A corn intelligent fertilization system based on soil microbial regulation

Abstract

The application relates to the technical field of intelligent fertilization control, and discloses a corn intelligent fertilization system based on soil microbial regulation. The system realizes robust identification and interpretable quantification of the microbial regulation relationship by constructing a microbial-nutrient cascade state diagram, performing cascade state modeling, combining with a plug-in type Jacobian influence matrix obtained through automatic differentiation and a Neyman orthogonal debiasing mechanism. On this basis, a credible regulation path is screened, a microbial regulation coefficient and a nutrient response coefficient are generated, and a micro-area variable topdressing prescription is constructed by combining with crop fertilizer demand constraints, so that synergistic optimization control of nitrogen fertilizer, microbial agents and inhibitors is realized. The method improves the precision and interpretability of fertilization decision, reduces the risk of nitrogen loss, and enhances the dynamic adaptability of variable fertilization.

Description

Technical Field

[0001] This invention relates to the field of intelligent fertilization control technology, and in particular to an intelligent fertilization system for corn based on soil microbial regulation. Background Technology

[0002] As a crucial food crop, maize's yield is significantly influenced by topdressing during the jointing stage. However, existing fertilization technologies largely rely on average soil nutrient levels or empirical models for uniform or coarse-grained regional fertilization, failing to reflect the spatial heterogeneity and dynamic changes of the rhizosphere environment within the field. Existing variable fertilization systems primarily rely on static or apparent characteristics such as soil physicochemical indicators or vegetation indices for decision-making, neglecting the regulatory role of soil microbial functional groups in nitrogen fixation, mineralization, nitrification, and denitrification processes, resulting in an incomplete characterization of nitrogen transformation mechanisms. Furthermore, existing data-driven fertilization models often employ black-box prediction methods, which, while possessing some fitting ability, lack a structured expression of the causal relationship between microorganisms and nutrients, limiting model interpretability and cross-regional generalization. Moreover, existing variable fertilization control strategies typically make decisions based on single-moment states, lacking a comprehensive assessment of the temporal evolution of nitrogen transformation processes and the risk of nitrogen loss, making it difficult to achieve synergistic optimization between crop nutrient requirements and environmental risk control. Therefore, it is necessary to propose an intelligent fertilization system that integrates soil microbial regulation mechanisms and multi-source temporal state modeling to improve the accuracy, dynamic adaptability, and eco-friendly nature of topdressing during the jointing stage of maize. Summary of the Invention

[0003] This invention aims to provide a smart maize fertilization system based on soil microbial regulation, addressing the shortcomings of existing variable fertilization technologies, such as a lack of characterization of rhizosphere microbial regulation mechanisms, insufficient modeling of nitrogen transformation processes, and weak interpretability of fertilization decisions. The core of this invention lies in constructing a microbial-nutrient cascade state modeling and causal regulation analysis framework. This framework unifies and abstracts maize rhizosphere microbial functional groups, soil nutrient pools, crop absorption status, and nitrogen loss risk status, and forms cascaded transfer samples based on multi-time-stack observation data to achieve dynamic characterization of rhizosphere nitrogen transformation processes. Based on this, a cascaded state transition function is introduced to model the system evolution relationship, and structured sensitive relationships between state variables are obtained through automatic differentiation, forming a plug-in Jacobian influence matrix. This enables an interpretable and quantitative expression of the influence of microorganisms on nutrient transformation and crop absorption. Simultaneously, addressing the causal confounding and estimation bias problems of traditional data-driven models, a bias correction mechanism based on Neyman orthogonality is introduced, combined with Riesz representation terms to correct prediction residuals, obtaining a robust bias-corrected regulation matrix and edge confidence assessment results, thereby improving the reliability of microbial regulation relationship identification. Based on the screening results of credible regulatory relationships, the microbial regulatory effect is transformed into regulatory coefficients and nutrient response coefficients. Combined with the target fertilizer requirement constraints and fertilization safety constraints of maize, a micro-region-level variable topdressing prescription is generated to achieve differentiated and precise fertilization control in multiple rhizosphere micro-regions. This enables the transformation from experience-driven to intelligent variable fertilization decision-making driven by microbial causal mechanisms.

[0004] This invention proposes a smart fertilization system for maize based on soil microbial regulation. This system is used to correct variable topdressing amounts in multiple rhizosphere topdressing management micro-zones within the topdressing window during the maize jointing stage. The system includes:

[0005] The rhizosphere state construction module is used to divide the target maize field into multiple rhizosphere topdressing management micro-regions. Before the start of the current topdressing window, it acquires the microbial nitrogen regulation observation data and field covariates corresponding to each rhizosphere topdressing management micro-region, and constructs micro-region-level time-series samples based on the microbial nitrogen regulation observation data and field covariates. It also constructs a microbial-nutrient cascade state diagram based on the micro-region-level time-series samples, and pairs micro-region-level time-series samples at adjacent observation times as cascaded transfer samples.

[0006] The microbial regulation impact recovery module trains a first-level cascade transfer function estimator based on cascade transfer samples. Then, it performs Jacobian calculation using the trained first-level cascade transfer function estimator to obtain a plug-in Jacobian impact matrix. Based on the prediction residuals output by the first-level cascade transfer function estimator and the Riesz characterization terms corresponding to each candidate regulatory edge in the microbial-nutrient cascade state graph, it performs Neyman orthogonal debiasing correction on the plug-in Jacobian impact matrix to obtain the debiased Jacobian regulation matrix and edge confidence.

[0007] The reliable control edge parsing module is used to screen reliable control edges from the microbial-nutrient hierarchical association state diagram based on the de-biased Jacobi control matrix and edge reliability, and generate the microbial control coefficient and nutrient response coefficient corresponding to each rhizosphere topdressing management micro-zone based on the reliable control edges.

[0008] The topdressing action generation module is used to construct a set of candidate topdressing actions based on the microbial regulation coefficient, nutrient response coefficient, maize target topdressing requirements and topdressing safety constraints, and to determine the target topdressing action from the set of candidate topdressing actions to generate a micro-region variable topdressing prescription for maize at the jointing stage.

[0009] The variable fertilization execution module is used to control the variable fertilization equipment to perform topdressing operations in the corresponding rhizosphere topdressing management micro-zone according to the variable topdressing prescription for the corn jointing stage micro-zone.

[0010] The feedback update module is used to obtain feedback observation data after topdressing operations, add the feedback observation data as a new cascade transfer sample, and update the topdressing prescription for the micro-regional variable of maize jointing stage in the next topdressing window.

[0011] By adopting the above solution, the beneficial effects achieved by the present invention are as follows:

[0012] This invention constructs a microbial-nutrient cascade state diagram and introduces cascade state transition modeling and microbial functional group regulation analysis mechanisms to achieve dynamic characterization and interpretable expression of nitrogen transformation processes in the rhizosphere of maize. This improves upon the insufficient response of traditional fertilization techniques to rhizosphere microecological processes, transforming fertilization decisions from static experience-driven to dynamic regulation-driven based on microbial causal mechanisms. Consequently, it significantly improves the accuracy and adaptability of topdressing during the jointing stage of maize, solves the problems of existing technologies being unable to accurately characterize micro-regional differences within the field and unclear nitrogen transformation pathways, and enhances the stability and transferability of variable fertilization models in complex field environments.

[0013] This invention introduces a class-connected state transition function and a plug-in Jacobian influence matrix to achieve a structured and quantitative expression of the influence of microbial functional groups on nutrient pools and crop absorption status. It also incorporates a Neyman orthogonal correction mechanism to robustly correct causal regulatory relationships, effectively reducing the risk of misjudgment caused by causal confounding and estimation bias in traditional data-driven models. This improves the accuracy and reliability of identifying microbial regulatory relationships, solves the problems of poor interpretability and insufficient cross-regional generalization ability of black-box models in existing technologies, and enhances the scientific basis and reliability of fertilization decisions.

