM intelligent agent interaction module receives numerical results from the intelligent fusion decision module, combines them with assimilation logs to generate a natural language decision report, and feeds it back to the user.

[0008] The external interface module is specifically used to perform automated calls and data routing to external heterogeneous databases; it receives data acquisition instructions from the LLM agent interaction module, and calls external databases, such as the ERA5-Land meteorological database, the HWSD soil database, and the remote sensing image database, through network requests; it transmits the acquired and cleaned structured meteorological forcing data and basic soil data to the dual-drive computing engine, and transmits the acquired soil moisture observation values ​​and leaf area index observation values ​​to the assimilation correction module.

[0009] The dual-drive computing engine is used to generate prior simulation values ​​of crop growth status and yield trend prediction values ​​in parallel, driven by localized parameters and meteorological data. The dual-drive computing engine includes two sub-modules: a mechanism model library and a deep learning library.

[0010] The mechanism model library pre-configures crop mechanism models including WOFOST, DSSAT, and AquaCrop; it receives localized parameter configurations generated by the LLM agent interaction module, generates set members using the Monte Carlo method, performs open-loop inference driven by meteorological data, and outputs a unified set of prior state vectors through the state mapping interface, including leaf area index, soil moisture, and biomass of each organ for each mechanism model.

[0011] The deep learning library pre-configures an LSTM model initialized based on regional historical data; it is used to mine crop growth trends from a statistical perspective; the deep learning library receives multi-dimensional historical weather sequences within a sliding window as input tensors, extracts deep temporal features through the network's internal gating mechanism, and directly outputs the predicted yield trend value at the current moment after passing through a fully connected layer and inverse normalization processing; the deep learning library is responsible for calculating the validation error of the predicted value within the past window period, outputting the basic confidence weights of the data-driven model, and transmitting them to the intelligent fusion decision module.

[0012] The assimilation correction module receives the set of prior state vectors output from the mechanistic model library in the dual-drive computing engine, and combines it with remote sensing observation data obtained from the external interface module to execute a bivariate asynchronous assimilation algorithm with allometric growth constraints. The assimilation correction module has built-in ensemble Kalman filter algorithm components and nonlinear observation operators, and dynamically performs state updates based on the availability of daily observation data. When leaf area index updates are involved, the allometric growth constraint subroutine is forcibly triggered, and the biomass of roots and stems is adjusted synchronously according to the dry matter distribution coefficient of the crop development stage, outputting a posterior state vector that eliminates physiological imbalances and a simulated yield value.

[0013] The intelligent fusion decision module is used to aggregate the simulated yield values ​​of the mechanistic models after assimilation correction and the trend prediction values ​​output by the deep learning library, and perform a two-level cascade fusion to output the final yield prediction result. The intelligent fusion decision module uses the Bayesian model averaging (BMA) algorithm to perform internal optimization based on the innovation (residual) covariance of each mechanistic model in the assimilation correction module, and generates the mechanistic set prediction value. The intelligent fusion decision module uses the LLM agent to calculate the climate anomaly index (CAI) of the current growing season and constructs an adaptive sigmoid weight activation function. Based on the adaptive sigmoid weight activation function, the fusion weights of the mechanistic set prediction value and the deep learning trend prediction value are dynamically allocated, and finally the crop yield prediction value and confidence interval are calculated and output.

[0014] Therefore, the crop yield prediction method and system based on big data and big language mechanisms described above have the following beneficial effects: (1) By using the principle of allometric growth, while assimilating the leaf area index, the biomass of the root system and stem is adjusted proportionally according to the crop development stage, which effectively solves the problem of root-to-shoot ratio imbalance caused by traditional assimilation methods, ensures that the mechanism model simulation conforms to the laws of plant physiology, and improves the stability of medium- and long-term yield prediction. (2) Through the dual-drive of data and mechanism and the LLM climate perception architecture, the data-driven model is used to ensure high accuracy in normal years, while in years of extreme high temperature or drought, the agent automatically increases the weight of the mechanism model and uses biological laws to constrain the prediction boundary, thereby achieving robust prediction under the background of climate change. (3) By using a bivariate asynchronous assimilation strategy, the microwave soil moisture data and the optical leaf area index data affected by cloud interference are organically complementary; even in the case of no optical images for a long time, the crop growth trajectory can still be corrected by soil moisture assimilation, thus ensuring the continuity of monitoring throughout the growing season. (4) By using the large language model agent as the semantic parsing hub, the user's unstructured agricultural description can be directly transformed into accurate mechanism model parameters, which solves the problem of difficulty in obtaining localized parameters of traditional models and greatly reduces the application threshold of professional systems.

