Crop growth simulation platform integration method and crop growth large model construction method

Abstract

The invention provides a crop growth simulation platform integration method and a crop growth large model construction method, and relates to the technical field of agricultural information prediction, and the method comprises the steps: obtaining the crop growth state, crop variety characteristics, climate factors, soil characteristics and field management information of each crop in a whole cycle; and obtaining crop growth variables of the crops in the whole period through semantic analysis. Automatically dividing crop growth variables to obtain a dynamic covariable, a static covariable and a time sequence process variable of crop growth, and performing embedding to obtain embedded vectors corresponding to the dynamic covariable, the static covariable and the time sequence process variable; and performing full-parameter training on the initial large language model according to the embedded vector to obtain a crop growth large model. The method improves the prediction precision of the model, enhances the generalization ability and adaptability of the model, enables the model to better cope with variable environmental conditions and different crop varieties, and improves the use effect of the crop growth model in crop growth state prediction.

Description

Technical Field

[0001] The present invention relates to the technical field of agricultural information prediction, and in particular to a crop growth simulation platform integration method and a crop growth large model construction method. Background Art

[0002] In order to understand the growth and development process of crops, crop growth models are usually regarded as the core research content in the field of agricultural informatics. Crop growth models are designed to dynamically simulate the growth and development process of crops and the interactions between climatic factors, soil properties and field management techniques, and express them as physiological characteristics throughout the life cycle using mathematical equations, providing scientific quantitative tools for crop productivity prediction, early warning and effect evaluation under different conditions.

[0003] In the prior art, due to the fixed internal formulas of crop growth models for different crops, and the common reference of variables and inconsistent dimensions between different crop growth models, it is impossible to directly achieve multi-crop, cross-regional, and pan-variety crop growth prediction. In addition, the long construction cycle, high development cost, and portability of crop growth models also hinder the practical application of crop growth models. Summary of the invention

[0004] The problem solved by the present invention is how to improve the generalization ability and use effect of the crop growth model in crop growth state prediction.

[0005] In order to solve the above problems, the present invention provides a crop growth simulation platform integration method and a crop growth large model construction method.

[0006] In a first aspect, the present invention provides a crop growth simulation platform integration method, comprising:

[0007] Obtain crop growth status, crop variety characteristics, climate factors, soil characteristics and field management information for each crop throughout its entire cycle;

[0008] Through semantic analysis, the crop growth status, the crop variety characteristics, the climate factors, the soil characteristics and the field management information are integrated with the input data and output data of the crop growth model corresponding to the crop in the crop growth simulation platform to obtain the crop growth variables of the crop in the whole cycle.

[0009] Optionally, the crop growth status, the crop variety characteristics, the climate factors, the soil characteristics and the field management information are integrated with the input data and output data of the crop growth model corresponding to the crop in the crop growth simulation platform through semantic analysis to obtain the crop growth variables of the crop in the whole cycle, including:

[0010] By semantic analysis, it is determined whether there are variables pointing to the same meaning between the crop growth state, the crop variety characteristics, the climate factors, the soil characteristics, and the field management information and the input data and the output data corresponding to the crop growth model in the crop growth simulation platform;

[0011] If so, the variables with the same meaning are coreference resolved, and the variables after the coreference resolution are dimensionally aligned, and then the variables after the dimension alignment are integrated with other variables to obtain the crop growth variable.

[0012] The crop growth simulation platform integration method of the present invention ensures that the model can take into account all key factors affecting crop growth by obtaining the growth status, climate factors, soil characteristics and field management information of each crop in the whole cycle. Then, using semantic analysis technology, the above information is integrated with the input data and output data of the crop growth model in the crop growth simulation platform to unify the dimensions of variables with common references, reduce the prediction deviation caused by inconsistent dimensions, thereby improving the accuracy of the prediction, and at the same time, by unifying the dimensions, the transferability of the model in different environments and conditions is enhanced, thereby realizing the integration and unification of cross-crop growth models.

[0013] In a second aspect, the present invention provides a method for constructing a large crop growth model, comprising:

[0014] Acquire crop growth variables of the crop in the whole cycle obtained by the crop growth simulation platform integration method described in any one of the above items;

[0015] Automatically dividing the crop growth variables to obtain dynamic covariates, static covariates and time series process variables of crop growth;

[0016] Embedding the dynamic covariate, the static covariate and the time series process variable to obtain embedding vectors corresponding to the dynamic covariate, the static covariate and the time series process variable respectively;

[0017] The initial large language model is trained according to the embedding vector to obtain a large crop growth model.

[0018] Optionally, the automatically dividing the crop growth variables to obtain dynamic covariates, static covariates and time series process variables of crop growth includes:

[0019] Determining the value of each covariate in the crop growth variable in the digital dictionary according to the crop growth covariate through semantic analysis;

[0020] All the variables are divided into the dynamic covariates, the static covariates and the time series process variables according to the types and time series of the variables.

[0021] Optionally, embedding the dynamic covariate, the static covariate and the time series process variable to obtain embedding vectors corresponding to the dynamic covariate, the static covariate and the time series process variable respectively includes:

[0022] By using an attention mechanism, according to the temporal features corresponding to the dynamic covariates, an attention score corresponding to each of the dynamic covariates is assigned;

[0023] Dynamically embedding is performed according to the attention score of the dynamic covariate, and the dynamic covariate is converted into the embedding vector.

