A farmland straw returning to field decision optimization method, system and electronic equipment

Abstract

The present application belongs to the field of agricultural resource utilization and artificial intelligence cross technology, and provides a farmland straw returning decision optimization method, system and electronic equipment, the method comprising: obtaining multi-source data of farmland crops, and preprocessing to obtain a standardized data set; training a multi-task deep learning model based on the standardized data set to predict target data under a given farmland management measure; dividing the target area into multiple spatial units, and simulating target data prediction values under different farmland management scenarios in each spatial unit based on the multi-task deep learning model to form a scenario result database for each spatial unit; for the scenario result database of each spatial unit, a multi-objective optimization algorithm is used for screening to obtain a farmland straw returning decision scheme representing the best trade-off relationship between the target data. The present application realizes the synergies of yield improvement, ammonia emission reduction and greenhouse gas emission reduction of the planting system.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of agricultural resource utilization and artificial intelligence, specifically relating to a decision optimization method for returning farmland straw to the field, as well as a decision optimization system and equipment for returning farmland straw to the field. Background Technology

[0002] Straw is a major agricultural byproduct, rich in organic matter, nitrogen, phosphorus, potassium, and trace elements. It can replace some chemical nitrogen fertilizers in farmland, improving soil properties and increasing resource utilization efficiency. However, practice shows that unreasonable straw return strategies can lead to a series of problems such as nitrogen fixation, increased pests and diseases, and impaired seedling growth, potentially resulting in reduced yields and environmental risks. Therefore, achieving precise and efficient straw return to the field is a key challenge facing my country's green agricultural transformation.

[0003] Currently, the development of straw return technology is mainly reflected in two aspects: In terms of single-point technologies, a large number of studies and patents have focused on specific aspects of straw return operations. Examples include crop rotation design aimed at emission reduction and efficiency improvement (see the publication text of Chinese patent application CN120898694A) and innovations in straw return machinery (see the publication text of Chinese patent application CN120615502A). At the level of intelligent decision-making, examples include corn straw return fertilization methods and systems (see the publication text of Chinese patent application CN120782587A) and methods for building straw return rate models based on straw image processing (see the publication text of Chinese patent application CN120431500A), which have promoted the intelligentization of the straw return process.

[0004] In traditional field trials, extensive research has explored the effects of straw return amount and method on yield, greenhouse gas emissions, and ammonia volatilization at the field scale. However, due to varying climatic conditions and soil properties, coupled with the interactions between different management practices, optimized straw return models suited to specific regions cannot be directly generalized to a national scale. Nevertheless, the large amount of in-situ observational data accumulated in field trials provides a solid foundation for subsequent data integration and the construction of intelligent models.

[0005] In the field of agricultural system technology integration, crop growth models relying on empirical formulas struggle to accurately capture the complex interactions between diverse factors such as climate, soil, and management. Furthermore, although artificial intelligence has been applied to the construction and effect prediction of straw return models, most are single-task prediction models, typically used only to predict yield or single pollutant emissions. They cannot jointly capture the relationships between yield and pollutants, or between different pollutants, thus limiting their ability to provide precise optimization techniques for synergistic yield improvement and emission reduction.

[0006] With the increasing demand for precision agricultural management, existing straw return technology models can no longer meet the needs of refined, integrated, and multi-objective coordinated development. How to precisely formulate comprehensive technical solutions that couple straw return amounts, return methods, water and fertilizer management, and cultivation techniques for different climate-soil-crop systems at a national macro-scale, in order to maximize the potential of synergistic effects between "increased production" and "reduced emissions," remains one of the key challenges that urgently needs to be overcome.

[0007] This section provides background information related to this application, which is not necessarily prior art. Summary of the Invention

[0008] One objective of this application is to solve or at least mitigate some or all of the aforementioned problems. To this end, one objective of this application is to provide a method for optimizing the decision-making process of returning crop straw to the field. This method uses a multi-task deep learning model to predict the synergistic effects of straw return on yield, ammonia volatilization, and greenhouse gas emissions, thereby achieving a synergistic effect of increased crop yield, reduced ammonia volatilization, and reduced greenhouse gas emissions in the crop planting system.

[0009] In a first aspect, the present invention provides a method for optimizing the decision-making process of returning farmland straw to the field, comprising: Obtain multi-source data on farmland crops and preprocess them to obtain a standardized dataset; A multi-task deep learning model is trained on a standardized dataset to predict target data under given farmland management measures. The target data includes at least crop yield, ammonia volatilization loss, and global warming potential. The multi-task deep learning model includes: a shared feature encoder for receiving preprocessed input features and extracting shared feature representations; multiple task-specific adapters for generating task-sensitive features corresponding to the target based on the shared feature representations; and multiple prediction output layers, each connected to a task-specific adapter, for outputting the predicted value for the corresponding target. The target area is divided into multiple spatial units. Within each spatial unit, a multi-task deep learning model is used to simulate the target data prediction values ​​under different farmland management scenarios, forming a scenario result database for each spatial unit. For the scenario results database of each spatial unit, a multi-objective optimization algorithm is used for screening to obtain a decision scheme for returning farmland straw to the field that represents the best trade-off relationship between the target data.

[0010] In some embodiments, multi-source data includes climate data, soil data, management practice data, and target data; wherein, Climate data includes average temperature and cumulative precipitation during the crop growing season; Soil data include soil pH, organic carbon content, total nitrogen content, bulk density, and clay content; Management data include chemical nitrogen fertilizer application rate, straw nitrogen substitution rate, nitrogen fertilizer type, water management method, tillage method, straw return amount and method; The target data include crop yield, ammonia volatilization loss, and global warming potential.