[0015] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0016] Figure 1 This is an overall flowchart of the crop yield prediction method based on big language and data-mechanism of the present invention; Figure 2 This is a schematic diagram of the bivariate asynchronous assimilation principle in an embodiment of the present invention; Figure 3This is a schematic diagram of the system structure of the crop yield prediction method based on big language and data-mechanism of the present invention. Detailed Implementation

[0017] The following detailed description of embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0018] Please see Figure 1-2 A crop yield prediction method based on big language and data mechanisms includes the following steps: S1. Multi-source data semantic analysis and parameter initialization: Obtain multi-source data of the target plot, perform semantic parsing on the multi-source data through a large language model agent, extract environmental feature entities, and map to generate a local parameter set; S11. Construct a spatiotemporal data acquisition pipeline to convert the target plot input by the user into coordinates. , Longitude Latitude; the spatiotemporal data variable set is obtained by calling the ERA5-Land reanalysis dataset via the RESTful API. The meteorological values ​​of the target plot were calculated by correcting the coordinate deviation between the ERA5 grid and the target plot using bilinear interpolation. For basic soil properties, the soil physical properties of the target plot are retrieved from HWSD v2.0 (Harmonized World Soil Database), resulting in the corresponding physical property vector for the target plot. ; Collection of spatiotemporal data variables , , Indicates the daily maximum temperature. Indicates the daily minimum temperature. Represents solar shortwave radiation. Indicates daily precipitation. Indicates wind speed. Indicates relative humidity; Calculate the meteorological values ​​of the target plot. The expression is: ; In the formula, Indicates the coordinates of the target plot; This represents the coordinates of the ERA5 grid points. This indicates the x-coordinate index of the ERA5 grid point. Indicates the y-coordinate index of the ERA5 grid point; This represents the meteorological values ​​at the ERA5 grid points; Physical attribute vector corresponding to the target plot , , Indicates clay particles, Indicates sand grains, Indicates the powder content (%). Indicates soil bulk density ( ), Indicates organic carbon content (%); S12. The user's natural language description is converted into physical parameter correction coefficients of the mechanistic model through the Large Language Model Agent (LLM Agent). (User input text) For example, the sentence "This land is sandy soil with poor water retention; the wheat suffered severe lodging last year" is used. The large language model agent, based on the Transformer architecture, extracts features and outputs a set of environmental feature entities. The large language model agent searches a pre-defined semantic-parameter mapping rule base based on the set of environmental feature entities to correct basic soil and crop parameters; for the set of environmental feature entities... and The soil saturated hydraulic conductivity is updated using a modified formula for the soil transformation function. and field water holding capacity The localized parameter set is obtained and serialized into a standard JSON format initialization file, simulation_config.json, which is used to drive the mechanism model in subsequent steps. Set of environmental feature entities satisfy: ; In the formula, Indicates soil texture; Indicates sandy land; Indicates soil hydraulic properties; This indicates poor water conductivity; Indicates historical disasters; Indicates that the person has fallen down; The updated expressions for soil saturated hydraulic conductivity and field capacity are as follows: ; ; In the formula, This indicates the saturated hydraulic conductivity of the soil in the target plot; Indicates the field holding capacity of the target plot; This represents the confidence correction factor for hydraulic conductivity output by the large language model agent. This represents the confidence correction factor for the output of the large language model agent regarding field water holding capacity; for the description of "hourglass land", LLM sets... , This means that the hydraulic conductivity increases by 20%, while the water holding capacity decreases by 15%. For sets of environmental feature entities (Lodging), adjust the maximum plant height threshold. and conversion factor This is to simulate the physiological characteristics of weaker crop stems in the target plot; , Indicates the default conversion factor; S13, Initialization of transfer learning for data-driven models; A pre-trained LSTM model based on regional statistical data is provided, and the weight matrix of the LSTM model is set as follows. If the user does not input the historical measured yield of the target plot, the yield of the target plot will be predicted using the weight matrix of the LSTM model; if the user inputs a small number of historical measured yields of the target plot (e.g., the past 3 years), the yield will be predicted. Then directly construct the transfer learning loss function. The LSTM base model is adjusted to obtain an initialized LSTM model as the data-driven model; Transfer learning loss function The expression is: ; In the formula, Indicates the number of samples; Indicates the sample index; This indicates that the sample label represents the actual output. This represents the forward propagation function of the LSTM; Input representing meteorological characteristics of historical years; This indicates the fine-tuning weights to be optimized; Represents the regularization constraint term; This represents the regularization coefficient, with a value of 0.01.

