A pre-training-fine-tuning-based general crop growth simulation method and system

Abstract

The application provides a universal crop growth simulation method and system based on pre-training-fine-tuning, which comprises the following steps: constructing a universal crop growth dataset; introducing a LoRA low-rank fine-tuning mechanism to fine-tune a specified growth index prediction model after pre-training, so that it is adapted to a target region and scene; using the specified growth index prediction model adapted to the target region and scene to obtain the predicted value of the specified growth index of the target crop in the upper and lower block periods; wherein the specified pre-training step comprises: constructing a JEPA cross-crop high-dimensional representation pre-training framework, training the JEPA pre-training framework based on the universal crop growth dataset, and obtaining a cross-crop high-dimensional representation predictor after the training is completed; specifying a target crop growth index prediction task, combining the cross-crop high-dimensional representation predictor with a decoder constructed based on cross-attention to construct a specified growth index prediction model, and training the model based on the universal crop growth dataset.

Description

Technical Field

[0001] This invention belongs to the field of crop growth simulation, specifically relating to a general crop growth simulation method and system based on pre-training and fine-tuning. Background Technology

[0002] Crop growth simulation plays a crucial role in agricultural production management, crop yield estimation, disaster early warning, and optimization of agricultural decision-making. However, crop growth is influenced by a variety of factors, including varietal characteristics, meteorological conditions, soil water and fertilizer status, and agricultural management practices, exhibiting significant dynamism, regionality, and complexity. Therefore, developing a method capable of accurately simulating and predicting crop growth processes across different crops, regions, and management scenarios is a critical issue that urgently needs to be addressed in the development of smart agriculture.

[0003] Currently, crop growth simulation mainly relies on traditional mechanistic methods and data-driven approaches. Traditional mechanistic methods are typically based on equations governing crop physiological and ecological processes, water and fertilizer transport, and canopy energy exchange, offering strong interpretability. However, these methods usually require users to possess expertise in crop physiology, soil water and fertilizer management, and model parameterization, and necessitate extensive data on crop varieties, soil conditions, meteorological parameters, and agricultural management practices. Therefore, the barrier to entry is high for ordinary agricultural producers, and parameter acquisition is difficult, calibration costs are high, and applicability is limited across different regions, varieties, and management conditions.

[0004] In recent years, data-driven methods have been increasingly applied to tasks such as crop growth forecasting and yield estimation. These methods can learn the complex relationships between crop growth and environmental and management conditions from historical data. However, these methods typically rely on data trained on specific crops, in specific regions, or under specific management scenarios, and the learned feature representations and predictive patterns often lack cross-crop, cross-region, and cross-scenario generality. Furthermore, because some data-driven methods exhibit black-box or gray-box characteristics, their internal growth pattern representations are difficult to interpret, further limiting their application in actual agricultural production. When these methods are applied to new crop types, variety conditions, planting areas, or agricultural management practices, their predictive accuracy and robustness often decrease significantly, making it difficult to meet the complex and ever-changing crop growth simulation needs of actual agricultural production. Summary of the Invention

[0005] Given the difficulty in obtaining complete crop parameters, limited user expertise, and complex and ever-changing crop growth scenarios, existing methods struggle to meet the agricultural production demand for a universal crop growth simulation method that is low-threshold, transferable, and applicable to multiple crops. To address this technical problem, this invention provides a universal crop growth simulation method and system based on pre-training and fine-tuning. Employing a pre-training and fine-tuning architecture, it utilizes growth data from multiple crops and varieties under different environments and management scenarios for training, extracting cross-species universal representations of crop growth processes. In the downstream application stage, sparse observations are used for fine-tuning to activate these universal growth representations and absorb new crop growth knowledge not covered in the pre-training stage. This enables the simulation and prediction of crop growth processes under new crops, new varieties, new planting areas, or new management scenarios not covered by existing modeling data.

[0006] According to one aspect of the present invention, a general crop growth simulation method based on pre-training-fine-tuning is provided, comprising:

[0007] Construct a general crop growth dataset;

[0008] A measured crop growth dataset for the target region and scenario is constructed, and the LoRA low-rank fine-tuning mechanism is introduced to fine-tune the pre-trained prediction model of the specified growth index, so as to obtain a prediction model of the specified growth index adapted to the target region and scenario.

[0009] The environmental information of the target crop in the target area and the scene within the upper and lower time periods is input into the specified growth index prediction model adapted to the target area and the scene to obtain the specified growth index prediction value of the target crop within the upper and lower time periods.

[0010] The pre-training steps of the specified growth index prediction model include:

[0011] A pre-training framework for JEPA cross-crop high-dimensional representation is constructed. The JEPA pre-training framework is trained based on a general crop growth dataset. After training, a cross-crop high-dimensional representation predictor is obtained. The cross-crop high-dimensional representation predictor uses the environmental information of the target crop in the previous and next time blocks as input to predict the shared high-dimensional potential representation of the crop in the target time block.

[0012] For the task of predicting the growth indicators of a specified target crop, a prediction model for the specified growth indicators is constructed by combining a cross-crop high-dimensional representation predictor with a decoder built based on cross-attention, and trained on a general crop growth dataset.

[0013] As a further technical solution, the JEPA pre-training framework is trained based on a general crop growth dataset, and the steps include:

[0014] The environmental information in the general crop growth dataset is processed through an embedding layer, and the initial token representation set is output.

[0015] The JEPA context block and JEPA target block are constructed based on the initial token representation set. The JEPA context block is encoded by the first encoder in the dual encoder to obtain the context block representation, and the JEPA target block is encoded by the second encoder in the dual encoder to obtain the target block representation.

[0016] The predictor outputs the predicted target block representation based on the context block representation output by the encoder, combines the context block representation position encoding with the JEPA target block position encoding, and calculates the upstream JEPA training loss function value based on the target block representation predicted by the predictor and the target block representation encoded by the encoder, and then backpropagates to update the trainable parameters.

