A hierarchical bias decoupling and gated fusion prediction method for carbon flux of mountain tea garden

Abstract

The present application relates to a kind of mountain tea garden carbon fluxes of layering deviation decoupling and gate fusion prediction method, belong to data processing field, method includes: mechanism priori branch, multiscale dynamic residual branch and structured deviation calibration branch are established, so that three branches are responsible for the processing task of different error sources;Then through state perception gate fusion layer, according to current environmental state, ecological process state, time sequence dynamic state and structured background state, dynamically allocate the contribution of each branch, and limit the final prediction result through physical consistency constraint and smoothing constraint, to form the error source layering, branch task clear, fusion weight variable and the output process consistent mountain tea garden carbon flux prediction mechanism.The present application reduces the coupling of different bias components in the same parameter space, improves the processing capacity of the prediction architecture for different error sources, and enables the formation of the final prediction result to be traced back through the output of each branch and its contribution value.

Description

Technical Field

[0001] This invention relates to the field of data processing, and in particular to a hierarchical bias decoupling and gating fusion prediction method for carbon flux in mountain tea gardens. Background Technology

[0002] With the development of agricultural carbon sink monitoring, smart agriculture, and digital management of the ecological environment, carbon flux prediction for economic crop ecosystems such as tea gardens is gradually shifting from single manual surveys or periodic statistics to continuous monitoring and model-based analysis driven by multi-source data. Existing industry technologies typically include the following implementation paths. The first category is carbon cycle modeling methods based on ecological process mechanisms. These methods usually establish models based on ecological processes such as vegetation photosynthesis, vegetation respiration, soil respiration, and carbon pool transfer, using variables such as meteorology, soil, and vegetation growth to estimate carbon flux changes. Their advantage lies in their good ecological interpretability, capable of describing carbon flux change patterns from the perspective of vegetation-soil-atmosphere carbon exchange processes. However, in mountainous tea gardens, topographic factors such as slope, aspect, altitude, and slope position can cause significant small-scale spatial differences. Agricultural activities such as harvesting, fertilization, pruning, and irrigation can also alter tea garden biomass, soil nutrients, and respiration intensity in a short period, making it difficult for general mechanism models to fully express complex local disturbances and easily leading to systematic prediction biases. The second category is data-driven carbon flux prediction methods. This type of method typically uses historical carbon flux observation data, meteorological data, soil data, remote sensing data, etc., to train machine learning models or deep learning models to learn the nonlinear relationship between multiple variables and carbon flux. Its advantage lies in its ability to capture complex temporal changes using large amounts of observational data, and it has a strong fitting ability for short-term fluctuations and nonlinear responses. However, this type of method usually relies on historical sample distributions and lacks ecological process boundary constraints, which may result in output results with good numerical prediction effects but unreasonable ecological significance; it is also prone to insufficient generalization ability when the monitored plots, slope conditions, or management intensity change. The third category is multi-source data fusion or multi-model fusion methods. This type of method typically performs time alignment, spatial matching, and feature fusion on data from different sources, or performs weighted averaging, result splicing, or stacking of prediction results from mechanistic models, statistical models, and machine learning models to improve overall prediction accuracy. These methods improve the information deficiency of single models to some extent, but common fusion methods tend to focus on fixed-weight combinations or empirical integration at the result level, lacking structured decomposition of different error sources. They struggle to clearly answer questions such as: which module handles biases generated by mechanistic models, short-term dynamic disturbances, and differences in topography and management? Under different conditions such as continuous rainfall, high humidity, agricultural events, or predictions of plots at different slopes, which type of model output should be trusted more? And how can we avoid the fusion results deviating from the carbon cycle process? The fourth category is a systematic solution for carbon sink accounting or digital twin management. This type of solution typically combines flux observation, plot surveys, remote sensing imagery, model calculations, and visualization platforms to achieve carbon source and sink analysis, dynamic accounting, or status display of tea gardens or regional ecosystems.Such solutions can improve the efficiency of data organization and management, but their technical focus is often on monitoring data access, accounting process management, remote sensing-assisted estimation or system display. They usually lack finer-grained technical processing for issues such as how to perform hierarchical modeling of carbon flux error sources in mountain tea gardens, how different branches divide the work to handle different types of deviations, and how to dynamically adjust the fusion weights according to the tea garden status.

[0003] Therefore, in the field of carbon flux prediction in mountain tea plantations, existing technologies share common problems: single-mechanism models offer interpretability but lack local adaptability; single-data-driven models have strong fitting capabilities but lack ecological constraints; and simple fusion models can integrate multiple types of information but struggle to distinguish error sources and dynamically adjust branch reliability. For scenarios like mountain tea plantations, which simultaneously experience topographical heterogeneity, meteorological disturbances, and high-frequency agricultural management impacts, how to construct a carbon flux prediction method that can take into account ecological process constraints, short-term dynamic responses, and structured scenario differences remains a technical problem that needs further resolution in this field.

[0004] While existing technologies can collect, process, and predict meteorological, soil, remote sensing, topographic, agricultural management, and carbon flux observation data, their technical focus is mainly on multi-source data aggregation, single-model prediction, or simple fusion of multiple prediction results. They do not classify carbon flux prediction errors in mountain tea gardens into mechanistic prior biases, dynamic residuals, and structured biases and process them separately. They have not established specialized processing mechanisms for different sources of carbon flux prediction errors in mountain tea gardens, nor have they dynamically determined the reliability of various bias correction results based on the current state of mountain tea gardens. This results in different error sources being coupled together in the same model parameters or fixed fusion weights. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of the prior art and provide a stratified bias decoupling and gating fusion prediction method for carbon flux in mountain tea gardens, thus solving the deficiencies of the prior art.

[0006] The objective of this invention is achieved through the following technical solution: a stratified bias decoupling and gating fusion prediction method for carbon flux in mountain tea gardens, the prediction method comprising:

[0007] S1. Collect on-site data of mountain tea gardens through data acquisition terminals according to preset time steps, write collection timestamp, plot identifier and data source identifier for each data, generate historical carbon flux observation data within the same time range by carbon flux observation equipment, and send the original monitoring record of the mountain tea garden on-site data and historical carbon flux observation data to the prediction server.

[0008] S2. Record agricultural management event information such as harvesting, fertilizing, pruning and irrigation through the management input terminal to form agricultural management records including plot identification, event type, event occurrence time, event duration and event intensity;

[0009] S3. The prediction server receives terrain data and remote sensing data and rasterized terrain data of corresponding plots, as well as data sent by data acquisition terminals and management input terminals. It simultaneously undertakes the processing tasks of different error sources through mechanism prior branches, multi-scale dynamic residual branches and structured bias calibration branches. Then, through the state-aware gating fusion layer, the contribution of each branch is dynamically allocated according to the current environmental state, ecological process state, temporal dynamic state and structured background state. The final prediction result is limited by physical consistency constraints and smoothing constraints. Finally, the final prediction result is displayed through the result output module.

[0010] Before executing the prediction method, the time synchronization, spatial matching and state vector construction of data from different sources for the same plot need to be completed through the data access and state construction module before the mechanism prior branch, multi-scale dynamic residual branch and structured bias calibration branch can enter the calculation stage.

[0011] The data access and status construction module is comprised of a data acquisition terminal that generates or accesses on-site monitoring data, a management input terminal that generates agricultural management records, and a prediction server that performs time synchronization, spatial matching, unit unification, anomaly verification, and status construction.

[0012] The data acquisition terminal periodically generates raw monitoring records. The raw monitoring records include environmental monitoring records, soil monitoring records and carbon flux observation records according to variable type. Each raw monitoring record includes at least one of the following: collection timestamp, data source identifier, variable name, variable value, plot identifier and spatial coordinates.

[0013] The management input terminal generates separate agricultural management records. These records are not part of the original monitoring records. Each agricultural management record must include at least the plot identifier, event type, event occurrence time, event duration, and event intensity.

[0014] The prediction server associates data from different sources with the corresponding tea garden plots based on the plot identifiers, spatial coordinates, and the correspondence between plot boundaries.