[0019] S2. Dual-drive parallel simulation: The set of driving mechanism models is used to simulate the crop growth process and output a set of prior states; the data-driven model generates the predicted yield trend value of the crop. S21. Construct a set of mechanism models and perform stochastic inference; S211. Construct a set of mechanism models, including the WOFOST model, DSSAT-CERES model and AquaCrop model, and read the baseline values ​​of parameters in the initialization file simulation_config.json; S212. Apply Gaussian perturbation to the key sensitive parameters of each mechanism model in the mechanism model set, so that each mechanism model generates... The set members yielded a total of 150 model instances; the mechanism model was obtained through calculation. The Key sensitive parameters of each set member , ; Indicates the index of the model instance; The expression for calculating the key sensitivity parameters is: ; In the formula, Represented as the first Baseline parameters for each model instance; This indicates that the expression follows a pattern with a mean of 0 and a variance of . The normally distributed random disturbance term; Represents a normal distribution; S213. Perform open-loop simulation: Drive 150 model instances, input meteorological data from the ERA5-Land reanalysis dataset, and run forward at a step size of [missing information]. Through the state mapping interface, the private variables of each mechanism model are mapped to standard state vectors. Output the first The set of prior states of the day ; The expression for the standard state vector is: ; In the formula, Leaf area index (LAI) represents the aboveground morphology of crops. The root zone weighted average soil volumetric water content represents the underground environment for crop growth. Indicates the dry weight of the leaves; Indicates the dry weight of the stem; Indicates the dry matter mass of the root; Indicates the dry matter mass of the stored organ; Indicates transpose; The expression for the set of prior states is: ; In the formula, This represents the prior state vector simulated by the first model instance; This represents the prior state vector simulated by the second model instance; This represents the prior state vector simulated by the 3N(150)th model instance; S22. Extracting temporal features using a data-driven model; S221. Construct the current moment Historical meteorological characteristics input tensor Set the length of the sliding window Slide window to retrieve 30 days The expression for the input tensor of historical meteorological features is: ; In the formula, Indicates the first A multidimensional meteorological vector for the day, including temperature, radiation, precipitation, etc. This represents the previous observable multidimensional meteorological vector; Indicates the next observable multidimensional meteorological vector; S222. Input historical meteorological features into the tensor input data-driven model, extract deep temporal features through forget gate, input gate, and output gate operations, and output the predicted yield trend value at the current moment through a fully connected layer. The expression is: ; ; In the formula, Indicates the current hidden state; This represents a composite nonlinear transformation operator within a long short-term memory network unit; This represents the hidden state vector of the previous time step (i.e., day t-1); This represents the Sigmoid activation function; Indicates the output layer weights; Indicates the output layer bias; This indicates the theoretical upper limit of the target plot's output. S223. Calculate the validation error of the current predicted output trend value over the past window period. The reciprocal of the validation error is used as the basic confidence weight for the data-driven model. .