[0014] This invention generates microbial regulation coefficients and nutrient response coefficients based on the results of reliable control edge screening, and constructs variable topdressing prescriptions by combining maize target fertilizer requirements and fertilization safety constraints. This achieves synergistic optimization control of nitrogen fertilizer, microbial agents, and inhibitors, enabling the fertilization strategy to effectively reduce the risk of nitrate nitrogen leaching and nitrous oxide emissions while meeting crop growth needs. This achieves a synergistic improvement in yield efficiency and environmental safety, solving the problem of low nutrient utilization and high environmental load in existing fertilization methods. It also significantly enhances the engineering feasibility and closed-loop adaptive capability of the system in actual farmland variable fertilization operations. Attached Figure Description

[0015] Figure 1 This is a schematic diagram of the module structure of a smart corn fertilization system based on soil microbial regulation in an embodiment of the present invention;

[0016] Figure 2 This is a schematic diagram of the plug-in Jacobian influence matrix proposed in Embodiment 7 of the present invention; Figure 2 In the matrix, the vertical axis represents source nodes, and the horizontal axis represents target nodes. Source nodes include nitrogen fixation, mineralization, nitrification, denitrification, ammonium nitrogen, and nitrate nitrogen. Target nodes include ammonium nitrogen, nitrate nitrogen, organic carbon, root uptake, loss risk, and functional groups. Nitrogen fixation, mineralization, nitrification, and denitrification represent the nitrogen fixation functional group, organic matter mineralization functional group, nitrification functional group, and denitrification functional group, respectively. Loss risk represents the nitrogen loss risk status. Cells in the matrix represent the initial influence relationship between the corresponding source node and the corresponding target node; "+" indicates a promoting initial influence, and "-" indicates an inhibiting initial influence. The intensity of the cell color indicates the strength of the influence; darker colors indicate stronger influences. Blank cells indicate weak influence relationships or relationships not identified as significant. Detailed Implementation

[0017] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0018] Example 1: System operating environment and module structure:

[0019] like Figure 1As shown in this embodiment, a smart fertilization system for maize based on soil microbial regulation is provided for correcting the variable topdressing amount in multiple rhizosphere topdressing management microzones within the topdressing window during the maize jointing stage. This system can be deployed on an agricultural management server, a field edge computing device, a vehicle-mounted control terminal of a variable fertilization device, or an agricultural IoT platform composed of the above devices.

[0020] In this embodiment, the system includes a rhizosphere state construction module, a microbial regulation influence recovery module, a reliable regulation edge analysis module, a topdressing action generation module, a variable fertilization execution module, and a feedback update module. The rhizosphere state construction module acquires and organizes microbial nitrogen regulation observation data, field covariates, and topdressing regulation inputs for each rhizosphere topdressing management micro-zone. The microbial regulation influence recovery module trains a cascaded transfer function estimator, acquires a plug-in Jacobian influence matrix, and performs Neyman orthogonal debiasing correction on the plug-in Jacobian influence matrix. The reliable regulation edge analysis module filters reliable regulation edges and generates microbial regulation coefficients and nutrient response coefficients. The topdressing action generation module generates variable topdressing prescriptions for maize micro-zones at the jointing stage. The variable fertilization execution module controls the variable fertilization equipment to perform topdressing operations. The feedback update module appends feedback observation data after topdressing operations as new cascaded transfer samples.

[0021] Specifically, the target cornfield is divided into four rhizosphere fertilization management micro-zones: the first, second, third, and fourth. In this embodiment, the four rhizosphere fertilization management micro-zones are divided according to the operation path of the variable fertilization equipment; specifically, the target cornfield is divided into four continuous operation areas along the operation direction of the variable fertilization equipment, with each continuous operation area corresponding to one rhizosphere fertilization management micro-zone. Each rhizosphere fertilization management micro-zone is equipped with one root zone soil sampling point and one soil temperature and humidity collection point, and a corresponding variable fertilization operation path is set. In other embodiments, the number of root zone soil sampling points and soil temperature and humidity collection points can be increased according to the area of ​​the rhizosphere fertilization management micro-zone, soil heterogeneity, and sampling accuracy requirements; alternatively, the target cornfield can be divided according to differences in soil fertility within the field, irrigation zones, differences in corn growth, or the distribution of rhizosphere sampling points.

[0022] In this embodiment, the current topdressing window is set to the topdressing management period from the early to mid-jointing stage of maize, used to generate and execute the micro-regional variable topdressing prescription for this round of maize jointing stage. The observation time is determined according to a field sampling cycle of once every 5 days, that is, the rhizosphere state construction module acquires microbial nitrogen regulation observation data and field covariates of each rhizosphere topdressing management micro-region every 5 days, and uses the data corresponding to two adjacent observation times to form cascade transfer samples. In other embodiments, the current topdressing window can also be set to other topdressing management periods within the jointing stage according to maize variety, local climate conditions, soil type, and field management system, and the field sampling cycle can also be set to 3 days, 7 days, or other time intervals suitable for forming micro-regional time-series samples.

[0023] Example 2: Microbial nitrogen regulation observation data, field covariates, and topdressing regulation inputs:

[0024] In this embodiment, the microbial nitrogen regulation observation data includes the abundance of microbial functional groups, microbial enzyme activity, available soil nutrients, maize rhizosphere absorption status, and nitrogen loss risk status. The abundance of microbial functional groups is used to characterize the relative abundance of different microbial functional groups in the corresponding rhizosphere topdressing management micro-zone. The microbial functional groups include nitrogen fixation functional groups, nitrification functional groups, denitrification functional groups, and organic matter mineralization functional groups.

[0025] The abundance of nitrogen fixation functional groups reflects the potential ability of the corresponding rhizosphere topdressing management microregion to convert atmospheric nitrogen or other nitrogen sources into ammonium nitrogen pools; the abundance of nitrification functional groups reflects the potential ability of ammonium nitrogen pools to convert into nitrate nitrogen pools; the abundance of denitrification functional groups reflects the potential risk of nitrate nitrogen being converted into gaseous nitrogen loss under low oxygen or high water content conditions; and the abundance of organic matter mineralization functional groups reflects the potential ability of soil organic matter decomposition to supply soluble organic carbon pools and ammonium nitrogen pools. The abundance of microbial functional groups was obtained through high-throughput sequencing.

[0026] Microbial enzyme activity includes urease activity, nitrate reductase activity, dehydrogenase activity, and cellulase activity. Microbial enzyme activity is used to supplement and characterize the actual activity level of microbial functional groups. For example, when the abundance of nitrification functional groups is high and the related enzyme activities are strong, the nitrification conversion capacity of the corresponding rhizosphere fertilization management micro-zone may be stronger.

[0027] The nutrient pools corresponding to available nutrients in soil include ammonium nitrogen, nitrate nitrogen, and soluble organic carbon. The ammonium nitrogen pool characterizes the content of ammonium nitrogen in the rhizosphere soil that can be absorbed or undergo further nitrification; the nitrate nitrogen pool characterizes the content of nitrate nitrogen in the rhizosphere soil that can be absorbed by maize but is also easily leached; and the soluble organic carbon pool characterizes the state of available carbon sources that can influence microbial activity and denitrification. These nutrient pools are determined from soil sample analysis data.

[0028] The root absorption status of maize corresponds to the chlorophyll index, the normalized normalized vegetation index (NMR) of the canopy, and the root zone water content. The chlorophyll index, obtained by a chlorophyll meter, reflects the nitrogen nutrition status of maize; the NMR is obtained by a ground canopy sensor and reflects the growth of the maize canopy; the root zone water content is obtained by a soil moisture sensor and is used to determine the water conditions for assessing the risk of nitrogen absorption and leaching.

[0029] Nitrogen loss risk status includes nitrate nitrogen leaching risk status and nitrous oxide emission risk status. Nitrate nitrogen leaching risk status is determined based on nitrate nitrogen pool, rainfall, irrigation amount, soil moisture content, and soil texture information; nitrous oxide emission risk status is determined based on denitrification functional group abundance, nitrate nitrogen pool, soluble organic carbon pool, soil moisture content, soil temperature, and pH value information. In this embodiment, nitrogen loss risk status can be normalized to a risk level between 0 and 1, with higher values ​​indicating higher risk.