[0015] The prediction server pre-establishes a land parcel spatial index table to complete land parcel spatial matching. The land parcel spatial index table includes the land parcel identifier ID of the i-th tea garden and the land parcel boundary polygon Ω corresponding to the land parcel identifier.

[0016] For environmental monitoring records, soil monitoring records, carbon flux observation records, and agricultural management records that already carry plot identifiers, the prediction server directly completes the association according to the plot identifier. When one of these four records also carries spatial coordinates, it determines whether the spatial coordinates are located within the plot boundary polygon corresponding to the plot identifier. If the plot identifier carried is inconsistent with the plot identifier determined according to the spatial coordinates, the record is marked as a spatial matching anomaly record.

[0017] For point monitoring records that do not carry plot identifiers but carry spatial coordinates, the prediction server performs a spatial inclusion relationship determination between the spatial coordinates carried by the point monitoring record and the plot boundary polygon Ω. When the spatial coordinates uniquely fall into the plot boundary polygon Ω, the corresponding plot identifier ID is written into the point monitoring record. The point monitoring records include environmental monitoring records, soil monitoring records, and carbon flux observation records that do not carry plot identifiers but carry spatial coordinates.

[0018] For remote sensing data and rasterized terrain data, the prediction server calculates the spatial intersection relationship between each raster cell and the plot boundary polygon Ω, and then weights and summarizes the raster cells that intersect with the plot boundary polygon Ω according to the intersection area, and then associates the summary result with the corresponding plot identifier ID.

[0019] For point monitoring records that cannot uniquely identify the corresponding plot, the prediction server marks them as spatially mismatched records; for environmental monitoring records, soil monitoring records, carbon flux observation records, or agricultural management records whose plot identifiers are inconsistent with the plot identifiers determined based on their spatial coordinates, the prediction server marks them as spatially mismatched records.

[0020] After completing spatial matching, the prediction server establishes a device parameter table, which includes at least the record data source identifier, variable name, and lower limit of the measurement range. and upper limit of range When the variables collected by data source s Variable values ​​at target time step t satisfy or When the variable value is found to be abnormal, the environmental monitoring record, soil monitoring record, or carbon flux observation record to which it belongs shall be marked as an abnormal equipment range record, and the writing of the environmental monitoring record, soil monitoring record, or carbon flux observation record to which the variable value belongs into the standardized status record table shall be stopped as valid data.

[0021] For environmental monitoring records, soil monitoring records, or carbon flux observation records within the equipment's measurement range, the prediction server uses a set of historical valid records from the same site and the same variable, as well as the same seasonal or phenological stage. Calculate the median and median absolute deviation And calculate the robustness deviation of the target record. When the robustness deviation is greater than the set deviation threshold, the corresponding record is marked as a candidate abnormal state record.

[0022] For candidate state anomaly records, the prediction server verifies their adjacent time changes, concurrent meteorological events, and agricultural management events within a preset verification time window. When a candidate state anomaly record exhibits continuous changes in the same direction across multiple adjacent time steps, or is accompanied by any of the following concurrent events: rainfall, persistent high humidity, irrigation, fertilization, harvesting, or pruning, and the corresponding environmental variables or management status provide consistent support, the candidate state anomaly record is determined to correspond to a real meteorological disturbance or a real agricultural management response. The candidate state anomaly record is then written into the standardized state record table as valid data and an actual disturbance identifier is attached. When a candidate state anomaly record exhibits an isolated jump, and there is no valid monitoring record or related environmental variable or agricultural management record of the same variable type in adjacent time steps to support it, it is marked as an invalid record and not used as valid input. Valid monitoring records of the same variable type as the candidate state anomaly record include, for example, if the current anomaly record is a soil monitoring record, the corresponding monitoring record is the soil monitoring record in the adjacent time step; if the current anomaly record is a carbon flux observation record, the corresponding monitoring record is the carbon flux observation record in the adjacent time step.

[0023] For consecutive missing lengths not exceeding the preset number of missing steps Furthermore, valid records exist before and after the missing interval, and the prediction server uses the valid time step data before and after the missing interval to interpolate and complete the data; for consecutive missing intervals exceeding a certain length... If there are no valid records before or after the missing interval, automatic completion will not be performed, and a missing test mark will be added;

[0024] A standardized state record table with plot identifier-time step as the primary key is formed, which is called by the mechanism prior branch, multi-scale dynamic residual branch, structured bias calibration branch, and state-aware gating fusion layer. For the target time step t, the prediction server constructs the state vector of the mountain tea garden. ,in, This represents the state vector of the mountain tea garden at the target time step t. Indicates the environmental state. Indicates soil condition. Indicates terrain condition, Indicates the management status. This indicates the historical carbon flux status.

[0025] The mechanism prior branch is executed by the mechanism calculation submodule in the prediction server, which receives the environmental state corresponding to the target time step. Soil condition Management status Based on the basic process parameters related to the target plot, and according to the process relationships between vegetation photosynthesis, vegetation respiration, soil respiration, and carbon pool transfer in the mountain tea garden, the environmental state at the target time step is calculated. Soil condition Management status The mapping relationship between basic process parameters related to the target site and prior values ​​of mechanistic carbon flux and ecological process states is expressed as follows: ,in, This represents the prior value of the mechanistic carbon flux. Indicates the state of ecological processes. This represents the a priori calculation process of the mechanism. The set of process parameters used to represent the prior branches of the mechanism;

[0026] Ecological process state Used to characterize the photosynthetic-related state, respiration-related state, carbon pool transfer state, carbon exchange direction, permissible change boundary, or a combination of these states of the current tea garden system;

[0027] Mechanistic carbon flux prior value It is not used directly as the sole final carbon flux prediction, but rather as a unified process benchmark for subsequent dynamic residual correction and structured bias calibration.

[0028] Ecological process state The results are then sent to the state-aware gating fusion layer and the physical consistency constraint submodule, respectively, to participate in gating weight generation and fusion result constraint.

[0029] The multi-scale dynamic residual branch is executed by the temporal residual learning submodule in the prediction server. The prediction server extracts a historical time window of length L from the standardized state record table with the target time step t as the endpoint, and arranges the historical carbon flux state, environmental state sequence, soil state sequence and management state sequence within the historical time window in chronological order to form a temporal input tensor. ,in, L represents the temporal input tensor corresponding to the target time step t, and L represents the length of the history window. Indicates time step The corresponding combined record of historical carbon flux, environmental conditions, soil conditions, and management conditions;

[0030] The temporal residual learning submodule, for multiple preset time scales, uses resampling or decomposition operations corresponding to each time scale to obtain sequences at each scale, and then performs feature extraction on the sequences at each scale to obtain temporal representations at different time scales. ,in, Indicates the first Each time scale corresponds to a resampling operation or decomposition operation. Indicates the number of time scales. Time series representations at different time scales;

[0031] The temporal dynamics representation obtained after multi-scale mixing of the temporal representations at each scale is as follows: The dynamic residual correction amount is obtained by mapping the residual output. ,in, Represents the dynamic representation of time series. This represents the dynamic residual correction amount. This represents multi-scale temporal feature extraction and hybrid operations. This represents the output operation that maps the time-series dynamic representation to residual correction values. The parameters represent the multi-scale dynamic residual branch.

[0032] The structured bias calibration branch is executed by the structured bias calculation submodule in the prediction server. The prediction server extracts topographic state, soil background state, management event intensity, and derived statistical features from the mountain tea garden state vector and standardized state record table to form a structured feature vector. ,in, This represents the structured feature vector corresponding to the target time step. Indicates the background state of the soil. Indicates the intensity of current or historical agricultural management events. Indicates the duration of continuous high humidity. This indicates the cumulative rainfall within the preset statistical window. This indicates the historical fertilization intensity within the preset statistical window. This indicates the harvesting frequency within the preset statistics window;

[0033] The structured deviation calculation submodule generates structured deviation calibration values ​​and structured scene representations based on structured feature vectors. ,in, This indicates the amount of structured deviation calibration. Represents a structured scene representation. ( ) represents the structured deviation mapping process. The parameters represent the structured deviation calibration branch.