[0024] Optionally, embedding the dynamic covariate, the static covariate and the time series process variable to obtain embedding vectors corresponding to the dynamic covariate, the static covariate and the time series process variable respectively includes:

[0025] According to the time series characteristics corresponding to the static covariates, adjust the weight and bias corresponding to each of the static covariates;

[0026] Through a fully connected network, static embedding is performed according to the weight and the bias corresponding to the static covariate, and the static covariate is converted into the embedding vector.

[0027] Optionally, embedding the dynamic covariate, the static covariate and the time series process variable to obtain embedding vectors corresponding to the dynamic covariate, the static covariate and the time series process variable respectively includes:

[0028] Complementing the time series process variable with zero values ​​to obtain the time series process variable having the same dimension as the dynamic covariate and the static covariate;

[0029] Through the Transformer encoder, the time series process variables with the same dimension are time-series embedded, and the time series process variables are converted into the embedding vector.

[0030] Optionally, performing full parameter training on the initial large language model according to the embedding vector to obtain a large crop growth model includes:

[0031] Concatenate the embedding vectors corresponding to the dynamic covariate, the static covariate and the time series process variable to obtain a fusion vector of the crop in the whole cycle;

[0032] The initial large language model is fully trained according to the fusion vector of the crop in the full cycle to obtain the large model of crop growth.

[0033] Optionally, performing full-parameter training on the initial large language model according to the fusion vector of the crop in the full cycle to obtain the large crop growth model includes:

[0034] Dividing the fused vector of the full period into a first fused vector and a second fused vector according to a preset period ratio, wherein the time period corresponding to the first fused vector is earlier than the time period corresponding to the second fused vector;

[0035] Using the first fused vector as an input parameter of the initial large language model and using the second fused vector as an output parameter of the initial large language model for training;

[0036] The initial large language model after the full parameter training is completed is used as the crop growth large model.

[0037] Optionally, the step of using the initial large language model after all-parameter training as the large crop growth model includes:

[0038] When the initial large language model is undergoing full parameter training, obtaining an error value between the prediction data of the initial large language model and the second fusion vector according to a preset frequency;

[0039] When the error value is less than a preset error threshold, it is determined that the training of the initial large language model is completed.

[0040] The method for constructing a large crop growth model of the present invention obtains the growth variables of the whole cycle of the crop through the crop growth simulation platform integration method. Since the growth variables output by the crop growth simulation platform are unified in dimension through the crop growth simulation platform integration method, it is ensured that data from different sources and types can be compared and integrated at the same scale, thereby eliminating the model prediction error caused by inconsistent dimensions. Such processing not only improves the efficiency of model training, because it reduces the complexity of data preprocessing, but also improves the prediction accuracy of the model, because the model can more accurately learn the relationship between different variables. In addition, the application of the growth variables obtained by the crop growth simulation platform integration method also enhances the generalization ability of the large crop growth model, so that the large crop growth model can be more easily migrated to different crops, regions and environmental conditions, thereby providing accurate growth prediction in a wider range of application scenarios. Then, by automatically dividing these variables into dynamic covariates, static covariates and time series process variables, the model can adopt different processing strategies for different types of variables, thereby more accurately capturing their impact on crop growth. Then, these variables are converted into embedded vectors using embedding technology, which not only retains the information of the original data, but also enhances the model's understanding and processing capabilities of these data. Finally, by training the initial large language model, we obtained a large crop growth model that can comprehensively consider the influence of various variables. While improving the prediction accuracy of the model, it also enhances the generalization ability and adaptability of the model, enabling it to better cope with changing environmental conditions and different crop varieties, and improve the use of crop growth models in crop growth status prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 This is one of the flow charts of the crop growth simulation platform integration method in one embodiment of the present invention;

[0042] Figure 2 This is a second flow chart of a crop growth simulation platform integration method in another embodiment of the present invention;

[0043] Figure 3 This is one of the flow charts of a method for constructing a large crop growth model in another embodiment of the present invention;

[0044] Figure 4 This is the second flow chart of the method for constructing a large crop growth model in another embodiment of the present invention. DETAILED DESCRIPTION

[0045] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific embodiments of the present invention are described in detail below in conjunction with the accompanying drawings. Although certain embodiments of the present invention are shown in the accompanying drawings, it should be understood that the present invention can be implemented in various forms and should not be interpreted as being limited to the embodiments described herein. On the contrary, these embodiments are provided to provide a more thorough and complete understanding of the present invention. It should be understood that the drawings and embodiments of the present invention are only for exemplary purposes and are not intended to limit the scope of protection of the present invention.

[0046] It should be understood that the various steps described in the method embodiments of the present invention may be performed in different orders and / or in parallel. In addition, the method embodiments may include additional steps and / or omit the steps shown. The scope of the present invention is not limited in this respect.

