A Crop Planting Management Method Based on a Deep Learning Agent Model

Abstract

This invention discloses a crop planting management method based on a deep learning agent model. It relates to the field of crop planting management technology and includes: acquiring crop parameters, soil parameters, and meteorological parameters and performing sensitivity analysis; calibrating a crop growth model based on the sensitivity analysis results to obtain optimal variety parameters; constructing various agricultural planting management scenarios and multi-objective optimization problems; simulating crop growth using the optimal variety parameters to obtain the Pareto optimal solution set of the management scheme; constructing a dataset using the input and output parameters of the calibrated crop model and training a deep learning agent model; and designing the TSPO-SU assimilation algorithm based on the trained deep learning agent model to assimilate remote sensing data with the deep learning agent model, inverting the environmental parameters of crop growth, and obtaining a refined planting management strategy. This invention can efficiently generate large-scale, refined, and dynamically adjustable crop planting management decision schemes.

Description

Technical Field

[0001] This invention relates to the field of crop planting management technology, and more specifically to a crop planting management method based on a deep learning agent model. Background Technology

[0002] In recent years, deep learning technology, with its powerful parallel computing capabilities, has provided a new way to accelerate the efficient assimilation of crop growth models and remote sensing data, and is expected to drive the development of precision planting management solutions towards a more efficient and practical direction.

[0003] Despite advancements in existing crop growth models, remote sensing, or a combination of both, planting management decision-making schemes still suffer from significant shortcomings in practical applications, hindering their large-scale, precision implementation. A detailed analysis follows: Deficiencies of existing technology: Remote sensing-based solutions rely heavily on empirical formulas to retrieve crop or soil parameters, failing to adequately reflect physical mechanisms. Furthermore, the retrieval process is susceptible to environmental factors such as changes in vegetation cover, resulting in limited reliability in estimating key soil conditions like moisture and nutrients, and an unstable decision-making foundation.

[0004] Real-time sensor-based solutions, while providing accurate location data, require the deployment of a large number of sensors to achieve wide coverage. This results in extremely high hardware costs, maintenance difficulties, and energy consumption, leading to poor economic efficiency and scalability, making them difficult to promote in ordinary farmland.

[0005] Schemes based on mechanistic crop growth models have a solid physical foundation, but the calibration of model parameters is complex and labor-intensive. Furthermore, their simulations are usually designed for homogeneous conditions, making it difficult to effectively characterize and respond to the inherent spatial heterogeneity in the field, resulting in insufficient universality of the generated schemes.

[0006] The scheme of assimilating crop growth models with remote sensing data: Although it can integrate model mechanisms and observational data and correct errors, traditional assimilation algorithms have extremely high computational complexity and low operating efficiency. When dealing with large-scale farmland areas, it is difficult to complete detailed and rapid simulations and decisions within the requirements of agricultural seasons, thus limiting its practicality.

[0007] Therefore, how to construct a technical system that can efficiently generate large-scale, refined, and dynamically adjustable crop planting management decision-making schemes, so as to achieve an order-of-magnitude improvement in computational efficiency while ensuring the rationality and accuracy of decision-making, and meet the requirements of timeliness and operability in actual production, is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0008] In view of this, the present invention provides a crop planting management method based on a deep learning surrogate model. The method uses a crop growth model to determine the optimal planting management strategy, and at the same time uses a deep learning algorithm to generate a surrogate model of the crop growth model, thereby accelerating the assimilation of the crop growth model with remote sensing data, retrieving soil conditions, and determining refined agricultural management measures to achieve the goal of optimizing production management.

[0009] To achieve the above objectives, the present invention adopts the following technical solution: A crop planting management method based on a deep learning agent model includes: Acquire crop parameters, soil parameters, and meteorological parameters and perform sensitivity analysis; Based on the results of sensitivity analysis, the crop growth model is calibrated to obtain the optimal variety parameters. Various agricultural planting management scenarios and multi-objective optimization problems are constructed. The crop growth is simulated using the optimal variety parameters to obtain the Pareto optimal solution set of the management scheme. Construct a deep learning agent model, generate a training dataset based on the input and output of a calibrated crop growth model, and train the deep learning agent model. Based on a trained deep learning agent model, the TSPO-SU assimilation algorithm is designed to assimilate remote sensing data with the deep learning agent model, invert environmental parameters for crop growth, determine the optimal planting management scheme from the Pareto solution set, and implement variable fertilization and variable irrigation measures based on the optimal planting management scheme to obtain a refined planting management strategy.

[0010] Preferably, the crop growth model is the WOFOST model. Based on the WOFOST model, a multidimensional parameter global sensitivity analysis mechanism is constructed by combining Latin hypercube sampling and Spearman rank correlation coefficient. The input parameters of the WOFOST model are analyzed to obtain the sensitivity ranking of each parameter. The results are compared with preset values ​​to obtain sensitive and insensitive parameters.

[0011] Preferably, the WOFOST model is dynamically calibrated using an adaptive particle swarm optimization algorithm.

[0012] Preferably, the process of obtaining the Pareto optimal solution set of the management scheme includes: By using crop growth models to simulate crop growth processes and yields under different management modes, yield data corresponding to each management scenario is obtained. Combined with local agricultural production constraints, a comprehensive benefit assessment of different management modes is conducted, a multi-objective optimization problem is constructed, and the Pareto optimal solution set of the management scheme is obtained by solving the problem. Based on the actual operating conditions of the farm, the optimal planting management scheme is determined from the Pareto solution set.