[0020] S3. Obtain leaf area index and soil moisture observation values ​​of the target plot, and assimilate the leaf area index and soil moisture observation values ​​into the mechanism model set through an asynchronous assimilation strategy to update the crop growth state variables. S31, Based on the current time Data availability defines the dynamic observation vector of the target plot. Including soil moisture observation values Leaf area index observations Soil moisture measurements were obtained from microwave remote sensing data, and leaf area index measurements were obtained from optical remote sensing data. S32. Using a day as a simulated step size, determine the following branch logic: When no observation data for the target plot is obtained: the prior state is used as the posterior state. , Indicates the posterior state. This indicates a priori state, and no state update operation is performed. When only soil moisture observations of the target plot are obtained: Construct a hydraulic conduction observation operator. To update soil moisture, the expression is: ; ; In the formula, Indicates the water conductivity coefficient; This represents the evaporation dissipation weighting coefficient; Indicates the amount of bare soil evaporation; Represents the gain matrix; When soil moisture and leaf area index (LAI) measurements of the target plot are acquired simultaneously: perform a full state update to obtain the preliminarily updated state vector. Then, constraint correction is triggered; the expression for the full state update is: ; The gain matrix is ​​calculated using the prior state covariance matrix and the Kalman filter algorithm. The expression is: ; ; In the formula, Represents the prior state covariance matrix; Indicates the number of members in the set; Indicates the index of the model instance; Indicates the first A model instance in The prior state vector at time t; Represents the average vector of a set; Represents the observation operator matrix; This represents the transpose of the observation operator matrix; Represents the observation noise covariance matrix; S33. Execution constraint correction: Calculate leaf biomass increment. Read the developmental stage DVS calculated by the current mechanistic model; query the preset crop physiological parameter table to obtain the allocation coefficients of photosynthetic products to roots, stems, and leaves at the current stage. , , Based on the principle of allometric growth in plant physiology, the biomass increment of each organ should be proportional to the allocation coefficient. Therefore, the forced correction of root renewal should be calculated. Stem renewal rate Output the root weight of the final constraint state and stem weight ; The expression for calculating the increase in leaf biomass is: ; In the formula, This indicates the a posteriori dry weight status of the blades after the initial update (unconstrained); This indicates that they represent the leaf dry matter weight independently deduced by the mechanistic model based on yesterday's conditions and today's meteorological data; The expressions for calculating the forced root renewal and stem renewal amounts are as follows: ; ; The expressions for calculating the root weight and leaf weight of the final constrained state are as follows: ; ; In the formula, , and Similarly, represent the root weight and stem weight simulated by the mechanistic model, respectively.

[0021] S4. Execute a two-level fusion strategy, evaluate the climate anomaly of the current environment through a large language model agent, dynamically adjust the fusion weight of the mechanism model set and the data-driven model, and calculate the final crop yield prediction value. S41, Perform Level 1 fusion; S411. Extract the innovation sequence of each mechanistic model within the assimilation window period. The expression is: ; In the formula, This represents the actual observation vector obtained by the system; Representation mechanism model The prior state vector; S412. Calculate the likelihood of the theoretical model using the Gaussian likelihood function. The expression is: ; In the formula, Indicates the assimilation window period; express Transpose of; Representation mechanism model exist The prior state error covariance matrix at time t; S413. Normalize the likelihood of the mechanism model to obtain the mechanism model. posterior weights Synthesize the first-level fusion result The system automatically selects the most robust combination of mechanistic models for the current growing season, expressed as: ; ; In the formula, Indicates the first The likelihood of each model; Representation Model The final simulated output value after assimilation correction; S42, Perform Level 2 fusion; S421. Obtain meteorological data for the current growing season through a large language model agent, compare it with the historical climate benchmark of the target plot over the past 30 years, and calculate the comprehensive climate anomaly index. If the comprehensive climate anomaly index is not lower than a set threshold, the fusion weight of the predicted values ​​from the mechanistic model set is increased; conversely, the fusion weight of the predicted values ​​from the data-driven model is increased. The expression is: ; In the formula, Indicates the current accumulated temperature; Indicates current precipitation; Indicates current radiation; This represents the historical average of accumulated temperature. This represents the historical average precipitation value. This represents the historical average of radiation. The standard deviation of accumulated temperature; This represents the standard deviation of precipitation; The standard deviation of radiation; S422. Construct an adaptive Sigmoid weighted activation function with the comprehensive climate anomaly index as the independent variable, and fuse the weights of the computer model. The expression is: ; In the formula, Represents the natural constant; Indicates the inflection point of the abnormal threshold, set This indicates that the weights are reversed when the climate deviates from more than two standard deviations. Indicates the slope coefficient. Control the steepness of the transition; when When the temperature is low, the climate is normal. When the value approaches 0, the system focuses on data-driven approaches; when... When it gets bigger, it's believed that extreme weather will occur. When the value approaches 1, the system switches to a focus on mechanism-driven approaches. S423. Calculate the final crop yield forecast. and prediction variance The confidence interval for the final crop yield forecast is obtained. The expression is: ; ; In the formula, This represents the prediction variance of the data-driven model; Representation mechanism model The final simulated output value after assimilation correction.