[0047] The term "including" and its variations used in this document are open inclusions, that is, "including but not limited to"; the term "based on" means "based at least in part on"; the term "one embodiment" means "at least one embodiment"; the term "another embodiment" means "at least one other embodiment"; the term "some embodiments" means "at least some embodiments"; the term "optionally" means "optional embodiments". The relevant definitions of other terms will be given in the following description. It should be noted that the concepts of "first", "second", etc. mentioned in the present invention are only used to distinguish different devices, modules or units, and are not used to limit the order or interdependence of the functions performed by these devices, modules or units.

[0048] It should be noted that the modifications of "one" and "plurality" mentioned in the present invention are illustrative rather than restrictive, and those skilled in the art should understand that, unless otherwise clearly indicated in the context, it should be understood as "one or more".

[0049] The names of the messages or information exchanged between multiple devices in the embodiments of the present invention are only used for illustrative purposes, and are not used to limit the scope of these messages or information.

[0050] Among the related technologies, crop growth models, as the core research content in the field of agricultural informatics, aim to simulate the crop growth and development process and its relationship with climate factors, soil characteristics and field management techniques, so as to effectively overcome the temporal and spatial limitations in agricultural production and research. Traditional crop growth models, due to their universal mechanism formulas, can accurately simulate various crop processes after parameter correction. However, there are usually problems such as common reference of variables and inconsistent dimensions between different crop growth models, which restrict the agricultural crop growth simulation platform services across crop growth models and multiple crop varieties. In addition, problems such as long construction cycle, high development cost, and poor portability of crop growth models have always hindered the actual production application of crop growth models.

[0051] In view of the problems existing in the above-mentioned related technologies, this embodiment provides a crop growth simulation platform integration method and a crop growth large model construction method.

[0052] Combination Figure 1 As shown, a crop growth simulation platform integration method provided by an embodiment of the present invention includes:

[0053] Obtain crop growth status, crop variety characteristics, climate factors, soil characteristics and field management information for each crop throughout its entire cycle.

[0054] Specifically, the prediction of the growth status of each crop in the whole cycle requires the crop growth status, crop variety characteristics, climate factors, soil characteristics and field management information. Among them, climate factors include given temperature, light, precipitation, etc., field management information includes the application amount of nitrogen, phosphorus and potassium fertilizers, crop variety characteristics include effective accumulated temperature, minimum germination temperature, etc., soil characteristics include field water holding capacity, saturated water content, saturated hydraulic conductivity, etc., and the crop growth status over a period of time is used to predict the time series of the whole cycle of crop growth.

[0055] Through semantic analysis, the crop growth status, the crop variety characteristics, the climate factors, the soil characteristics and the field management information are integrated with the input data and output data of the crop growth model corresponding to the crop in the crop growth simulation platform to obtain the crop growth variables of the crop in the whole cycle.

[0056] Specifically, semantic analysis is used to determine whether the variables in the input data and output data of the crop growth model in the crop growth simulation platform, crop growth status, crop variety characteristics, climate factors, soil characteristics, and field management information point to the same meaning. Among them, the input data are generally crop growth status, climate factors, soil characteristics, field management information, and crop variety characteristics; the output data is generally a full-process prediction of the crop growth status. If two variables point to the same meaning, deduplication is performed and the two variables are set to the same dimension. Therefore, based on the semantic understanding ability of the large language model, the integration and unification of input and output across crop growth models is achieved. The same crop characteristics, weather factors, soil characteristics, and field management can call different crop models in the big data service at the same time to alleviate the imbalance of the prediction ability of the single crop growth model. In a preferred embodiment of the present invention, weather factors are unified into six variables: daily irrigation amount, daily minimum temperature, daily maximum temperature, daily rainfall (including snow), daily wind speed and steam atmospheric pressure; field management is unified into eight variables: nitrogen, phosphorus, potassium, ammonium fertilizer, nitrate compound fertilizer, potassium oxide compound fertilizer, phosphate compound fertilizer, and phosphorus-potassium compound fertilizer; crop growth status is unified into 83 variables such as growth stage, leaf area index, and total weight of storage organs; crop characteristics and soil characteristics are unified into 63 variables such as field water holding capacity, soil moisture content, and basal temperature for emergence. A total of 160 variables. Next, the default configuration of the simulation platform can be selected through the visualization platform configuration management system, or the default configuration can be fine-tuned to correct the required parameters, including: crop characteristics, climate factors, soil characteristics, and field management techniques. The crop growth simulation platform uses three methods: downloading publicly released crop parameters, obtaining crop parameters through field real collection and correction, and collecting crop parameters on the Internet. It provides characteristic parameters of 48 crops such as soybeans, corn, wheat, and rice, as well as soil characteristic parameters of more than 100 plots. Weather parameters are obtained by calling NASA weather database, and field management technology is obtained through field surveys, real plans of the Academy of Agricultural Sciences, and national standards. Finally, users can use the simulation platform that integrates multiple crop growth models to view and obtain full-cycle predictions of crop growth status in 5 common and different ways, including: web, Android, Apple, mini-program, and direct API calls.

[0057] The crop growth simulation platform integration method of the present invention ensures that the model can take into account all key factors affecting crop growth by obtaining the growth status, climate factors, soil characteristics and field management information of each crop in the whole cycle. Then, using semantic analysis technology, the above information is integrated with the input data and output data of the crop growth model in the crop growth simulation platform to unify the dimensions of variables with common references, reduce the prediction deviation caused by inconsistent dimensions, thereby improving the accuracy of prediction, and at the same time, by unifying the dimensions, the transferability of the model in different environments and conditions is enhanced, thereby realizing the integration and unification of cross-crop growth models.