[0011] In some embodiments, preprocessing includes: Encode the type data as an integer; Perform Z-score standardization on numerical data; For data where some target variables are missing, a mask matrix is ​​generated to identify the locations of valid data.

[0012] In some embodiments, when training a multi-task deep learning model, a dynamic weight allocation mechanism based on task variance is used to calculate the total loss. The total loss is the average of the losses of each task weighted according to the inverse of their variances. When calculating the loss of each task, the loss terms corresponding to missing data are ignored in conjunction with the mask matrix.

[0013] In some embodiments, farmland management scenarios are constructed in the following ways: Define a technology package of multiple farmland management measures formed by categorical management variables; Set the value range and step size for continuous management variables to generate multiple value points; Each technology package is combined with the value points of all continuous management variables to form different farmland management scenarios.

[0014] In some embodiments, the multi-objective optimization algorithm is the third-generation non-dominated sorting genetic algorithm NSGA-III; the screening process includes: The crop yield, ammonia volatilization loss, and global warming potential forecasts for each farmland management scenario were normalized. Generate reference points that are uniformly distributed in the three-dimensional target space; Stratification of scenarios is based on non-dominated ordering; Based on the relationship between the scenario and the reference point, a balanced selection is made, and the decision on returning farmland straw to the field is iteratively filtered out.

[0015] In some embodiments, the method further includes generating a spatial distribution map of optimized management measures based on a selected farmland straw return decision scheme.

[0016] In some embodiments, the method further includes: generating a report on the potential for increased production and reduced emissions based on a comparison of the prediction results of the selected straw return decision scheme and the baseline scheme; the baseline scheme is other straw return strategies adopted by farmers and corresponding supporting management measures such as nitrogen fertilizer and tillage.

[0017] Secondly, this application also provides a decision optimization system for returning farmland straw to the field, including: The data acquisition module acquires multi-source data on farmland crops and preprocesses it to obtain a standardized dataset. The model prediction module trains a multi-task deep learning model based on a standardized dataset to predict target data under given farmland management measures. The target data includes at least crop yield, ammonia volatilization loss, and global warming potential. The multi-task deep learning model includes: a shared feature encoder, which receives preprocessed input features and extracts shared feature representations; multiple task-specific adapters, which generate task-sensitive features corresponding to the target based on the shared feature representations; and multiple prediction output layers, each connected to a task-specific adapter, to output the predicted value for the corresponding target. The unit prediction module divides the target area into multiple spatial units. Within each spatial unit, it simulates the target data prediction values ​​under different farmland management scenarios based on a multi-task deep learning model, forming a scenario result database for each spatial unit. The decision screening module uses a multi-objective optimization algorithm to screen the scenario result database for each spatial unit, and obtains a decision scheme for returning farmland straw to the field that represents the best trade-off relationship between the target data.

[0018] Thirdly, this application also provides an electronic device, which includes a processor and a memory. The memory stores at least one instruction, at least one program, code set, or instruction set. The processor loads and executes the at least one instruction, at least one program, code set, or instruction set to implement the above-described method for optimizing the decision of returning farmland straw to the field.

[0019] Fourthly, this application also provides a computer-readable storage medium storing at least one instruction, at least one program, code set, or instruction set, wherein the at least one instruction, at least one program, code set, or instruction set is loaded and executed by a processor to implement the above-described method for optimizing the decision of returning farmland straw to the field.

[0020] The beneficial effects of this invention are: 1) Through a multi-task deep learning model, this invention can predict the synergistic effects and trade-offs of straw return to the field on yield, ammonia volatilization and greenhouse gas emissions in a unified and accurate manner. It overcomes the shortcomings of single-objective models that fail to consider all aspects and achieves the synergistic effects of yield increase, ammonia volatilization reduction and greenhouse gas emission reduction in the three major staple crops of rice, wheat and corn. This provides a scientific basis for the synergistic optimization of "increased production and reduced emissions".

[0021] 2) Based on the national gridded simulation and intelligent optimization algorithm, a set of optimal technical solutions with balanced distribution can be automatically selected for different regions, promoting the leap of agricultural management from "extensive and general" to "spatial precision customization".

[0022] 3) It forms a full-process management system of prediction, optimization and visualization. Users can intuitively understand the trade-offs between objectives through the interactive interface, which significantly reduces the threshold for use and promotes the transformation and application of green agricultural technologies. Attached Figure Description

[0023] Figure 1 This is a flowchart of the decision optimization method for returning farmland straw to the field provided in the embodiments of this application. Figure 2 This is a schematic diagram illustrating the fitting effect of three crop multi-task learning models provided in the embodiments of this application (the red dashed line in the figure represents the ideal calibration line, and the green solid line represents the fitting trend line). Detailed Implementation

[0024] The present invention will be further described below with reference to the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solution of the present invention, and should not be used to limit the scope of protection of the present invention.

[0025] Before explaining any implementation of this application in detail, it should be understood that this application is not limited to its application to the structural details and component arrangements set forth in the following description or shown in the above drawings.

[0026] In this application, the terms "comprising," "including," "having," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0027] In this application, the term "and / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent three cases: A existing alone, A and B existing simultaneously, and B existing alone. Additionally, the character " / " in this application generally indicates that the preceding and following related objects have an "and / or" relationship.

