A method for simulating maize growth process based on a deep learning agent model
By using a deep learning-based proxy model to simulate the maize growth process, and by pre-training with synthetic data and fine-tuning with measured data, the time-consuming and data-dependent problems of existing maize growth simulation technologies are solved. This achieves efficient and accurate regional maize growth simulation that conforms to physiological laws.
Patent Information
- Application Number
- CN202611371409.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-09-04
- Publication Date
- 2026-11-10
AI Technical Summary
Existing maize growth simulation methods are time-consuming to calibrate parameters at the regional scale and rely on a large amount of high-cost field measurement data. Furthermore, machine learning methods lack constraints on crop growth mechanisms, leading to simulation results that deviate from physiological laws.
A deep learning-based agent model is adopted. By acquiring parameters of sensitive maize varieties, synthetic data is constructed for pre-training. The model is trained by combining supervised loss and physical constraint loss. Transfer fine-tuning is performed using a small amount of measured data to optimize variety parameters and achieve rapid and accurate regional maize growth simulation.
It reduces reliance on field measurement data, improves simulation efficiency and accuracy, ensures that simulation results conform to the physiological laws of maize growth, and achieves efficient parameter calibration and integrated simulation of growth processes at the regional scale.
Smart Images

Figure CN123069909A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of smart agriculture technology, specifically to a method for simulating the corn growth process based on a deep learning agent model. Background Technology
[0002] Existing maize growth simulation methods are mainly divided into two categories. One category is mechanistic crop growth models, such as APSIM and DSSAT, which simulate the growth and development dynamics of crops under meteorological, soil, and management conditions by describing physiological and ecological processes such as photosynthesis, transpiration, and nutrient cycling. The other category is machine learning methods, such as artificial neural networks, random forests, and support vector machines, which achieve rapid prediction by establishing nonlinear mapping relationships between meteorological, soil, management, and varietal characteristics and growth indicators. The two categories of methods have different focuses: the former emphasizes process mechanisms, while the latter focuses on data-driven efficient computation.
[0003] Crop modeling methods typically require calibration of variety parameters in actual maize production, a time-consuming process. When the application scale expands from a single point to a region, parameter calibration and model iteration need to be repeated for a large number of sample points, leading to an exponential increase in simulation workload. This makes it difficult to meet the needs of rapid simulation at the regional scale. Machine learning methods are purely data-driven and heavily rely on a large amount of high-quality field measurement data. However, maize field observations are costly, time-consuming, and have limited frequency, resulting in insufficient generalization ability under small sample sizes. Furthermore, the lack of constraints on crop growth mechanisms can easily lead to simulation results that are close to the observed values but whose growth processes deviate from physiological laws.
[0004] The information disclosed in the background section is only intended to enhance the understanding of the background of this disclosure, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention
[0005] The purpose of this invention is to provide a method for simulating the corn growth process based on a deep learning agent model, so as to solve the problems mentioned in the background art.
[0006] To achieve the above objectives, the present invention provides the following technical solution:
[0007] A method for simulating the corn growth process based on a deep learning agent model, characterized by the following specific steps:
[0008] S1. Obtain the parameters of sensitive maize varieties and their reasonable value ranges, perform random sampling to generate multiple combinations of variety parameters, combine each combination of variety parameters with meteorological conditions, soil conditions and field management conditions in different regions to construct maize growth simulation scenarios, input each simulation scenario into the APSIM crop model for batch simulation, and obtain synthetic data constrained by maize growth mechanism rules.
[0009] S2. The artificial intelligence agent model is pre-trained based on synthetic data. The artificial intelligence agent model is constructed using GRU layer, position encoding layer, Transformer encoder layer, adaptive pooling layer, shared feature extraction layer and multi-task output head. The input data in the synthetic data is used as the model input and the output data in the synthetic data is used as the model output. The pre-trained agent model is obtained by jointly optimizing the supervision loss and physical constraint loss.
[0010] S3. Obtain the measured field data of corn in the area to be simulated, and use the measured field data of corn to perform transfer fine-tuning on the pre-trained surrogate model to obtain the corrected surrogate model.