[0022] Please see Figure 3 The system for crop yield prediction based on big language and data-mechanism includes an external interface module, an LLM agent interaction module, a dual-drive computing engine, an assimilation correction module, and an intelligent fusion decision module.

[0023] The LLM intelligent agent interaction module is used to directly interact with the user and acts as the master node to schedule other modules. It receives unstructured agricultural description text and plot information from the user; performs semantic parsing using the built-in large language model intelligent agent to generate a localized parameter configuration file for the mechanistic model, and distributes the localized parameter configuration file to the dual-drive computing engine; the LLM intelligent agent interaction module sends data acquisition instructions to external interface modules based on the parsed spatiotemporal coordinates; at the end of the yield estimation process, the LLM intelligent agent interaction module receives numerical results from the intelligent fusion decision module, combines them with assimilation logs to generate a natural language decision report, and feeds it back to the user.

[0024] The external interface module is specifically used to perform automated calls and data routing to external heterogeneous databases; it receives data acquisition instructions from the LLM agent interaction module, and calls external databases, such as the ERA5-Land meteorological database, the HWSD soil database, and the remote sensing image database, through network requests; it transmits the acquired and cleaned structured meteorological forcing data and basic soil data to the dual-drive computing engine, and transmits the acquired soil moisture observation values ​​and leaf area index observation values ​​to the assimilation correction module.

[0025] The dual-drive computing engine is used to generate prior simulation values ​​of crop growth status and yield trend prediction values ​​in parallel, driven by localized parameters and meteorological data. The dual-drive computing engine includes two sub-modules: a mechanism model library and a deep learning library.

[0026] The mechanism model library pre-configures crop mechanism models including WOFOST, DSSAT, and AquaCrop; it receives localized parameter configurations generated by the LLM agent interaction module, generates set members using the Monte Carlo method, performs open-loop inference driven by meteorological data, and outputs a unified set of prior state vectors through the state mapping interface, including leaf area index, soil moisture, and biomass of each organ for each mechanism model.

[0027] The deep learning library pre-configures an LSTM model initialized based on regional historical data; it is used to mine crop growth trends from a statistical perspective; the deep learning library receives multi-dimensional historical weather sequences within a sliding window as input tensors, extracts deep temporal features through the network's internal gating mechanism, and directly outputs the predicted yield trend value at the current moment after passing through a fully connected layer and inverse normalization processing; the deep learning library is responsible for calculating the validation error of the predicted value within the past window period, outputting the basic confidence weights of the data-driven model, and transmitting them to the intelligent fusion decision module.

[0028] The assimilation correction module receives the set of prior state vectors output from the mechanistic model library in the dual-drive computing engine, and combines it with remote sensing observation data obtained from the external interface module to execute a bivariate asynchronous assimilation algorithm with allometric growth constraints. The assimilation correction module has built-in ensemble Kalman filter algorithm components and nonlinear observation operators, and dynamically performs state updates based on the availability of daily observation data. When leaf area index updates are involved, the allometric growth constraint subroutine is forcibly triggered, and the biomass of roots and stems is adjusted synchronously according to the dry matter distribution coefficient of the crop development stage, outputting a posterior state vector that eliminates physiological imbalances and a simulated yield value.