[0030] Field covariates include soil temperature, soil moisture content, pH, electrical conductivity, rainfall, irrigation amount, historical fertilization amount, and current maize growth stage. Soil temperature and soil moisture content are continuously collected by field sensors; pH and electrical conductivity are obtained by soil testing equipment or soil samples; rainfall and irrigation amount are obtained from weather stations; historical fertilization amount includes the amount of nitrogen fertilizer, basal fertilizer, topdressing fertilizer, or other nitrogen source input applied in the corresponding rhizosphere fertilization management micro-plot before the current observation time; the current maize growth stage is used to indicate whether it is in the early, middle, or late jointing stage.

[0031] The topdressing regulation inputs were constructed based on the nitrogen fertilizer application rate, microbial inoculant application rate, nitrification inhibitor application rate, and organic carbon source supplementation rate in the corresponding rhizosphere topdressing management micro-zone prior to the current observation time. The nitrogen fertilizer application rate characterizes the nitrogen input in the previous management cycle; the microbial inoculant application rate characterizes the exogenous regulation of specific microbial functional groups; the nitrification inhibitor application rate characterizes the inhibition and regulation of the conversion rate of ammonium nitrogen to nitrate nitrogen; and the organic carbon source supplementation rate characterizes the regulation of microbial activity and the carbon-nitrogen coupling relationship. These topdressing regulation inputs were obtained from agricultural operation records.

[0032] Example 3: Microbial Cascade State Vector and Microbial-Cultural Cascade State Diagram:

[0033] The rhizosphere state construction module organizes the abundance of microbial functional groups, microbial enzyme activities, available soil nutrients, maize rhizosphere absorption status, and nitrogen loss risk status of each rhizosphere topdressing management micro-region at the same observation time into a microbial cascade state vector. The microbial cascade state vector can be arranged in a fixed field order, for example, including the abundance of nitrogen fixation functional groups, nitrification functional groups, denitrification functional groups, organic matter mineralization functional groups, urease activity, nitrate reductase activity, ammonium nitrogen pool, nitrate nitrogen pool, soluble organic carbon pool, chlorophyll index, canopy normalized vegetation index, rhizosphere water content, nitrate nitrogen leaching risk status, and nitrous oxide emission risk status. The above field order is only an example; in actual implementation, it is sufficient to keep the fields consistent between the training and execution phases.

[0034] The rhizosphere state construction module also includes soil temperature, soil moisture content, pH, electrical conductivity, rainfall, irrigation amount, historical fertilizer application, and current maize growth stage tissues at the same observation time as field covariates. To facilitate model processing, the microbial cascade state vector and field covariates can undergo missing value imputation, outlier checking, unit unification, and normalization. For indicators with different dimensions, such as abundance, content, index, and risk level, normalization based on historical observation range or piecewise standardization based on agronomic thresholds can be used, respectively.

[0035] The system uses microbial functional groups, nutrient pools, root absorption status, and nitrogen loss risk status as graph nodes, and rhizosphere nitrogen transformation chains and spatial adjacency relationships of rhizosphere topdressing management micro-regions as candidate edges to construct a microbial-nutrient hierarchical association state graph. The graph nodes in the microbial-nutrient hierarchical association state graph include microbial functional group nodes, nutrient pool nodes, root absorption status nodes, and nitrogen loss risk status nodes.

[0036] The rhizosphere nitrogen transformation chain is constructed based on the nitrogen transformation direction between microbial functional groups, nutrient pools, root absorption states, and nitrogen loss risk states. In this embodiment, the rhizosphere nitrogen transformation chain includes a nitrogen fixation replenishment chain from the nitrogen fixation functional group to the ammonium nitrogen pool, an organic matter mineralization supply chain from the organic matter mineralization functional group to the soluble organic carbon pool and the ammonium nitrogen pool, a nitrification transformation chain from the ammonium nitrogen pool to the nitrate nitrogen pool via the nitrification functional group, a denitrification risk chain from the nitrate nitrogen pool to the nitrogen loss risk state via the denitrification functional group, a leaching risk chain from the nitrate nitrogen pool to the nitrogen loss risk state, and a nutrient absorption chain from the ammonium nitrogen pool and the nitrate nitrogen pool to the root absorption state.

[0037] Among them, the nitrogen fixation replenishment chain is used to characterize the replenishment effect of the nitrogen fixation functional group on the ammonium nitrogen pool; the organic matter mineralization supply chain is used to characterize the supply effect of the organic matter mineralization functional group on the soluble organic carbon pool and the ammonium nitrogen pool; the nitrification transformation chain is used to characterize the process of the ammonium nitrogen pool transforming into the nitrate nitrogen pool under the action of the nitrification functional group; the denitrification risk chain is used to characterize the process of the nitrate nitrogen pool increasing the risk of nitrogen loss under the action of the denitrification functional group; the leaching risk chain is used to characterize the process of the nitrate nitrogen pool increasing the risk of nitrate nitrogen leaching under rainfall or irrigation conditions; and the nutrient absorption chain is used to characterize the supply effect of the ammonium nitrogen pool and the nitrate nitrogen pool on the root absorption state.

[0038] The spatial adjacency relationship of rhizosphere fertilization management microregions is used to characterize the adjacency relationships between four rhizosphere fertilization management microregions. For example, the first rhizosphere fertilization management microregion is adjacent to the second rhizosphere fertilization management microregion, the second rhizosphere fertilization management microregion is adjacent to the third rhizosphere fertilization management microregion, and the third rhizosphere fertilization management microregion is adjacent to the fourth rhizosphere fertilization management microregion. The spatial adjacency relationship can be used to add candidate edges between microregions in the microbial-nutrient hierarchical association state diagram to characterize the influence of neighboring microregions caused by water migration, nutrient diffusion, or the continuity of fertilization machinery operations.

[0039] Example 4: Construction of Cascaded Transfer Samples:

[0040] The rhizosphere state construction module pairs micro-region-level time-series samples from adjacent observation times into cascaded transfer samples. Specifically, for the same rhizosphere topdressing management micro-region, the system uses the microbial cascade state vector, field covariates, and topdressing regulation input at the current observation time as transfer input, and the microbial cascade state vector at the next observation time as transfer output, forming a cascaded transfer sample.

[0041] For example, in the first rhizosphere topdressing management micro-region, the system acquires the microbial cascade state vector, field covariates, and topdressing regulation input at the first observation time during the jointing stage, and acquires the microbial cascade state vector at the next observation time during the second observation time during the jointing stage. The system uses the data from the first observation time as the transfer input and the microbial cascade state vector from the second observation time as the transfer output, forming a cascade transfer sample for the first rhizosphere topdressing management micro-region. For the second, third, and fourth rhizosphere topdressing management micro-regions, the system uses the same method to form corresponding cascade transfer samples.

[0042] In this embodiment, cascaded transfer samples are used to train a cascaded transfer function estimator. The topdressing regulation input in the transfer input enables the model to learn the effects of different nitrogen fertilizer application rates, microbial inoculant application rates, nitrification inhibitor application rates, and organic carbon source supplementation rates on the microbial-nutrient state at the next observation time. The transfer output serves as the actual observation target and is used to calculate the error between the model-generated predicted cascaded state vector and the actual observation target.

[0043] Example 5: Training a one-level class transition function estimator:

[0044] The microbial regulation impact recovery module trains a first-order cascade transfer function estimator based on cascade transfer samples. This first-order cascade transfer function estimator is used to predict the predicted cascade state vector for the next observation time, based on the current microbial cascade state vector, field covariates, and topdressing regulation input.