[0034] The state-aware gating fusion layer is executed by the gating computation submodule in the prediction server. The gating computation submodule receives the ecological process state output by the mechanism prior branch. Temporal dynamic representation of multi-scale dynamic residual branch output Structured scene representation output from the structured deviation calibration branch It also receives a current state summary formed from the current environmental state, soil state, and management state. The gated input vector is obtained as In this context, ";" indicates vector concatenation.

[0035] Generate unnormalized gating score The fusion weights were then obtained through normalization. Thus, the gating weight vector is obtained as follows: And satisfy and ,in, Represents the gated mapping weight matrix. This represents the gating bias vector. This indicates the unnormalized gating score. This represents the confidence weight of the prior branch of the mechanism. The confidence weights represent the multi-scale dynamic residual branches. Represents the confidence weight of the structured bias calibration branch;

[0036] Construct prior candidate values ​​for the respective mechanisms Candidate values ​​for dynamic residual correction And the structured bias correction candidate value ,get ,in, This represents the final carbon flux prediction at the target time step t. This represents the dynamic residual correction amount. This represents the prior value of the mechanistic carbon flux. Indicates the amount of structured deviation calibration;

[0037] To determine the relative dominance of the three candidate values ​​in the current target time step, the dominant candidate branch index is defined as follows: , Indicates the dominant candidate branch index, when ,satisfy and This indicates the prior candidate values ​​of the mechanism. It has the largest fusion coefficient among the weighted combinations of the three candidate values, when At 2, it indicates a candidate value for dynamic residual correction. It has the largest fusion coefficient when At 3, it indicates a candidate value for structured bias correction. The branch with the largest fusion coefficient is the candidate branch that is determined as the common dominant branch when two or more gating weights are equal and reach the maximum value.

[0038] Define the actual correction contribution of dynamic residuals as The actual correction contribution of structural bias is defined as The final carbon flux prediction is expressed as This allows us to determine the actual compositional relationship between the mechanistic process baseline, dynamic residual correction, and structured bias correction on the final predicted value.

[0039] The physical consistency constraint and smoothing constraint are executed by the loss calculation submodule in the prediction server. The loss calculation submodule calculates the prediction error loss, the physical consistency constraint loss and the smoothing constraint loss.

[0040] Prediction error loss This is used to measure the difference between the final carbon flux prediction and the observed carbon flux, where... Indicates the number of training samples. This represents the carbon flux observation at the target time step. This represents the loss due to prediction error;

[0041] The physical consistency constraint loss is expressed as ,in, Indicates the directional constraint coefficient. Indicates a stable state activation flag. Indicates the upper boundary. Indicates the lower boundary. This represents the loss due to physical consistency constraints;

[0042] Smoothing constraint loss is expressed as ,in, This represents the loss due to smoothing constraints.

[0043] The final total loss is expressed as ,in, , and These are the non-negative weighting coefficients for controlling prediction error loss, physical consistency constraint loss, and smoothness constraint loss, respectively. This represents the total loss.

[0044] The result output module is jointly executed by the result encapsulation submodule in the prediction server and the result display terminal. The result encapsulation submodule encapsulates the data corresponding to the same plot and the same target time step into a prediction result data object. ,in, This represents the data object representing the prediction result. Indicates the land parcel identification. Indicates the target time step. This represents the prior value of the mechanistic carbon flux. This represents the dynamic residual correction amount. This indicates the amount of structured deviation calibration. Represents the gating weight vector. This represents the final carbon flux prediction at the target time step t. This indicates explanatory information;

[0045] After receiving the prediction result data object, the results display terminal displays the carbon flux prediction curve, the three types of gating weight change curves, and the contribution value of each branch in the order of land parcel and time.

[0046] The present invention has the following advantages:

[0047] 1. A hierarchical processing relationship is established around mechanistic prior bias, dynamic residual, and structural bias. The mechanistic prior branch generates the mechanistic carbon flux prior value and the ecological process state, the multi-scale dynamic residual branch outputs the dynamic residual correction, and the structural bias calibration branch outputs the structural bias calibration. This ensures that different error sources have clear input data, processing logic, output results, and correction paths, thereby reducing the coupling of different bias components in the same parameter space, improving the prediction architecture's ability to specifically handle different error sources, and enabling the basis for the formation of the final prediction results to be traced through the outputs of each branch and their contribution values.

[0048] 2. The task of the mechanism prior branch is limited to providing the prior value of mechanism carbon flux, the direction of ecological processes, and the allowable change boundary. The task of the multi-scale dynamic residual branch is limited to learning the dynamic residual of carbon flux observations relative to the prior value of mechanism carbon flux. Through multi-scale time series processing, short-term changes and lag responses caused by continuous rainfall, sustained high humidity, temperature fluctuations, and agricultural management events are extracted. At the same time, the structured bias calibration branch compensates for the long-term or phased system offsets formed by different plots based on structured characteristics such as slope, aspect, altitude, slope position, soil background, and management intensity. This reduces the situation where short-term disturbances and plot background differences are jointly absorbed into the mechanism parameters or time series model shared parameters, so that the prediction process can adapt to short-term state changes and spatial heterogeneity of mountain tea gardens while retaining the carbon exchange process baseline.

[0049] 3. Through the state-aware gating fusion layer, the ecological process state output by the mechanism prior branch, the temporal dynamic representation output by the multi-scale dynamic residual branch, and the structured scene representation output by the structured bias calibration branch are used as the basis for generating gating weights. This dynamically determines the contribution of the mechanism carbon flux prior value, dynamic residual correction, and structured bias calibration to the final carbon flux prediction value. Therefore, when the environment is relatively stable and there are no significant agricultural management events, the final prediction result can retain more of the mechanism process baseline. During periods of continuous rainfall, persistent high humidity, or the impact of agricultural management events, the contribution of dynamic residual correction can be enhanced. When predicting across slope positions or plots, the contribution of structured bias calibration can be enhanced, thereby avoiding the same set of fixed weights being repeatedly applied to different prediction states and improving the matching degree between the output of each branch and the current state of the mountain tea garden.

[0050] 4. In the joint training process, prediction error loss, physical consistency constraint loss, and smoothing constraint loss are introduced. The ecological process state output by the mechanism prior branch is used to constrain the direction of change and allowable change boundary of the final carbon flux prediction value. The activation conditions of the smoothing constraint are determined according to the management state, environmental state, and soil state. Abnormal fluctuations lacking state change basis are suppressed only in the time period without agricultural management events and without significant environmental disturbances. Thus, without weakening the real carbon flux response caused by events such as continuous rainfall, harvesting, fertilization, pruning, and irrigation, the possibility of the final prediction result deviating from the carbon exchange process law of mountain tea garden is reduced, and the ecological process consistency and temporal stability of the continuous prediction results are improved. Attached Figure Description

[0051] Figure 1 This is a schematic diagram of the process of the present invention.

[0052] Figure 2 This is a schematic diagram of the spatial matching process for land parcels based on multi-source data. Detailed Implementation

[0053] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of the embodiments. The components of the embodiments of this application described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the detailed description of the embodiments of this application provided below with reference to the accompanying drawings is not intended to limit the scope of protection of the claimed application, but merely represents selected embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application. The present invention will be further described below with reference to the accompanying drawings.

[0054] like Figure 1As shown, this invention specifically relates to a hierarchical bias decoupling and gating fusion prediction method for carbon flux in mountain tea gardens. Instead of using multiple models to predict the same complete carbon flux separately and then applying fixed weights, this invention defines the task boundaries of each branch according to the formation mechanism of prediction bias. Specifically, a mechanistic prior branch provides the ecological process baseline; a multi-scale dynamic residual branch handles short-term disturbances and hysteretic responses caused by continuous rainfall, persistent high humidity, temperature fluctuations, and agricultural management events; a structured bias calibration branch handles long-term or phased system offsets caused by differences in slope, aspect, altitude, slope position, soil background, and management intensity; and a state-aware gating fusion layer dynamically determines the credibility of each branch based on the current sample state. The fusion result is constrained by prediction error loss, physical consistency constraint loss, and smoothing constraint loss, thus forming a complete prediction chain of "multi-source state construction—mechanistic prior generation—dynamic residual extraction—structured bias calibration—state-aware gating—physical constraint output." The core processing is executed by the prediction server, and data input and result output can be completed through data acquisition terminals, management input terminals, and result display terminals.