[0013] Preferably, the deep learning agent model adopts the KalmanNet architecture, specifically including: an input layer, an output layer, and a state prediction and update layer; the input layer includes two types of parameters: static parameters and dynamic time series parameters; the output layer includes the output leaf area index (LAI), growth and development stage (DVS), biomass (BIO), and soil moisture (SM) of the WOFOST model; the state prediction and update layer includes a state prediction submodule, an observation mapping submodule, an adaptive Kalman gain generation submodule, and a state update submodule.

[0014] Preferably, a dataset is generated using a calibrated WOFOST model. The dataset is divided into an input parameter dataset and an output parameter dataset. The input parameter training set is further divided into a static parameter dataset and a dynamic parameter dataset. The static parameter dataset contains the sensitivity parameters of the WOFOST model, and the dynamic parameter dataset contains the daily meteorological input parameters of the WOFOST model. The output parameter dataset contains the output generated by the model based on the input parameters. The dataset is then divided into a training set, a validation set, and a test set. KalmanNet is trained on the dataset, and the validation process employs 5-fold time series cross-validation.

[0015] Preferably, the TSPO-SU assimilation algorithm includes: A continuous data assimilation algorithm is used to initially optimize the static input parameters of KalmanNet in order to reduce the initial bias of the model. During the forward simulation of the model, a sequential data assimilation algorithm is introduced to update the crop state variables corresponding to the observation time. After completing one round of sequential state updates, a continuous data assimilation algorithm is used again to correct the static input parameters of KalmanNet, further suppressing error accumulation and improving the overall simulation consistency; crop growth environment parameters are obtained and crop growth state simulation is optimized.

[0016] Preferably, the continuous data assimilation algorithm is a four-dimensional variational data assimilation algorithm, which optimizes the static input parameters of KalmanNet by constructing an objective function to minimize the difference between the KalmanNet simulation results and the remote sensing observation results. The objective function expression for the four-dimensional variational data assimilation algorithm is as follows:

[0017] in: The objective function is... This represents the vector of sensitive parameters of the model to be optimized; This represents the background parameters obtained from model calibration; The background error covariance matrix; The observation error covariance matrix; Indicates the model at time... The simulated output; This represents the remote sensing inversion observation value at the corresponding time; when the objective function When the minimum value is obtained, the corresponding parameter These are the optimized model sensitivity parameters; The sequential data assimilation algorithm is a fast nutrient assimilation algorithm. When obtaining remote sensing observation data, it updates crop state variables hourly to improve KalmanNet's response to spatial heterogeneity and short-term changes.

[0018] As can be seen from the above technical solution, compared with the prior art, the present invention discloses a crop planting management method based on a deep learning agent model, with the following technical effects: Quantization performance improvement: The deep learning proxy model used in this invention can maintain more than 98% of the coefficient of determination (R²) with the original mechanism model. 2 Under the premise of ), the time efficiency of regional simulation is improved by more than 750 times, achieving an order-of-magnitude leap in computational efficiency, and the improvement in computational efficiency can be significantly improved with the performance of computer graphics processing unit (GPU).

[0019] Enhanced decision-making capabilities: The designed assimilation algorithm enables the system to integrate the latest remote sensing observations in near real-time, dynamically generate and adjust management prescriptions such as irrigation and fertilization, and improve the decision response speed from the "day" level to the "hour" level, significantly enhancing the adaptability and accuracy of management.

[0020] Expanded application scope: The breakthrough in efficiency makes it possible to conduct daily, high-resolution growth simulation and scheme generation for tens of thousands of acres of farmland, which solves the limitation of traditional methods that can only be used for single-point or small-scale studies, and powerfully promotes the transition of precision agriculture technology from theory to large-scale field application. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0022] Figure 1 The overall flowchart provided for this invention.

[0023] Figure 2 The KalmanNet structure diagram provided by this invention.

[0024] Figure 3 The flowchart of the TSPO-SU assimilation algorithm provided by this invention is shown.

[0025] Figure 4 A comparison chart of cotton biomass prediction results using the 4DVar algorithm for data assimilation based on the traditional crop model (WOFOST) (a) and the deep learning proxy model (DL-WF) (b) provided by this invention.

[0026] Figure 5 The figure shows a comparison of the prediction results of the RNA algorithm based on the traditional crop model (WOFOST) (a) and the biomass prediction results of the TSPO-SU algorithm based on the deep learning proxy model (DL-WF) (b) provided by this invention after data assimilation.

[0027] Figure 6 The image shows the prediction results of tobacco leaf weight, stem weight, and biomass based on the traditional crop model (WOFOST) and the 4DVar algorithm for data assimilation, as provided by this invention.

[0028] Figure 7 The image shows the prediction results of tobacco leaf weight, stem weight, and biomass using the 4DVar algorithm based on the deep learning agent model (DL-WF) provided by this invention.

[0029] Figure 8 The image shows the predicted results of tobacco leaf weight, stem weight, and biomass using the RNA algorithm for data assimilation in the traditional crop model (WOFOST) provided by this invention.