[0058] Optionally, the crop growth status, the crop variety characteristics, the climate factors, the soil characteristics and the field management information are integrated with the input data and output data of the crop growth model corresponding to the crop in the crop growth simulation platform through semantic analysis to obtain the crop growth variables of the crop in the whole cycle, including:

[0059] By semantic analysis, it is determined whether there are variables pointing to the same meaning between the crop growth state, the crop variety characteristics, the climate factors, the soil characteristics, and the field management information and the input data and the output data corresponding to the crop growth model in the crop growth simulation platform;

[0060] If so, the variables with the same meaning are coreference resolved, and the variables after the coreference resolution are dimensionally aligned, and then the variables after the dimension alignment are integrated with other variables to obtain the crop growth variable.

[0061] Specifically, through advanced semantic analysis technology, we can deeply understand the crop growth status, crop variety characteristics, climate factors, soil characteristics and field management information, and match this information with the input and output data of the crop growth model in the crop growth simulation platform. In this process, semantic analysis technology can identify and judge which of these variables refer to the same meaning, for example, different terms representing "average daily temperature" in different data sources are identified as the same variable. Once these co-referencing variables are identified, co-referencing resolution is performed next, that is, the repeated representation of these variables is eliminated to avoid data redundancy. Subsequently, the dimensions of these co-referencing variables are aligned to ensure that the variables are consistent in numerical scale and unit, such as converting all temperature data to degrees Celsius. Finally, these dimensionally aligned variables are integrated with other variables to form a complete set of crop growth variables, providing comprehensive and consistent input data for the crop growth model. The multi-crop simulation platform integrates the characteristic parameters of 48 crops such as soybeans, corn, wheat, and rice, as well as the soil characteristic parameters of more than 100 plots. These parameters are uniformly processed and integrated through the above steps to support the construction and prediction of crop growth models. In a preferred embodiment of the present invention, Figure 2 As shown in the figure, the coreference problem in the data is identified and resolved by judging the existence of coreference based on a large language model to ensure the consistency and accuracy of the data. Next, the dimension unification and data integration are carried out to integrate the data from network data crawling, NASA database, field surveys, and the data of the real scheme of the Academy of Agricultural Sciences and national standards. Then, the visualization platform is used to configure and manage crop characteristics, soil characteristics, weather factors, and field management. After completing these steps, the integrated crop model can be called to access and visualize data through multiple methods such as API interface, web page, Android, Apple, and WeChat applet, and finally generate full-cycle data on crop growth status. This process embodies an end-to-end solution from data preparation to model application, providing accurate crop growth prediction and decision support.

[0062] In the embodiments of the present invention, by eliminating the problems of co-referencing variables and inconsistent dimensions, the consistency and comparability of data are greatly improved, so that data from different sources and different formats can be used in the same model, thereby expanding the training data set of the model and improving the generalization ability of the model. In addition, the efficiency of data processing is also improved, because the unified data processing process reduces the workload of data preprocessing, making model training more efficient.

[0063] like Figure 3 As shown, the present invention also provides a method for constructing a large crop growth model, comprising:

[0064] The crop growth variables of the crop in the whole cycle are obtained by the crop growth simulation platform integration method as described above.

[0065] Specifically, obtaining crop growth variables throughout the entire crop cycle is achieved through the above-mentioned crop growth simulation platform integration method. This step involves collecting and integrating data on various factors that affect crop growth, including crop growth status, climate factors, soil properties, field management information, and crop variety characteristics. These data are used as input to the model to predict crop growth. Through co-reference resolution and dimensional alignment techniques, it is ensured that data from different sources and types can be compared and analyzed at the same scale, thereby providing high-quality input data for building an accurate crop growth model.

[0066] The crop growth variables are automatically divided to obtain dynamic covariates, static covariates and time series process variables of crop growth.

[0067] Specifically, once comprehensive crop growth variables are obtained, the next step is to automatically divide these variables into dynamic covariates, static covariates, and time series process variables. Dynamic covariates include factors that change over time, such as daily temperature and precipitation; static covariates include factors that are relatively stable during the crop growth cycle, such as soil physical properties and crop genetic characteristics; and time series process variables involve the continuous state of crop growth, such as leaf area index and biomass.

[0068] The dynamic covariate, the static covariate and the time series process variable are embedded to obtain embedding vectors corresponding to the dynamic covariate, the static covariate and the time series process variable respectively.

[0069] Specifically, after the variables are divided, the next step is to embed the dynamic covariates, static covariates, and time series process variables and convert them into embedded vectors. This step uses the attention mechanism and neural network technology to assign a vector representation in a high-dimensional space to each covariate, which can capture the complex relationships and interactions between covariates. This embedding representation enables the model to handle different data types more flexibly and better capture patterns and trends in the data.

[0070] The initial large language model is trained with all parameters according to the embedding vector to obtain a large crop growth model.