[0017] The process iterates until the preset accuracy is met or the preset maximum number of training rounds is reached. The embedding layer, the first encoder in the dual encoder, and the predictor in the trained JEPA pre-trained framework are retained to obtain a cross-crop high-dimensional representation predictor.

[0018] As a further technical solution, the embedding layer includes embedding branches for three types of environmental information: meteorology, agricultural management, and crop growth, which process the corresponding types of environmental information respectively.

[0019] The embedded branch takes the day as the smallest processing unit, and embeds and aggregates each variable in the corresponding type of environmental information to obtain the branch variable aggregation token representation.

[0020] Add time position encoding and branch position encoding to the aggregated token representation of the branch variable to obtain the initial token representation.

[0021] As a further technical solution, the steps for constructing the JEPA context block and the JEPA target block include:

[0022] The JEPA context block contains the initial tokens corresponding to the daily meteorological and agricultural management environmental information for the entire time period of the upper and lower blocks and the target block, as well as the initial tokens corresponding to the crop growth environmental information for sparse dates within the context block; the JEPA target block contains the initial tokens corresponding to the daily crop growth environmental information for the target block.

[0023] As a further technical solution, the steps for training based on a general crop growth dataset include:

[0024] By specifying growth indicators, the target crop dataset is extracted from the general crop dataset, and the target crop dataset is processed using a cross-crop high-dimensional representation predictor to obtain the shared high-dimensional potential representation of the target crop within the target block time period.

[0025] The decoder performs cross-attention-based decoding based on specified growth indicators and the shared high-dimensional latent representation of the target crop within the target block time period to obtain the predicted values ​​of each specified growth indicator.

[0026] The reconstruction loss is calculated based on the predicted values ​​and observed values ​​of each specified growth indicator, and the decoder parameters are updated via backpropagation.

[0027] The process is iterated until the preset accuracy is met or the preset maximum number of training rounds is reached, resulting in a trained prediction model for the specified growth index.

[0028] As a further technical solution, the decoder performs cross-attention-based decoding based on specified growth indicators and the shared high-dimensional latent representation of the target crop within the target block time period to obtain the predicted values ​​of each specified growth indicator. The steps include:

[0029] Based on the shared high-dimensional latent representation of the target block time period and the query generated by the target block time period, the value is generated based on the shared high-dimensional latent representation of the target block time period, and a special task key is generated based on the specified growth indicators and the target block time period. Cross attention is used to obtain the indicator-specific representation of each specified growth indicator of the target crop within the target block time period. Finally, through regression head processing, the predicted values ​​of each specified growth indicator of the target crop within the target block time period are obtained.

[0030] As a further technical solution, the steps of introducing the LoRA low-rank fine-tuning mechanism to fine-tune the trained prediction model for a specified growth index include:

[0031] A low-rank LoRA matrix is ​​introduced into all linear layers of the decoder of the pre-trained specified growth index prediction model.

[0032] The trained specified growth index prediction model, after introducing the LoRA low-rank matrix, is fine-tuned based on the measured crop growth dataset, and finally outputs a specified growth index prediction model that is adapted to the target area and the scene. During the fine-tuning process, the original parameters of the trained specified growth index prediction model are frozen, and only the LoRA low-rank matrix introduced in all linear layers is updated.

[0033] According to another aspect of this specification, a pre-training-fine-tuning-based general crop growth simulation system is provided, comprising:

[0034] The general data preparation module is used to build a general crop growth dataset;

[0035] The adaptation and fine-tuning module is used to construct a measured crop growth dataset for the target region and scenario, and introduces the LoRA low-rank fine-tuning mechanism to fine-tune the pre-trained prediction model of the specified growth index, so as to obtain a prediction model of the specified growth index adapted to the target region and scenario.

[0036] The crop growth simulation module is used to input the environmental information of the target crop in the target area and the scene within the upper and lower time periods into a specified growth index prediction model adapted to the target area and the scene, and obtain the specified growth index prediction value of the target crop within the upper and lower time periods.

[0037] The pre-training steps of the specified growth index prediction model include:

[0038] A pre-training framework for JEPA cross-crop high-dimensional representation is constructed. The JEPA pre-training framework is trained based on a general crop growth dataset. After training, a cross-crop high-dimensional representation predictor is obtained. The cross-crop high-dimensional representation predictor uses the environmental information of the target crop in the previous and next time blocks as input to predict the shared high-dimensional potential representation of the crop in the target time block.

[0039] For the task of predicting the growth indicators of a specified target crop, a prediction model for the specified growth indicators is constructed by combining a cross-crop high-dimensional representation predictor with a decoder built based on cross-attention, and trained on a general crop growth dataset.

[0040] According to another aspect of this specification, an electronic device is provided, including a memory and a processor, the memory storing program instructions that are executed by the processor, the processor invoking the program instructions to execute a pre-training-fine-tuning-based general crop growth simulation method.

[0041] According to another aspect of this specification, a non-transitory computer-readable storage medium is provided, the non-transitory computer-readable storage medium storing computer instructions that cause the computer to execute a pre-training-fine-tuning-based general crop growth simulation method.

[0042] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0043] (1) This invention utilizes a large dataset of crop growth data for pre-training, enabling the model to learn the common dynamic laws in the growth process of different crops, and to develop transferable general crop growth representations.

[0044] (2) The present invention only requires a small amount of actual observation data to fine-tune and correct the pre-trained model, so that the model can quickly adapt to the target crop growth simulation task in specific application scenarios.

[0045] (3) The present invention does not require users to have professional knowledge of crop growth simulation. Only a small amount of sparse crop observation information is required, and the model can automatically infer and simulate the crop growth status.