[0028] In this application, the terms "connection," "combination," "coupling," and "installation" can refer to direct connection, combination, coupling, or installation, or indirect connection, combination, coupling, or installation. For example, a direct connection refers to two parts or components being connected together without the need for an intermediary, while an indirect connection refers to two parts or components each being connected to at least one intermediary, with the connection achieved through the intermediary. Furthermore, "connection" and "coupling" are not limited to physical or mechanical connections or couplings, but can also include electrical connections or couplings.

[0029] In this application, those skilled in the art will understand that relative terms (e.g., “about,” “approximately,” “basically,” etc.) used in conjunction with quantities or conditions are to include the values ​​and have the meaning indicated by the context. For example, such relative terms include at least the degree of error associated with the measurement of a particular value, tolerances associated with the particular value due to manufacturing, assembly, use, etc. Such terms should also be considered as disclosing a range defined by the absolute values ​​of the two endpoints. Relative terms may refer to a certain percentage (e.g., 1%, 5%, 10% or more) of the indicated value. Numerical values ​​that do not use relative terms should also be disclosed as specific values ​​with tolerances. Furthermore, “basically” when expressing relative angular relationships (e.g., substantially parallel, substantially perpendicular) may refer to a certain degree (e.g., 1 degree, 5 degrees, 10 degrees or more) added to or subtracted from the indicated angle.

[0030] In this application, those skilled in the art will understand that the function performed by a component can be performed by one component, multiple components, one part, or multiple parts. Similarly, the function performed by a part can also be performed by one part, one component, or a combination of multiple parts.

[0031] In this application, the directional terms "upper," "lower," "left," "right," "front," and "rear" are used to describe the orientation and positional relationships shown in the accompanying drawings and should not be construed as limiting the embodiments of this application. Furthermore, in the context, it should be understood that when an element is mentioned as being connected "upper" or "lower" to another element, it can be directly connected to the other element "upper" or "lower," or indirectly connected through an intermediate element. It should also be understood that directional terms such as upper side, lower side, left side, right side, front side, and rear side not only represent positive orientation but can also be understood as lateral orientation. For example, "below" can include directly below, lower left, lower right, lower front, and lower rear.

[0032] In this application, the terms "controller," "processor," "central processing unit," "CPU," and "MCU" are used interchangeably. When using the unit "controller," "processor," "central processing unit," "CPU," or "MCU" to perform a specific function, unless otherwise stated, these functions may be performed by a single or multiple of the aforementioned units.

[0033] In this application, the terms "device," "module," or "unit" are used to describe devices that can be implemented in hardware or software to perform a specific function.

[0034] In this application, the terms “calculation,” “judgment,” “control,” “determine,” “identify,” etc., refer to the operation and process of a computer system or similar electronic computing device (e.g., controller, processor, etc.).

[0035] The benefits, other advantages, and solutions to problems will be described below with reference to specific embodiments. However, these benefits, advantages, solutions to problems, and any features that may lead to or make any benefit, advantage, or solution appear or become more significant should not be construed as key, necessary, or essential features of any or all claims.

[0036] This application provides a method and system for optimizing the decision-making process of returning farmland straw to the field, so as to achieve synergistic effects of increasing yield, reducing ammonia volatilization emissions, and reducing greenhouse gas emissions in the planting system of the three major staple crops of rice, wheat, and corn, and fully leveraging their comprehensive production and ecological potential.

[0037] Figure 1 and Figure 2 The diagram shown is a flowchart and schematic diagram of a decision optimization method for returning farmland straw to the field provided in an embodiment of this application. The following is a detailed explanation. Figure 1 and Figure 2 The technical solution of this application is described in detail below. It should be noted that this specification provides the operational steps of the methods described in the embodiments or flowcharts, but based on conventional or non-inventive labor, more or fewer operational steps may be included. The order of steps listed in the embodiments is merely one possible execution order among many steps and does not represent the only execution order.

[0038] This invention provides a method for optimizing the decision-making process of returning farmland straw to the field. This method is adapted for deployment on mobile terminals / edge devices and specifically includes the following steps: S10: Obtain multi-source data on farmland crops and preprocess them to obtain a standardized dataset.

[0039] This module is responsible for aggregating, cleaning, encoding, and standardizing raw data from field experiments, meteorological observations, soil surveys, and farmer surveys across the country, creating a standardized dataset that can be directly learned by machine learning models. This module is fundamental to ensuring the scientific rigor, accuracy, and generalization ability of the models.

[0040] Specifically, the following steps are included: Step S11: Obtain multi-source data on farmland crops For the three major staple crops of rice, wheat and corn, multi-source data are collected, including climate data, soil data, management measure data and target data.

[0041] Specifically, multi-source data includes the following: Climate data: including average temperature (MAT) and cumulative precipitation (MAP) during the crop growing season.

[0042] Soil data: including soil pH, organic carbon (SOC), total nitrogen (TN), bulk density (BD), and clay content (Clay).

[0043] Management measures data include: chemical nitrogen fertilizer application rate, straw nitrogen substitution ratio, nitrogen fertilizer type (urea / enhanced fertilizer / other / no nitrogen fertilizer application), water management method (rain-fed / irrigated for wheat and corn, intermittent / continuous for rice), tillage method (tillage / no-till), straw return to field amount, and straw return to field method (surface mulching / deep application / no straw return).

[0044] Based on the amount of chemical nitrogen fertilizer used and the amount of straw returned to the field, the nitrogen substitution ratio of straw is calculated as follows: Ratio = N s / (N f +N s ) Where, N s This refers to the amount of nitrogen (N) from straw returned to the field. f This refers to the amount of chemical nitrogen fertilizer applied, and Ratio is the ratio of nitrogen substitution by straw.