[0011] S4. Based on the calibration surrogate model, optimize the variety parameters of the maize to be simulated. Generate candidate variety parameter combinations in the variety parameter value space. Combine each candidate variety parameter combination with the meteorological data, soil data and field management data of the corresponding field as the model input. Input the model in batches for forward inference to obtain the simulation output corresponding to each candidate variety parameter combination. Construct a joint optimization objective function and search for the variety parameter combination that minimizes the joint optimization objective function to obtain the optimal variety parameters.
[0012] S5. Input the selected variety parameters and regional scale environmental data into the calibration surrogate model, perform forward inference, and obtain the regional maize growth simulation results.
[0013] S6. Output the simulation results of maize growth in the output area.
[0014] Furthermore, in S1, the sensitive variety parameters include accumulated temperature from emergence to the end of panicle differentiation, accumulated temperature from flag leaf to flowering, accumulated temperature from flowering to maturity, photoperiod sensitivity slope, potential grain weight, light energy utilization rate, leaf mortality rate slope, and maximum grain number coefficient. The field management conditions include low nitrogen, normal, and high nitrogen fertilizer treatments, as well as low density, normal density, and dense planting density. The synthetic data input includes meteorological data, soil data, field management data, and variety parameters. The synthetic data output includes leaf area index time series, aboveground biomass time series, growth period time series, and final yield.
[0015] Furthermore, the GRU layer is used to extract local temporal features between adjacent time steps during maize growth, the position encoding layer is used to inject temporal position information into the feature sequences extracted by GRU, the Transformer encoder layer is used to establish long-distance temporal dependencies between different growth stages, the adaptive pooling layer is used to extract global feature vectors of fixed dimensions, the shared feature extraction layer is used to extract shared deep growth features, and the multi-task output head includes a yield head, a leaf area index temporal head, a aboveground biomass temporal head, and a growth period temporal head. The yield head outputs the final yield scalar, and each temporal head outputs the daily time series of the corresponding growth state variables.
[0016] Furthermore, the total loss of model pre-training is a weighted sum of the supervision loss and the physical constraint loss. The supervision loss is a weighted combination of four losses: yield, growth period, leaf area index, and aboveground biomass. The yield loss is calculated using mean squared error.
[0017] Furthermore, the physical constraint loss is a weighted sum of the leaf area index senescence period decline constraint loss, the growth period monotonically increasing constraint loss, and the aboveground biomass monotonically increasing constraint loss. The leaf area index senescence period decline constraint is used to penalize the predicted result of abnormal rebound after the peak of the leaf area index. The growth period monotonically increasing constraint is used to penalize the predicted result of regression in the growth period stage. The aboveground biomass monotonically increasing constraint is used to penalize the predicted result of negative biomass growth.
[0018] Furthermore, the transfer fine-tuning adopts a layered freeze fine-tuning strategy, freezing the parameters of the GRU layer, Transformer encoder layer, positional encoding layer and adaptive pooling layer in the pre-trained model, and only updating the parameters of the shared feature extraction layer and the output heads of each task. The fine-tuning loss function is consistent with the pre-training loss function in S2.
[0019] Furthermore, in S4, the joint optimization objective function is a weighted sum of leaf area index, aboveground biomass, and the mean square error of simulated and measured values of the growth period. By traversing the value space of variety parameters, the optimal variety parameters that minimize the value of the joint optimization objective function are selected.
[0020] Compared with the prior art, the beneficial effects of the present invention are:
[0021] 1. This invention utilizes crop growth models to generate synthetic data with mechanistic constraints in batches, and pre-trains artificial intelligence proxy models, thereby significantly reducing the dependence on large-scale field measurement data and solving the problems of high cost and long cycle of field data acquisition. It constructs a joint optimization objective function based on leaf area index, aboveground biomass, and multi-phenotypic time series of growth period, improves forward inference efficiency, and achieves efficient and accurate calibration of variety parameters, solving the problem of time-consuming parameter calibration at the regional scale. The trained proxy model can transform the complex calculations of traditional crop growth models into millisecond-level forward inference. When applied at the regional scale, it is not necessary to repeat the complete crop model calculation, effectively reducing computational overhead and realizing integrated simulation of growth process and final yield.