[0071] Specifically, finally, these embedded vectors are used to perform full-parameter training on the initial large language model to obtain the final large crop growth model. This training process involves adjusting the parameters of the model to minimize the difference between the predicted value and the actual value. Through full-parameter training, crop growth prediction for multiple crops, across regions, and across varieties is achieved. In addition, the model can also adapt to extreme climatic conditions through fine-tuning, further improving the prediction accuracy and practical application capabilities of the model. In this way, the present invention not only improves the accuracy of crop growth prediction, but also enhances the adaptability of the model to different environmental conditions.

[0072] The method for constructing a large crop growth model of the present invention obtains the growth variables of the crop full cycle through the crop growth simulation platform integration method. Since the growth variables output by the crop growth simulation platform are unified in dimension through the crop growth simulation platform integration method, it is ensured that data from different sources and types can be compared and integrated at the same scale, thereby eliminating the model prediction error caused by inconsistent dimensions. Such processing not only improves the efficiency of model training, because it reduces the complexity of data preprocessing, but also improves the prediction accuracy of the model, because the model can more accurately learn the relationship between different variables. In addition, the application of the growth variables obtained by the crop growth simulation platform integration method also enhances the generalization ability of the large crop growth model, so that the large crop growth model can be more easily migrated to different crops, regions and environmental conditions, thereby providing accurate growth prediction in a wider range of application scenarios. Then, by automatically dividing these variables into dynamic covariates, static covariates and time series process variables, the model can adopt different processing strategies for different types of variables, thereby more accurately capturing their impact on crop growth. Then, these variables are converted into embedded vectors using embedding technology, which not only retains the information of the original data, but also enhances the model's understanding and processing capabilities of these data. Finally, by training the initial large language model, we obtained a large crop growth model that can comprehensively consider the influence of various variables. While improving the prediction accuracy of the model, it also enhances the generalization ability and adaptability of the model, enabling it to better cope with changing environmental conditions and different crop varieties, and improve the use of crop growth models in crop growth status prediction.

[0073] Optionally, the automatically dividing the crop growth variables to obtain dynamic covariates, static covariates and time series process variables of crop growth includes:

[0074] Determining the value of each variable in the digital dictionary according to the crop growth variables through semantic analysis;

[0075] All the variables are divided into the dynamic covariates, the static covariates and the time series process variables according to the types and time series of the variables.

[0076] Specifically, the process of automatically dividing crop growth variables is realized through semantic analysis technology, which involves in-depth analysis of crop growth variables to determine the specific classification of each variable in the digital dictionary. Specifically, the system first uses the semantic understanding ability of the large language model to conduct a comprehensive semantic analysis of the integrated crop growth status, climate factors, soil characteristics and field management information, so as to identify and understand the intrinsic meaning of various variables and their role in crop growth. In a preferred embodiment of the present invention, the multi-crop simulation platform integrates the characteristic parameters of 48 crops such as soybeans, corn, wheat, and rice, as well as the soil characteristic parameters of more than 100 plots. These parameters are accurately classified into dynamic covariates (such as daily temperature and precipitation), static covariates (such as soil water holding capacity and saturated water content) and time series process variables (such as crop leaf area index and biomass) through semantic analysis. This automated division method not only improves the efficiency of data processing, but also ensures that different variables are correctly classified through precise semantic matching, providing high-quality input data for subsequent model training and prediction.

[0077] In the embodiment of the present invention, the variables are automatically divided through semantic analysis, which greatly improves the speed and accuracy of data processing and reduces the errors and biases of manual classification. Secondly, this method ensures the quality and consistency of the model input data, thereby improving the accuracy and reliability of the crop growth prediction model.

[0078] Optionally, embedding the dynamic covariate, the static covariate and the time series process variable to obtain embedding vectors corresponding to the dynamic covariate, the static covariate and the time series process variable respectively includes:

[0079] By using an attention mechanism, according to the temporal features corresponding to the dynamic covariates, an attention score corresponding to each of the dynamic covariates is assigned;

[0080] Dynamically embedding is performed according to the attention score of the dynamic covariate, and the dynamic covariate is converted into the embedding vector.

[0081] Specifically, dynamic covariates, static covariates, and time-series process variables are embedded to ensure that the model can capture the unique characteristics of each variable and its relative impact on crop growth. First, the attention mechanism is used to analyze the time-series characteristics of dynamic covariates, such as daily temperature and precipitation, which change over time and have a direct and immediate impact on crop growth. The attention mechanism assigns a corresponding attention score according to the importance of each dynamic covariate, which means that covariates with a greater impact on crop growth will receive a higher attention score. In a preferred embodiment of the present invention, the crop growth model may identify temperature and precipitation at key growth stages as important dynamic covariates and give them higher attention scores. Then, the dynamic covariates are dynamically embedded according to these attention scores and converted into an embedding vector, which can encode the time-series information of the covariates and their relative importance to crop growth. This embedding method enables the model to take into account not only the numerical size of the covariates, but also the changing trends over time and the dynamic impact on crop growth.

[0082] In the embodiment of the present invention, through the attention mechanism and dynamic embedding technology, the model can more accurately capture and represent the impact of dynamic covariates on crop growth. This method enables the model to not only identify the most critical environmental factors for crop growth in a specific period, but also adapt to the changes of these factors over time.