[0045] Target data: These are the core indicators that need to be predicted and optimized by the model. They include: crop yield per unit area (Y, in kg / ha); ammonia volatilization loss (NH3, in kg / ha); and global warming potential (GWP, in kg CO2-eq / ha).

[0046] In some embodiments, the various types of data can be added and determined according to the specific application scenario. For example, the target data may also include net income, nitrogen leaching, nitrogen runoff loss, soil organic carbon change, water use efficiency, etc.

[0047] Based on measured or estimated emissions of methane (CH4) and nitrous oxide (N2O), and using the global warming potential parameters on a 100-year timescale (CH4 = 28, N2O = 265) stipulated by the Intergovernmental Panel on Climate Change (IPCC), the emissions are uniformly converted into carbon dioxide equivalents (CO2-eq, kg / ha), as follows: GWP = CH4 × 28 + N2O × 265 It should be noted that when calculating GWP, given that CH4 emissions from wheat and maize cultivation systems are extremely low, only N2O is included in the calculation; however, for rice cultivation systems, both CH4 and N2O emissions must be calculated together.

[0048] Step S12: Encode and impute missing values ​​for multi-source data. Integer encoding for categorical data: For all categorical management variables (such as fertilizer type, return method, etc.), to avoid the problem of drastic dimensionality increase caused by one-hot encoding and to allow the model to potentially learn the order relationship between categories, an integer encoding strategy is used for conversion. For example, "conventional tillage" and "no-tillage" are encoded as 0 and 1, respectively.

[0049] The processing results are shown in the table below.

[0050]

[0051] Missing value imputation: For missing values ​​in input features (climate, soil, management measures), the average value of the feature in all valid samples or the background value is used to impute them, ensuring the integrity of the dataset.

[0052] Step S13: Perform target-specific standardization processing on numerical input features and target data. Feature standardization: All numerical input features (such as MAT, MAP, pH, chemical nitrogen fertilizer application rate, etc.) are Z-score standardized to have a mean of 0 and a standard deviation of 1. This process eliminates differences in the units and ranges of different features, accelerating model convergence. The calculation is as follows: z = (x - μ) / σ Where x is the original feature value, μ is the mean of all samples of the feature, σ is the standard deviation of all samples of the feature, and z is the standardized value.

[0053] Independent standardization of target variables: Z-score standardization was performed independently on the three target variables Y, NH3, and GWP.

[0054] This step ensures that target variables of different dimensions and orders of magnitude are at a similar scale during model learning, which is a key prerequisite for balanced optimization in multi-task learning. The parameters of the standardizer (mean μ and standard deviation σ) will be saved for subsequent destandardization of the prediction results.

[0055] Step S14: Dynamic Mask Generation For data where some target variables are missing, a mask matrix is ​​generated to identify the locations of valid data.

[0056] To address the common issue of missing target variables in agricultural observation data (such as some experiments measuring only yield but not gas emissions), the system generates a mask matrix that is identical in shape to the target matrix. Specifically, for each target in each sample, if a valid observation exists, the corresponding mask position is marked as 1; if data is missing, it is marked as 0. This mask guides the calculation of the loss function during model training, ensuring that the model learns only from valid data, thus fully utilizing the incomplete dataset.

[0057] S20 is a multi-task deep learning model trained on a standardized dataset. The model is used to collaboratively predict crop yield, ammonia volatilization loss, and global warming potential under given farmland management measures.

[0058] This step is responsible for building a neural network model that can learn multiple related tasks simultaneously and collaboratively, aiming to achieve collaborative prediction and optimization of key indicators such as production, ammonia volatilization and greenhouse gas emissions through a unified framework.

[0059] Multi-task deep learning models include: A shared feature encoder is used to receive preprocessed input features and extract shared feature representations; Multiple task-specific adapters are connected to a shared feature encoder. Each adapter corresponds to a prediction target and is used to generate task-sensitive features for the corresponding target based on the shared feature representation. Multiple prediction output layers, each connected to a task-specific adapter, are used to output the predicted value for the corresponding target.

[0060] The specific process of training a multi-task deep learning model based on a standardized dataset includes: Step S21: Input Layer and Shared Feature Encoder First, a shared feature encoding network is constructed. This network serves as the basic feature extraction layer of the model, receiving standardized 13-dimensional input data (including two climate factors: annual mean temperature (MAT) and annual cumulative precipitation (MAP); five soil properties: pH, organic carbon (SOC), total nitrogen (TN), soil bulk density (BD), and clay content (Clay); and six field management measures: chemical nitrogen fertilizer application rate, straw nitrogen substitution ratio, nitrogen fertilizer type, straw return method, tillage method, and water management method). High-order abstract features are extracted step-by-step through the following hierarchical structure: (1) First fully connected layer and nonlinear transformation: The 13-dimensional input is mapped to a 256-dimensional feature space, and then a smooth nonlinearity is introduced through the SiLU activation function.

[0061] (2) Feature normalization and regularization: Layer normalization is performed on the 256-dimensional features to stabilize training, and then random deactivation is applied with a probability of 0.4 to randomly block some neuron connections to prevent overfitting.

[0062] (3) Second fully connected layer and quadratic transformation: The 256-dimensional features are compressed to 128-dimensional, and then activated by SiLU and normalized by layer, and random deactivation is applied with a probability of 0.3. The final output is a 128-dimensional shared feature vector.

[0063] The Shared Feature Encoding Network is a deep neural network that learns to extract high-order feature representations applicable to all targets from a uniform input (climate, soil, management measures).