[0029] The intelligent fusion decision module is used to aggregate the simulated yield values ​​of the mechanistic models after assimilation correction and the trend prediction values ​​output by the deep learning library, and perform a two-level cascade fusion to output the final yield prediction result. The intelligent fusion decision module uses the Bayesian model averaging (BMA) algorithm to perform internal optimization based on the innovation (residual) covariance of each mechanistic model in the assimilation correction module, and generates the mechanistic set prediction value. The intelligent fusion decision module uses the LLM agent to calculate the climate anomaly index (CAI) of the current growing season and constructs an adaptive sigmoid weight activation function. Based on the adaptive sigmoid weight activation function, the fusion weights of the mechanistic set prediction value and the deep learning trend prediction value are dynamically allocated, and finally the crop yield prediction value and confidence interval are calculated and output.

[0030] Therefore, this invention adopts the above-mentioned crop yield prediction method and system based on big language and data-mechanism. It realizes semantic parsing and automatic parameter configuration of multi-source agricultural data through big language model, solves the physiological consistency problem of model simulation by using bivariate asynchronous assimilation technology and allometric growth constraint mechanism, and realizes adaptive fusion of mechanism model and data-driven model through climate anomaly perception. Thus, while ensuring the rationality of biological mechanism, it significantly improves the prediction accuracy and robustness of crop yield.

[0031] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A crop yield prediction method based on big language and data-mechanism, characterized in that, Includes the following steps: S1. Multi-source data semantic analysis and parameter initialization: Obtain multi-source data of the target plot, perform semantic parsing on the multi-source data through a large language model agent, extract environmental feature entities, and map to generate a local parameter set; S2. Dual-drive parallel simulation: The set of driving mechanism models is used to simulate the crop growth process and output a set of prior states; the data-driven model generates the predicted yield trend value of the crop. S3. Obtain leaf area index and soil moisture observation values ​​of the target plot, and assimilate the leaf area index and soil moisture observation values ​​into the mechanism model set through an asynchronous assimilation strategy to update the crop growth state variables. S4. Implement a two-level fusion strategy, use a large language model agent to assess the climate anomaly of the current environment, dynamically adjust the fusion weights of the mechanism model set and the data-driven model, and calculate the final crop yield prediction.

2. The crop yield prediction method based on big language and data-mechanism as described in claim 1, characterized in that, In S1: S11. Construct a spatiotemporal data acquisition pipeline to convert the target plot input by the user into coordinates. , Longitude For latitude; call the ERA5-Land reanalysis dataset to obtain the spatiotemporal data variable set. The meteorological values ​​of the target plot were calculated by correcting the coordinate deviation between the ERA5 grid and the target plot using bilinear interpolation. Obtain the soil physical properties of the target plot and obtain the corresponding physical property vector. ; S12. The user's natural language description is converted into physical parameter correction coefficients by a large language model intelligent agent, based on the user's input text. The large language model agent extracts features based on the Transformer architecture and outputs a set of environmental feature entities. For the set of environmental characteristic entities, the soil saturated hydraulic conductivity is updated using a modified formula for the soil transformation function. and field water holding capacity This process obtains the localized parameter set and serializes it into a standard JSON format initialization file, simulation_config.json. S13, Initialization of transfer learning for data-driven models; A base LSTM model is pre-defined, and the weight matrix of the LSTM model is set as follows. ; If the user does not input the historical measured yield of the target plot, the yield of the target plot is predicted using the weight matrix of the LSTM model; if the user inputs the historical measured yield of the target plot... Then directly construct the transfer learning loss function. The LSTM base model is adjusted to obtain an initialized LSTM model, which serves as the data-driven model.

3. The crop yield prediction method based on big language and data-mechanism as described in claim 2, characterized in that, In S12, the expressions for updating soil saturated hydraulic conductivity and field capacity are as follows: ； ； In the formula, This indicates the saturated hydraulic conductivity of the soil in the target plot; Indicates the field holding capacity of the target plot; This represents the confidence correction factor for hydraulic conductivity output by the large language model agent. This represents the confidence correction factor for field water holding capacity output by the large language model agent.