[0030] Figure 9 The image shows the prediction results of tobacco leaf weight, stem weight, and biomass based on the TSPO-SU algorithm with deep learning agent model (DL-WF) provided by this invention.

[0031] Figure 10 Soil available nitrogen prediction map provided by the present invention.

[0032] Figure 11 This invention provides a soil available phosphorus prediction map.

[0033] Figure 12 This invention provides a prediction map of available potassium in soil.

[0034] Figure 13 The fertilizer prescription diagram for the Nishimura base provided by this invention.

[0035] Figure 14 Yield distribution maps of prescription and control fields at the Nishimura base provided for this invention.

[0036] Figure 15 The method flowchart provided by the present invention. Detailed Implementation

[0037] 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, and 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.

[0038] In a specific embodiment of the present invention, such as Figure 1 As shown, crop growth models are used to determine optimal planting and management strategies. Simultaneously, deep learning algorithms are used to generate surrogate models of the crop growth models, accelerating the assimilation of crop growth models with remote sensing data and retrieving soil conditions to determine refined agricultural management measures, thereby optimizing production management. Specifically, this includes: S1: Localize the crop growth model. Perform sensitivity analysis to determine sensitive crop and soil parameters. Simultaneously, collect physiological parameters (such as biomass, leaf area index, growth stage, etc.) of the crop variety in previous years and experimental surveys, as well as soil and meteorological parameters of the sample plots. Use the collected parameters to calibrate the crop variety parameters of the crop growth model. S2: Determine the optimal planting strategy using a model. Combining local soil conditions and recent weather conditions, while also considering the prices of irrigation, fertilizer, pesticides, and labor, various management strategies are digitized. Then, the environmental and profit changes under different management strategies are simulated ergonomically to obtain high-yield, environmentally friendly crop planting strategies. These strategies include, but are not limited to, optimal planting date, planting density, planting depth, irrigation volume and frequency, fertilizer volume and frequency, pesticide volume and frequency, and crop rotation. See the implementation example for details on the specific planting management methods developed by the model. S3: Generate a training set using the calibrated crop growth model, and train a deep learning model agent using this dataset. The training set consists of three parts: static input parameters, dynamic input parameters, and output parameters. Static parameters are sensitive soil parameters, including but not limited to available nitrogen, available phosphorus, and available potassium in the soil; dynamic parameters are daily meteorological parameters, including but not limited to radiation, maximum temperature, minimum temperature, rainfall, and average wind speed; output parameters include but not limited to soil moisture, leaf area index, growth stage, and biomass. See the example for details on the generation of the training set and the selection and training of the model. S4: Algorithm design based on deep learning model proxy and remote sensing data assimilation: Since it is difficult to change the intermediate variables in the process of deep learning model operation, the data assimilation algorithm of traditional crop growth model is not suitable for deep learning model. Therefore, we designed the TSPO-SU (Two Stage Parameter Optimization–State Update) algorithm by combining two algorithms: continuous data assimilation algorithm and sequential data assimilation algorithm. The specific implementation of the algorithm is shown in the embodiment. S5: During the crop growth and development stage, remote sensing data is used to retrieve time series of ground parameters such as leaf area index (LAI) and soil moisture (SM). The remote sensing data is then assimilated with a deep learning model agent to retrieve the soil available nutrient content (SAN), which is highly sensitive to crop growth. This enables rapid monitoring of crop growth status and surface conditions, thereby achieving precise irrigation and fertilization.

[0039] like Figure 8 As shown in the figure, this invention discloses a crop planting management method based on a deep learning agent model, including: Acquire crop parameters, soil parameters, and meteorological parameters and perform sensitivity analysis; Based on the results of sensitivity analysis, the crop growth model is calibrated to obtain the optimal variety parameters. Various agricultural planting management scenarios and multi-objective optimization problems are constructed. The crop growth is simulated using the optimal variety parameters to obtain the Pareto optimal solution set of the management scheme. Construct a deep learning agent model, generate a training dataset based on the input and output of a calibrated crop growth model, and train the deep learning agent model. Based on a trained deep learning agent model, the TSPO-SU assimilation algorithm is designed to assimilate remote sensing data with the deep learning agent model, invert environmental parameters for crop growth, determine the optimal planting management scheme from the Pareto solution set, and implement variable fertilization and variable irrigation measures based on the optimal planting management scheme to obtain a refined planting management strategy.

[0040] In a specific embodiment of the present invention, the parameter set collection takes a cotton field base in Kashgar City, Kashgar Prefecture, Xinjiang Uygur Autonomous Region (76.19°E, 39.41°N) as an example. This base covers an area of ​​1260 mu (approximately 84 hectares) and is suitable for remote sensing technology monitoring and application at the field scale. First, the conditions of the base itself are understood. The base is located in a subtropical monsoon climate, with hot and rainy summers and mild and dry winters. Through field surveys and historical database retrieval, the following parameters are obtained: Meteorological parameters: Daily data for the past 5 years and forecasts for the next 3 years (minimum / maximum temperature, rainfall, solar radiation, etc.); Soil parameters: soil mechanical composition, field capacity (FC), available nitrogen (SAN), available phosphorus (SAP), available potassium (SAK), and soil moisture, measured in layers (0-30cm, 30-60cm, 60-100cm); Crop physiological and biochemical parameters: Leaf area index (LAI), aboveground biomass (BIO), and growth stage (DVS) data from multiple field surveys over the past three years (each at least one month apart).