[0045] During training, the microbial regulation impact recovery module acquires transfer inputs and outputs from the cascade transfer samples. Transfer inputs include the microbial cascade state vector at the current observation time, field covariates, and topdressing regulation inputs; transfer outputs include the microbial cascade state vector actually observed at the next observation time. The system loads the transfer inputs into a cascade transfer function estimator, generating predicted cascade state vectors through a state index generation unit and a node type mapping unit.

[0046] In this embodiment, the microbial cascade state vector includes 14 graph node states, namely 4 microbial functional group node states, 2 microbial enzyme activity node states, 3 nutrient pool node states, 3 root absorption state node states, and 2 nitrogen loss risk state node states. Field covariates include 8 covariates: soil temperature, soil moisture content, pH value, electrical conductivity, rainfall, irrigation amount, historical fertilizer application, and current maize growth stage. Topdressing regulation inputs include 4 regulation quantities: nitrogen fertilizer application rate, microbial inoculant application rate, nitrification inhibitor application rate, and organic carbon source supplementation rate. Therefore, the single training input of the first-order cascade transfer function estimator consists of the 14-dimensional microbial cascade state vector at the current observation time, 8-dimensional field covariates, and 4-dimensional topdressing regulation inputs, with the training objective being the 14-dimensional microbial cascade state vector at the next observation time.

[0047] The state index generation unit is used to weight and combine the microbial cascade state vector, field covariates, and topdressing regulation inputs at the current observation time to generate the state index value corresponding to each node in the microbial-nutrient cascade state diagram. The state index value can be understood as the intermediate prediction value of each node before entering the next observation time, after considering the current microbial-nutrient state, field environmental conditions, and previous topdressing regulation inputs.

[0048] The node type mapping unit is used to map state index values ​​using different link functions based on the data type and value range of the graph nodes, generating a predicted cascaded state vector. For microbial functional group nodes and nutrient pool nodes, since the abundance of microbial functional groups and the content of nutrient pools should not be negative in an agricultural observation sense, a non-negative link function is used for mapping. In this embodiment, the ReLU function is used to ensure that the predicted abundance and predicted nutrient content values ​​are not less than zero. For nitrogen loss risk state nodes, since the nitrate nitrogen leaching risk state and nitrous oxide emission risk state are represented by a normalized risk level between 0 and 1 in this embodiment, a Sigmoid link function is used for mapping so that the predicted risk value falls within the 0 to 1 range. For the root absorption status node, since the chlorophyll index, canopy normalized vegetation index, and root zone water content are continuous observations, an identity link function is used for mapping in this embodiment. In other embodiments, when there is a significant nonlinear response relationship between the root absorption status and the microbial functional groups and nutrient pools, a multilayer perceptron link function with a hidden layer can also be used for mapping, and the number of hidden units in the hidden layer can be set to 16 or 32.

[0049] The microbial regulation impact recovery module constructs a regularized prediction loss based on the error between the predicted cascade state vector and the transition output. The regularized prediction loss includes a prediction error term, a sparse penalty term for the regulation weight matrix, and an agronomic consistency constraint term. The prediction error term constrains the predicted cascade state vector to closely approximate the observed transition output; the sparse penalty term for the regulation weight matrix encourages the model to retain only stable and necessary regulatory relationships, reducing the interference of meaningless node relationships on subsequent Jacobian calculations; the agronomic consistency constraint term constrains the prediction results to meet basic agronomic relationships, such as non-negative abundance of microbial functional groups and nutrient pool content, risk states not exceeding the normalized range, and the direction of nutrient response change after topdressing not conflicting with basic nitrogen supply logic.

[0050] In this embodiment, the weights of the prediction error term, the sparsity penalty term of the control weight matrix, and the agronomic consistency constraint term are set to 1, 0.001, and 0.1, respectively. The agronomic consistency constraint term includes the following constraints: the predicted values ​​of the microbial functional group node and the nutrient pool node are not less than 0; the predicted values ​​of the nitrogen loss risk state node are between 0 and 1; when the nitrate nitrogen leaching risk state is higher than 0.7, a penalty is imposed on the prediction result that leads to a further increase in the nitrate nitrogen pool without improving root absorption; when the nitrous oxide emission risk state is higher than 0.7, a penalty is imposed on the prediction result that leads to a further increase in denitrification risk. The above weights and risk thresholds are only parameter settings in this embodiment. In other embodiments, they can be adjusted according to the historical field sample size, risk control requirements, and model validation results.

[0051] The microbial regulation impact recovery module updates the model parameters of the first-order enjoint transfer function estimator based on the regularized prediction loss. In this embodiment, the microbial regulation impact recovery module uses an adaptive moment estimator optimizer to update the model parameters of the first-order enjoint transfer function estimator. Specifically, it calculates the gradient of the model parameters based on the regularized prediction loss and adjusts the update step size of the model parameters based on the first-order and second-order moment estimates of the gradient to reduce the impact of differences in the state dimensions of different graph nodes on the stability of parameter updates. When the preset maximum number of iterations is reached, the prediction error on the validation set no longer decreases significantly, or the regularized prediction loss meets the convergence condition, the system saves the current model parameters, obtaining the trained first-order enjoint transfer function estimator.

[0052] In this embodiment, the initial learning rate of the adaptive moment estimation optimizer is set to 0.001, the first-order moment decay coefficient is set to 0.9, the second-order moment decay coefficient is set to 0.999, the training batch size is set to 16, and the maximum number of iterations is set to 200 rounds. During training, the microbial regulation influence recovery module calculates the prediction error on the validation subset after each iteration. When the prediction error of the validation subset does not decrease for 20 consecutive rounds, or the number of training iterations reaches 200 rounds, or the regularized prediction loss meets the preset convergence condition, training stops, and the model parameters corresponding to the lowest prediction error of the validation subset are saved, resulting in the trained first-order enjoined transfer function estimator. In other embodiments, gradient descent, stochastic gradient descent, or other parameter update methods suitable for machine learning model training can also be used to update the model parameters of the first-order enjoined transfer function estimator.

[0053] Example 6: Generation of cascaded state vectors for model execution and prediction after training:

[0054] During the model execution phase, the system no longer updates the model parameters of the first-level cascade transfer function estimator through cascade transfer samples. Instead, it loads the current transfer input corresponding to the current topdressing window into the trained first-level cascade transfer function estimator. The current transfer input includes the microbial cascade state vector at the current observation time, field covariates, and topdressing regulation input.

[0055] The trained first-order cascade transfer function estimator generates state index values ​​corresponding to each node in the microbial-nutrient cascade state graph through a state index generation unit, and maps the state index values ​​to the predicted cascade state vector for the next observation time through a node type mapping unit. The predicted cascade state vector includes the predicted abundance of microbial functional groups, the predicted nutrient pool state, the predicted root absorption state, and the predicted nitrogen loss risk state for the next observation time.

[0056] The predicted cascade state vector differs from the transition output in the cascade transition samples. The predicted cascade state vector is the model prediction result output by the first-stage cascade transition function estimator after training; the transition output is the microbial cascade state vector actually observed at the next observation time. During the training phase, the difference between the two is used to construct the prediction residual and prediction error term; during the execution phase, the predicted cascade state vector is used to perform Jacobian calculation and provides the state prediction basis for the generation of subsequent topdressing actions.

[0057] Example 7: Obtaining the Jacobian Influence Matrix Using a Plug-in Approach:

[0058] The microbial regulation impact recovery module uses the state of each graph node in the microbial cascade state vector at the current observation time as the derivative variable, and performs automatic differentiation on the partial derivative relationship of each graph node state in the predicted cascade state vector at the next observation time with respect to the derivative variable to obtain the plug-in partial derivative number between nodes.

[0059] Inter-node plug-in partial derivatives are used to represent the direction and magnitude of the predicted state change of a target node at the next observation time when the state of a source node changes by one unit at the current observation time. The microbial regulation influence recovery module arranges the inter-node plug-in partial derivatives according to the source and target nodes in the microbial-nutrient hierarchical state diagram to generate a plug-in Jacobian influence matrix.