[0046] (4) The present invention can activate the general crop representation learned in the pre-training stage through actual test data and absorb new crop growth knowledge, so that it has the ability to transfer across crops, adaptive correction and continuous learning. Attached Figure Description

[0047] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the accompanying drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0048] Figure 1 A flowchart illustrating a general crop growth simulation method based on pre-training and fine-tuning provided in an embodiment of the present invention;

[0049] Figure 2 This is a schematic diagram comparing the prediction performance of pre-trained datasets of different sizes in an embodiment of the present invention;

[0050] Figure 3 This is a schematic diagram comparing the response of the model-predicted growth results to environmental and agricultural factors under the unadjusted conditions in this embodiment of the invention.

[0051] Figure 4 This is a schematic diagram comparing the prediction effects of whether or not the pre-training data contains rice samples in an embodiment of the present invention.

[0052] Figure 5 This is a comparison chart of the predicted curve and the actual observation points in an embodiment of the present invention;

[0053] Figure 6 This is a schematic diagram comparing the attention matrices of different input factors for predicting different growth indicators in an embodiment of the present invention.

[0054] Figure 7 A schematic diagram of the structure of a general crop growth simulation system based on pre-training and fine-tuning provided in an embodiment of the present invention;

[0055] Figure 8 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0056] It should be noted that:

[0057] The terms “comprising” and “having”, and any variations thereof, in the specification, claims, and accompanying drawings of this invention are intended to cover a non-exclusive inclusion, such as a process, method, system, product, or apparatus that includes a series of steps or units, not necessarily limited to those explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0058] The block diagrams shown in the accompanying drawings are merely functional entities and do not necessarily correspond to physically independent entities. That is, these functional entities can be implemented in software, in one or more hardware modules or integrated circuits, or in different network and / or processor devices and / or microcontroller devices. The flowcharts shown in the accompanying drawings are merely illustrative and do not necessarily include all content and operations / steps, nor do they necessarily have to be performed in the described order. For example, some operations / steps can be decomposed, while others can be combined or partially combined; therefore, the actual execution order may change depending on the specific circumstances.

[0059] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. In addition, the technical features of the various embodiments or individual embodiments provided by the present invention can be arbitrarily combined to form new technical solutions. Such combinations are not bound by the order of steps and / or structural composition patterns, but must be based on the ability of those skilled in the art to implement them. When the combination of technical solutions is contradictory or cannot be implemented, it should be considered that such a combination of technical solutions does not exist and is not within the scope of protection claimed by the present invention.

[0060] like Figure 1 As shown, this embodiment of the invention provides a general crop growth simulation method based on pre-training and fine-tuning, including the following steps:

[0061] Step 1: Construct a measured crop growth dataset for the target region and scenario, and introduce the LoRA low-rank fine-tuning mechanism to fine-tune the pre-trained prediction model of the specified growth index, so as to obtain a prediction model of the specified growth index that is adapted to the target region and scenario.

[0062] Step 2: Input the environmental information of the target crop in the upper and lower time blocks of the target area and scene into the specified growth index prediction model adapted to the target area and scene to obtain the specified growth index prediction value of the target crop in the upper and lower time blocks.

[0063] The pre-training of the specified growth index prediction model described in this invention further includes:

[0064] Step S1: Construct a general crop growth dataset;

[0065] Step S2: Construct a pre-training framework for JEPA cross-crop high-dimensional representation. Train the JEPA pre-training framework based on a general crop growth dataset. After training, obtain a cross-crop high-dimensional representation predictor. The cross-crop high-dimensional representation predictor uses the environmental information of the target crop in the previous and next time blocks as input to predict the shared high-dimensional potential representation of the crop in the target time block.

[0066] Step S3: Specify the target crop growth index prediction task, jointly construct the specified growth index prediction model with the cross-crop high-dimensional representation predictor and the decoder based on cross-attention, and train it based on a general crop growth dataset.

[0067] In step S1, the general crop dataset contains daily growth and environmental information for several types of crops, without the need for labeled data. The environmental information, as the raw input data, consists of three categories: meteorological, agricultural management, and crop growth data.

[0068] Meteorological environmental information includes, but is not limited to: daily maximum temperature, daily minimum temperature, daily cumulative radiation, daily cumulative rainfall, wind speed, and daily average water vapor pressure difference; agricultural management environmental information includes, but is not limited to: cumulative crop fertilizer application; crop growth includes, but is not limited to: crop growth stage, ear quality, stem quality, leaf quality, leaf nitrogen content, and plant area index.

[0069] In one specific implementation, the general crop dataset selected daily growth and environmental information for 10 types of crops, including rice, wheat, corn, soybean, potato, rapeseed, cotton, sugarcane, sugar beet, barley, and sorghum.

[0070] In step S2, a pre-trained framework for JEPA (Joint Embedding Prediction Architecture) is trained based on a general crop growth dataset. The steps include:

[0071] Step S2-1: Process the environmental information in the general crop growth dataset through the embedding layer and output the initial token representation set;

[0072] Step S2-2: Construct the JEPA context block and JEPA target block based on the initial token representation set. Encode the JEPA context block using the first encoder in the dual encoder to obtain the context block representation, and encode the JEPA target block using the second encoder in the dual encoder to obtain the target block representation.

[0073] Step S2-3: The predictor outputs the predicted target block representation based on the context block representation output by the encoder, combines the context block representation position encoding with the JEPA target block position encoding, and calculates the upstream JEPA training loss function value based on the target block representation predicted by the predictor and the target block representation encoded by the encoder, and backpropagates to update the trainable parameters.

[0074] Steps S2-4: Iterate until the preset accuracy is met or the preset maximum number of training rounds is reached. Retain the embedding layer, the first encoder in the dual encoder, and the prediction layer in the trained JEPA pre-trained framework to obtain a cross-crop high-dimensional representation predictor.

[0075] The embedding layer includes embedding branches for three types of environmental information: meteorology, agricultural management, and crop growth. Each branch processes the corresponding type of environmental information, with the day as the smallest processing unit. It embeds and aggregates the variables in the corresponding type of environmental information to obtain a branch variable aggregation token representation. Time location encoding and branch location encoding are added to the branch variable aggregation token representation to obtain an initial token representation. Optionally, the embedding layer uses a fully connected neural network.