[0083] Optionally, embedding the dynamic covariate, the static covariate and the time series process variable to obtain embedding vectors corresponding to the dynamic covariate, the static covariate and the time series process variable respectively includes:

[0084] According to the time series characteristics corresponding to the static covariates, adjust the weight and bias corresponding to each of the static covariates;

[0085] Through a fully connected network, static embedding is performed according to the weight and the bias corresponding to the static covariate, and the static covariate is converted into the embedding vector.

[0086] Specifically, the system first identifies static covariates, such as the physical properties of the soil and the genetic characteristics of the crop, which are relatively stable during the crop growth cycle but have a lasting impact on the long-term growth of the crop. Then, based on the characteristics of these static covariates, the system adjusts the corresponding weights and biases in the network to ensure that the embedding vector accurately reflects the specific impact of each static covariate on crop growth. Static embedding is performed through a fully connected network and converted into an embedding vector. In this process, the network structure is able to learn the specific effects of each crop and soil type on growth and encode these effects into the embedding vector. Through a fully connected network and learnable weights and biases, the model is able to generate a unique embedding vector for each static covariate, which not only contains the original information of the static covariate, but also captures its specific impact on crop growth.

[0087] In the embodiment of the present invention, the fully connected network and learnable weights and biases enable the model to more accurately simulate the growth of crops under different soil and genetic conditions, thereby improving the accuracy and reliability of the prediction.

[0088] Optionally, embedding the dynamic covariate, the static covariate and the time series process variable to obtain embedding vectors corresponding to the dynamic covariate, the static covariate and the time series process variable respectively includes:

[0089] Complementing the time series process variable with zero values ​​to obtain the time series process variable having the same dimension as the dynamic covariate and the static covariate;

[0090] Through the Transformer encoder, the time series process variables with the same dimension are time-series embedded, and the time series process variables are converted into the embedding vector.

[0091] Specifically, first, the time series process variables are filled with zero values. This step is to align the time series of the data so that the time series process variables match the dynamic covariates and static covariates in the time dimension to form a complete data set. For example, a large crop growth model may process data from the entire growth cycle of a crop from sowing to harvesting, including time series process variables such as leaf area index and biomass, which are adjusted to the same time dimension as other variables such as daily climate data and soil properties through zero-filling. Subsequently, the system uses the Transformer encoder to perform deep learning on these aligned time series process variables. The Transformer encoder can capture long-term dependencies and short-term patterns, and convert the time series process variables into an embedding vector that contains rich time series information and dynamic changes in crop growth. This time series embedding method allows the model to take into account the continuity and stage changes in the growth process when predicting the growth status of crops.

[0092] In the embodiment of the present invention, through the application of zero-value completion and Transformer encoder, the model can more accurately capture and represent the temporal information of the crop growth process. This method enables the model to not only identify the key moments of crop growth, but also understand the interdependence and evolution trends between these moments.

[0093] Optionally, the training of the initial large language model according to the embedding vector to obtain the large crop growth model includes:

[0094] Concatenate the embedding vectors corresponding to the dynamic covariate, the static covariate and the time series process variable to obtain a fusion vector of the crop in the whole cycle;

[0095] The initial large language model is trained according to the fusion vector of the crop in the full cycle to obtain the crop growth large model.

[0096] Specifically, the training of the large crop growth model is completed by splicing the embedded vectors of dynamic covariates, static covariates and time series process variables to form a fusion vector, and then training the initial large language model. This step makes full use of the rich data obtained from the multi-crop simulation platform integration method, which contains characteristic parameters of various crops and various soil characteristic parameters. These covariates of different sources and types are converted into vector form through embedding representation, and then the embedded vectors of dynamic covariates, static covariates and time series process variables are effectively spliced ​​according to their role and characteristics in crop growth to form a fusion vector containing comprehensive information. This fusion vector integrates various influencing factors of crop growth and can provide the model with detailed information on the entire crop cycle. In a preferred embodiment of the present invention, the embedded vectors of daily temperature and precipitation (dynamic covariates), soil water holding capacity and saturated water content (static covariates), and crop leaf area index and biomass (time series process variables) are spliced ​​to form a comprehensive fusion vector, which can fully reflect the complex dynamic process of crop growth. Subsequently, this fusion vector is used to train the initial large language model, and by adjusting the model parameters to minimize the difference between the predicted value and the actual value, a large crop growth model that can accurately predict the crop growth status is finally obtained.

[0097] In the embodiment of the present invention, by concatenating and fusing the embedded vectors of different variables, the model can comprehensively consider various influencing factors of crop growth, thereby improving the accuracy and reliability of prediction.

[0098] Optionally, the training of the initial large language model according to the fusion vector of the crop in the full cycle to obtain the large crop growth model includes:

[0099] Dividing the fused vector of the full period into a first fused vector and a second fused vector according to a preset period ratio, wherein the time period corresponding to the first fused vector is earlier than the time period corresponding to the second fused vector;

[0100] Using the first fused vector as an input parameter of the initial large language model and using the second fused vector as an output parameter of the initial large language model for training;

[0101] The initial large language model after training is used as the crop growth large model.