4. The crop yield prediction method based on big language and data-mechanism as described in claim 3, characterized in that, In S2: S21. Construct a set of mechanism models and perform stochastic inference; S211. Construct a set of mechanism models, including the WOFOST model, DSSAT-CERES model and AquaCrop model, and read the baseline values ​​of parameters in the initialization file simulation_config.json; S212. Apply Gaussian perturbation to the key sensitive parameters of each mechanism model in the mechanism model set, so that each mechanism model generates... The set members were obtained in total. One model instance; the mechanism model is obtained through calculation. The Key sensitive parameters of each set member , , Indicates the index of the model instance; S213, Perform open-loop simulation: drive For each model instance, input meteorological data from the ERA5-Land reanalysis dataset, and run forward at a step size of [missing value]. Through the state mapping interface, the private variables of each mechanism model are mapped to standard state vectors. Output the first The set of prior states of the day ; The expression for the standard state vector is: ； In the formula, Leaf area index (LAI) represents the aboveground morphology of crops. The root zone weighted average soil volumetric water content represents the underground environment for crop growth. Indicates the dry weight of the leaves; Indicates the dry weight of the stem; Indicates the dry matter mass of the root; Indicates the dry matter mass of the stored organ; Indicates transpose; S22. Extracting temporal features using a data-driven model; S221. Construct the current moment Historical meteorological characteristics input tensor Set the length of the sliding window ; S222. Input historical meteorological features into the tensor input data-driven model, extract deep temporal features through forget gate, input gate, and output gate operations, and output the predicted yield trend value at the current moment through a fully connected layer. ; S223. Calculate the validation error of the current predicted output trend value over the past window period. The reciprocal of the validation error is used as the basic confidence weight for the data-driven model. .

5. The crop yield prediction method based on big language and data-mechanism as described in claim 4, characterized in that, In S3: S31. Define the dynamic observation vector of the target plot. Including soil moisture observation values Leaf area index observations ; S32. Using a day as a simulated step size, determine the following branch logic: When no observation data for the target plot is obtained: the prior state is used as the posterior state. , Indicates the posterior state. This indicates a priori state, and no state update operation is performed. When only soil moisture observations of the target plot are obtained: Construct a hydraulic conduction observation operator. To update soil moisture, the expression is: ； ； In the formula, Indicates the water conductivity coefficient; This represents the evaporation dissipation weighting coefficient; Indicates the amount of bare soil evaporation; Represents the gain matrix; When soil moisture and leaf area index (LAI) measurements of the target plot are acquired simultaneously: perform a full state update to obtain the preliminarily updated state vector. Then, constraint correction is triggered; the expression for the full state update is: ； S33. Execution constraint correction: Calculate leaf biomass increment. Read the developmental stage DVS calculated by the current mechanistic model; query the preset crop physiological parameter table to obtain the allocation coefficients of photosynthetic products to roots, stems, and leaves at the current stage. , , ; Calculate the amount of forced root system updates Stem renewal rate Output the root weight of the final constraint state and stem weight .

6. The crop yield prediction method based on big language and data-mechanism as described in claim 5, characterized in that, In S33: The expressions for calculating the forced root renewal and stem renewal amounts are as follows: ； ； The expressions for calculating the root weight and leaf weight of the final constrained state are as follows: ； ； In the formula, Indicates the original root weight; Indicates the original stem weight.

7. The crop yield prediction method based on big language and data-mechanism as described in claim 6, characterized in that, In S4: S41, Perform Level 1 fusion; S411. Extract the innovation sequence of each mechanistic model within the assimilation window period. ; S412, Computational theoretical model using Gaussian likelihood function likelihood ; S413. Normalize the likelihood of the mechanism model to obtain the mechanism model. posterior weights Synthesize the first-level fusion result It automatically selects the combination of mechanistic models that performs most robustly in the current growing season; S42, Perform Level 2 fusion; S421. Obtain meteorological data for the current growing season through a large language model agent, compare it with the historical climate benchmark of the target plot, and calculate the comprehensive climate anomaly index. ; S422. Construct an adaptive Sigmoid weighted activation function with the comprehensive climate anomaly index as the independent variable, and fuse the weights of the computer model. ; S423. Calculate the final crop yield forecast. and prediction variance The confidence interval for the final crop yield forecast is obtained. .