[0041] Historical database: Production and disaster records for the past five years.

[0042] Specifically, the crop growth model is the WOFOST model. Based on the WOFOST model, a multidimensional parameter global sensitivity analysis mechanism is constructed by combining Latin hypercube sampling and Spearman rank correlation coefficient. The input parameters of the WOFOST model are analyzed to obtain the sensitivity ranking of each parameter. The results are compared with preset values ​​to obtain sensitive and insensitive parameters.

[0043] Specifically, the WOFOST model is dynamically calibrated using the Adaptive Particle Swarm Optimization (APSO) algorithm. This includes: obtaining crop parameters; using the APSO algorithm to adjust the difference (DIFF) between the WOFOST model simulation results and the crop parameter data to obtain the optimal variety parameters, as shown in the following formula: (1) Wherein, Field_BIO represents the BIO from the field survey, Model_BIO represents the BIO result from the model simulation, Field_DVS represents the DVS from the field survey, Model_DVS represents the DVS result from the WOFOST model simulation, Field_LAI represents the LAI from the field survey, Model_LAI represents the LAI result from the WOFOST model simulation, n represents the number of quadrats, and i represents the number of surveys.

[0044] In one specific embodiment of the present invention, WOFOST is a popular and well-developed crop growth model, mainly simulating changes in crop environmental water balance, crop canopy development, and biomass of various components under various management conditions. The WOFOST model can simulate cotton growth under various management conditions and has been widely used in climate change, different cotton-growing areas, and different management methods.

[0045] Before model calibration, a parameter sensitivity analysis was conducted to identify key parameters that significantly impact the output results, thereby reducing the complexity of subsequent calibration work. Since the WOFOST model contains 80 input parameters across three categories—soil, management, and crop—a multidimensional global sensitivity analysis mechanism was constructed using a combination of Latin hypercube sampling (LHS) and Spearman rank correlation coefficient (SPRC) to systematically screen sensitive parameters. Based on the sensitivity ranking of each parameter in the analysis results, parameters with high cumulative SPRC contributions and absolute SPRC values ​​greater than 0.3 were identified as sensitive parameters, while the rest were classified as insensitive parameters. ① Conduct field investigations on sensitive soil parameters (such as porosity, saturated hydraulic conductivity, etc.) and management parameters (planting / harvesting dates, irrigation / fertilization amounts and times, etc.); ② The sensitive crop parameters are dynamically calibrated using the Adaptive Particle Swarm Optimization (APSO) algorithm; ③ For non-sensitive parameters, the Chinese Soil Survey Dataset (soil parameters) and the model default values ​​(management and crop parameters) are called respectively.

[0046] The process for determining crop sensitivity parameters is as follows: ① First, physiological and biochemical parameters of cotton throughout its entire growth cycle were obtained through multi-gradient field trials: Two high-yield, two medium-yield, and two low-yield fields were selected based on yield levels, and physiological indicators of cotton during key growth stages (early, mid, and maturity) were systematically monitored. The observation system adopted a two-level nested sampling design: three evenly distributed quadrats (10m×10m) were set up in each field, and observation points were set up within each quadrat using a three-point sampling method to obtain parameters such as leaf area index (LAI), soil moisture (SM), leaf weight (LM), and aboveground biomass (BIO). The average value of the five points represents the physiological and biochemical parameter values ​​of each quadrat.

[0047] ② After obtaining the physiological and biochemical parameters of cotton throughout its entire growth period, the APSO algorithm is used to continuously adjust the difference between the model simulation results and the physiological and biochemical parameter data from the experimental survey (DIFF in Formula 1) to obtain the variety parameters that best match the local cotton growth and development.

[0048] Specifically, the process of obtaining the Pareto optimal solution set of the management scheme includes: Crop growth models are used to simulate crop growth processes and yields under different management modes, thereby obtaining yield data corresponding to each management scenario. Combined with local agricultural production costs, policy subsidies, and environmental constraints, a comprehensive benefit assessment of different management modes is conducted. A multi-objective optimization problem is constructed, and the Pareto optimal solution set of the management scheme is obtained by solving it. Based on the actual operating conditions of the farm, the optimal planting management scheme is determined from the Pareto solution set. The calculation formula is as follows: (2) (3) in, : Comprehensive evaluation indicators for agricultural planting management plans; T: Evaluation period or planting years; r: Discount rate, used to reflect the time value of money; G t : Comprehensive annual benefits in year t; C e : The cost of ecological restoration or environmental remediation after the assessment period ends; : No. The unit market price of annual crops; Under the corresponding management model, the first... Annual output; Policy subsidies or incentive benefits related to this management model; Direct production costs; Regulatory compliance costs; Environmental costs.

[0049] In this embodiment, based on historical meteorological data and local soil physicochemical parameters, various agricultural planting management scenarios are constructed. A crop growth model is used to simulate the cotton growth process and yield under different management models, thereby obtaining yield data corresponding to each management scenario. Based on this, a comprehensive benefit assessment of different management models is conducted, taking into account local agricultural production costs, policy subsidies, and environmental constraints. A multi-objective optimization problem is constructed, and the Pareto optimal solution set of the management scheme is obtained by solving it. Then, based on the actual operating conditions of the farm, a suitable planting management scheme is selected from the Pareto solution set.