[0060] like Figure 2 As shown, the rows of the plug-in Jacobian influence matrix correspond to source nodes, and the columns correspond to target nodes. Positive matrix elements indicate that the source node has a promoting initial influence on the target node; negative matrix elements indicate that the source node has a suppressive initial influence on the target node; the larger the absolute value of the matrix elements, the stronger the corresponding influence relationship. It should be noted that the plug-in Jacobian influence matrix is ​​obtained directly by differentiating the trained one-stage enjoined transfer function estimator, and therefore represents the initial influence strength result. Further bias correction using prediction residuals and the Riesz representation term is still required.

[0061] For ease of understanding, Table 1 lists the agricultural implications of some matrix elements in the plug-in Jacobian influence matrix:

[0062] Table 1

[0063] ;

[0064] Table 1 is only used to illustrate the agricultural meaning of some matrix elements in the plug-in Jacobian influence matrix, and does not limit the matrix to include only the node relationships mentioned above. In actual operation, the plug-in Jacobian influence matrix can cover all partial derivative relationships between candidate source nodes and candidate target nodes in the microbial-nutrient hierarchical association state diagram.

[0065] Example 8: Neyman orthogonal de-biasing correction, de-biasing Jacobian control matrix, and edge credibility:

[0066] Since the plug-in Jacobian influence matrix is ​​directly obtained by differentiating a first-order class transfer function estimator, its elements may be affected by model prediction errors, micro-region sample differences, field environmental disturbances, and sample size fluctuations. To reduce the sensitivity of the simple model differentiation results to prediction errors, this embodiment uses the prediction residuals output by the first-order class transfer function estimator and the Riesz characterization terms corresponding to each candidate regulatory edge in the microbial-nutrient class linkage state diagram to perform Neyman orthogonal debiasing correction on the plug-in Jacobian influence matrix, obtaining the debiased Jacobian regulation matrix and edge confidence.

[0067] The prediction residual is used to characterize the difference between the predicted cascaded state vector and the transition output. For example, if a cascaded transition function estimator predicts that the ammonium nitrogen pool state at the next observation time in the first rhizosphere topdressing management micro-region is 18, but the actual transition output shows an ammonium nitrogen pool state of 20, then the prediction residual corresponding to that node can reflect the degree to which the model underestimates the ammonium nitrogen pool state. The Riesz representation term is used to characterize the weighting effect of the corresponding candidate control edge in the residual correction. That is, the system uses the Riesz representation term corresponding to the candidate control edge to weight the prediction residual to form the residual correction amount oriented towards the candidate control edge.

[0068] Specifically, the microbial regulation impact recovery module extracts the plug-in partial derivative corresponding to each candidate regulatory edge in the plug-in Jacobian impact matrix. The candidate regulatory edges come from the microbial-nutrient hierarchical association state graph and have clear source nodes and target nodes. For example, there are candidate regulatory edges from nitrogen fixation functional group nodes to ammonium nitrogen pool nodes, candidate regulatory edges from nitrification functional group nodes to nitrate nitrogen pool nodes, and candidate regulatory edges from denitrification functional group nodes to nitrogen loss risk state nodes.

[0069] Subsequently, the microbial regulation impact recovery module performs a weighted correction on the prediction residuals based on the Riesz characterization term corresponding to the candidate regulation edge, obtaining the residual correction amount for that candidate regulation edge. The residual correction amount is used to compensate for the bias introduced by the prediction error in the plug-in partial derivatives. The system combines the plug-in partial derivatives corresponding to the candidate regulation edge with the residual correction amount to form an orthogonal score value. The orthogonal score value is used to represent the influence strength of the candidate regulation edge after residual correction.

[0070] The microbial regulation impact recovery module aggregates the orthogonal score values ​​corresponding to the same candidate regulatory edge in multiple cascaded transfer samples to obtain the debiased regulatory intensity of that candidate regulatory edge. Aggregation can employ averaging, weighted averaging, or robust statistical aggregation methods. Subsequently, the system reorganizes the debiased regulatory intensity of each candidate regulatory edge according to the source and target node arrangement in the plug-in Jacobian influence matrix to obtain the debiased Jacobian regulatory matrix.

[0071] The microbial regulation impact recovery module further determines the standard error of a candidate regulatory edge based on the sample variance of the orthogonal score values ​​corresponding to the same candidate regulatory edge, and generates the edge confidence of the candidate regulatory edge based on the marginal regulation strength and standard error corresponding to the candidate regulatory edge in the departed Jacobian regulation matrix. Edge confidence is used to characterize the regulatory stability of the candidate regulatory edge in multiple cascade transfer samples or multiple rhizosphere topdressing management microregions. Generally, the edge confidence of a candidate regulatory edge is high when the marginal regulation strength is large and the standard error is small; the edge confidence of a candidate regulatory edge is low when the marginal regulation strength is weak or the standard error is large.

[0072] Example 9: Trusted Control Edge Screening

[0073] The reliable regulatory edge analysis module filters reliable regulatory edges from the microbial-nutrient hierarchical association state diagram based on the de-biased Jacobian regulatory matrix and edge reliability. The module first selects candidate regulatory edges from the candidate regulatory edges in the microbial-nutrient hierarchical association state diagram. These candidate edges have source nodes belonging to nitrogen fixation, nitrification, denitrification, or organic matter mineralization functional groups, and target nodes belonging to nutrient pool nodes, root absorption state nodes, or nitrogen loss risk state nodes. These are then designated as microbial candidate edges.

[0074] Microbial candidate edges correspond to rhizosphere nitrogen transformation chains. For example, microbial candidate edges pointing to the ammonium nitrogen pool from the nitrogen fixation functional group correspond to the nitrogen fixation replenishment chain; microbial candidate edges pointing to the nitrate nitrogen pool from the nitrification functional group correspond to the nitrification transformation chain; microbial candidate edges pointing to the nitrogen loss risk state from the denitrification functional group correspond to the denitrification risk chain; and microbial candidate edges pointing to the soluble organic carbon pool or ammonium nitrogen pool from the organic matter mineralization functional group correspond to the organic matter mineralization supply chain. By first defining the source node and the target node, the system can avoid directly using candidate edges unrelated to microbial nitrogen regulation for topdressing decisions.

[0075] For each microbial candidate edge, the reliable regulatory edge parsing module extracts the marginal regulatory strength corresponding to that microbial candidate edge from the departed Jacobian regulatory matrix and extracts the reliability value corresponding to that microbial candidate edge from the edge reliability. When the marginal regulatory strength is higher than the positive regulatory threshold and the reliability value is higher than the preset reliability threshold, the system identifies that microbial candidate edge as a promoting reliable regulatory edge. A promoting reliable regulatory edge indicates that the microbial candidate edge has a relatively stable positive promoting effect, such as the nitrogen fixation functional group having a stable promoting effect on the ammonium nitrogen pool.

[0076] Specifically, the edge credibility is stored using the source node identifier and the target node identifier as indexes. The credibility control edge parsing module matches the corresponding credibility record in the edge credibility based on the source node identifier and the target node identifier of the microbial candidate edge, and uses the value in the matched credibility record as the credibility value corresponding to the microbial candidate edge. When the same microbial candidate edge corresponds to multiple credibility records, the values ​​in the multiple credibility records are averaged to obtain the credibility value corresponding to the microbial candidate edge. For example, if the source node of a candidate edge for a microorganism is identified as a "nitrogen fixation functional group node" and the target node is identified as an "ammonium nitrogen pool node", the trusted control edge parsing module searches for a trust record in the edge trust value where the source node is identified as a "nitrogen fixation functional group node" and the target node is identified as an "ammonium nitrogen pool node". If a trust record is found and the value in the trust record is 0.86, then 0.86 is determined as the trust value corresponding to the candidate edge for the microorganism. If three trust records are found and their values ​​are 0.82, 0.86 and 0.88 respectively, then the average of the three is calculated to obtain the trust value corresponding to the candidate edge for the microorganism as 0.85.