[0022] 2. Based on pre-training, physical constraints such as monotonically increasing biomass, monotonically increasing growth period, and decreasing leaf area during senescence are introduced to ensure that the output of the surrogate model conforms to the physiological laws of maize growth. This effectively reduces problems such as abnormal growth period sequence and unreasonable growth state variables, and improves the process rationality of the simulation results. By using a small amount of field measurement data, the pre-trained model is fine-tuned through stratified freeze-transfer, so that the surrogate model can retain the universal growth laws of the crop model while learning the deviations that the crop model could not describe under real production conditions, thus taking into account both mechanism and adaptability. Attached Figure Description
[0023] Figure 1 This is a schematic diagram of the overall framework structure of the present invention;
[0024] Figure 2 This is a simulated scatter plot of aboveground biomass (AGB) in the independent validation set of this invention.
[0025] Figure 3 A scatter plot of simulated leaf area index (LAI) for independent verification of this invention;
[0026] Figure 4 This is a scatter plot of the reproductive stage simulation for the independent validation set of this invention.
[0027] Figure 5 This is a simulated scatter plot of the final yield of the independent validation set of this invention.
[0028] Figure 6 This is a dynamic comparison diagram of the reproductive stages of a single sample according to the present invention;
[0029] Figure 7 This is a dynamic comparison chart of leaf area index for a single sample according to the present invention;
[0030] Figure 8 This is a dynamic comparison diagram of aboveground biomass of a single sample according to the present invention. Detailed Implementation
[0031] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments.
[0032] Example:
[0033] Please see Figure 1-8 The present invention provides a technical solution:
[0034] A method for simulating the corn growth process based on a deep learning agent model, comprising the following steps:
[0035] S1. Obtain the parameters of sensitive maize varieties and their reasonable value ranges. Randomly sample within the reasonable value range to generate multiple combinations of variety parameters. Combine each combination of variety parameters with meteorological conditions, soil conditions and field management conditions in different regions to construct maize growth simulation scenarios covering multiple growth environments. Input each simulation scenario into the APSIM crop model (Agricultural Production Systems Simulator) for batch simulation to obtain synthetic data constrained by maize growth mechanism rules.
[0036] By utilizing the APSIM crop model, simulated data covering a wide range of growth scenarios can be quickly generated under known parameter space and environmental conditions. This provides sufficient training samples with physical prior knowledge for the pre-training of subsequent deep learning models, thereby reducing the dependence on the amount of measured data.
[0037] Specifically, measured maize data from various parks and farmers' fields over the years were collected. This measured maize growth data was used to calibrate maize variety parameters and screen for sensitive varietal parameters that significantly affect growth and yield. Sensitive varietal parameters can include accumulated temperature from emergence to the end of ear differentiation, accumulated temperature from flag leaf to flowering, accumulated temperature from flowering to maturity, photoperiod sensitivity slope, potential grain weight, light energy utilization rate, leaf mortality rate slope, and maximum grain number coefficient, but are not limited to these data. These parameters reflect the genetic characteristics of different maize varieties. For example, accumulated temperature parameters determine the length of the growth period, and light energy utilization rate affects the rate of dry matter accumulation. Simultaneously, reasonable value ranges for each sensitive varietal parameter were determined based on actual conditions. Random sampling was conducted within the determined range of sensitive varietal parameter values to generate multiple sets of varietal parameter combinations. Specifically, the Latin hypercube sampling method was used to generate 500 parameter combinations. Latin hypercube sampling is a stratified random sampling method that can uniformly cover the entire parameter range in a multidimensional parameter space with fewer sample points, improving sampling efficiency. By combining meteorological, soil, and field management conditions in different regions, a maize growth simulation scenario covering multiple growth environments was constructed. Field management conditions included nitrogen fertilizer treatments such as low nitrogen, normal, and high nitrogen, as well as planting densities such as low density, normal density, and dense planting. Each simulation scenario was input into the APSIM crop model for batch simulation to obtain corresponding maize growth process simulation data. The simulation data included meteorological data (daily maximum and minimum temperatures, precipitation, and radiation), soil data (bulk density, field water holding capacity, wilting coefficient, pH value, saturated water content, and organic carbon content), field management data (density, nitrogen application rate, sowing and harvesting dates), and variety parameters at the input end, as well as maize growth status variables at the output end. Output variables included the leaf area index time series representing the photosynthetic process, the aboveground biomass time series representing dry matter distribution, the growth period time series representing the maize growth cycle, and the final yield.