[0076] The parameters of the second encoder are obtained by updating the parameters of the first encoder using exponential smoothing.

[0077] In step S2-2, the steps for constructing the JEPA context block and target block include:

[0078] The JEPA context block contains the initial tokens corresponding to the daily meteorological and agricultural management environmental information for the entire period, as well as the initial tokens corresponding to the crop growth environmental information for sparse dates within the observation period, and the initial tokens corresponding to the crop growth environmental information for sparse dates within the observation period. The JEPA target block contains the initial tokens corresponding to the daily crop growth environmental information within the prediction period.

[0079] As a specific implementation method, in step S2, the first... The mathematical representation of the original input data for the day is as follows:

[0080] (1)

[0081] In the formula, Indicates the first The original input set of the day; For the first A collection of meteorological and environmental information for each day; For the first A collection of environmental information on agricultural management types for each day; For the first A collection of environmental information on crop growth type over the day; These represent weather type, agricultural management type, and crop growth type, respectively.

[0082] In step S2-1, the steps of processing environmental information in the general crop growth dataset through the embedding layer and outputting the initial token representation set include:

[0083] The embedded branches process the environmental information in the general crop growth dataset on a daily basis, and obtain the daily branch variable aggregate token representation;

[0084] Add time position encoding and branch position encoding to the daily branch variable aggregate token representation output by each embedded branch to obtain the initial token representation set.

[0085] Specifically, with the first The first branch Taking a single variable as an example, through embedding and aggregation, we obtain the branch variable aggregation token representation, which is mathematically represented as:

[0086] (2)

[0087] (3)

[0088] (4)

[0089] In equation (2), For branch index, For variable indexing; Indicates the first Heavenly The first branch Embedding vectors of variables; For layer normalization operation; For the first The variable projection matrix of each branch is composed of trainable parameters. constitute: These represent the trainable parameters in the variable projection matrices of the meteorological type branch, the agricultural management type branch, and the crop growth type branch, respectively. Indicates the first Heavenly The first branch The original data of each variable; For the first The first branch Variable position embedding of variables.

[0090] In equation (3), Indicates the first Heavenly The first branch Attention weights for each variable; Indicates the first Trainable attention vectors for each branch; Represents an exponential function; Indicates the first Index of any variable in each branch; Indicates the first Heavenly All variables in each branch; Indicates the first Heavenly The first branch Embedding vectors of variables;

[0091] In equation (4), Indicates the first Heavenly The branch variables of each branch are aggregated into a token.

[0092] Furthermore, by adding time position encoding and branch position encoding to the branch variable aggregate token representation, the initial token that needs to be input to the encoder is obtained, mathematically represented as:

[0093] (5)

[0094] In the formula, For the first All variables in the branch at the _th ... The initial token for input Encoder is pending; Indicates the first Date position code of the day Indicates the first Branch type encoding for each branch.

[0095] Secondly, the dual encoders include a context block encoder and a target block encoder, which encode the constructed Joint Embedded Prediction Architecture (JEPA) context block and target block respectively. The parameters of the target block encoder are obtained by updating the parameters of the context block encoder using exponential smoothing.

[0096] In step S2-2, optionally, the period from the start of crop growth to the last crop growth observation date can be used. The observation period is defined as the period from which sparse anchor point observation days are selected. Based on the last crop growth observation day Until the last day when crop growth needs to be predicted For the prediction period The entire cycle is Then the context block The mathematical representation of is:

[0097] (6)

[0098] In the formula, S represents the initial set of tokens. This represents the initial token set for the entire meteorological cycle. This represents the initial token set for full-cycle agricultural management. The initial set of tokens representing the crop growth type corresponding to the sparse anchor observation days.

[0099] target block The mathematical representation of is:

[0100] (7)

[0101] In the formula, The initial set of tokens representing the crop growth type for the prediction cycle.

[0102] Furthermore, regarding the context block With target block The encoding result is obtained through dual encoders, and its mathematical representation is as follows:

[0103] (8)

[0104] (9)

[0105] (10)

[0106] In the formula, The token obtained by encoding the context block; For the first encoder, These are the parameters for the first encoder; The token obtained by encoding the target block. For the second encoder, For the second encoder parameters; This represents the momentum coefficient.

[0107] In steps S2-3, the predictor (transformer architecture) represents the context block based on the encoder output. Joint context block characterization position encoding Encoding the target block representation position The target block representation is predicted. Mathematically, this is expressed as:

[0108] (11)

[0109] In the formula, Indicates a predictor; This represents the updatable parameters in the predictor.

[0110] In steps S2-4, the upstream JEPA training loss function is constrained by both SmoothL1 loss and dynamic consistency loss, mathematically expressed as:

[0111] (12)

[0112] (13)

[0113] (14)

[0114] In the formula, This indicates JEPA's projected losses; Loss due to robust regression; , These represent the predicted first and second halves of the series. Crop growth characteristics Token and the first Target crop growth characteristic token for the day; To stop the backpropagation operation; Indicates dynamically consistent loss; For parameters Composition of dynamic prediction head; A valid set of adjacent date transfers; , and Let these represent the aggregated tokens of the branch variables for the meteorological branch, the agricultural management branch, and the crop growth branch, respectively, corresponding to the tokens mentioned earlier. ; The total loss function of the upstream pre-trained Encoder; The weights are for dynamic consistency loss.

[0115] Update parameters using backpropagation with the total loss function:

[0116] (15)

[0117] In the formula, This represents the set of all trainable parameters that need to be updated during the upstream pre-training phase. This represents the updatable parameters in the predictor; Represents the gradient; Indicates the learning rate; This indicates an assignment operation, meaning that the variable on the left will be replaced by the result of the calculation on the right.

[0118] The cross-crop high-dimensional representation predictor obtained in step S2 is used to achieve unified latent representation learning for crop growth processes.