[0050] The planting management plan includes, but is not limited to, any combination of the following agricultural management measures: planting date, planting density, fertilization amount and time, irrigation amount and time, pesticide application amount and time, crop rotation type and rotation method, etc.

[0051] By conducting multi-objective evaluation and optimization analysis of different agricultural management models, and under the premise of meeting environmental constraints and regulatory requirements, we can maximize comprehensive benefits or achieve multi-objective trade-offs, thereby providing a quantitative basis for refined agricultural management and planting decisions.

[0052] Specifically, such as Figure 2As shown, the deep learning agent model adopts the KalmanNet architecture, specifically including: an input layer, an output layer, and a state prediction and update layer; the input layer includes two types of parameters: static parameters and dynamic time series parameters; the output layer includes the output leaf area index (LAI), growth and development stage (DVS), biomass (BIO), and soil moisture (SM) of the WOFOST model; the state prediction and update layer includes a state prediction submodule, an observation mapping submodule, an adaptive Kalman gain generation submodule, and a state update submodule.

[0053] Specifically, a dataset is generated using a calibrated WOFOST model. The dataset is divided into an input parameter dataset and an output parameter dataset. The input parameter training set is further divided into a static parameter dataset and a dynamic parameter dataset. The static parameter dataset contains the sensitivity parameters of the WOFOST model, while the dynamic parameter dataset contains the daily meteorological input parameters of the WOFOST model. The output parameter dataset contains the outputs generated by the model based on the input parameters. The dataset is then divided into a training set, a validation set, and a test set. KalmanNet is trained on this dataset, and the validation process employs 5-fold time series cross-validation.

[0054] In a specific embodiment of the present invention, the KalmanNet architecture is adopted. Compared with LSTM-based proxy models, KalmanNet typically has lower computational complexity and fewer parameters because it learns adaptive Kalman gain only within a constrained state space framework, rather than modeling the dynamics of the entire system. The architecture design is as follows: ① Input layer, including time series data with both static and dynamic parameters: 1) Static parameters: Sensitive soil parameters such as initial available nitrogen content (AN), initial available phosphorus content (AP), initial available potassium content (AK), soil saturated water content (SAT), and field capacity (FC); 2) Dynamic parameters: Daily meteorological data such as maximum temperature, minimum temperature, humidity, wind speed, rainfall, and radiation; ② Output layer: The main outputs of the WOFOST model include LAI, DVS, BIO, and SM.

[0055] ③ State prediction and update layer (KalmanNet core structure): In this embodiment, the state prediction and update layer includes, in sequence: The state prediction submodule, under static parameter constraints, combines dynamic driving input and the updated state from the previous time step, and uses a recurrent neural network to construct a state transition model to recursively generate the prior predicted state of the crop system at each time step. The observation mapping submodule is used to map the predicted state to the observation space to obtain the corresponding predicted observation values; The adaptive Kalman gain generation submodule learns a time-varying Kalman gain matrix through a neural network, which is used to characterize the reliable weight relationship between model predictions and observation information. The state update submodule, based on the Kalman gain and the observation innovation term, corrects the predicted state to obtain the updated crop growth state.

[0056] Furthermore, the model training and optimization process includes: Loss function: The joint loss function balances absolute error and relative error (Formula 4): L=0.7 RMSE +0.3 MAPE (4) Optimization strategy: Optimizer: Adam (learning rate 0.001, decay rate β1=0.9, β2=0.999); Batch size: 32; Training cycles: 1000 (early stopping method to monitor validation set loss, tolerance 80 cycles).

[0057] Data preparation and validation include: 1) Dataset Construction: The static and dynamic parameters are the input parameter datasets, and the main output of the WOFOST model is the output parameter dataset; at the same time, the dataset is randomly divided into training set (70%), validation set (15%), and test set (15%).

[0058] 2) Verification method: Time series cross-validation: Employing a 5-fold sliding window to prevent future information leakage; This example utilizes the input and output results of a crop growth model to construct a training dataset. The dataset is divided into a training set and a test set. The training set is used for model training, and the test set is used to verify the model's generalization ability. The validation process employs 5-fold time-series cross-validation. The system deployment utilizes NVIDIA A100 GPUs for parallel computing, which significantly shortens the single-sample inversion time and enables large-area, high-precision, and rapid monitoring of crop growth status using remote sensing imagery.

[0059] Specifically, such as Figure 3 As shown, the TSPO-SU assimilation algorithm includes: A continuous data assimilation algorithm is used to initially optimize the static input parameters of KalmanNet in order to reduce the initial bias of the model. During the forward simulation of the model, a sequential data assimilation algorithm is introduced to update the crop state variables corresponding to the observation time. After completing one round of sequential state updates, a continuous data assimilation algorithm is used again to correct the static input parameters of KalmanNet, further suppressing error accumulation and improving the overall simulation consistency; crop growth environment parameters are obtained and crop growth state simulation is optimized.