[0038] By randomly sampling multiple sets of parameters within a reasonable parameter space and using crop growth models for batch simulation, a large-scale synthetic dataset constrained by maize growth mechanism rules is constructed. The synthetic dataset serves as pre-training data for subsequent artificial intelligence models. By using crop models to generate synthetic data, the problem of pure data-driven methods being highly dependent on small sample data and having poor generalization ability is solved, providing a low-cost, large-scale physical prior knowledge base for subsequent model training.
[0039] S2. The AI agent model is pre-trained based on synthetic data. The AI agent model is constructed using GRU layer, position encoding layer, Transformer encoder layer, adaptive pooling layer, shared feature extraction layer and multi-task output head. The input data in the synthetic data is used as the model input and the output data in the synthetic data is used as the model output. The pre-trained agent model is obtained by jointly optimizing the supervision loss and physical constraint loss.
[0040] In order to obtain an artificial intelligence agent model that can quickly and accurately simulate the corn growth process, the model needs to learn the mapping relationship in the synthetic data generated by the crop model, and also needs to incorporate the physical laws of corn growth to ensure the physiological rationality of the simulation results.
[0041] The model structure is designed as follows: The GRU layer is used to extract local temporal features between adjacent time steps during maize growth, capturing the continuous dynamic impact of environmental factors such as weather and soil on crop growth in the short term. The positional encoding layer injects temporal positional information into the feature sequences extracted by GRU, enabling the model to distinguish inputs from different growth dates. The Transformer encoder layer receives the GRU output with positional information and uses a self-attention mechanism to establish long-distance temporal dependencies between different growth stages, capturing stage transitions throughout the entire growth period. The adaptive pooling layer performs adaptive average pooling on the temporal features output by the Transformer, extracting a fixed-dimensional global feature vector. The shared feature extraction layer consists of multiple fully connected layers and activation functions, performing nonlinear transformations on the global feature vector to extract shared deep growth features. The multi-task output head includes a yield head and three temporal heads for leaf area index, aboveground biomass, and growth period. The yield head outputs the final yield scalar at the end of the entire growth cycle, and each temporal head outputs the daily time series of the corresponding growth state variables throughout the year. By combining GRU and Transformer, the maize growth process is modeled.
[0042] To enable the model to simultaneously learn the implicit process knowledge inherent in the crop growth model and the explicit constraints of maize growth patterns, a multi-task weighted loss function is adopted, with a total loss of... It is composed of a weighted average of monitoring loss and physical constraint loss: .
[0043] The supervision loss adopts a multi-task weighted loss function. The total supervision loss is composed of a weighted average of the yield loss and the time-series losses of the three growth phenotypes. The formula for calculating the supervision loss is as follows: .in, For the total supervisory loss, For production loss, For the loss of reproductive years, For leaf area index loss, For aboveground biomass loss, , , , These are the loss weights for the corresponding output variables. Yield is the final simulation target, so it is assigned the highest weight of 1. Aboveground biomass is directly related to yield, so it is assigned a weight of 0.8. Leaf area index is an important canopy indicator, but slightly lower than yield and biomass, so it is assigned a weight of 0.7. Growth period is relatively easy to learn and is an ordered stage variable, so it is assigned a weight of 0.5. The weights are set based on the importance and learning difficulty of each phenotype in crop growth simulation. Yield and biomass directly reflect the final output, so they are assigned higher weights to guide the model to prioritize learning these key indicators.
[0044] For output quantities (scalars), mean square error is used for calculation: Where N is the batch sample size. Let i be the predicted output value for the i-th sample. Let be the actual output value of the i-th sample. The mean squared error measures the squared difference between the predicted value and the actual value. The smaller the value, the more accurate the prediction.
[0045] For the three time series outputs—growth period, leaf area index, and aboveground biomass—the mask mean square error is used for calculation. Losses are calculated only for valid time steps within the growing season, ignoring invalid values from outside the growing season. Taking leaf area index as an example, the loss calculation formula is as follows: Where N is the batch sample size. Let be the set of growing season time steps for the i-th sample, and t be the index of the time step within the growing season. The total number of effective time steps during the growing season. Let i be the predicted leaf area index value for the i-th sample at time t. Let be the true value of the leaf area index of the i-th sample at time step t. The masking mechanism avoids the interference of null values during non-growing seasons on model training, allowing the model to focus on learning the growth process that has practical significance.