8. The crop yield prediction method based on big language and data-mechanism as described in claim 7, characterized in that, In S422, the fusion weights of the computer model The expression is: ； In the formula, Represents the natural constant; Indicates the inflection point of the abnormal threshold; This represents the slope coefficient.

9. The crop yield prediction method based on big language and data-mechanism as described in claim 8, characterized in that, In S423, the final crop yield forecast is calculated. and prediction variance The expression is: ； ； In the formula, This represents the prediction variance of the data-driven model; Representation mechanism model The final simulated output value after assimilation correction.

10. A system for crop yield prediction based on big language and data-mechanism as described in any one of claims 1-9, characterized in that, include: The LLM intelligent agent interaction module is used to interact directly with the user and acts as the master node to schedule other modules; The LLM agent interaction module is responsible for receiving unstructured agricultural description text and plot information input by the user; performing semantic parsing through the built-in large language model agent to generate a localized parameter configuration file for the mechanism model, and distributing the localized parameter configuration file to the dual-drive computing engine; the LLM agent interaction module sends data acquisition instructions to the external interface module based on the parsed spatiotemporal coordinates; the LLM agent interaction module receives numerical results from the intelligent fusion decision module, combines them with assimilation logs to generate a natural language decision report and feeds it back to the user; The external interface module is specifically designed for automated calls and data routing to external heterogeneous databases. It receives data acquisition instructions from the LLM agent interaction module and calls an external database via network request; it transmits the acquired and cleaned structured meteorological forcing data and basic soil data to the dual-drive computing engine, and transmits the acquired soil moisture observation values ​​and leaf area index observation values ​​to the assimilation correction module. The dual-drive computing engine is used to generate prior simulation values ​​of crop growth status and yield trend prediction values ​​in parallel, driven by localized parameters and meteorological data. The dual-drive computing engine includes two sub-modules: a mechanism model library and a deep learning library. Mechanism model library, pre-set mechanism models, are used to receive local parameter configurations generated by the LLM agent interaction module, generate set members using the Monte Carlo method, perform open-loop inference driven by meteorological data, and output the prior state set through the state mapping interface; A deep learning library with a pre-initialized LSTM model; it receives a multi-dimensional historical weather sequence within a sliding window as an input tensor, extracts temporal features through a gating mechanism within the network, and outputs the predicted yield trend value at the current moment after passing through a fully connected layer and inverse normalization. Calculate the validation error of the predicted value within the past window period, output the basic confidence weight of the data-driven model, and transmit it to the intelligent fusion decision module; The assimilation correction module receives the prior state set output from the mechanism model library in the dual-drive computing engine, combines it with remote sensing observation data obtained from the external interface module, and executes an asynchronous assimilation algorithm. The assimilation correction module has built-in ensemble Kalman filter algorithm components and nonlinear observation operators, and dynamically performs state updates based on the availability of daily observation data. When leaf area index updates are involved, the allometric growth constraint subroutine is triggered, and the biomass of roots and stems is adjusted synchronously according to the dry matter distribution coefficient of the crop development stage, and the posterior state vector and simulated yield value are output. The intelligent fusion decision module is used to combine the assimilated and corrected mechanistic model output simulation values ​​with the trend prediction values ​​output by the deep learning library, perform two-level cascade fusion, and output the final output prediction result. The intelligent fusion decision-making module uses the Bayesian model averaging algorithm to perform internal optimization based on the information covariance of each mechanism model in the assimilation correction module, and generates the mechanism set prediction value. The intelligent fusion decision module uses an LLM agent to calculate the climate anomaly index for the current growing season and constructs an adaptive Sigmoid weighted activation function. Based on the adaptive Sigmoid weighted activation function, it dynamically allocates the fusion weights of the mechanism set prediction value and the deep learning trend prediction value, and finally calculates and outputs the crop yield prediction value and confidence interval.

Images