[0060] Specifically, the continuous data assimilation algorithm is a four-dimensional variational data assimilation algorithm. By constructing an objective function, it optimizes the static input parameters of KalmanNet to minimize the difference between the KalmanNet simulation results and the remote sensing observation results. The objective function expression for the four-dimensional variational data assimilation algorithm is as follows:

[0061] in: The objective function is... This represents the vector of sensitive parameters of the model to be optimized; This represents the background parameters obtained from model calibration; The background error covariance matrix; The observation error covariance matrix; Indicates the model at time... The simulated output; This represents the remote sensing inversion observation value at the corresponding time; when the objective function When the minimum value is obtained, the corresponding parameter These are the optimized model sensitivity parameters; The sequential data assimilation algorithm is a fast nutrient assimilation algorithm. When obtaining remote sensing observation data, it updates crop state variables hourly to improve the model's response to spatial heterogeneity and short-term changes.

[0062] In a specific embodiment of the present invention, a TSPO-SU (Two-Stage Parameter Optimization and Sequential Update) assimilation algorithm is proposed to fuse remote sensing observation data with crop growth models or their deep learning surrogate models, thereby achieving dynamic correction of key crop state variables and related parameters. Unlike traditional continuous data assimilation methods that perform a one-time global optimization of model parameters within a single time window, the TSPO-SU assimilation algorithm combines continuous parameter optimization with a sequential state update mechanism, introducing observation information in stages over time, thus improving the model's stability and adaptability under long-term growth conditions.

[0063] Specifically, the TSPO-SU assimilation algorithm includes the following steps: First, a continuous data assimilation algorithm is used to initially optimize the static input parameters of KalmanNet in order to reduce the initial bias of the model; Secondly, during the forward simulation of the model, a sequential data assimilation algorithm is introduced to update the crop state variables corresponding to the observation time, so that the model state can absorb remote sensing observation information in real time. Subsequently, after completing a round of sequential state updates, a continuous data assimilation algorithm is used again to make secondary corrections to the initial state or key parameters of the model, in order to further suppress error accumulation and improve the overall simulation consistency.

[0064] Through the collaborative design of "two-stage parameter optimization + sequential state update" described above, the continuous assimilation and update of model state or parameters can be achieved.

[0065] ① Design of continuous data assimilation algorithm

[0066] In this embodiment, the continuous data assimilation algorithm is preferably a four-dimensional variational data assimilation algorithm (4DVar). Its basic principle is to construct an objective function to optimize the sensitive parameters of the model, thereby minimizing the difference between the model simulation results and the remote sensing observation results.

[0067] The objective function expression for 4DVar is shown in formula (5): ② Design of Sequential Data Assimilation Algorithm In this embodiment, the sequential data assimilation algorithm is preferably the Rapid Nutrient Assimilation Algorithm (RNA). This algorithm updates crop state variables hourly when remote sensing observation data is obtained, thereby improving the model's responsiveness to spatial heterogeneity and short-term changes.

[0068] Taking the update of leaf area index (LAI) as an example, its assimilation update formula is: (6) in: The updated assimilation results are used for model simulation at the next time step; This represents the simulation results of the crop model at the previous time step; The LAI value is obtained from remote sensing inversion or empirical regression model; For the study area The average value; For the study area The average value; and They are preset Maximum and minimum values; This is an adjustment coefficient used to adjust the degree of influence of remote sensing observations on spatial heterogeneity correction.

[0069] The sequential assimilation and renewal mechanism of soil moisture (SM) is the same as the LAI renewal principle described above, and will not be repeated here.

[0070] Furthermore, by introducing the TSPO-SU assimilation algorithm, this embodiment has at least the following technical effects: Through the collaborative mechanism of "two-stage parameter optimization + sequential state update", the initial bias and error accumulation of the model are effectively reduced. Improve the stability and adaptability of crop models or their surrogate models under long growing seasons and sparse observation conditions; Enhance the model's responsiveness to remote sensing information and spatial heterogeneity; Based on the assimilation results, crop yield prediction and soil available nutrient content estimation can be obtained, providing technical support for precision fertilization management and agricultural decision-making.

[0071] After establishing the overall management strategy, refined management can be implemented in cotton fields. In this example, Sentinel 1 and Sentinel 2 satellite remote sensing data were selected. Correlation was established between the downloaded remote sensing image data and the surface LAI and SM, and the surface LAI and SM were retrieved. Specifically, the LAI was retrieved using Sentinel 2 visible light remote sensing imagery, and the SM was retrieved using Sentinel 1 microwave remote sensing imagery.

[0072] For field fertilization, the amount of fertilizer to be applied to each plot is: (7) Among them, Fer A This indicates the amount of fertilizer applied to the target plot unit (which can be nitrogen, potassium, or phosphorus fertilizer). SAN indicates the content of available nutrients in the soil, which can be AN, AP, or AK, representing the average content of available nutrients in the soil. Fer indicates the amount of fertilizer applied according to the optimal fertilization strategy.

[0073] For field irrigation, the irrigation amount for each plot unit is: (8) Where IA represents the irrigation amount of the target plot unit, FC represents the field water holding capacity, SM represents the soil moisture content of the target plot unit, and I represents the irrigation amount determined by the optimal management strategy. Compared with the traditional irrigation strategy, this irrigation strategy takes into account both FC and SM, that is, it takes into account the effective water use range of crops. Therefore, the allocation of irrigation water will be more reasonable.

[0074] Furthermore, the comparison of the proxy performance of the deep learning model is shown in Table 1: Table 1. Performance of KalmanNet cotton deep learning model in surrogate simulation and comparison with other mainstream deep learning models.