[0064] Step S22: Task-Specific Adapter Based on shared features, lightweight task adapters with identical structures but independent parameters are established for the three targets: NH3, GWP, and production.

[0065] Each adapter is a small, fully connected network: (1) Feature transformation: Receive the 128-dimensional features output by the shared encoder and transform them into 64-dimensional features through a fully connected layer.

[0066] (2) Task-sensitive processing: SiLU activation and layer normalization are performed on the transformed 64-dimensional features in sequence to form a feature representation that is more sensitive to the specific task.

[0067] Lightweight dedicated network branches are provided for the three targets of yield, NH3, and GWP, and fine-tuned based on shared features to capture the unique response patterns of each target.

[0068] Step S23: Predict the output layer Each adapter connects to a single-neuron output layer (64→1 dimension), whose weight parameters are sampled from a narrow normal distribution with a standard deviation of 0.1, and whose bias parameters are initialized to zero. This initialization strategy ensures a smooth start and fast convergence of the training process, and ultimately directly outputs the predicted values ​​of the three targets: NH3, GWP, and yield.

[0069] During model training, systematic training and optimization are performed to ensure optimal performance. This includes the following optimizations: 1) For the input layer and the shared feature encoder, during model training, a dynamic weight allocation mechanism based on task variance (intra-batch loss variance) is adopted, automatically assigning learning weights to each task according to the following formula: Weight = 1.0 / (Task Variance + ε) ε is a constant used for numerical stability, and is 1e-6. Tasks with smaller variance will receive higher learning priority. This mechanism, combined with a mask matrix, effectively eliminates the interference of missing data on the training process and calculates the average of all effective weighted loss values ​​to form the final training signal that guides the updating of model parameters.

[0070] Understandably, when training a multi-task deep learning model, a dynamic weight allocation mechanism based on task variance is used to calculate the total loss, which is the average of the losses of each task weighted according to the inverse of their variances; when calculating the loss of each task, the loss terms corresponding to missing data are ignored in conjunction with the mask matrix.

[0071] A dynamic training mechanism is adopted for multi-task deep learning models: a dynamic loss weight allocation and mask learning mechanism based on task variance are introduced to enable the model to balance the learning progress of different tasks and effectively utilize incomplete datasets that are common in reality and have missing some target data.

[0072] 2) Early Stopping Mechanism and Optimal Model. Before training, the preprocessed data was divided into training and test sets in an 80%:20% ratio, and a fixed random seed was set to ensure reproducibility of the experiments. The AdamW optimizer was used for parameter optimization, with an initial learning rate of 1e-3 and a weight decay coefficient of 1e-4 to prevent overfitting. During training, the ReduceLROnPlateau scheduler was used to monitor the validation set loss. When the loss did not decrease for 5 consecutive rounds, the learning rate was automatically halved; gradient pruning was also implemented, with the upper limit of the gradient norm set to 1.0 to ensure training stability. Finally, an early stopping criterion was adopted: if the validation set loss did not reach a new low for 15 consecutive rounds, training was terminated and the model with the best validation performance at that time was saved as the final output.

[0073] The multi-task deep learning model proposed in this application can simultaneously and jointly output predicted values ​​for three targets. The fitting effect diagrams for the multi-task learning model on three crops—rice, wheat, and corn—are shown below. Figure 2 As shown.

[0074] S30 divides the target area into multiple spatial units, and simulates the target data prediction values ​​under different farmland management scenarios based on a multi-task deep learning model in each spatial unit, forming a scenario result database for each spatial unit.

[0075] This step utilizes a pre-trained multi-task deep learning model to simulate a massive combination of techniques on a target region-scale grid, evaluating the potential for multi-objective collaborative optimization at each grid scale.

[0076] Simulation of nationwide grid-scale management measures, specifically including the following processes: Step S31: Constructing spatial simulation units and background database The target area is determined based on the analysis objectives and can be the entire country, a specific province, a city, a county, or a district. The target area is divided into regular geographic grid units to form multiple spatial units. This embodiment uses the entire country as an example for decision-making simulation.

[0077] The major rice, wheat, and corn producing areas across the country were divided into regular geographic grids (e.g., grids of 10 km × 10 km). Each grid cell was equipped with background feature data, including the grid's multi-year average MAT and MAP, soil pH, SOC, TN, BD, and Clay content. This data constituted the "background" input for each grid simulation.

[0078] Step S32: Define the set of controllable management scenarios For each grid (i.e., spatial unit), the system will simulate all possible combinations of farmland management measures that can be decided by the user to form a farmland management scenario.

[0079] The specific process includes: Classified management measures combination: Based on actual agronomic practices, all possible combinations of farmland management measures are combined, such as: fertilizer type (4 types) × straw return method (3 types) × water management method (2 types) × tillage method (2 types), forming a technical package of 48 farmland management measures.

[0080] Setting the range of continuous management variables: Reasonable value ranges and step sizes are set for continuous variables. For example, the application rate of chemical nitrogen fertilizer ranges from 0 to 400 kg / ha, with a step size of 20 kg / ha; the straw nitrogen replacement ratio ranges from 0% to 100%, with a step size of 10%. By combining the technical package of 48 farmland management measures with all possible values ​​of continuous variables, and eliminating illogical combinations, tens of thousands of specific farmland management measures to be simulated can be generated for each grid, forming farmland management scenarios.