[0055] The data acquisition terminal is deployed at the monitoring site in the mountain tea garden. It is used to collect environmental and soil condition data such as air temperature, air humidity, rainfall, solar radiation, wind speed, soil temperature, and soil moisture according to a preset time step. It also writes a collection timestamp, plot identifier, and data source identifier for each data point. The carbon flux observation equipment generates historical carbon flux observation data within the same time range. The data from the mountain tea garden site and the historical carbon flux observation data are combined to form the original monitoring record and sent to the prediction server.

[0056] The management input terminal is used to record the occurrence time, event type, affected plot, and event intensity of agricultural management events such as harvesting, fertilizing, pruning, and irrigation.

[0057] The prediction server receives remote sensing data, topographic data, agricultural management data, environmental status data, soil status data, and historical carbon flux observation data, and sequentially performs time synchronization, spatial matching, unit unification, state vector construction, mechanism prior calculation, multi-scale dynamic residual calculation, structured bias calibration, state-aware gating fusion, and physical consistency constraints.

[0058] The results display terminal is used to receive and display the final carbon flux prediction, the prior value of mechanistic carbon flux, the dynamic residual correction, the structured bias calibration, the gating weight, and the contribution value of each branch.

[0059] Furthermore, the prediction server uses the plot identifier and time step as a unified index for multi-source data, converts heterogeneous data from the same plot and within the same time range into mountain tea garden state vectors, and sends the mountain tea garden state vectors to the mechanism prior branch, the multi-scale dynamic residual branch, and the structured bias calibration branch, respectively.

[0060] The mechanism prior branch generates a prior value of mechanistic carbon flux and the state of ecological processes based on the carbon cycle process in the tea garden. The multi-scale dynamic residual branch generates a dynamic residual correction amount relative to the prior value of mechanistic carbon flux and its time-series dynamic representation based on historical carbon flux trajectories and their synchronous environmental, soil, and agricultural management time-series data. The structured deviation calibration branch generates a structured deviation calibration amount reflecting the long-term or phased systematic deviation of the plot and its structured scene representation based on topographic state, soil background state, management intensity, and derived structured features. The state-aware gating fusion layer further generates normalized gating weights based on the state of ecological processes, time-series dynamic representation, and structured scene representation, and uses the gating weights to determine the deviation correction method that should be adopted for the current prediction sample to obtain the final carbon flux prediction value.

[0061] Furthermore, before executing the prediction method of this invention, the data access and state construction module must complete the time synchronization, spatial matching, and state vector construction of data from different sources for the same plot. Only then can the mechanistic prior branch, multi-scale dynamic residual branch, and structured bias calibration branch enter the calculation stage. The data access and state construction module is a functional module at the front end of the prediction link. The data acquisition terminal is responsible for generating or accessing field monitoring data, the management and input terminal is responsible for generating agricultural management records, and the prediction server is responsible for time synchronization, spatial matching, unit unification, anomaly verification, and state construction.

[0062] The data acquisition terminal periodically generates raw monitoring records, which are classified by variable type as environmental monitoring records, soil monitoring records, and carbon flux observation records. Each raw monitoring record includes at least one of the following: collection timestamp, data source identifier, variable name, variable value, plot identifier, and spatial coordinates.

[0063] The management input terminal generates separate agricultural management records, which are not part of the original monitoring records. Each agricultural management record must include at least the plot identifier, event type, event occurrence time, event duration, and event intensity. Environmental monitoring records, soil monitoring records, and carbon flux observation records that do not carry plot identifiers but carry spatial coordinates are collectively referred to as point monitoring records; "point" means that the record has acquisition coordinates that can be represented as a spatial point. Remote sensing data and rasterized topographic data must include at least raster cell values ​​and their spatial positioning information. The prediction server associates data from different sources with the corresponding tea garden plots based on the correspondence between plot identifiers, spatial coordinates, and plot boundaries.

[0064] After receiving the above data, the prediction server first performs resampling according to a preset uniform time step, converting data with different sampling frequencies to the same time base, and then organizes the data that has completed time synchronization according to the target time step.

[0065] like Figure 2 As shown, the prediction server pre-builds a land parcel spatial index table, which includes at least the first... Each tea garden plot has a plot identifier ID and the corresponding plot boundary polygon Ω. For environmental monitoring records, soil monitoring records, carbon flux observation records, and agricultural management records that already carry plot identifiers, the prediction server directly completes the association based on the plot identifier. When any of the above four records also carries spatial coordinates, it further determines whether the spatial coordinates are located within the plot boundary polygon corresponding to the plot identifier. If the plot identifier carried is inconsistent with the plot identifier determined based on the spatial coordinates, the record is marked as a spatial matching anomaly.

[0066] For point monitoring records that do not carry land parcel identifiers but carry spatial coordinates, the prediction server performs a spatial inclusion relationship determination between the spatial coordinates carried by the point monitoring record and the land parcel boundary polygon Ω; when the spatial coordinates uniquely fall into the land parcel boundary polygon Ω, the corresponding land parcel identifier ID is written into the point monitoring record.

[0067] For remote sensing data and rasterized terrain data, the prediction server calculates the spatial intersection relationship between each raster cell and the plot boundary polygon Ω, and then weights and summarizes the raster cells that intersect with the plot boundary polygon according to the intersection area, and then associates the summary result with the corresponding plot identifier ID.

[0068] Specifically, the prediction server reads the spatial boundary of the raster cells and their raster cell values, uses spatial overlay operation to calculate the area of ​​the intersection region between each raster cell and the target plot boundary polygon Ω, and uses the ratio of the area of ​​each intersection region to the sum of the areas of all intersection regions within the target plot as the weight of the corresponding raster cell. The values ​​of each raster cell are multiplied by their corresponding weights and then summed to obtain the summary value of the target plot, and the summary value is associated with the plot identifier ID of the target plot.

[0069] For point monitoring records that cannot uniquely identify the corresponding plot, the prediction server marks them as spatially mismatched records; for environmental monitoring records, soil monitoring records, carbon flux observation records, or agricultural management records whose plot identifiers are inconsistent with the plot identifiers determined based on their spatial coordinates, the prediction server marks them as spatially mismatched records.

[0070] Through the above processing, different data sources will ultimately use the land parcel identifier as the unified spatial association primary key.

[0071] After spatial matching is completed, the prediction server performs unit conversions for similar variables from different data sources according to a pre-defined unit mapping table. The prediction server pre-establishes a device parameter table, which records at least the data source identifier, variable name, and lower limit of the measurement range. and upper limit of range When the variables collected by data source s Variable values ​​at target time step t satisfy or When the variable value is found to be abnormal, the environmental monitoring record, soil monitoring record, or carbon flux observation record to which it belongs is marked as an equipment range abnormality record, and the writing of the environmental monitoring record, soil monitoring record, or carbon flux observation record to which the variable value belongs into the standardized status record table is stopped as valid data.

[0072] For environmental monitoring records, soil monitoring records, or carbon flux observation records within the equipment's measurement range, the prediction server uses a set of historical valid records from the same site and the same variable, as well as the same seasonal or phenological stage. Calculate the median and median absolute deviation :

[0073] ,

[0074] ,

[0075] And calculate the robustness deviation of the target record according to the following formula:

[0076] ,

[0077] in, Indicates a plot of land. Represents variables, Indicates seasonal or phenological stages. To prevent the default positive number with a denominator of zero. When robust deviation... Greater than the preset deviation threshold When this occurs, the corresponding record is marked as a candidate abnormal state record; preferably, a preset deviation threshold is used. It is 3.5.