[0119] Step S3, the training steps based on the general crop growth dataset, include:

[0120] Step S3-1: Specify the growth index, extract the target crop dataset from the general crop dataset, and process the target crop dataset using the cross-crop high-dimensional representation predictor to obtain the shared high-dimensional potential representation of the target crop within the target block time period.

[0121] Step S3-2: The decoder performs cross-attention-based decoding based on the specified growth indicators and the shared high-dimensional potential representation of the target crop within the target block time period to obtain the predicted values ​​of each specified growth indicator.

[0122] Step S3-3: Calculate the reconstruction loss based on the predicted values ​​and observed values ​​of each specified growth indicator, and update the decoder parameters through backpropagation.

[0123] Step S3-4: Iterate until the preset accuracy is met or the preset maximum number of training rounds is reached to obtain the trained prediction model for the specified growth index.

[0124] Specifically, in step S3-1, a specific variety dataset is extracted from the daily general crop dataset, and the crop growth indicators to be predicted are defined. With target block time period Mathematically, this is expressed as:

[0125] (16)

[0126] In the formula, This is a collection of growth indicators to be predicted. Corresponding to each specified growth indicator, Specify the number of growth indicators; The target block time period.

[0127] Correspondingly, the shared high-dimensional latent representation of the target crop within the target block time period output by the cross-crop high-dimensional representation predictor is mathematically represented as:

[0128] (17)

[0129] In step S3-2, the decoder performs cross-attention-based decoding based on the specified growth indicators and the shared high-dimensional latent representation of the target crop within the target block time period to obtain the predicted values ​​of each specified growth indicator. The steps include:

[0130] Based on the shared high-dimensional latent representation of the target block time period and the query generated by the target block time period, the value is generated based on the shared high-dimensional latent representation of the target block time period, and a special task key is generated based on the specified growth indicators and the target block time period. Cross attention is used to obtain the indicator-specific representation of each specified growth indicator of the target crop within the target block time period. Finally, through regression head processing, the predicted values ​​of each specified growth indicator of the target crop within the target block time period are obtained.

[0131] Specifically, in step S3-2, the steps for constructing the query, value, and special task key include:

[0132] In order to realize the shared high-dimensional potential representation of the target crop within the target block time period Mapped to a specified growth indicator, in the decoder, the query and value are based on... Generation, mathematically represented as:

[0133] (18)

[0134] In the formula, For query; Value; and Trainable projection matrix; Encode the time period location of the target block.

[0135] Special task key Generated from different specified growth indicators and target block time periods, mathematically represented as:

[0136] (19)

[0137] In the formula, For the first A special task key corresponding to a specified growth indicator; For the first Key projection matrix corresponding to a specified growth indicator; For the first A learnable task embedding for a specified growth metric.

[0138] Furthermore, in step S3-2, for the first segment to be predicted in the target block time period... Day, No. The first crop growth indicator, the decoder uses the first... The query is for the first time segment in the target block. Attention is calculated on the Key components corresponding to the specified growth indicators to obtain the first... Day, No. The daily cross-attention weights of each crop growth indicator within the target block time period:

[0139] (20)

[0140] In the formula, For the target block time period The daily query component is... Middle element; For the target block Japan The Key component of each crop growth indicator is: Middle element; For attention space dimensions; For the first to be predicted day The crop growth indicators are in the target block time period. Cross-attention weights for each day.

[0141] Subsequently, based on the calculated attention weights, specific decoding representations of each growth index were calculated. And obtain the first by a regression head day Predicted values ​​of crop growth indicators Mathematically, this is expressed as:

[0142] (twenty one)

[0143] (twenty two)

[0144] In the formula, for The Middle The Value component corresponding to the day; For the first day Specific decoding and characterization of crop growth indicators; For the first day Predicted values ​​for crop growth indicators; For the first The regression head of a crop growth index.

[0145] Specifically, in step S3-3, the following steps are set: For the prediction period day The true value of each crop growth indicator. Given the corresponding mask flag, the Decoder's reconstruction loss function... Mathematically, this is expressed as:

[0146] (twenty three)

[0147] In the formula, For the first Loss weights for each crop growth indicator.

[0148] Through the loss function The backpropagation training of the decoder parameters is mathematically represented as follows:

[0149] (twenty four)

[0150] In the formula, This represents the set of decoder parameters to be trained.

[0151] Step S3 is essentially the downstream prediction stage. The decoder uses the growth indicator to be predicted and the target time token as queries, selects and aggregates effective information from the shared high-dimensional representation through the cross-attention mechanism, and finally outputs the predicted value of the target growth indicator (crop growth period, dry matter, leaf area or nitrogen status, etc.).

[0152] Steps S1-2 are essentially the JEPA self-supervised learning process. In a specific implementation, step S1 establishes a large database (a general crop dataset) containing growth process information for various crops (such as rice, corn, wheat, sugar beets, sugarcane, tobacco, potatoes, soybeans, etc.). This dataset includes weather, agricultural management, and crop growth data, totaling 18 variables and continuous records of three types of variables. Then, the embedding layer expands the dimensions of each variable to a high-dimensional space of 192. The encoder then aggregates the 18 192-dimensional variables into three 192-dimensional aggregated high-dimensional variables (represented by the branch variable aggregation token). Figure 1 (Embedding). Crop growth information is masked according to a time scale, and the model is required to use the uncovered area to predict the masked area through the predictor. This self-supervised pre-training framework allows the encoder to learn the common growth patterns of all crops.