[0046] The physical constraint loss aims to ensure that the model output conforms to the physiological and ecological laws of maize growth, and consists of the following three components: .
[0047] in, The total loss due to physical constraints. The loss is constrained by the decline in leaf area index during senescence. The monotonically increasing constraint loss during the reproductive period For the monotonically increasing biomass constraint loss, , , These are the weight coefficients for the corresponding physical constraint terms. Based on the complexity of the three phenotypic curves, they are assigned values of 0.05, 0.04, and 0.03, respectively. The weights are set relatively small to apply soft constraints to the model without significantly affecting the dominance of the supervised loss, thus guiding its output to conform to physical laws.
[0048] The formula for calculating the loss due to the senescence constraint of leaf area index is: .in, Let T be the time step in which the leaf area index reaches its peak in the i-th sample, and T be the total number of time steps. , are the predicted leaf area index values for the i-th and i+1-th samples at time step t, respectively. A penalty is applied to time steps where the leaf area index rebounds after the peak. After reaching the peak, the leaf area index should decrease monotonically with canopy aging. If the model predicts a rebound, a positive penalty term is generated to force the model to learn the correct aging trend.
[0049] The formula for calculating the monotonically increasing constraint loss during the reproductive period is: . , These are the predicted fertility period values for the i-th and i+1-th samples at time step t, respectively. This function penalizes time steps where the fertility period regresses. From seedling emergence to maturity, the fertility period of corn is strictly monotonically increasing and cannot regress, ensuring that the model output conforms to physiological laws.
[0050] The formula for calculating the loss under the monotonically increasing biomass constraint is: . , are the predicted aboveground biomass values for the i-th and i+1-th samples at time step t, respectively. This function penalizes time steps where aboveground biomass shows negative growth. Throughout the entire growth period of maize, aboveground biomass continuously accumulates with photosynthesis and should theoretically always maintain a monotonically increasing trend. This constraint prevents the model from generating unreasonable biomass declines.
[0051] The GRU-Transformer hybrid architecture is designed to balance the temporal modeling capabilities of local dynamics and global dependencies. Physical constraint loss is introduced, and the physiological laws of maize growth are added as explicit constraints to the training objective. This solves the problem that the simulation results of pure data-driven models do not conform to physiological laws. A multi-task learning framework is adopted to predict multiple key growth phenotypes and yields simultaneously, realizing the integrated simulation of the growth process and the final yield.
[0052] S3. Obtain field-measured data of corn in the area to be simulated, and use the field-measured data of corn to perform transfer fine-tuning on the pre-trained surrogate model to obtain the corrected surrogate model.
[0053] Using a small amount of real field observation data, the pre-trained model is calibrated to adapt to biases that crop models cannot fully describe in real production environments, such as the effects of unmodeled factors like pests, diseases, weeds, and extreme weather.
[0054] Specifically, the field-measured data of maize in the area to be simulated are obtained, including time-series observations of leaf area index, aboveground biomass, growth period, and final yield at different growth stages, as well as corresponding meteorological, soil, and field management data. At the same time, the variety parameters corresponding to the maize variety to be simulated are obtained, and the variety parameters are initially calibrated using the historical measured phenotypic data of the field. The calibrated variety parameters, together with the measured meteorological, soil, and management data, constitute the model input for the fine-tuning stage.
[0055] The pre-trained model weights saved in S2 are loaded and used as the initial parameters for fine-tuning. Considering the limited amount of measured data, a layered freeze fine-tuning strategy is adopted: the parameters of the lower-level feature extraction modules such as the GRU layer, Transformer encoder layer, positional encoding layer, and adaptive pooling layer in the pre-trained model are frozen, and only the parameters of the upper-level shared feature extraction layer and the output heads of each task are updated. This can completely preserve the general temporal dynamics of crop growth and cross-variety mechanism priors learned in the pre-training stage, and only allow the top-level network to adapt to the real phenotypic distribution of specific varieties and plots. The loss function for fine-tuning is consistent with the loss function of pre-training in S2.
[0056] By adopting a stratified freezing transfer learning strategy, even with only a small amount of measured data, we can utilize the rich prior knowledge learned by the pre-trained model and effectively correct the systematic bias of the model for specific fields, thus solving the problem of insufficient generalization ability of pure data-driven methods under small sample conditions.