[0078] For candidate state anomaly records, the prediction server verifies their adjacent time changes, concurrent meteorological events, and agricultural management events within a preset verification time window. When a candidate state anomaly record exhibits continuous changes in the same direction across multiple adjacent time steps, or is accompanied by any of the following concurrent events: rainfall, persistent high humidity, irrigation, fertilization, harvesting, or pruning, and the corresponding environmental variables or management status provide consistent support, the candidate state anomaly record is determined to correspond to a real meteorological disturbance or a real agricultural management response. The candidate state anomaly record is then written into the standardized state record table as valid data and an actual disturbance identifier is attached. When a candidate state anomaly record exhibits an isolated jump, and there is no valid monitoring record or related environmental variable or agricultural management record of the same variable type in adjacent time steps to support it, it is marked as an invalid record and not used as valid input. Valid monitoring records of the same variable type as the candidate state anomaly record include, for example, if the current anomaly record is a soil monitoring record, the corresponding monitoring record is the soil monitoring record in the adjacent time step; if the current anomaly record is a carbon flux observation record, the corresponding monitoring record is the carbon flux observation record in the adjacent time step.

[0079] An isolated jump refers to a candidate state anomaly record that meets the aforementioned criteria for a candidate state anomaly record only in a single target time step, and the valid monitoring records of the same variable in the adjacent time steps before and after it do not show continuous changes in the same direction. When a candidate state anomaly record shows an isolated jump, and the valid monitoring records, related environmental variables, and agricultural management records of the same variable type as the candidate state anomaly record in the adjacent time steps cannot provide consistent support, it is marked as an invalid record and is not used as a valid input.

[0080] For consecutive missing lengths not exceeding the preset number of missing steps Furthermore, valid records exist before and after the missing interval, and the prediction server uses the valid time step data before and after the missing interval to interpolate and complete the data; for consecutive missing intervals exceeding [a certain length], [the server performs interpolation]. Alternatively, if there are no valid records before or after the missing interval, automatic completion will not be performed, and a missing measurement marker will be added. After completing the above processing, the prediction server generates a standardized state record table with "plot identifier - time step" as the primary key, which can be called by the mechanism prior branch, multi-scale dynamic residual branch, structured bias calibration branch, and state-aware gating fusion layer.

[0081] For the target time step t, the prediction server constructs the state vector of the mountain tea garden:

[0082] ,

[0083] in, This represents the state vector of the mountain tea garden at the target time step t. Describing environmental conditions includes at least one or more of the following: temperature, humidity, rainfall, solar radiation, and wind speed. To indicate soil condition, it must include at least one or more of the following: soil temperature, soil moisture, soil organic carbon, total nitrogen, and pH value. Describing terrain features, including at least one or more of the following: slope, aspect, elevation, slope position, and parcel identifier. This indicates the management status and is used to characterize the event type, occurrence time, duration, and intensity of agricultural management events such as harvesting, fertilizing, pruning, and irrigation. The historical carbon flux status is represented by a sequence of carbon flux observations within a preset historical window prior to the target time step. The carbon flux observation sequence is generated by carbon flux observation records received or collected by the data acquisition terminal and sorted according to the plot identifier and collection timestamp.

[0084] The data access and state construction modules output the current state object, the historical time series input tensor, and the structured feature vector, respectively. The current state object is used by the mechanism prior branch, the historical time series input tensor is used by the multi-scale dynamic residual branch, and the structured feature vector is used by the structured bias calibration branch.

[0085] The three branches perform calculations based on the input object constructed from the same plot and the same target time step, but they read different state fields and historical ranges: the mechanism prior branch mainly reads the current environmental state, the current soil state, and the current management state; the multi-scale dynamic residual branch reads the environmental state sequence, soil state sequence, management state sequence, and historical carbon flux state within the preset historical window; and the structured bias calibration branch reads the topographic state, soil background state, management intensity, and derived statistical features.

[0086] The three branches can run in parallel after the state construction is complete. (Mechanism prior branch output) and Multi-scale dynamic residual branch output and Structured deviation calibration branch output and The state-aware gating fusion layer must be executed after the intermediate representation outputs have been completed in all three branches, utilizing... and Generate gate weights The final carbon flux prediction value is then generated by the fusion calculation submodule. .

[0087] Furthermore, the mechanism prior branch is executed by the mechanism computation submodule in the prediction server; the mechanism computation submodule receives the environmental state corresponding to the target time step. Soil condition Management status Based on the basic process parameters related to the target plot, and according to the process relationships between vegetation photosynthesis, vegetation respiration, soil respiration, and carbon pool transfer in the mountain tea garden, the environmental state at the target time step is calculated. Soil condition Management status The mapping relationship between basic process parameters related to the target site and prior values ​​of mechanistic carbon flux and ecological process states is expressed as follows:

[0088] ,

[0089] in, This represents the prior value of the mechanistic carbon flux. Indicates the state of ecological processes. This represents the a priori calculation process of the mechanism. This represents the set of process parameters used in the prior branch of the mechanism.

[0090] Ecological process state Used to characterize the photosynthetic-related state, respiration-related state, carbon pool transfer state, carbon exchange direction, permissible change boundary, or a combination of the above states of the current tea garden system.

[0091] Mechanistic carbon flux prior value It is not used directly as the sole final carbon flux prediction, but rather as a unified process benchmark for subsequent dynamic residual correction and structured bias calibration.

[0092] Ecological process state The results are then sent to the state-aware gating fusion layer and the physical consistency constraint submodule, respectively, to participate in gating weight generation and fusion result constraint.

[0093] The task boundary of the mechanism prior branch is limited to generating ecological process benchmarks and providing ecological process states. It does not use mechanism parameters to uniformly compensate for the following local biases: short-term dynamic biases caused by continuous rainfall, sustained high humidity, and temperature fluctuations; event response biases caused by agricultural management events such as harvesting, fertilization, pruning, and irrigation; and structured biases caused by different slope positions, different plots, and different soil backgrounds. These local biases are handled by the multi-scale dynamic residual branch and the structured bias calibration branch, respectively, thereby avoiding the back-coupling of dynamic disturbance errors and plot difference errors to mechanism parameters.

[0094] Furthermore, the multi-scale dynamic residual branch is executed by the temporal residual learning submodule in the prediction server. The prediction server extracts a historical time window of length L from the standardized state record table, with the target time step t as the endpoint, and arranges the historical carbon flux state, environmental state sequence, soil state sequence, and management state sequence within the historical time window in chronological order to form a temporal input tensor:

[0095] ,

[0096] in, L represents the temporal input tensor corresponding to the target time step t, and L represents the length of the history window. Indicates time step The corresponding combined record of historical carbon flux, environmental conditions, soil conditions, and management conditions.

[0097] The temporal residual learning submodule, for multiple preset time scales, uses resampling or decomposition operations corresponding to each time scale to obtain sequences at each scale, and then performs feature extraction on the sequences at each scale to obtain temporal representations at different time scales:

[0098] ,

[0099] in, Indicates the first Each time scale corresponds to a resampling operation or decomposition operation. This indicates the number of time scales. Finer time scales are used to preserve localized rapid changes after the onset of rainfall, sudden temperature changes, irrigation operations, or harvesting operations, while coarser time scales are used to extract sustained high humidity, post-event recovery processes, and trend changes over longer time periods.

[0100] The temporal dynamics representation is obtained by mixing the temporal representations at each scale with multiple scales.

[0101] .

[0102] The dynamic residual correction amount is obtained by mapping the residual output:

[0103] ,

[0104] in, Represents the dynamic representation of time series. This represents the dynamic residual correction amount. This represents multi-scale temporal feature extraction and hybrid operations. This represents the output operation that maps the time-series dynamic representation to residual correction values. The parameters represent the multi-scale dynamic residual branch.

[0105] During the training phase, the carbon flux observation at the target time step is denoted as... The prediction server first calculates the residual benchmark:

[0106] ,

[0107] in, This represents the residual between observed carbon flux and prior mechanistic carbon flux. The multi-scale dynamic residual branch uses historical time windows containing the effects of continuous rainfall, sustained high humidity, temperature fluctuations, or agricultural management events as the main training samples, thus adjusting the output dynamic residual. This branch approximates the dynamic residual components in the above samples. It only outputs the correction amount and the time-series dynamic characterization, and does not directly output the complete carbon flux prediction value, thereby avoiding repeated learning of the basic ecological processes that the prior branch of the mechanism can already express.