[0075] Note: DVS (Crop Growth Stage), LAI (Leaf Area Index), BIO (Biomass), SM (Soil Moisture).

[0076] As can be seen from Table 1, the performance of the deep learning model proxy is very close to that of the original model, and its running time is nearly 750 times faster than that of the original model (776s required to run 3000 times in the same environment). Its overall performance is the best among all models.

[0077] like Figure 4 and Figure 5 As shown, the deep learning proxy model using TSPO-SU assimilation achieves the best assimilation effect and is suitable for deep learning model assimilation.

[0078] The refined management method has been applied and promoted in tobacco planting at the Xicun base in Mianchi County, Sanmenxia, ​​Henan Province. Using the same method, a KalmanNet proxy model for tobacco was constructed. This embodiment adds two output variables, stem weight and leaf weight, to the cotton instance. The remaining steps and principles are the same, and there are no differences in other steps and principles. The proxy effect comparison is shown in Table 2: Table 2. Proxy simulation performance of the KalmanNet tobacco leaf deep learning model and its comparison with other mainstream deep learning models.

[0079] As can be seen from Table 2, the performance of the deep learning model proxy is very close to that of the original model, and its running time is improved by nearly 920 times compared with the original model. Its overall performance is the best among all models.

[0080] like Figure 6 and 7 As shown, it can be seen that the KalmanNet proxy model (DL-WF) using the 4D-Var algorithm has a similar data assimilation effect to the WOFOST model.

[0081] from Figure 8-9 It can be seen that the deep learning proxy model using TSPO-SU assimilation also achieves the best assimilation effect, accurately simulating the leaf weight, stem weight, and biomass of tobacco leaves in the region, and is suitable for deep learning model assimilation.

[0082] Figure 10-12 By comparing the simulated available nitrogen, available phosphorus, and available potassium in the soil with the measured values, it can be seen that the method can accurately simulate the nutrient status of the ground.

[0083] Figure 13 and Figure 14 Fertilizer prescription diagram and effect diagram for variables.

[0084] As can be seen from the figure, the fields using variable fertilizer prescriptions achieved higher yields with the same amount of fertilizer, thus achieving the goal of "increasing yield without increasing fertilizer use".

[0085] Furthermore, regarding alternatives to deep learning agent model architectures.

[0086] The deep learning agent model mentioned in this invention is not limited to a single network architecture. In addition to the architectures described, convolutional neural networks (CNN), recurrent neural networks (RNN), long short-term memory networks (LSTM), gated recurrent unit (GRU) networks, Transformer encoder structures, or hybrid models thereof can all be used as alternatives, as long as they can achieve high-precision fitting of the input-output mapping relationship of complex mechanism models.

[0087] The training strategy for the model can also be replaced. In addition to supervised training, when there is sufficient data, semi-supervised learning or self-supervised learning strategies can be used to build surrogate models to reduce the dependence on precisely labeled data (i.e., a large number of simulation results of traditional models).

[0088] Alternative solutions for specific implementations of assimilation algorithms

[0089] In the specific implementation of the TSPO-SU assimilation algorithm, the iterative framework of "two-stage optimization" and "state update" can be equivalently implemented using a variety of mainstream optimization and filtering algorithms. Among them, the continuous data assimilation stage can choose global optimization methods such as simulated annealing, particle swarm optimization, and genetic algorithm; while the sequential data assimilation stage can choose mainstream sequential assimilation techniques such as ensemble Kalman filtering, ensemble root mean square filtering, and particle filtering.

[0090] Furthermore, the stage division logic of "coarse calibration" and "fine calibration" in this algorithm also has flexible alternative implementation methods. Besides the method described above, equivalent technical effects can be achieved by adjusting the ratio of iterations between the two stages, setting differentiated convergence thresholds, or introducing multi-scale assimilation strategies (such as quickly completing preliminary calibration in low-resolution space and then gradually improving to the actual field resolution for fine optimization). These alternative approaches are all within the core framework and can maintain the overall performance of the algorithm in terms of efficiency and accuracy.

[0091] Alternatives to the integrated decision support system process

[0092] Sensitivity analysis: This invention employs a Latin hypercube sampling method combined with Spearman correlation coefficients. Alternative methods may include, but are not limited to: Sobol global sensitivity analysis, the EFAST method, and the variance-based Morris screening method. All these methods can achieve the goal of identifying key sensitive parameters from a large number of parameters.

[0093] Model calibration: In addition to the methods described above, other automated parameter optimization frameworks can be used, such as combining simulated annealing algorithms, Markov chain Monte Carlo methods, and other intelligent optimization algorithms to automate the calibration process and improve efficiency.

[0094] Strategy generation stage: When formulating planting strategies, the optimization objectives can be expanded or replaced. In addition to the economic and environmental benefits considered in this invention, other objectives (such as environmental priority) can be introduced as multi-objective optimization constraints, thereby generating a management plan with environmental protection as the primary focus.

[0095] Alternative solutions for system integration and data input

[0096] Alternative Remote Sensing Data Sources: The remote sensing data that this invention can utilize is not limited to specific satellite or UAV platforms. Multispectral, hyperspectral, thermal infrared, and synthetic aperture radar images, as long as they can extract effective information related to crop growth or soil conditions (such as leaf area index and soil moisture), can all be used as input to the assimilation algorithm.