[0077] When the marginal regulation strength is below the negative regulation threshold and the confidence value is above the preset confidence threshold, the system identifies the microbial candidate edge as an inhibitory reliable regulation edge. An inhibitory reliable regulation edge indicates that the microbial candidate edge has a relatively stable negative inhibitory effect; for example, a certain microbial regulation relationship may have a stable inhibitory effect on the nitrogen loss risk state. When the confidence value is not higher than the preset confidence threshold, or the marginal regulation strength is between the negative and positive regulation thresholds, the system marks the microbial candidate edge as an uncertain regulation edge. Uncertain regulation edges are not used as the primary aggregation objects for subsequent microbial regulation coefficients and nutrient response coefficients.

[0078] In this embodiment, the positive regulation threshold, negative regulation threshold, and preset confidence threshold can be set based on historical sample statistical results, expert experience, or validation set effects. For example, the positive and negative regulation thresholds can be set based on the quantiles of the marginal regulation intensity distribution, and the preset confidence threshold can be set based on the stability of candidate regulation edges in multiple rhizosphere topdressing management micro-regions. These thresholds are system parameters and can be adjusted under different regions, different maize varieties, or different fertilization regimes.

[0079] Example 10: Microbial Regulation Coefficient and Nutrient Response Coefficient

[0080] The reliable regulatory edge parsing module eliminates uncertain regulatory edges and identifies promoting and inhibiting reliable regulatory edges as effective reliable regulatory edges. These effective reliable regulatory edges are used to generate microbial regulation coefficients and nutrient response coefficients.

[0081] In the effective and reliable regulatory edges, the reliable regulatory edge parsing module selects reliable regulatory edges whose source node is a nitrogen-fixing functional group and whose target node is an ammonium nitrogen pool or root absorption state, as nitrogen-fixing contribution edges. The system performs weighted aggregation based on the marginal regulatory strength and edge reliability of each nitrogen-fixing contribution edge to obtain the nitrogen-fixing contribution coefficient. The nitrogen-fixing contribution coefficient is used to characterize the stable promoting ability of the nitrogen-fixing functional group on the available nitrogen supply and maize absorption state within the corresponding rhizosphere topdressing management micro-region.

[0082] In the effective and reliable control edges, the reliable control edge parsing module selects promoting reliable control edges whose source nodes are nitrification functional groups and whose target nodes are nitrate nitrogen pools or nitrogen loss risk states, as nitrification conversion edges. The system performs weighted aggregation based on the marginal control strength and edge reliability of each nitrification conversion edge to obtain the nitrification conversion coefficient. The nitrification conversion coefficient is used to characterize the microbial driving strength of ammonium nitrogen to nitrate nitrogen conversion in the corresponding rhizosphere topdressing management microzone. When the nitrification conversion coefficient is high, the system can pay more attention to the risk of nitrate nitrogen accumulation and leaching in subsequent topdressing actions.

[0083] In the effective and reliable control edges, the reliable control edge parsing module selects promoting reliable control edges whose source node is a denitrification functional group and whose target node is a nitrogen loss risk state, as denitrification risk edges. The system performs weighted aggregation based on the marginal control strength and edge reliability corresponding to each denitrification risk edge to obtain the denitrification risk coefficient. The denitrification risk coefficient is used to characterize the degree to which the denitrification process in the corresponding rhizosphere topdressing management micro-zone promotes the stability of the nitrogen loss risk state.

[0084] In the effective and reliable regulatory edges, the reliable regulatory edge parsing module selects effective and reliable regulatory edges whose source nodes are nitrogen fixation functional groups, nitrification functional groups, denitrification functional groups, or organic matter mineralization functional groups, and whose target nodes are in a nitrogen loss risk state and belong to the inhibitory reliable regulatory edge category. These edges are designated as risk inhibition edges. The system performs negative weighted aggregation based on the marginal regulatory strength and edge reliability corresponding to each risk inhibition edge to obtain the risk inhibition correction coefficient. The risk inhibition correction coefficient is used to characterize the microbial regulatory relationship that has a stable inhibitory effect on the nitrogen loss risk state.

[0085] The reliable regulatory edge analysis module uses the nitrogen fixation contribution coefficient, nitrification conversion coefficient, denitrification risk coefficient, and risk inhibition correction coefficient as microbial regulation coefficients. These microbial regulation coefficients are then combined with the current nutrient pool state to form the nutrient response coefficient. The current nutrient pool state includes the node states of the ammonium nitrogen pool, nitrate nitrogen pool, and soluble organic carbon pool within the rhizosphere topdressing management micro-zone corresponding to the current observation time. The nutrient response coefficient characterizes the expected response capability to topdressing input under the combined effects of the current nutrient base and microbial regulation relationship.

[0086] For example, when the nitrogen fixation contribution coefficient of the first rhizosphere topdressing management microzone is high, the current ammonium nitrogen pool is at a moderate level, and the nitrogen loss risk is low, the system can determine that the demand for nitrogen fertilizer topdressing in this microzone is relatively mild. When the nitrification conversion coefficient and denitrification risk coefficient of the second rhizosphere topdressing management microzone are both high, and the nitrate nitrogen leaching risk is high, the system can reduce the amount of nitrogen fertilizer applied to this microzone or increase the control weight of nitrification inhibitors. When the nitrogen fixation contribution coefficient of the third rhizosphere topdressing management microzone is low and the root absorption is weak, the system can appropriately increase the amount of nitrogen fertilizer applied or combine it with the application of microbial agents. The above examples are used to illustrate the usage of nutrient response coefficients and do not limit specific values.

[0087] Example explanation:

[0088] For example, in the first micro-region of topdressing management, the system screened out the following edges, as shown in Table 2:

[0089] Table 2

[0090] ;

[0091] Edge 7 is an uncertain control edge, so it is removed first and will not participate in subsequent calculations.

[0092] 1. Confirmation of nitrogen fixation contribution coefficient: Source node = nitrogen fixation functional group; Target node = ammonium nitrogen pool or root absorption state; Type = promoting reliable regulatory edge; Edge 1 and edge 2 are selected as nitrogen fixation contribution edges;

[0093] The nitrogen fixation contribution coefficient is 0.60 × 0.90 + 0.30 × 0.80 = 0.78, indicating that the nitrogen fixation contribution of this micro-region is relatively significant.

[0094] 2. Confirmation of nitration conversion coefficient: Source node = nitration functional group; Target node = nitrate nitrogen pool or nitrogen loss risk state; Type = promoting reliable control edge; Edges 3 and 4 are selected as nitration conversion edges;

[0095] The nitrification conversion coefficient = 0.50 × 0.85 + 0.40 × 0.75 = 0.725; this indicates that the nitrification conversion in this micro-area is relatively strong, and attention should be paid to the risk of nitrate nitrogen accumulation and leaching when applying topdressing later.

[0096] 3. Confirmation of denitrification risk coefficient: Source node = denitrification functional group; Target node = nitrogen loss risk state; Type = promoting reliable control edge; Edge 5 is selected as the denitrification risk edge;

[0097] The denitrification risk coefficient is 0.70 × 0.90 = 0.63, indicating that the denitrification risk in this micro-region is relatively significant.

[0098] 4. Confirmation of risk inhibition correction coefficient: Target node = nitrogen loss risk state; Type = inhibitory credible regulatory edge; Source node = microbial functional group; Edge 6 is selected as the risk inhibition edge; The marginal regulatory strength of edge 6 is negative, indicating that it has an inhibitory effect.

[0099] The risk mitigation correction coefficient is -0.25 × 0.80 = -0.20, indicating that it has a certain mitigating effect on the risk of nitrogen loss.

[0100] The system combines these coefficients:

[0101] Microbial regulation coefficient = {Nitrogen fixation contribution coefficient: 0.78, Nitrification conversion coefficient: 0.725, Denitrification risk coefficient: 0.63, Risk inhibition correction coefficient: -0.20};

[0102] The system then checks the current nutrient pool status. The first micro-region currently shows: Ammonium nitrogen pool: Medium; Nitrate nitrogen pool: High; Soluble organic carbon pool: Medium.