[0108] Furthermore, the structured bias calibration branch is executed by the structured bias calculation submodule in the prediction server. The prediction server extracts topographic state, soil background state, management event intensity, and derived statistical features from the mountain tea garden state vector and standardized state record table to form a structured feature vector:

[0109] ,

[0110] in, This represents the structured feature vector corresponding to the target time step. Indicates terrain condition, Indicates the background state of the soil. Indicates the intensity of current or historical agricultural management events. Indicates the duration of continuous high humidity. This indicates the cumulative rainfall within the preset statistics window. This indicates the historical fertilization intensity within the preset statistical window. This indicates the harvesting frequency within the preset statistics window.

[0111] The above fields can be selected based on the actual availability of existing project data, and it is not required that all fields exist simultaneously.

[0112] The structured deviation calculation submodule generates structured deviation calibration values ​​and structured scene representations based on structured feature vectors.

[0113] ,

[0114] in, Indicates the amount of structured deviation calibration; Represents a structured scene representation; ( ) represents the structured deviation mapping process; The parameters represent the structured deviation calibration branch.

[0115] The structured deviation calibration branch focuses on learning long-term or phased systemic offsets caused by differences in slope, aspect, altitude, slope position, background soil conditions, and management intensity, and does not undertake the main task of learning high-frequency fluctuations within continuous time steps. The structured deviation calibration branch can use a tree model suitable for handling nonlinear structured variables and variable interactions, or other models that can output continuous deviation calibration quantities based on structured feature vectors; the specific algorithm type does not constitute a limitation on the core technical solution of this invention.

[0116] Furthermore, the state-aware gating fusion layer is executed by the gating computation submodule in the prediction server. The gating computation submodule receives the ecological process states output from the mechanism prior branch. Temporal dynamic representation of multi-scale dynamic residual branch output Structured scene representation output from the structured deviation calibration branch It can further receive a current state summary formed by the current environmental state, soil state, and management state. .

[0117] The gated input vector is represented as:

[0118] ,

[0119] Here, ";" indicates vector concatenation. The gating calculation submodule first generates the unnormalized gating score:

[0120] ,

[0121] in, Represents the gated mapping weight matrix. This represents the gating bias vector. This represents the unnormalized gating score. The fusion weights are then obtained through normalization.

[0122] ,

[0123] This yields the gating weight vector:

[0124] ,

[0125] And it satisfies:

[0126] ,

[0127] in, This represents the confidence weight of the prior branch of the mechanism. The confidence weights represent the multi-scale dynamic residual branches. This represents the confidence weight of the structured bias calibration branch.

[0128] To maintain consistency between the prior value of mechanistic carbon flux and the physical meaning of the two types of bias corrections during the fusion process, candidate values ​​for mechanistic priors, dynamic residual corrections, and structured bias corrections are constructed respectively:

[0129] ,

[0130] The final carbon flux forecast is calculated as follows:

[0131] ,

[0132] Substituting the three candidate values ​​and combining them with the constraint that the sum of the weights is 1, the above equation is equivalent to:

[0133] ,

[0134] in, This represents the final carbon flux prediction at the target time step t. To determine the relative dominance of the three candidate values ​​in the current target time step, a dominant candidate branch index is defined:

[0135] ,

[0136] when ,satisfy and This indicates the prior candidate values ​​of the mechanism. It has the largest fusion coefficient among the weighted combinations of the three candidate values; when At time 2, it indicates a candidate value for dynamic residual correction. It has the largest fusion coefficient; when At 3, it indicates a candidate value for structured bias correction. It has the largest fusion coefficient; when two or more gating weights are equal and reach the maximum value together, the corresponding candidate branch is determined as the common dominant branch.

[0137] Furthermore, the actual correction contribution of dynamic residuals is defined as... The actual correction contribution of structural bias is defined as Then the final carbon flux prediction value at the target time step t is expressed as: This allows us to determine the actual compositional relationship between the mechanistic process baseline, dynamic residual correction, and structured bias correction on the final predicted value.

[0138] Furthermore, the physical consistency and smoothness constraint module is executed by the loss calculation submodule in the prediction server, and is mainly used in the joint training and fine-tuning stages of the mechanistic prior branch, multi-scale dynamic residual branch, structured bias calibration branch, and state-aware gated fusion layer. The loss calculation submodule calculates the prediction error loss, physical consistency constraint loss, and smoothness constraint loss.

[0139] Prediction error loss is used to measure the difference between the final carbon flux prediction and the observed carbon flux, and is preferably expressed as:

[0140] ,

[0141] in, Indicates the number of training samples. This represents the observed carbon flux at the target time step.

[0142] In a preferred embodiment, the mechanistic prior branch is based on the state of the ecological process. Output the lower boundary of the carbon flux allowed at the target time step. Upper boundary and the direction of process change ,in This can represent whether the predicted value should increase, decrease, or remain relatively stable relative to the previous time step. The physical consistency constraint loss is expressed as:

[0143] ,

[0144] in, This is a penalty item for exceeding the upper boundary. This is a penalty for exceeding the lower boundary. The process change direction conflict penalty term; the upper boundary over-limit penalty term and the lower boundary over-limit penalty term are used to penalize the final carbon flux prediction value exceeding the allowable boundary of the mechanism process, and the process change direction conflict penalty term is used to... The direction of temporal variation of the final carbon flux prediction penalty during activation conflicts with the direction of variation of the mechanistic process. Indicates the directional constraint coefficient; This indicates an activation flag indicating a stable state.

[0145] Stable state activation indicator Determined based on current management status and changes in environmental status:

[0146] ,

[0147] in, and These represent the threshold values ​​for judging significant disturbances in environmental and soil conditions, respectively. These threshold values ​​can be set based on historical ranges of change or existing ecological process parameters.

[0148] The smoothing constraint loss is expressed as:

[0149] ,

[0150] This loss is only effective during periods without significant meteorological disturbances and agricultural management events, and is used to suppress high-frequency oscillations in the final forecast result that lack evidence of state change; when events such as continuous rainfall, sudden temperature changes, harvesting, fertilization, pruning, or irrigation occur, Setting it to 0 prevents the use of smoothing constraints to forcibly eliminate the changes in real carbon flux caused by events.

[0151] The total loss of the joint training of the mechanism prior branch, the multi-scale dynamic residual branch, the structured bias calibration branch, and the state-aware gated fusion layer is expressed as:

[0152] ,

[0153] in, , and These are the non-negative weight coefficients for controlling the prediction error loss, physical consistency constraint loss, and smoothness constraint loss, respectively, and are used to control the contributions of prediction error, physical consistency, and smoothness in the joint training process.

[0154] Furthermore, the result output module is jointly executed by the result encapsulation submodule in the prediction server and the result display terminal. The result encapsulation submodule encapsulates the following data corresponding to the same plot and the same target time step into prediction result data objects:

[0155] ,

[0156] in, Indicates land parcel identification; Indicates the target time step; This represents the prior value of the mechanistic carbon flux; This represents the dynamic residual correction amount; Indicates the amount of structured deviation calibration; Represents the gating weight vector; This represents the final carbon flux prediction for the target time step t; This indicates explanatory information, including one or more of the following: branch contribution values, management event logs, and contributions from major structured features.

[0157] After receiving the prediction result data object, the results display terminal displays the carbon flux prediction curve, the three types of gating weight change curves, and the contribution values ​​of each branch in order of plot and time. The final carbon flux prediction value can also be used as a pre-input for dynamic carbon sink accounting, or as update data for the carbon exchange status of the corresponding plot of the digital twin tea garden.