[0081] Step S33: Batch Simulation and Result Calculation The background features of each grid cell are concatenated with the adjustable features of each management scenario to form a complete model input. A trained multi-task deep learning model is then used for high-throughput batch prediction to obtain the predicted NH3, GWP, and yield values ​​for each scenario. Finally, a results database containing tens of thousands of records is generated for each grid cell nationwide, with each record representing a management scenario and its corresponding multi-objective output of "NH3-GWP-yield".

[0082] S40: For the scenario result database of each spatial unit, a multi-objective optimization algorithm is used to filter and obtain a decision scheme for returning farmland straw to the field that represents the best trade-off relationship between the target data.

[0083] The core task of this step is to automatically select the optimal combination of technologies that can best coordinate multiple objectives (such as simultaneously pursuing high yield, low NH3, and low GWP) from the massive number of simulated farmland management scenarios in each grid.

[0084] This invention employs the third-generation non-dominated sorting genetic algorithm (NSGA-III) to handle this complex multi-objective optimization problem. The screening process using the multi-objective optimization algorithm includes: normalizing the crop yield, ammonia volatilization, and global warming potential predictions for each farmland management scenario; generating reference points uniformly distributed in a three-dimensional objective space; stratifying the scenarios based on non-dominated sorting; and iteratively selecting the farmland straw return decision based on the correlation between the scenario and the reference points.

[0085] In one embodiment, the screening process using NSGA-III specifically includes the following steps: Step S41: Target Normalization Target value normalization: Since the three targets "yield", "NH3" and "GWP" have different dimensions and optimization directions (the larger the yield, the better, and the smaller the NH3 and GWP, the better), the three target values ​​of all simulation scenarios in each grid are first normalized to scale their range to [0, 1].

[0086] The production target is calculated as follows: Output: X' = (X - Xmin) / (Xmax - Xmin) Where: X' represents the normalized crop yield value, X represents the target crop yield value, Xmin represents the minimum crop yield value, and Xmax represents the maximum crop yield value.

[0087] The environmental emission indicators (GWP and NH3) are calculated as follows to ensure that all targets are maximized: Environmental emissions target: X' = (Xmax - X) / (Xmax - Xmin) Where: X' represents the normalized environmental emission index value, X represents the target value of the environmental emission index, Xmin represents the minimum value of the environmental emission index, and Xmax represents the maximum value of the environmental emission index.

[0088] Step S42: NSGA-III Optimization Process Reference Point Generation: The core of the third-generation non-dominated sorting genetic algorithm NSGA-III is to provide a set of reference points uniformly distributed in the target space to guide the population towards a complete Pareto front. For a problem with M objectives, the reference points are systematically placed on an M-1 dimensional unit hyperplane (i.e., all coordinates sum to 1). Based on the number of objectives (M=3) and the number of partitions (8 partitions, 100 population size, 50 selection rounds), the system generates a set of structured reference points on the unit hyperplane using the systematic methods of Das and Dennis: (1, 0, 0), (0, 1, 0), (0, 0, 1), (0.5, 0.5, 0), (0.5, 0, 0.5), (0, 0.5, 0.5). These points are uniformly distributed on an equilateral triangular plane formed by the three-dimensional target space of "yield-NH3-GWP".

[0089] Non-dominated ranking: All individuals in the population (i.e., technological situations) are compared pairwise and ranked non-dominatedly based on their performance on the three objectives. Individuals that are not "beaten" by any other individual on all objectives (i.e., superior on at least one objective) are designated as the first frontier (Pareto Rank 1), then removed. The second frontier is then identified from the remaining individuals, and so on. The ranking reflects the overall superiority or inferiority of the individuals.

[0090] Reference-Point-Based Balanced Screening: To ensure that the final selected technical solutions achieve optimal performance while comprehensively covering different objective trade-offs, this module employs a reference-point-based intelligent balanced screening mechanism. The core of this mechanism lies in "filling in the gaps," using a forced algorithm to explore all preset optimization angles. This ensures that the final output "best solution set" not only performs excellently but also exhibits significant differences and a wide distribution, fully showcasing all possible excellent trade-offs between "high output, low emissions, and low cost," providing decision-makers with a diverse selection menu.

[0091] Iteration and Convergence: In the application of this invention, since the set of technological scenarios is discrete and pre-simulated, the crossover and mutation operations of the genetic algorithm are simplified. The core process is an iterative screening based on the aforementioned non-dominated sorting and reference point association selection. Starting from the initial entire set (population), through multiple rounds of selection, dominated or overly densely distributed individuals are gradually eliminated. The iteration continues until a preset number of screening rounds is reached, or the optimal frontier members no longer change (convergence) for several consecutive times. The final output is a set of non-dominated solutions (Pareto optimal solution set). Any technological scenario in this set cannot improve one objective without harming at least one other objective. This provides decision-makers with a clear, comprehensive, and balanced set of optimal technology combinations, avoiding the outcome clustering at a single extreme, such as pursuing only the highest yield or only the lowest emissions.

[0092] S50 outputs decision support information, generating a spatial distribution map of optimized management measures based on the final solution selected from the Pareto optimal solution set.

[0093] This module transforms the optimized filtering results into understandable and actionable decision support products tailored to different users.

[0094] Step S51: Extraction and mapping of the optimal technical solution For each grid cell, from the Pareto optimal solution set selected by the NSGA-III algorithm, a final solution can be selected based on the decision-maker's preferences (e.g., the optimal solution that can synergistically achieve increased yield and reduced NH3 and GWP). The system then extracts the corresponding specific management measure combinations, including precise total nitrogen application rates, straw nitrogen substitution rates, straw return methods, tillage methods, and water management methods, and generates a spatial distribution map of optimization measures at the national or regional scale, visually demonstrating the spatial heterogeneity of the technologies.