[0097] Field sensor data fusion: In addition to being used directly to drive models or assimilation, IoT sensor data deployed in the field can serve as an independent verification data source or be used to build local calibration models, thereby enhancing the reliability and accuracy of results at the system level.

[0098] All the above alternative solutions are equivalent or modified means to achieve the core idea of ​​this invention—that is, to achieve large-scale precision agricultural decision-making through efficient proxy simulation, intelligent data assimilation and system integration—and should all be included within the scope of the claims of this patent.

[0099] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to the method section.

[0100] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A crop planting management method based on a deep learning agent model, characterized in that, include: Acquire crop parameters, soil parameters, and meteorological parameters and perform sensitivity analysis; Based on the results of sensitivity analysis, the crop growth model is calibrated to obtain the optimal variety parameters. Various agricultural planting management scenarios and multi-objective optimization problems are constructed. The crop growth is simulated using the optimal variety parameters to obtain the Pareto optimal solution set of the management scheme. Construct a deep learning agent model, generate a training dataset based on the input and output of a calibrated crop growth model, and train the deep learning agent model. Based on a trained deep learning agent model, the TSPO-SU assimilation algorithm is designed to assimilate remote sensing data with the deep learning agent model, invert environmental parameters for crop growth, determine the optimal planting management scheme from the Pareto solution set, and implement variable fertilization and variable irrigation measures based on the optimal planting management scheme to obtain a refined planting management strategy.

2. The crop planting management method based on a deep learning agent model according to claim 1, characterized in that, The crop growth model is the WOFOST model. Based on the WOFOST model, a multidimensional parameter global sensitivity analysis mechanism is constructed by combining Latin hypercube sampling and Spearman rank correlation coefficient. The input parameters of the WOFOST model are analyzed to obtain the sensitivity ranking of each parameter. The results are compared with preset values ​​to obtain sensitive and insensitive parameters.

3. The crop planting management method based on a deep learning agent model according to claim 2, characterized in that, The WOFOST model is dynamically calibrated using an adaptive particle swarm optimization algorithm.

4. The crop planting management method based on a deep learning agent model according to claim 3, characterized in that, The process of obtaining the Pareto optimal solution set of the management scheme includes: By using crop growth models to simulate crop growth processes and yields under different management modes, yield data corresponding to each management scenario is obtained. Combined with local agricultural production constraints, a comprehensive benefit assessment of different management modes is conducted, a multi-objective optimization problem is constructed, and the Pareto optimal solution set of the management scheme is obtained by solving the problem. Based on the actual operating conditions of the farm, the optimal planting management scheme is determined from the Pareto solution set.

5. A crop planting management method based on a deep learning agent model according to claim 4, characterized in that, The deep learning proxy model adopts the KalmanNet architecture, specifically including: an input layer, an output layer, and a state prediction and update layer; the input layer includes two types of parameters: static parameters and dynamic time series parameters; the output layer includes the output of the WOFOST model: leaf area index (LAI), growth and development stage (DVS), biomass (BIO), and soil moisture (SM); the state prediction and update layer includes a state prediction submodule, an observation mapping submodule, an adaptive Kalman gain generation submodule, and a state update submodule.

6. The crop planting management method based on a deep learning agent model according to claim 5, characterized in that, The dataset is generated using the calibrated WOFOST model. The dataset is divided into an input parameter dataset and an output parameter dataset. The input parameter training set is further divided into a static parameter dataset and a dynamic parameter dataset. The static parameter dataset contains the sensitivity parameters of the WOFOST model, and the dynamic parameter dataset contains the daily meteorological input parameters of the WOFOST model. The output parameter dataset is the output generated by the model based on the input parameters; the dataset is divided into training set, validation set and test set, and KalmanNet is trained. The validation process implements 5-fold time series cross-validation.

7. The crop planting management method based on a deep learning agent model according to claim 1, characterized in that, The TSPO-SU assimilation algorithm includes: A continuous data assimilation algorithm is used to initially optimize the static input parameters of KalmanNet in order to reduce the initial bias of the model. During the forward simulation of the model, a sequential data assimilation algorithm is introduced to update the crop state variables corresponding to the observation time. After completing one round of sequential state updates, a continuous data assimilation algorithm is used again to correct the static input parameters of KalmanNet, further suppressing error accumulation and improving the overall simulation consistency; crop growth environment parameters are obtained and crop growth state simulation is optimized.

8. A crop planting management method based on a deep learning agent model according to claim 7, characterized in that, The continuous data assimilation algorithm is a four-dimensional variational data assimilation algorithm. By constructing an objective function, it optimizes the static input parameters of KalmanNet to minimize the difference between the KalmanNet simulation results and the remote sensing observation results. The objective function expression for the four-dimensional variational data assimilation algorithm is as follows: in: The objective function is... This represents the vector of sensitive parameters of the model to be optimized; This represents the background parameters obtained from model calibration; The background error covariance matrix; The observation error covariance matrix; Indicates the model at time... The simulated output; This represents the remote sensing inversion observation value at the corresponding time; when the objective function When the minimum value is obtained, the corresponding parameter These are the optimized model sensitivity parameters; The sequential data assimilation algorithm is a fast nutrient assimilation algorithm. When obtaining remote sensing observation data, it updates crop state variables hourly to improve KalmanNet's response to spatial heterogeneity and short-term changes.

Images