[0103] The system can determine that: nitrogen fixation contributes significantly; nitrification conversion is strong; denitrification risk is high; nitrate nitrogen levels are already high; although there is some risk suppression effect, it is insufficient to completely offset the risk.

[0104] Therefore, the nutrient response in the first microzone may be as follows: it is not suitable to continue to increase nitrogen fertilizer significantly; the amount of nitrogen fertilizer applied should be controlled; it is advisable to use nitrification inhibitors; the subsequent topdressing action generation module should avoid selecting candidate topdressing actions that would lead to a further increase in nitrate nitrogen and a further increase in the risk of leaching.

[0105] Example 11: Candidate Topdressing Action Set and Micro-region Variable Topdressing Prescription:

[0106] The topdressing action generation module constructs a set of candidate topdressing actions based on microbial regulation coefficients, nutrient response coefficients, target topdressing requirements for maize, and topdressing safety constraints. Target topdressing requirements for maize can be determined based on maize variety, target yield, current growth stage, historical fertilizer application rates, chlorophyll index, normalized canopy vegetation index, and local agronomically recommended fertilization standards. Topdressing safety constraints may include maximum nitrogen fertilizer application thresholds, minimum nitrogen fertilizer application thresholds, nitrate nitrogen leaching risk thresholds, nitrous oxide emission risk thresholds, soil moisture constraints, variable fertilizer application equipment discharge capacity constraints, and operation speed constraints.

[0107] The candidate topdressing action set includes multiple candidate topdressing actions. Each candidate topdressing action may include nitrogen fertilizer application rate, microbial inoculant application rate, nitrification inhibitor application rate, and organic carbon source supplementation rate. The topdressing action generation module evaluates the impact of candidate topdressing actions on the nutrient pool status, root absorption status, and nitrogen loss risk status at the next observation time for each rhizosphere topdressing management microzone using the nutrient response coefficient, and eliminates candidate topdressing actions that do not meet the topdressing safety constraints.

[0108] In this embodiment, if a candidate topdressing action causes the predicted nitrate nitrogen leaching risk state to exceed the nitrate nitrogen leaching risk threshold, or causes the predicted nitrous oxide emission risk state to exceed the nitrous oxide emission risk threshold, the topdressing action generation module removes the candidate topdressing action from the candidate topdressing action set. If the nitrogen fertilizer application amount corresponding to the candidate topdressing action exceeds the maximum fertilizer discharge capacity of the variable fertilizer application equipment, or is less than the minimum fertilizer discharge amount required to maintain stable fertilizer discharge by the equipment, it can also be removed or corrected.

[0109] The topdressing action generation module determines the target topdressing action from the remaining candidate topdressing actions. The system can use rule scoring, weighted scoring, constraint optimization, or other known decision-making methods to determine the target topdressing action.

[0110] The system generates variable topdressing prescriptions for maize micro-regions during the jointing stage based on the target topdressing actions corresponding to the first to fourth rhizosphere topdressing management micro-regions. These prescriptions may include the micro-region number, operation boundary, target nitrogen fertilizer application rate, target microbial inoculant application rate, target nitrification inhibitor application rate, target organic carbon source supplementation rate, fertilization path, recommended operation speed, and execution time. These prescriptions can be output as prescription tables, prescription layers, operation files readable by variable fertilization equipment, or instructions from the agricultural machinery control terminal.

[0111] For example, a micro-zone topdressing prescription for maize at the jointing stage could indicate the following: the first rhizosphere topdressing management micro-zone uses the first nitrogen fertilizer application rate and the first inoculant pumping rate; the second rhizosphere topdressing management micro-zone uses the second nitrogen fertilizer application rate and the second nitrification inhibitor pumping rate; the third rhizosphere topdressing management micro-zone uses the third nitrogen fertilizer application rate and the third organic carbon source supplementation rate; and the fourth rhizosphere topdressing management micro-zone uses the fourth nitrogen fertilizer application rate. The specific values ​​for different micro-zones are determined based on the nutrient response coefficient and topdressing safety constraints of the corresponding micro-zone.

[0112] Example 12: Variable Fertilization Implementation and Feedback Updates:

[0113] The variable-rate fertilization execution module includes an onboard positioning unit, a prescription analysis unit, a variable-rate fertilizer dispensing control unit, and an operation recording unit. The onboard positioning unit is used to determine the current rhizosphere topdressing management micro-zone where the variable-rate fertilization equipment is located. The onboard positioning unit may include satellite positioning equipment, field base station positioning equipment, inertial navigation equipment, or an operation path identification module.

[0114] The prescription parsing unit is used to read the variable topdressing prescription for the maize jointing stage micro-zone corresponding to the current rhizosphere topdressing management micro-zone. The prescription parsing unit can parse information such as micro-zone number, operation boundary, fertilizer discharge rate, microbial agent pumping rate, inhibitor pumping rate, and operation speed in the prescription file, and transmit the parsing results to the variable fertilizer discharge control unit.

[0115] The variable-rate fertilizer dispensing control unit is used to control the fertilizer dispensing rate, microbial agent pumping rate, and inhibitor pumping rate based on the variable-rate fertilizer prescription for micro-zones during the corn jointing stage. The variable-rate fertilizer dispensing control unit can control electric fertilizer applicators, liquid pumps, spray valves, solenoid valves, or fertilizer controllers, enabling the variable-rate fertilizer application equipment to perform different fertilizer dispensing actions in different rhizosphere fertilizer management micro-zones. When the onboard positioning unit determines that the variable-rate fertilizer application equipment has moved from the first rhizosphere fertilizer management micro-zone to the second rhizosphere fertilizer management micro-zone, the variable-rate fertilizer dispensing control unit adjusts the fertilizer dispensing and pumping rates according to the prescription parameters corresponding to the second rhizosphere fertilizer management micro-zone.

[0116] The operation recording unit is used to record the actual amount of fertilizer applied, operation speed, spraying pressure, and operation time, and the recorded results are used as part of the feedback observation data.

[0117] The feedback update module acquires feedback observation data after topdressing, adds this data as a new cascade transfer sample, and uses it to update the micro-variable topdressing prescription for the maize jointing stage in the next topdressing window. By continuously adding feedback observation data, the system can update the cascade transfer function estimator, plug-in Jacobian influence matrix, departitioned Jacobian regulation matrix, edge confidence, microbial regulation coefficient, and nutrient response coefficient in subsequent topdressing windows, thereby gradually adapting the variable topdressing prescription to the actual microbial nitrogen regulation state of the target maize field.

[0118] The present invention and its embodiments have been described above. This description is not restrictive. The accompanying drawings are only one embodiment of the present invention, and the actual structure is not limited thereto. In short, if a person skilled in the art is inspired by this description and designs a similar structure and embodiment without departing from the spirit of the present invention, such design should fall within the protection scope of the present invention.