[0153] The predictor output obtained in step S2 is only a high-dimensional representation of the crop growth prediction result (i.e., a shared high-dimensional potential representation of the predicted crop during the target block period, with a very high dimension of 192 dimensions, but in reality, only 6 true variables are predicted, which is 6 dimensions; step S3 is set to reconstruct the high-dimensional representation back to the real space). Step S3 designs a cross-attention mechanism to construct corresponding key vectors for the 6 true variables ( Figure 1 (6 small blocks), and at the same time, a total key query and value vector will be generated by the 192-dimensional high-dimensional representation. Figure 1 (Two separate large blocks). Through this query, key, and value approach, and after a cross-attention calculation, a high-dimensional representation is mapped back to six real-world crop variables (growth period, stem, leaf, and ear weight, plant area index, and leaf nitrogen content). It is important to note that the dataset in step S3 can only contain one crop, but is not limited to any particular variety. The trained model predicting the specified growth indicators will map the input back to the specific indicators of the specific crop in the real-world dimension. Figure 1 (The example in this case is a rice variety).

[0154] Step 1, which introduces the LoRA low-rank fine-tuning mechanism to fine-tune the trained prediction model for a specified growth metric, includes the following steps:

[0155] A low-rank LoRA matrix is ​​introduced into all linear layers of the decoder of the pre-trained specified growth index prediction model.

[0156] The trained specified growth index prediction model, after introducing the LoRA low-rank matrix, is fine-tuned based on the measured crop growth dataset, and finally outputs a specified growth index prediction model that is adapted to the target area and the scene. During the fine-tuning process, the original parameters of the trained specified growth index prediction model are frozen, and only the LoRA low-rank matrix introduced in all linear layers is updated.

[0157] Specifically, step 1 introduces the LoRA low-rank training mechanism. LoRA fine-tuning does not disrupt the previously learned crop growth patterns and can incorporate complex environmental factors that are difficult to fully represent in a particular application scenario. Furthermore, it has a small number of parameters and a lightweight structure, making it suitable for rapid deployment. This enables the model to be quickly and lightweightly adapted to the target region and scenario. The steps include the following sub-steps:

[0158] Step 1-1: Introduce a low-rank LoRA matrix into the linear layer of the decoder:

[0159] The original linear layer computation in the Decoder can be mathematically represented as:

[0160] (25)

[0161] In the formula, For the first Input to a linear layer; No. The output of a linear layer; and These are its original weight matrix and bias term, respectively.

[0162] Introducing a LoRA low-rank matrix into all linear layers of the decoder, mathematically represented as:

[0163] (26)

[0164] In the formula, The decoder weights after introducing the LoRA low-rank matrix; The original weights of the decoder are frozen; The low-rank increment introduced for LoRA; Let be the rank of the LoRA method; To achieve sparse scaling; and It is a trainable low-rank matrix.

[0165] All linear layers (total) The output of the layer becomes:

[0166] (27)

[0167] In the decoder after the introduction of LoRA, its query ,value and special task key And the calculation of the projection layer, mathematically represented as:

[0168] (28)

[0169] (29)

[0170] (30)

[0171] In the formula, Parameter freezing; These are the trainable parameters for LoRA training. This is a cross-attention computation operation; This represents the first output of the decoder after introducing LoRA. day Predicted values ​​for crop growth indicators; This is an increment symbol.

[0172] Steps 1-2, LoRA downstream training objective function and parameter update: After introducing the LoRA low-rank matrix, the loss function required for training... Represented as:

[0173] (31)

[0174] (32)

[0175] In the formula, After introducing the LoRA low-rank matrix, the decoder calculates the supervised loss on the measured crop growth dataset to learn the personalized features of a specific site. To smooth out the constraint weights; Regularize the weights for LoRA incremental parameters; is the set of linear layers; is the second-order difference smoothing constraint for the predicted sequence.

[0176] Calculate the loss The corresponding LoRA low-rank matrix parameters can be trained and updated. :

[0177] (33)

[0178] Step 2 uses the model fine-tuned in Step 1 to input meteorological, agricultural, and sparse observation information to achieve continuous dynamic reasoning and prediction of future crop growth.

[0179] As a specific implementation, the measured crop growth dataset for the target area and scenario in step 1 is taken from the test dataset of Jianli City, Hubei Province, using Qingxiangyou 261 rice as an example for verification. This field contains six growth observation data points throughout its entire growth period. To test the model's dynamic prediction capability under different sparse observation conditions, this embodiment inputs the 1st, 2nd, 3rd, 4th, and 5th sparse observation data points respectively, constructing five test cases. The test results are as follows: Figure 2 , Figure 3 , Figure 4 , Figure 5 , Figure 6 As shown:

[0180] Figure 2 The prediction performance of pre-training datasets of different sizes was compared. In the figure, R² is the coefficient of determination, used to measure the model's ability to interpret data or the goodness of fit of predictions. Its value typically ranges from 0 to 1; the closer the value is to 1, the better the model's prediction performance. Figure 2 Data shows that the larger the scale of pre-training data, the better the model's predictive performance; Figure 3 The response of the model-predicted crop growth results to environmental and agricultural factors under the unadjusted condition was compared. The results showed that the model predictions conformed to the basic laws of the influence of accumulated temperature and fertilization on crop growth. Figure 4 The prediction performance was compared between the pre-training data and the data containing rice samples. Figure 4 The predictive performance of development stage (DVS), panicle weight (WSO), stem weight (WST), leaf weight (WLV), total plant area index (TPAI), and leaf nitrogen content (LNC) is shown. The results show that the model has good predictive ability regardless of whether rice samples are included in the pre-training data, indicating that it has learned general growth characteristics across crops. Figure 5 The results of comparing the predicted curves of this method with the actual observation points show that this method can simulate crop growth changes with relatively high accuracy. Figure 6 The attention matrices of different input factors for predicting different growth indicators were compared. The results show that the proposed method can reflect the influence of different input variables and observation information on each output variable and has good interpretability.

[0181] Each of the above test cases includes meteorological and agricultural records and sparse crop growth observation data from the start of crop growth to the last growth observation, as well as meteorological forecast data and agricultural management plan data for the forecast period. This data is processed through a trained embedding layer to obtain the initial token representation of the context block. Then, the forward computation of the model is performed sequentially:

[0182] (34)

[0183] In the formula, For the first segment within the target block day Predicted values ​​for crop growth indicators; The decoder has been trained. The completed predictor; This is the first encoder that has been trained.