[0158] Furthermore, the training and inference phases employ different execution sequences. The training phase first completes historical data preprocessing and training sample construction, then calibrates the mechanism prior branch, followed by training the multi-scale dynamic residual branch and the structured bias calibration branch, and finally introduces a state-aware gating fusion layer for joint fine-tuning. The inference phase is triggered after the state construction of new target time step data is completed; when the management input terminal adds harvesting, fertilizing, pruning, or irrigation events, the prediction server updates the management status of the corresponding plot and regenerates the target time step prediction results affected by the events.

[0159] In different application scenarios, the gating weights are automatically generated from the input state and branch representations, rather than being manually set. During periods of continuous rainfall or persistent high humidity, the multi-scale dynamic residual branch extracts strong dynamic disturbances and hysteresis representations from the historical window, and the state-aware gating fusion layer correspondingly increases the credibility weight of the dynamic residual branch. When making predictions between different slope positions or different plots, the differences in topography and soil background reflected by the structured scene representation increase, and the state-aware gating fusion layer correspondingly increases the credibility weight of the structured bias calibration branch. When the environmental state is relatively stable and there are no agricultural management events, the credibility of the ecological process state and the mechanism process benchmark is high, and the state-aware gating fusion layer correspondingly increases the credibility weight of the mechanism prior branch. Within the impact window after harvesting, fertilizing, or pruning events, the management state simultaneously enters the multi-scale dynamic residual branch and the structured bias calibration branch, enabling the gating fusion layer to simultaneously utilize short-term response information after the event and the phased offset information caused by management intensity.

[0160] The above description is merely a preferred embodiment of the present invention. It should be understood that the present invention is not limited to the forms disclosed herein and should not be construed as excluding other embodiments. It can be used in various other combinations, modifications, and improvements, and can be altered within the scope of the concept described herein through the above teachings or related technologies or knowledge. Modifications and variations made by those skilled in the art that do not depart from the spirit and scope of the present invention should be within the protection scope of the appended claims.

Claims

1. A stratified bias decoupling and gating fusion prediction method for carbon flux in mountain tea gardens, characterized in that: The prediction method includes: S1. Collect on-site data of mountain tea gardens through data acquisition terminals according to preset time steps, write collection timestamp, plot identifier and data source identifier for each data, generate historical carbon flux observation data within the same time range by carbon flux observation equipment, and send the original monitoring record of the mountain tea garden on-site data and historical carbon flux observation data to the prediction server. S2. Record agricultural management event information such as harvesting, fertilizing, pruning and irrigation through the management input terminal to form agricultural management records including plot identification, event type, event occurrence time, event duration and event intensity; S3. The prediction server receives terrain data and remote sensing data and rasterized terrain data of corresponding plots, as well as data sent by data acquisition terminals and management input terminals. It simultaneously undertakes the processing tasks of different error sources through mechanism prior branches, multi-scale dynamic residual branches and structured bias calibration branches. Then, through the state-aware gating fusion layer, the contribution of each branch is dynamically allocated according to the current environmental state, ecological process state, temporal dynamic state and structured background state. The final prediction result is limited by physical consistency constraints and smoothing constraints. Finally, the final prediction result is displayed through the result output module.

2. The method for predicting carbon flux in mountain tea gardens by stratified bias decoupling and gating fusion according to claim 1, characterized in that: Before executing the prediction method, the time synchronization, spatial matching and state vector construction of data from different sources for the same plot need to be completed through the data access and state construction module before the mechanism prior branch, multi-scale dynamic residual branch and structured bias calibration branch can enter the calculation stage. The data access and status construction module is comprised of a data acquisition terminal that generates or accesses on-site monitoring data, a management input terminal that generates agricultural management records, and a prediction server that performs time synchronization, spatial matching, unit unification, anomaly verification, and status construction. The data acquisition terminal periodically generates raw monitoring records. The raw monitoring records include environmental monitoring records, soil monitoring records and carbon flux observation records according to variable type. Each raw monitoring record includes at least one of the following: collection timestamp, data source identifier, variable name, variable value, plot identifier and spatial coordinates. The management input terminal generates separate agricultural management records, which are not part of the original monitoring records; The prediction server associates data from different sources with the corresponding tea garden plots based on the plot identifiers, spatial coordinates, and the correspondence between plot boundaries.

3. The method for predicting carbon flux in mountain tea gardens by stratified bias decoupling and gating fusion according to claim 2, characterized in that: The prediction server pre-establishes a land parcel spatial index table to complete land parcel spatial matching. The land parcel spatial index table includes the land parcel identifier ID of the i-th tea garden and the land parcel boundary polygon Ω corresponding to the land parcel identifier. For environmental monitoring records, soil monitoring records, carbon flux observation records, and agricultural management records that already carry plot identifiers, the prediction server directly completes the association according to the plot identifier. When one of these four records also carries spatial coordinates, it determines whether the spatial coordinates are located within the plot boundary polygon corresponding to the plot identifier. If the plot identifier carried is inconsistent with the plot identifier determined according to the spatial coordinates, the record is marked as a spatial matching anomaly record. For point monitoring records that do not carry plot identifiers but carry spatial coordinates, the prediction server performs a spatial inclusion relationship determination between the spatial coordinates carried by the point monitoring record and the plot boundary polygon Ω. When the spatial coordinates uniquely fall into the plot boundary polygon Ω, the corresponding plot identifier ID is written into the point monitoring record. The point monitoring records include environmental monitoring records, soil monitoring records, and carbon flux observation records that do not carry plot identifiers but carry spatial coordinates. For remote sensing data and rasterized terrain data, the prediction server calculates the spatial intersection relationship between each raster cell and the plot boundary polygon Ω, and then weights and summarizes the raster cells that intersect with the plot boundary polygon Ω according to the intersection area, and then associates the summary result with the corresponding plot identifier ID. For point monitoring records that cannot uniquely identify the corresponding plot, the prediction server marks them as spatially mismatched records; for environmental monitoring records, soil monitoring records, carbon flux observation records, or agricultural management records whose plot identifiers are inconsistent with the plot identifiers determined based on their spatial coordinates, the prediction server marks them as spatially mismatched records.

4. The stratified bias decoupling and gating fusion prediction method for carbon flux in mountain tea gardens according to claim 3, characterized in that: After completing spatial matching, the prediction server establishes a device parameter table, which includes at least the record data source identifier, variable name, and lower limit of the measurement range. and upper limit of range ; When data source s collects variables Variable values ​​at target time step t satisfy or When the variable value is found to be abnormal, the environmental monitoring record, soil monitoring record, or carbon flux observation record to which it belongs shall be marked as an abnormal equipment range record, and the writing of the environmental monitoring record, soil monitoring record, or carbon flux observation record to which the variable value belongs into the standardized status record table shall be stopped as valid data. For environmental monitoring records, soil monitoring records, or carbon flux observation records within the equipment's measurement range, the prediction server uses a set of historical valid records from the same site and the same variable, as well as the same seasonal or phenological stage. Calculate the median and median absolute deviation And calculate the robustness deviation of the target record. When the robustness deviation is greater than the set deviation threshold, the corresponding record is marked as a candidate abnormal state record. For candidate state anomaly records, the prediction server verifies their adjacent time changes, concurrent meteorological events, and agricultural management events within a preset verification time window. When the candidate state anomaly record shows continuous changes in the same direction across multiple adjacent time steps, or when any of the following events occurs concurrently: rainfall, sustained high humidity, irrigation, fertilization, harvesting, or pruning, and the corresponding environmental variables or management status provide consistent support, the candidate state anomaly record is determined to correspond to a real meteorological disturbance or a real agricultural management response. The candidate state anomaly record is then written into the standardized state record table as valid data and an actual disturbance identifier is attached. When the candidate state anomaly record shows an isolated jump, and there are no valid monitoring records or related environmental variables or agricultural management records of the same variable type as the candidate state anomaly record in adjacent time steps, it is marked as an invalid record and is not used as valid input. For consecutive missing lengths not exceeding the preset number of missing steps Furthermore, valid records exist before and after the missing interval, and the prediction server uses the valid time step data before and after the missing interval to interpolate and complete the data; for consecutive missing intervals exceeding a certain length... If there are no valid records before or after the missing interval, automatic completion will not be performed, and a missing test mark will be added; A standardized state record table with plot identifier-time step as the primary key is formed, which is called by the mechanism prior branch, multi-scale dynamic residual branch, structured bias calibration branch, and state-aware gating fusion layer. For the target time step t, the prediction server constructs the state vector of the mountain tea garden. ,in, This represents the state vector of the mountain tea garden at the target time step t. Indicates the environmental state. Indicates soil condition. Indicates terrain condition, Indicates the management status. This indicates the historical carbon flux status.