[0057] S4. Based on the calibration surrogate model, the variety parameters of the maize to be simulated are optimized. Candidate variety parameter combinations are generated in the variety parameter value space. Each candidate variety parameter combination is combined with the meteorological data, soil data and field management data of the corresponding field as the model input. The data are then batch-input into the calibration surrogate model for forward inference to obtain the simulation output corresponding to each candidate variety parameter combination. A joint optimization objective function based on leaf area index, aboveground biomass and multi-phenotypic time series of growth period is constructed. The variety parameter combination that minimizes the joint optimization objective function is searched to obtain the optimal variety parameters.
[0058] By utilizing a trained surrogate model, the parameters of maize varieties can be calibrated efficiently, significantly improving the efficiency of parameter calibration. The specific process is as follows: Within the value space of the variety parameters, multiple candidate variety parameter combinations are generated using Latin hypercube sampling. Each candidate variety parameter combination is then concatenated with the corresponding meteorological data, soil data, and field management data of the field as the model input. This input is then batch-input into the calibration surrogate model obtained from S3, and forward inference yields the simulated output corresponding to each candidate variety parameter.
[0059] The formula for calculating the objective function of multi-phenotype joint optimization is as follows:
[0060]
[0061] in, Combination of variety parameters The corresponding multi-phenotype joint optimization objective value, For the combination of varieties to be optimized, , , These are the measured leaf area index, aboveground biomass, and growth phenotype sequences during the growth period, respectively. , , Combining variety parameters The simulation output is obtained after inputting the meteorological data, soil data, and field management data of the corresponding field into the calibration proxy model. , , These are the weighting coefficients for each phenotypic task and the weighting of aboveground biomass. Assign a value of 1 to the leaf area index weight. Assigned a value of 0.7, reproductive option weight The weight is assigned a value of 0.5, and the weight setting is consistent with the monitoring loss in S2. The aboveground biomass directly reflects the accumulation of dry matter and has the strongest correlation with the final yield, so it is given the highest weight.
[0062] The parameter optimization process involves searching within the parameter value space Θ to achieve the overall objective. Minimal parameter combination: .in, To optimize the variety parameters, this embodiment uses the NSGA-II algorithm for multi-objective optimization search.
[0063] Using a surrogate model for parameter calibration greatly shortens the simulation time and improves simulation efficiency. A joint optimization objective based on three key growth phenotypes—leaf area index, aboveground biomass, and growth period—was constructed, which simultaneously constrains the calibration of variety parameters by three physiological processes: canopy formation, dry matter accumulation, and growth period, thereby improving the reliability and accuracy of the calibration.
[0064] S5. Input the selected variety parameters and regional-scale environmental data into the calibration surrogate model, perform forward inference, and obtain the regional maize growth simulation results.
[0065] The calibrated variety parameters are applied to maize growth simulation at the regional scale. The optimized variety parameters and regional meteorological, soil and management data are input into the calibration surrogate model. The model performs forward inference and quickly outputs daily leaf area index, aboveground biomass, growth period time series and final yield.
[0066] S6. Output the simulation results of maize growth in the output region.
[0067] Construct a regional corn production monitoring, yield forecasting, and agricultural digital twin system.
[0068] This invention was validated using an independent dataset. The results showed that the R² values for total leaf area index, aboveground biomass, growth period, and yield were 0.99, 0.71, 0.93, and 0.54, respectively; the RMSE values were 32.05, 0.98, 0.64, and 94.25, respectively; and the relative errors (RE) were 23.96%, 5.12%, 7.44%, and 13.81%, respectively. Compared with traditional crop models, this invention significantly improves simulation efficiency.
[0069] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.
[0070] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented in software, the above embodiments can be implemented, in whole or in part, as a computer program product. Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented by electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution.
[0071] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment, depending on actual needs.
[0072] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.