[0184] The implementation of the various embodiments of the present invention is based on programmed processing through a system with processor functionality. Therefore, in practical engineering, the technical solutions and functions of the various embodiments of the present invention are encapsulated into various modules. Based on this reality, and building upon the above embodiments, the embodiments of the present invention provide a pre-training-fine-tuning-based general crop growth simulation system, which is used to execute a pre-training-fine-tuning-based general crop growth simulation method from the above method embodiments.

[0185] See Figure 7 The system includes:

[0186] The general data preparation module is used to build a general crop growth dataset;

[0187] The adaptation and fine-tuning module is used to construct a measured crop growth dataset for the target region and scenario, and introduces the LoRA low-rank fine-tuning mechanism to fine-tune the pre-trained prediction model of the specified growth index, so as to obtain a prediction model of the specified growth index adapted to the target region and scenario.

[0188] The crop growth simulation module is used to input the environmental information of the target crop in the target area and the scene within the upper and lower time periods into a specified growth index prediction model adapted to the target area and the scene, and obtain the specified growth index prediction value of the target crop within the upper and lower time periods.

[0189] The pre-training steps of the specified growth index prediction model include:

[0190] A pre-training framework for JEPA cross-crop high-dimensional representation is constructed. The JEPA pre-training framework is trained based on a general crop growth dataset. After training, a cross-crop high-dimensional representation predictor is obtained. The cross-crop high-dimensional representation predictor uses the environmental information of the target crop in the previous and next time blocks as input to predict the shared high-dimensional potential representation of the crop in the target time block.

[0191] For the task of predicting the growth indicators of a specified target crop, a prediction model for the specified growth indicators is constructed by combining a cross-crop high-dimensional representation predictor with a decoder built based on cross-attention, and trained on a general crop growth dataset.

[0192] It should be noted that the system embodiments provided by the present invention are used not only to implement the methods in the above method embodiments, but also to implement the methods in other method embodiments provided by the present invention. The only difference is that corresponding functional modules are set. The principle is basically the same as that of the above system embodiments provided by the present invention. As long as those skilled in the art can improve the modules in the above system embodiments by referring to the specific technical solutions in other method embodiments and combining technical features to obtain corresponding technical means and technical solutions composed of these technical means, on the basis of the above system embodiments, and on the premise of ensuring the practicality of the technical solutions, they can obtain corresponding system-like embodiments for implementing the methods in other method-like embodiments.

[0193] The method in this embodiment of the invention is implemented using an electronic device; therefore, it is necessary to introduce the relevant electronic device. For this purpose, embodiments of the present invention provide an electronic device, such as... Figure 8 As shown, the electronic device includes: at least one processor, a communication interface, at least one memory, and a communication bus, wherein the at least one processor, the communication interface, and the at least one memory communicate with each other via the communication bus. The at least one processor invokes logical instructions stored in the at least one memory to execute all or part of the steps of the methods provided in the foregoing method embodiments.

[0194] Furthermore, when the logical instructions in at least one of the aforementioned memories are implemented as software functional units and sold or used as independent products, they are stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or a part of the technical solution, is embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (a personal computer, server, or network device) to execute all or part of the steps of the methods described in the various method embodiments of the present invention. The aforementioned storage medium includes: USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks—various media for storing program code.

[0195] The system embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, located in one place, or distributed across multiple network units. The purpose of this embodiment is achieved by selecting some or all of the modules according to actual needs. Those skilled in the art will understand and implement this without any inventive effort.

[0196] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0197] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0198] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0199] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0200] Based on the same technical concept as the foregoing embodiments, the present invention provides a non-transitory computer-readable storage medium that stores computer instructions that cause the computer to execute a method.

[0201] 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 the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the technical solutions of the embodiments of the present invention.

Claims

1. A general crop growth simulation method based on pre-training and fine-tuning, characterized in that, include: A measured crop growth dataset for the target region and scenario is constructed, and the LoRA low-rank fine-tuning mechanism is introduced to fine-tune the pre-trained prediction model of the specified growth index, so as to obtain a prediction model of the specified growth index adapted to the target region and scenario. The environmental information of the target crop in the target area and the scene within the upper and lower time periods is input into the specified growth index prediction model adapted to the target area and the scene to obtain the specified growth index prediction value of the target crop within the upper and lower time periods. The pre-training steps of the specified growth index prediction model include: Construct a general crop growth dataset; A pre-training framework for JEPA cross-crop high-dimensional representation is constructed. The JEPA pre-training framework is trained based on a general crop growth dataset. After training, a cross-crop high-dimensional representation predictor is obtained. The cross-crop high-dimensional representation predictor uses the environmental information of the target crop in the previous and next time blocks as input to predict the shared high-dimensional potential representation of the crop in the target time block. For the task of predicting the growth indicators of a specified target crop, a prediction model for the specified growth indicators is constructed by combining a cross-crop high-dimensional representation predictor with a decoder built based on cross-attention, and trained on a general crop growth dataset. The steps for training the JEPA pre-training framework based on a general crop growth dataset include: The environmental information in the general crop growth dataset is processed through an embedding layer, and the initial token representation set is output. The JEPA context block and JEPA target block are constructed based on the initial token representation set. The JEPA context block is encoded by the first encoder in the dual encoder to obtain the context block representation, and the JEPA target block is encoded by the second encoder in the dual encoder to obtain the target block representation. The predictor outputs the predicted target block representation based on the context block representation output by the encoder, combines the context block representation position encoding with the JEPA target block position encoding, and calculates the upstream JEPA training loss function value based on the target block representation predicted by the predictor and the target block representation encoded by the encoder, and then backpropagates to update the trainable parameters. The process iterates until the preset accuracy is met or the preset maximum number of training rounds is reached. The embedding layer, the first encoder in the dual encoder, and the predictor in the trained JEPA pre-trained framework are retained to obtain a cross-crop high-dimensional representation predictor.