[0102] Specifically, the training process of the large crop growth model involves dividing the fusion vector into a first fusion vector and a second fusion vector according to a preset period ratio, wherein the first fusion vector represents the input parameter and the second fusion vector represents the output parameter. Specifically, the system first divides the fusion vector into two parts according to the entire cycle of crop growth: the first fusion vector contains data from the early stage of crop growth, such as initial climate conditions, soil properties and crop growth status, while the second fusion vector contains data from the late stage of crop growth, such as climate conditions, soil properties and crop growth status before harvest. In a preferred embodiment of the present invention, crop growth prediction can be regarded as a function f that converts the actual growth conditions of crops in the previous L days into actual growth conditions. Dynamic Covariates and static covariates Mapping for accurate predictions of the future The dynamic covariate It includes m dimensions such as temperature, light, precipitation, nitrogen, phosphorus, and potassium fertilizer application on the i-th day, and the static covariate It includes p dimensions such as field water holding capacity, saturated water content, saturated hydraulic conductivity, etc., and the crop growth status It includes q dimensions such as leaf area index, total yield, root depth, etc. on the i-th day. Therefore, H is the length of time to be predicted, and L+H represents the total length of the growth cycle. For example, the growth cycles of corn, rice, soybeans, and wheat among the main food crops are 5 months, 5 months, 4 months, and 4 months, respectively. H represents the length of prediction, which is a basic definition of time series prediction. When predicting, it is necessary to know the growth status of the crop for L days, basic crop information, basic soil information, etc., and predict the crop growth status for the next H days.

[0103] In the embodiment of the present invention, by dividing the fusion vector into two parts, input and output, for training, the crop growth model can learn the dynamic change process of crop growth and predict the later growth results based on the early growth conditions. This method enables the model to not only capture the immediate state of crop growth, but also predict future growth trends, providing more comprehensive and forward-looking decision support for agricultural production.

[0104] Optionally, the using the trained initial large language model as the crop growth large model includes:

[0105] When the initial large language model is being trained, obtaining an error value between the prediction data of the initial large language model and the second fusion vector according to a preset frequency;

[0106] When the error value is less than a preset error threshold, it is determined that the training of the initial large language model is completed.

[0107] Specifically, the training completion judgment of the crop growth model is achieved by monitoring the prediction error of the initial large language model during the training process. Specifically, during the training process, the system automatically obtains the prediction data of the initial large language model according to the preset frequency, and calculates the error value between these prediction data and the second fusion vector (i.e., the actual growth data). In a preferred embodiment of the present invention, the accuracy of the prediction will be measured by the mean square error or the mean absolute error. The mean square error and the mean absolute error are very common ways to measure the accuracy of the prediction, and are a measure of the average difference between the predicted crop growth state and the actual crop growth state. Among them, the mean square error is the subtraction of the variables corresponding to the predicted crop growth state and the actual crop growth state at the corresponding time, and then the square is summed and averaged, which represents the average of the square sum of the prediction errors of all variables within the prediction H days. The mean absolute error is the subtraction of the variables corresponding to the predicted crop growth state and the actual crop growth state at the corresponding time, and then the absolute value is calculated and then the sum is averaged (to prevent positive and negative cancellation), which represents the average of the absolute values ​​of the prediction errors of all variables within the prediction H days. In addition, the mean square error can also reflect the variance of the predicted value to the true value, while the mean absolute error reflects the absolute size of the predicted value to the true value. The smaller the two errors are, the more accurate the prediction is.

[0108] In the embodiment of the present invention, by real-time monitoring and evaluating the accuracy of model predictions, it can be ensured that the large crop growth model stops training after reaching a predetermined accuracy, thereby improving the efficiency and effect of model training and ensuring the accuracy and reliability of the model in practical applications.

[0109] In summary, in a preferred embodiment of the present invention, combined with Figure 4As shown in the figure, the dynamic covariates, static covariates and time series process variables are first standardized, and then the properties of these variables can be automatically decomposed through the data dictionary. Subsequently, the dynamic covariates are embedded using an attention-based mechanism, the static covariates are embedded in a fully connected network, and the time series process variables are embedded using a Transformer encoder. After these embedding vectors are represented by an attention-based embedding fusion, they are input into the network structure of the large language model, and finally the full-cycle prediction of crop growth status is achieved, including pre-training and fine-tuning of crop data under extreme weather conditions, so as to improve the prediction accuracy and adaptability of the model under different environmental conditions.

[0110] A person of ordinary skill in the art can understand that all or part of the processes in the above-mentioned embodiment method can be completed by instructing the relevant hardware through a computer program, and the program can be stored in a computer-readable storage medium. When the program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, the storage medium can be a disk, an optical disk, a read-only memory (ROM) or a random access memory (RAM), etc. In the present application, the unit described as a separate component may or may not be physically separated, and the component displayed as a unit may or may not be a physical unit, that is, it may be located in one place, or it may be distributed on multiple network units. Some or all of the units can be selected according to actual needs to achieve the purpose of the embodiment of the present invention. In addition, each functional unit in each embodiment of the present invention can be integrated in a processing unit, or each unit can exist physically separately, or two or more units can be integrated in one unit. The above-mentioned integrated unit can be implemented in the form of hardware or in the form of a software functional unit.