Claims

1. A smart fertilization system for maize based on soil microbial regulation, characterized in that, The system includes: The rhizosphere state construction module divides the target maize field into rhizosphere topdressing management micro-regions, obtains microbial nitrogen regulation observation data and field covariates corresponding to each rhizosphere topdressing management micro-region, and constructs microbial-nutrient cascade state map and cascade transfer samples. The microbial regulation impact recovery module trains a first-level cascade transfer function estimator based on cascade transfer samples. Then, it performs Jacobian calculation using the trained first-level cascade transfer function estimator to obtain a plug-in Jacobian impact matrix. Based on the prediction residuals output by the first-level cascade transfer function estimator and the Riesz characterization terms corresponding to each candidate regulatory edge in the microbial-nutrient cascade state graph, it performs Neyman orthogonal debiasing correction on the plug-in Jacobian impact matrix to obtain the debiased Jacobian regulation matrix and edge confidence. Specifically, the module includes: for each candidate regulatory edge in the plug-in Jacobian influence matrix, the microbial regulation impact recovery module extracts the plug-in partial derivative corresponding to the candidate regulatory edge; based on the Riesz characterization term corresponding to the candidate regulatory edge, the prediction residual is weighted and corrected to obtain the residual correction amount corresponding to the candidate regulatory edge; the residual correction amount is used to compensate for the bias introduced by the prediction error in the plug-in partial derivative; the system combines the plug-in partial derivative corresponding to the candidate regulatory edge and the residual correction amount to form an orthogonal score value; the orthogonal score values ​​corresponding to the same candidate regulatory edge in multiple cascaded transfer samples are aggregated to obtain the debiased regulatory strength of the candidate regulatory edge; subsequently, the system reorganizes the debiased regulatory strength of each candidate regulatory edge according to the arrangement of the source node and target node in the plug-in Jacobian influence matrix to obtain the debiased Jacobian regulatory matrix; the microbial regulation impact recovery module also determines the standard error of the candidate regulatory edge based on the sample variance of the orthogonal score value corresponding to the same candidate regulatory edge, and generates the edge confidence of the candidate regulatory edge based on the marginal regulatory strength and standard error corresponding to the candidate regulatory edge in the debiased Jacobian regulatory matrix; The reliable regulatory edge parsing module filters reliable regulatory edges from the microbial-nutrient hierarchical association state diagram based on the de-biased Jacobi regulatory matrix and edge reliability, and generates the microbial regulation coefficient and nutrient response coefficient corresponding to each rhizosphere topdressing management micro-zone based on the reliable regulatory edges. The topdressing action generation module determines the target topdressing action based on the microbial regulation coefficient and nutrient response coefficient, and generates a micro-region variable topdressing prescription for maize during the jointing stage. The variable fertilization execution module executes topdressing operations based on the micro-regional variable topdressing prescription during the corn jointing stage. The feedback update module acquires the feedback observation data after the topdressing operation and adds the feedback observation data as a new cascade transfer sample. The microbial nitrogen regulation observation data includes the abundance of microbial functional groups, microbial enzyme activity, available soil nutrients, maize rhizosphere absorption status, and nitrogen loss risk status. Based on the nitrogen transformation direction between the microbial functional groups corresponding to the abundance of microbial functional groups, the nutrient pool corresponding to the available nutrients in the soil, the root absorption state corresponding to the rhizosphere absorption state of maize, and the nitrogen loss risk state, a rhizosphere nitrogen transformation chain is constructed. The process of constructing a microbial-nutrient cascade state diagram and cascade transfer samples includes the following steps: Step R1: Organize the microbial nitrogen regulation observation data corresponding to the same observation time of each micro-region under the topdressing management into a microbial cascade state vector; Step R2: Using microbial functional groups, nutrient pools, root absorption status, and nitrogen loss risk status as graph nodes, and rhizosphere nitrogen transformation chains and the spatial adjacency relationships of rhizosphere topdressing management micro-regions as candidate edges, construct a microbial-nutrient hierarchical association state graph. Step R3: Based on the amount of nitrogen fertilizer applied, microbial agent applied, nitrification inhibitor applied, and organic carbon source supplemented in the corresponding rhizosphere topdressing management micro-zone before the current observation time, construct the topdressing regulation input; use the microbial cascade state vector, field covariates, and topdressing regulation input at the current observation time as the transfer input, and use the microbial cascade state vector at the next observation time as the transfer output to form a cascade transfer sample.

2. The intelligent maize fertilization system based on soil microbial regulation according to claim 1, characterized in that: The abundance of microbial functional groups corresponds to the following microbial functional groups: nitrogen fixation, nitrification, denitrification, and organic matter mineralization. The nutrient pools corresponding to readily available nutrients in soil include ammonium nitrogen, nitrate nitrogen, and soluble organic carbon. The root absorption status corresponding to the rhizosphere absorption status of maize includes chlorophyll index, canopy normalized vegetation index, and root zone water content.

3. The intelligent maize fertilization system based on soil microbial regulation according to claim 2, characterized in that: The process of obtaining the plug-in Jacobian influence matrix includes the following steps: Step S41: Obtain the current transfer input; Step S42: Load the current transition input into the trained one-level enjoint transition function estimator, which includes a state index generation unit and a node type mapping unit; Step S43: Through the state index generation unit, the microbial cascade state vector, field covariates and topdressing regulation input at the current observation time are weighted and combined to generate the state index value corresponding to each node in the microbial-nutrient cascade state diagram. Step S44: Through the node type mapping unit, based on the data type and value range corresponding to the graph node, different link functions are used to map the state index value to generate the predicted concatenated state vector for the next observation time. Step S45: Using the state of each graph node in the microbial cascade state vector at the current observation time as the derivative variable, perform automatic differentiation on the partial derivative relationship of each graph node state in the predicted cascade state vector at the next observation time with respect to the derivative variable to obtain the plug-in partial derivatives between nodes. Step S46: Arrange the plug-in partial derivatives between nodes according to the source and target nodes in the microbial-nutrient hierarchical association state diagram to generate a plug-in Jacobian influence matrix.

4. The intelligent maize fertilization system based on soil microbial regulation according to claim 3, characterized in that: The different link functions are as follows: the graph nodes corresponding to microbial functional groups and nutrient pools use non-negative link functions, the graph nodes corresponding to nitrogen loss risk states use sigmoid link functions, and the graph nodes corresponding to root absorption states use identity link functions or multilayer perceptron link functions.

5. The intelligent maize fertilization system based on soil microbial regulation according to claim 2, characterized in that: The process of selecting reliable control edges includes: From the candidate regulatory edges in the microbial-nutrient hierarchical association state diagram, source nodes that are nitrogen fixation functional groups, nitrification functional groups, denitrification functional groups, or organic matter mineralization functional groups are screened as microbial candidate edges; positive and negative regulatory thresholds are set. For each microbial candidate edge, the marginal regulation intensity corresponding to the microbial candidate edge is extracted from the debiased Jacobi regulation matrix, and the confidence value corresponding to the microbial candidate edge is extracted from the edge confidence. When the marginal regulation intensity is higher than the positive regulation threshold and the confidence value is higher than the preset confidence threshold, the microbial candidate edge is identified as a promoting confidence regulation edge. When the marginal regulation intensity is lower than the negative regulation threshold and the confidence value is higher than the preset confidence threshold, the microbial candidate edge is identified as an inhibitory confidence regulation edge.

6. The intelligent maize fertilization system based on soil microbial regulation according to claim 5, characterized in that: The generation process of microbial regulation coefficient and nutrient response coefficient includes: The promoting and inhibiting credible control edges are defined as effective credible control edges. In the effective and reliable regulatory edges, promoting reliable regulatory edges with source nodes of nitrogen fixation functional groups and target nodes of ammonium nitrogen pool or root absorption state are selected as nitrogen fixation contribution edges, and weighted aggregation is performed to obtain nitrogen fixation contribution coefficients. In the effective and reliable control edges, promoteable reliable control edges with source nodes of nitration functional groups and target nodes of nitrate nitrogen pools or nitrogen loss risk states are selected as nitration conversion edges, and weighted aggregation is performed to obtain the nitration conversion coefficient. In the effective and reliable control edges, promoteable reliable control edges with source nodes in the denitrification functional group and target nodes in the nitrogen loss risk state are selected as denitrification risk edges, and weighted aggregation is performed to obtain the denitrification risk coefficient. Among the effective and reliable control edges, candidate control edges that are in the nitrogen loss risk state and belong to the suppression type of reliable control edges are selected as risk suppression edges, and negative weighted aggregation is performed to obtain the risk suppression correction coefficient. The nitrogen fixation contribution coefficient, nitrification conversion coefficient, denitrification risk coefficient, and risk inhibition correction coefficient are used as microbial regulation coefficients. These microbial regulation coefficients are then combined with the current nutrient pool to form nutrient response coefficients.

Images