Claims
1. A method for simulating the growth process of maize based on a deep learning agent model, characterized in that, The specific steps include: S1. Obtain the parameters of sensitive maize varieties and their reasonable value ranges, perform random sampling to generate multiple combinations of variety parameters, combine each combination of variety parameters with meteorological conditions, soil conditions and field management conditions in different regions to construct maize growth simulation scenarios, input each simulation scenario into the APSIM crop model for batch simulation, and obtain synthetic data constrained by maize growth mechanism rules. S2. The artificial intelligence agent model is pre-trained based on synthetic data. The artificial intelligence agent model is constructed using GRU layer, position encoding layer, Transformer encoder layer, adaptive pooling layer, shared feature extraction layer and multi-task output head. The input data in the synthetic data is used as the model input and the output data in the synthetic data is used as the model output. The pre-trained agent model is obtained by jointly optimizing the supervision loss and physical constraint loss. S3. Obtain the measured field data of corn in the area to be simulated, and use the measured field data of corn to perform transfer fine-tuning on the pre-trained surrogate model to obtain the corrected surrogate model. S4. Based on the calibration surrogate model, optimize the variety parameters of the maize to be simulated. Generate candidate variety parameter combinations in the variety parameter value space. Combine each candidate variety parameter combination with the meteorological data, soil data and field management data of the corresponding field as the model input. Input the model in batches for forward inference to obtain the simulation output corresponding to each candidate variety parameter combination. Construct a joint optimization objective function and search for the variety parameter combination that minimizes the joint optimization objective function to obtain the optimal variety parameters. S5. Input the selected variety parameters and regional scale environmental data into the calibration surrogate model, perform forward inference, and obtain the regional maize growth simulation results. S6. Output the simulation results of maize growth in the output area.
2. The method for simulating the corn growth process based on a deep learning agent model according to claim 1, characterized in that, In S1, the sensitive variety parameters include accumulated temperature from emergence to the end of panicle differentiation, accumulated temperature from flag leaf to flowering, accumulated temperature from flowering to maturity, photoperiod sensitivity slope, potential grain weight, light energy utilization rate, leaf mortality rate slope, and maximum grain number coefficient. The field management conditions include low nitrogen, normal, and high nitrogen fertilizer treatments, as well as low density, normal density, and dense planting density. The synthetic data input includes meteorological data, soil data, field management data, and variety parameters. The synthetic data output includes leaf area index time series, aboveground biomass time series, growth period time series, and final yield.
3. The method for simulating the corn growth process based on a deep learning agent model according to claim 1, characterized in that, The GRU layer is used to extract local temporal features between adjacent time steps during maize growth. The position encoding layer is used to inject temporal position information into the feature sequences extracted by GRU. The Transformer encoder layer is used to establish long-distance temporal dependencies between different growth stages. The adaptive pooling layer is used to extract global feature vectors of fixed dimensions. The shared feature extraction layer is used to extract shared deep growth features. The multi-task output head includes a yield head, a leaf area index temporal head, a aboveground biomass temporal head, and a growth period temporal head. The yield head outputs the final yield scalar, and each temporal head outputs the daily time series of the corresponding growth state variables.
4. The method for simulating the corn growth process based on a deep learning agent model according to claim 1, characterized in that, The total loss of the model pre-training is a weighted sum of the supervision loss and the physical constraint loss. The supervision loss is a weighted combination of four losses: yield, growth period, leaf area index, and aboveground biomass. The yield loss is calculated using mean squared error.
5. The method for simulating the corn growth process based on a deep learning agent model according to claim 1, characterized in that, The physical constraint loss is a weighted sum of the leaf area index senescence period decline constraint loss, the growth period monotonically increasing constraint loss, and the aboveground biomass monotonically increasing constraint loss. The leaf area index senescence period decline constraint is used to penalize the predicted result of abnormal rebound after the peak of the leaf area index. The growth period monotonically increasing constraint is used to penalize the predicted result of regression in the growth period. The aboveground biomass monotonically increasing constraint is used to penalize the predicted result of negative biomass growth.
6. The method for simulating the maize growth process based on a deep learning agent model according to claim 1, characterized in that, The transfer fine-tuning adopts a hierarchical freeze fine-tuning strategy, freezing the parameters of the GRU layer, Transformer encoder layer, positional encoding layer and adaptive pooling layer in the pre-trained model, and only updating the parameters of the shared feature extraction layer and the output heads of each task. The fine-tuning loss function is consistent with the pre-training loss function in S2.
7. The method for simulating the corn growth process based on a deep learning agent model according to claim 1, characterized in that, In step S4, the joint optimization objective function is a weighted sum of leaf area index, aboveground biomass, and the mean square error of simulated and measured values during the growth period. By traversing the value space of variety parameters, the optimal variety parameters that minimize the value of the joint optimization objective function are selected.