[0111] Although the present invention is disclosed as above, the protection scope of the present invention is not limited thereto. Those skilled in the art may make various changes and modifications without departing from the spirit and scope of the present invention, and these changes and modifications will fall within the protection scope of the present invention.

Claims

1. A crop growth simulation platform integration method, characterized in that: include: Obtain crop growth status, crop variety characteristics, climate factors, soil characteristics and field management information for each crop throughout its entire cycle; Through semantic analysis, the crop growth status, the crop variety characteristics, the climate factors, the soil characteristics and the field management information are integrated with the input data and output data of the crop growth model corresponding to the crop in the crop growth simulation platform to obtain the crop growth variables of the crop in the whole cycle.

2. The crop growth simulation platform integration method according to claim 1, characterized in that: The crop growth status, the crop variety characteristics, the climate factors, the soil characteristics and the field management information are integrated with the input data and output data of the crop growth model corresponding to the crop in the crop growth simulation platform through semantic analysis to obtain the crop growth variables of the crop in the whole cycle, including: By semantic analysis, it is determined whether there are variables pointing to the same meaning between the crop growth state, the crop variety characteristics, the climate factors, the soil characteristics, and the field management information and the input data and the output data corresponding to the crop growth model in the crop growth simulation platform; If so, the variables with the same meaning are coreference resolved, and the variables after the coreference resolution are dimensionally aligned, and then the variables after the dimension alignment are integrated with other variables to obtain the crop growth variable.

3. A method for constructing a large crop growth model, characterized in that: include: Acquire crop growth variables of the crop in the whole cycle obtained by the crop growth simulation platform integration method described in any one of claims 1 to 2; Automatically dividing the crop growth variables to obtain dynamic covariates, static covariates and time series process variables of crop growth; Embedding the dynamic covariate, the static covariate and the time series process variable to obtain embedding vectors corresponding to the dynamic covariate, the static covariate and the time series process variable respectively; The initial large language model is trained according to the embedding vector to obtain a large crop growth model.

4. The method for constructing a large crop growth model according to claim 3, characterized in that: The automatic division of the crop growth variables to obtain dynamic covariates, static covariates and time series process variables of crop growth includes: Determining the value of each variable in the digital dictionary according to the crop growth variables through semantic analysis; All the variables are divided into the dynamic covariates, the static covariates and the time series process variables according to the types and time series of the variables.

5. The method for constructing a large crop growth model according to claim 3, characterized in that: The embedding of the dynamic covariate, the static covariate and the time series process variable to obtain the embedding vectors corresponding to the dynamic covariate, the static covariate and the time series process variable respectively includes: By using an attention mechanism, according to the temporal features corresponding to the dynamic covariates, an attention score corresponding to each of the dynamic covariates is assigned; Dynamically embedding is performed according to the attention score of the dynamic covariate, and the dynamic covariate is converted into the embedding vector.

6. The method for constructing a large crop growth model according to claim 3, characterized in that: The embedding of the dynamic covariate, the static covariate and the time series process variable to obtain the embedding vectors corresponding to the dynamic covariate, the static covariate and the time series process variable respectively includes: According to the time series characteristics corresponding to the static covariates, adjust the weight and bias corresponding to each of the static covariates; Through a fully connected network, static embedding is performed according to the weight and the bias corresponding to the static covariate, and the static covariate is converted into the embedding vector.

7. The method for constructing a large crop growth model according to claim 3, characterized in that: The embedding of the dynamic covariate, the static covariate and the time series process variable to obtain the embedding vectors corresponding to the dynamic covariate, the static covariate and the time series process variable respectively includes: Complementing the time series process variable with zero values ​​to obtain the time series process variable having the same dimension as the dynamic covariate and the static covariate; Through the Transformer encoder, the time series process variables with the same dimension are time-series embedded, and the time series process variables are converted into the embedding vector.

8. The method for constructing a large crop growth model according to claim 3, characterized in that: The method of performing full parameter training on the initial large language model according to the embedding vector to obtain a large crop growth model includes: Concatenate the embedding vectors corresponding to the dynamic covariate, the static covariate and the time series process variable to obtain a fusion vector of the crop in the whole cycle; The initial large language model is trained according to the fusion vector of the crop in the full cycle to obtain the crop growth large model.

9. The method for constructing a large crop growth model according to claim 8, characterized in that: The step of performing full parameter training on the initial large language model according to the fusion vector of the crop in the full cycle to obtain the large crop growth model includes: Dividing the fused vector of the full period into a first fused vector and a second fused vector according to a preset period ratio, wherein the time period corresponding to the first fused vector is earlier than the time period corresponding to the second fused vector; Using the first fused vector as an input parameter of the initial large language model and using the second fused vector as an output parameter of the initial large language model for training; The initial large language model after the full parameter training is completed is used as the crop growth large model.

10. The method for constructing a large crop growth model according to claim 9, characterized in that: The step of using the initial large language model after training as the large crop growth model comprises: When the initial large language model is being trained, obtaining an error value between the prediction data of the initial large language model and the second fusion vector according to a preset frequency; When the error value is less than a preset error threshold, it is determined that the training of the initial large language model is completed.

Images