[0095] Step S52: Potential Assessment and Decision Report Generation Optimization potential calculation: The predicted results (NH3, GWP, and yield) of each grid after adopting the optimized scheme are compared with the simulation results of the current conventional management practices for that grid (i.e., the predicted values ​​of the baseline scheme). The percentages of increased yield potential, increased income potential, and emission reduction potential are calculated, and a total potential assessment report for the region and the whole country is generated. The baseline scheme is based on other straw return strategies commonly adopted by farmers and corresponding nitrogen fertilizer and tillage management measures, and the data comes from farmer surveys or census data.

[0096] Structured Decision Report Output: The system can generate detailed, structured decision reports for specified locations (provinces, cities, counties, or specific latitudes and longitudes). The report content includes: the location's climate and soil conditions, estimated effects from current practices, details of recommended optimization techniques, the expected multi-objective benefits of implementing the proposed solutions, and the improvement compared to the current situation.

[0097] Step S53: Interactive Decision Support The system can be deployed as an online platform or desktop application, providing interactive functionality. Users can click on any location on the map, and the system instantly retrieves the Pareto optimal solution set for that location's grid, dynamically displaying the trade-offs between "production, NH3, and GWP" in the form of a 3D scatter plot or parallel coordinate graph. Users can set their preference weights for each objective via sliders (e.g., prioritizing production or emissions reduction), and the system recommends the most suitable technical solution from the Pareto solution set in real time, achieving personalized, interactive, and intelligent decision support.

[0098] This application employs a multi-task deep learning model, which has the following advantages: (1) Achieving collaborative prediction: The model can output the predicted values ​​of the three objectives at once and jointly, which inherently describes the complex relationship and trade-off between them. This is the fundamental premise for multi-objective collaborative optimization.

[0099] (2) Improve data utilization efficiency and model generalization ability: Shared learning reduces the risk of overfitting, and the masking mechanism makes full use of all effective observation data, enabling the model to be trained robustly even under data scarcity conditions, and enhancing its generalization ability in different regions across the country.

[0100] (3) High prediction accuracy: Deep networks can capture high-dimensional nonlinear relationships that are difficult to describe by traditional empirical models, thus providing more accurate predictions of "management measures-multi-objective response".

[0101] This application provides a high-precision, comprehensive, and multi-objective synergistic strategy for straw return to the field. Specific improvements include: (1) Rasterized scenario analysis: The main production areas in the country are divided into 10km×10km geographical grids. Baseline data (climate, soil) are integrated into each grid and combined with all feasible and detailed management variables (such as total nitrogen application, straw nitrogen substitution rate, straw return to field method, water management, etc.) to generate tens of thousands of specific management scenarios and build a nationwide “scenario-outcome” simulation database.

[0102] (2) Integrated NSGA-III optimization screening: For the simulation results library of each grid, the NSGA-III algorithm is applied for screening. The core of the algorithm includes: non-dominated sorting to ensure the optimality of the solution; based on the association of structured reference points and niche preservation strategy, the algorithm is forced to explore each trade-off direction uniformly in the three-dimensional target space of "yield-NH3-GWP".

[0103] The corresponding advantages are: (1) Achieve precise spatial recommendation: Based on a large amount of field observation data, it breaks the limitation of regional technical recommendation mode and can tailor the best straw return program for any grid in the country.

[0104] (2) Provides a comprehensive and balanced set of optimal solutions: The final output is not a single "optimal solution", but a set of Pareto optimal solutions. The solutions in this set achieve the best trade-off state in all three objectives, which cannot be improved simultaneously, and are widely and evenly distributed in the objective space, fully demonstrating the best technology recommendations from "emphasizing high production" to "emphasizing emission reduction" to "synergistic production increase and emission reduction".

[0105] (3) Automated and optimized decision-making process: The complex multi-objective optimization problem is transformed into a computable automated process, which ensures the scientific nature and global optimality of the screening results.

[0106] Secondly, embodiments of this application also provide a decision-making system for returning farmland straw to the field, including: The data acquisition module acquires multi-source data on farmland crops and preprocesses it to obtain a standardized dataset. The model prediction module trains a multi-task deep learning model on a standardized dataset to predict target data under given farmland management measures. The target data includes crop yield, ammonia volatilization, and global warming potential. The multi-task deep learning model includes: a shared feature encoder, which receives preprocessed input features and extracts shared feature representations; multiple task-specific adapters, which generate task-sensitive features corresponding to the target based on the shared feature representations; and multiple prediction output layers, each connected to a task-specific adapter, to output the predicted value for the corresponding target. The unit prediction module divides the target area into multiple spatial units. Within each spatial unit, it simulates the target data prediction values ​​under different farmland management scenarios based on a multi-task deep learning model, forming a scenario result database for each spatial unit. The decision screening module uses a multi-objective optimization algorithm to screen the scenario result database for each spatial unit, and obtains a decision scheme for returning farmland straw to the field that represents the best trade-off relationship between the target data.

[0107] Those skilled in the art to which this application pertains will clearly understand that, for the sake of convenience and brevity, the specific working process of the system, device, module, or unit described above can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0108] Thirdly, embodiments of this application also provide an electronic device, which includes a processor and a memory. The memory stores at least one instruction, at least one program, code set, or instruction set. The processor loads and executes the at least one instruction, at least one program, code set, or instruction set to implement the farmland straw return decision optimization method in embodiments of this application.