5. The stratified bias decoupling and gating fusion prediction method for carbon flux in mountain tea gardens according to claim 1, characterized in that: The mechanism prior branch is executed by the mechanism calculation submodule in the prediction server, which receives the environmental state corresponding to the target time step. Soil condition Management status Based on the basic process parameters related to the target plot, and according to the process relationships between vegetation photosynthesis, vegetation respiration, soil respiration, and carbon pool transfer in the mountain tea garden, the environmental state at the target time step is calculated. Soil condition Management status The mapping relationship between basic process parameters related to the target site and prior values ​​of mechanistic carbon flux and ecological process states is expressed as follows: ,in, This represents the prior value of the mechanistic carbon flux. Indicates the state of ecological processes. This represents the prior calculation process of the mechanism. The set of process parameters used to represent the prior branches of the mechanism; Ecological process state Used to characterize the photosynthetic-related state, respiration-related state, carbon pool transfer state, carbon exchange direction, permissible change boundary, or a combination of these states of the current tea garden system; Mechanistic carbon flux prior value It is not used directly as the sole final carbon flux prediction, but rather as a unified process benchmark for subsequent dynamic residual correction and structured bias calibration. Ecological process state The results are then sent to the state-aware gating fusion layer and the physical consistency constraint submodule, respectively, to participate in gating weight generation and fusion result constraint.

6. The stratified bias decoupling and gating fusion prediction method for carbon flux in mountain tea gardens according to claim 4, characterized in that: The multi-scale dynamic residual branch is executed by the temporal residual learning submodule in the prediction server. The prediction server extracts a historical time window of length L from the standardized state record table with the target time step t as the endpoint, and arranges the historical carbon flux state, environmental state sequence, soil state sequence and management state sequence within the historical time window in chronological order to form a temporal input tensor. ,in, L represents the temporal input tensor corresponding to the target time step t, and L represents the length of the history window. Indicates time step The corresponding combined record of historical carbon flux, environmental conditions, soil conditions, and management conditions; The temporal residual learning submodule, for multiple preset time scales, uses resampling or decomposition operations corresponding to each time scale to obtain sequences at each scale, and then performs feature extraction on the sequences at each scale to obtain temporal representations at different time scales. ,in, Indicates the first Each time scale corresponds to a resampling operation or decomposition operation. Indicates the number of time scales. Time series representations at different time scales; The temporal dynamics representation obtained after multi-scale mixing of temporal representations at each scale is as follows: The dynamic residual correction amount is obtained by mapping the residual output. ,in, Represents the dynamic representation of time series. This represents the dynamic residual correction amount. This represents multi-scale temporal feature extraction and hybrid operations. This represents the output operation that maps the time-series dynamic representation to residual correction values. The parameters represent the multi-scale dynamic residual branch.

7. The method for predicting carbon flux in mountain tea gardens by stratified bias decoupling and gating fusion according to claim 4, characterized in that: The structured bias calibration branch is executed by the structured bias calculation submodule in the prediction server. The prediction server extracts topographic state, soil background state, management event intensity, and derived statistical features from the mountain tea garden state vector and standardized state record table to form a structured feature vector. ,in, This represents the structured feature vector corresponding to the target time step. Indicates the background state of the soil. Indicates the intensity of current or historical agricultural management events. Indicates the duration of continuous high humidity. This indicates the cumulative rainfall within the preset statistics window. This indicates the historical fertilization intensity within the preset statistical window. This indicates the harvesting frequency within the preset statistics window; The structured deviation calculation submodule generates structured deviation calibration values ​​and structured scene representations based on structured feature vectors. ,in, This indicates the amount of structured deviation calibration. Represents a structured scene representation. ( ) represents the structured deviation mapping process. The parameters represent the structured deviation calibration branch.

8. The method for predicting carbon flux in mountain tea gardens by stratified bias decoupling and gating fusion according to claim 1, characterized in that: The state-aware gating fusion layer is executed by the gating computation submodule in the prediction server. The gating computation submodule receives the ecological process state output by the mechanism prior branch. Temporal dynamic representation of multi-scale dynamic residual branch output Structured scene representation output from the structured deviation calibration branch It also receives a current state summary formed by the current environmental state, soil state, and management state. The gated input vector is obtained as , where ";" indicates vector concatenation; Generate unnormalized gating score The fusion weights were then obtained through normalization. Thus, the gating weight vector is obtained as follows: And satisfy and ,in, Represents the gated mapping weight matrix. This represents the gating bias vector. This indicates the unnormalized gating score. This represents the confidence weight of the prior branch of the mechanism. The confidence weights represent the multi-scale dynamic residual branches. Represents the confidence weight of the structured bias calibration branch; Construct the prior candidate values ​​of the mechanism respectively Candidate values ​​for dynamic residual correction And the structured bias correction candidate value ,get ,in, This represents the final carbon flux prediction at the target time step t. This represents the dynamic residual correction amount. This represents the prior value of the mechanistic carbon flux. Indicates the amount of structured deviation calibration; To determine the relative dominance of the three candidate values ​​in the current target time step, the dominant candidate branch index is defined as... , Indicates the dominant candidate branch index, when ,satisfy and This indicates the prior candidate values ​​of the mechanism. It has the largest fusion coefficient among the weighted combinations of the three candidate values, when At time 2, it indicates a candidate value for dynamic residual correction. It has the largest fusion coefficient when At 3, it indicates a candidate value for structured bias correction. The branch with the largest fusion coefficient is the candidate branch that is determined as the common dominant branch when two or more gating weights are equal and reach the maximum value. Define the actual correction contribution of dynamic residuals as The actual correction contribution of structured bias is defined as The final carbon flux prediction is expressed as This allows us to determine the actual compositional relationship between the mechanistic process baseline, dynamic residual correction, and structured bias correction on the final predicted value.

9. The stratified bias decoupling and gating fusion prediction method for carbon flux in mountain tea gardens according to claim 8, characterized in that: The physical consistency constraint and smoothing constraint are executed by the loss calculation submodule in the prediction server. The loss calculation submodule calculates the prediction error loss, the physical consistency constraint loss and the smoothing constraint loss. Prediction error loss This is used to measure the difference between the final carbon flux prediction and the observed carbon flux, where... Indicates the number of training samples. This represents the carbon flux observation at the target time step. This represents the loss due to prediction error; The physical consistency constraint loss is expressed as ,in, Indicates the directional constraint coefficient. Indicates a stable state activation flag. Indicates the upper boundary. Indicates the lower boundary. This represents the loss due to physical consistency constraints; Smoothing constraint loss is expressed as ,in, This represents the loss due to smoothing constraints. The final total loss is expressed as ,in, , and These are the non-negative weighting coefficients for controlling prediction error loss, physical consistency constraint loss, and smoothness constraint loss, respectively. This represents the total loss.

10. The method for predicting carbon flux in mountain tea gardens by stratified bias decoupling and gating fusion according to claim 1, characterized in that: The result output module is jointly executed by the result encapsulation submodule in the prediction server and the result display terminal. The result encapsulation submodule encapsulates the data corresponding to the same plot and the same target time step into a prediction result data object. ,in, This represents the data object representing the prediction result. Indicates land parcel identification. Indicates the target time step. This represents the prior value of the mechanistic carbon flux. This represents the dynamic residual correction amount. This indicates the amount of structured deviation calibration. Represents the gating weight vector. This represents the final carbon flux prediction at the target time step t. This indicates explanatory information; After receiving the prediction result data object, the results display terminal displays the carbon flux prediction curve, the three types of gating weight change curves, and the contribution value of each branch in the order of land parcel and time.

Images