2. The general crop growth simulation method based on pre-training and fine-tuning as described in claim 1, characterized in that, The embedding layer contains embedding branches for three types of environmental information: meteorology, agricultural management, and crop growth, which process the corresponding types of environmental information respectively. The embedded branch takes the day as the smallest processing unit, and embeds and aggregates each variable in the corresponding type of environmental information to obtain the branch variable aggregation token representation. Add time position encoding and branch position encoding to the aggregated token representation of the branch variable to obtain the initial token representation.

3. The general crop growth simulation method based on pre-training and fine-tuning as described in claim 2, characterized in that, The steps for constructing the JEPA context block and the JEPA target block include: The JEPA context block contains the initial tokens corresponding to the daily meteorological and agricultural management environmental information for the entire time period of the upper and lower blocks and the target block, as well as the initial tokens corresponding to the crop growth environmental information for sparse dates within the context block; the JEPA target block contains the initial tokens corresponding to the daily crop growth environmental information for the target block.

4. The general crop growth simulation method based on pre-training and fine-tuning as described in claim 1, characterized in that, The steps for training based on a general crop growth dataset include: By specifying growth indicators, the target crop dataset is extracted from the general crop dataset, and the target crop dataset is processed using a cross-crop high-dimensional representation predictor to obtain the shared high-dimensional potential representation of the target crop within the target block time period. The decoder performs cross-attention-based decoding based on specified growth indicators and the shared high-dimensional latent representation of the target crop within the target block time period to obtain the predicted values ​​of each specified growth indicator. The reconstruction loss is calculated based on the predicted values ​​and observed values ​​of each specified growth indicator, and the decoder parameters are updated via backpropagation. The process is iterated until the preset accuracy is met or the preset maximum number of training rounds is reached, resulting in a trained prediction model for the specified growth index.

5. The general crop growth simulation method based on pre-training-fine-tuning as described in claim 4, characterized in that, The decoder performs cross-attention-based decoding based on specified growth indicators and the shared high-dimensional latent representation of the target crop within the target block time period to obtain the predicted values ​​of each specified growth indicator. The steps include: Based on the shared high-dimensional latent representation of the target block time period and the query generated by the target block time period, the value is generated based on the shared high-dimensional latent representation of the target block time period, and a special task key is generated based on the specified growth indicators and the target block time period. Cross attention is used to obtain the indicator-specific representation of each specified growth indicator of the target crop within the target block time period. Finally, through regression head processing, the predicted values ​​of each specified growth indicator of the target crop within the target block time period are obtained.

6. The general crop growth simulation method based on pre-training-fine-tuning as described in claim 1, characterized in that, The steps for fine-tuning a trained prediction model for a specified growth metric by introducing the LoRA low-rank fine-tuning mechanism include: A low-rank LoRA matrix is ​​introduced into all linear layers of the decoder of the pre-trained specified growth index prediction model. The trained specified growth index prediction model, after introducing the LoRA low-rank matrix, is fine-tuned based on the measured crop growth dataset, and finally outputs a specified growth index prediction model that is adapted to the target area and the scene. During the fine-tuning process, the original parameters of the trained specified growth index prediction model are frozen, and only the LoRA low-rank matrix introduced in all linear layers is updated.

7. A general crop growth simulation system based on pre-training and fine-tuning, characterized in that, include: The general data preparation module is used to build a general crop growth dataset; The adaptation and fine-tuning module is used to construct a measured crop growth dataset for the target region and scenario, and introduces the LoRA low-rank fine-tuning mechanism to fine-tune the pre-trained prediction model of the specified growth index, so as to obtain a prediction model of the specified growth index adapted to the target region and scenario. The crop growth simulation module is used to input the environmental information of the target crop in the target area and the scene within the upper and lower time periods into a specified growth index prediction model adapted to the target area and the scene, and obtain the specified growth index prediction value of the target crop within the upper and lower time periods. The pre-training steps of the specified growth index prediction model include: A pre-training framework for JEPA cross-crop high-dimensional representation is constructed. The JEPA pre-training framework is trained based on a general crop growth dataset. After training, a cross-crop high-dimensional representation predictor is obtained. The cross-crop high-dimensional representation predictor uses the environmental information of the target crop in the previous and next time blocks as input to predict the shared high-dimensional potential representation of the crop in the target time block. For the task of predicting the growth indicators of a specified target crop, a prediction model for the specified growth indicators is constructed by combining a cross-crop high-dimensional representation predictor with a decoder built based on cross-attention, and trained on a general crop growth dataset. The steps for training the JEPA pre-training framework based on a general crop growth dataset include: The environmental information in the general crop growth dataset is processed through an embedding layer, and the initial token representation set is output. The JEPA context block and JEPA target block are constructed based on the initial token representation set. The JEPA context block is encoded by the first encoder in the dual encoder to obtain the context block representation, and the JEPA target block is encoded by the second encoder in the dual encoder to obtain the target block representation. The predictor outputs the predicted target block representation based on the context block representation output by the encoder, combines the context block representation position encoding with the JEPA target block position encoding, and calculates the upstream JEPA training loss function value based on the target block representation predicted by the predictor and the target block representation encoded by the encoder, and then backpropagates to update the trainable parameters. The process iterates until the preset accuracy is met or the preset maximum number of training rounds is reached. The embedding layer, the first encoder in the dual encoder, and the predictor in the trained JEPA pre-trained framework are retained to obtain a cross-crop high-dimensional representation predictor.

8. An electronic device, characterized in that, The method includes a memory and a processor, wherein the memory stores program instructions that are executed by the processor, and the processor invokes the program instructions to execute a pre-training-fine-tuning general crop growth simulation method according to any one of claims 1 to 6.

9. A non-transitory computer-readable storage medium, characterized in that, The non-transitory computer-readable storage medium stores computer instructions that cause the computer to execute a pre-training-fine-tuning general crop growth simulation method as described in any one of claims 1 to 6.

Images