[0109] Fourthly, embodiments of this application also provide a computer-readable storage medium storing at least one instruction, at least one program, code set, or instruction set, wherein the at least one instruction, at least one program, code set, or instruction set is loaded and executed by a processor to implement the farmland straw return decision optimization method in embodiments of this application.

[0110] In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented in software, it can be implemented, in whole or in part, as a computer program product. A computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the flow or function according to the embodiments of this application is generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that integrates one or more available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium (e.g., a solid-state disk (SSD)).

[0111] The above are merely preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the technical principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for optimizing the decision-making process of returning farmland straw to the field, characterized in that, include: Obtain multi-source data on farmland crops and preprocess them to obtain a standardized dataset; A multi-task deep learning model is trained based on the standardized dataset to predict target data under given farmland management measures. The target data includes at least crop yield, ammonia volatilization loss, and global warming potential. The multi-task deep learning model comprises: a shared feature encoder for receiving preprocessed input features and extracting shared feature representations; multiple task-specific adapters for generating task-sensitive features corresponding to the target based on the shared feature representations; and multiple prediction output layers, each connected to one of the task-specific adapters, for outputting predicted values ​​for the corresponding target. The target area is divided into multiple spatial units. Within each spatial unit, the target data prediction values ​​under different farmland management scenarios are simulated based on the multi-task deep learning model, forming a scenario result database for each spatial unit. For the scenario result database of each spatial unit, a multi-objective optimization algorithm is used for screening to obtain a decision scheme for returning farmland straw to the field that represents the best trade-off relationship between the target data.

2. The method for optimizing the decision-making process of returning farmland straw to the field according to claim 1, characterized in that, The multi-source data includes climate data, soil data, management measure data, and target data; among which, The climate data includes average temperature and cumulative precipitation during the crop growing season; The soil data includes soil pH, organic carbon content, total nitrogen content, bulk density, and clay content; The management measures data include the amount of chemical nitrogen fertilizer applied, the proportion of nitrogen substitution in straw, the type of nitrogen fertilizer, the water management method, the tillage method, the amount of straw returned to the field, and the method of straw return to the field; The target data include crop yield, ammonia volatilization loss, and global warming potential.

3. The method for optimizing the decision-making process of returning farmland straw to the field according to claim 2, characterized in that, The preprocessing includes: Encode the type data as an integer; Perform Z-score standardization on numerical data; For data where some target variables are missing, a mask matrix is ​​generated to identify the locations of valid data.

4. The method for optimizing the decision-making process of returning farmland straw to the field according to claim 3, characterized in that, When training the multi-task deep learning model, a dynamic weight allocation mechanism based on task variance is used to calculate the total loss. The total loss is the average of the losses of each task weighted according to the reciprocal of their variances. When calculating the loss of each task, the loss terms corresponding to missing data are ignored in combination with the mask matrix.

5. The method for optimizing the decision-making process of returning farmland straw to the field according to claim 1, characterized in that, The farmland management scenario is constructed in the following way: Define a technology package of multiple farmland management measures formed by categorical management variables; Set the value range and step size for continuous management variables to generate multiple value points; Each of the aforementioned technology packages is combined with the value points of all continuous management variables to form the different farmland management scenarios.

6. The method for optimizing the decision-making process of returning farmland straw to the field according to claim 1, characterized in that, The multi-objective optimization algorithm is the third-generation non-dominated sorting genetic algorithm NSGA-III; the screening process using the multi-objective optimization algorithm includes: The crop yield, ammonia volatilization loss, and global warming potential forecasts for each farmland management scenario were normalized. Generate reference points that are uniformly distributed in the three-dimensional target space; Stratification of scenarios is based on non-dominated ordering; Based on the relationship between the scenario and the reference point, a balanced selection is made, and the decision on returning farmland straw to the field is iteratively filtered out.

7. The method for optimizing the decision-making process of returning farmland straw to the field according to claim 1, characterized in that, The method further includes generating a spatial distribution map of optimized management measures based on the selected farmland straw return decision scheme.

8. The method for optimizing the decision-making process of returning farmland straw to the field according to claim 1, characterized in that, The method also includes: generating a report assessing the potential for increased production and reduced emissions by comparing the prediction results of the selected farmland straw return decision scheme with the baseline scheme.

9. A decision optimization system for returning farmland straw to the field, characterized in that, include: The data acquisition module acquires multi-source data on farmland crops and preprocesses it to obtain a standardized dataset. The model prediction module trains a multi-task deep learning model based on the standardized dataset to predict target data under given farmland management measures. The target data includes at least crop yield, ammonia volatilization loss, and global warming potential. The multi-task deep learning model includes: a shared feature encoder for receiving preprocessed input features and extracting shared feature representations; multiple task-specific adapters for generating task-sensitive features corresponding to the target based on the shared feature representations; and multiple prediction output layers, each connected to one of the task-specific adapters, for outputting predicted values ​​for the corresponding target. The unit prediction module divides the target area into multiple spatial units, and simulates the target data prediction values ​​under different farmland management scenarios based on the multi-task deep learning model in each spatial unit, forming a scenario result database for each spatial unit. The decision screening module uses a multi-objective optimization algorithm to screen the scenario result database for each spatial unit to obtain a decision scheme for returning farmland straw to the field that represents the best trade-off relationship between the target data.

10. An electronic device, characterized in that, The electronic device includes a processor and a memory, wherein the memory stores at least one instruction, at least one program, a code set, or an instruction set, and the at least one instruction, the at least one program, the code set, or the instruction set is loaded and executed by the processor to implement the farmland straw return decision optimization method as described in any one of claims 1-8.

Images