A soil organic carbon dynamic monitoring system based on model fusion

Abstract

This invention discloses a model fusion-based dynamic monitoring system for soil organic carbon, belonging to the field of soil monitoring and ecological environment information technology. The system includes the following steps: S1, acquiring multi-source data related to soil organic carbon monitoring and preprocessing it to obtain a standardized dataset; S2, generating mechanistic features with both mechanistic and spatiotemporal correlation through a carbon cycle process model; S3, constructing an artificial intelligence prediction model to learn the mapping relationship between soil organic carbon and multi-source inputs to achieve rapid inversion of soil organic carbon; S4, fusing the carbon cycle process model and the artificial intelligence prediction model to obtain a fused model; S5, achieving long-term continuous monitoring through dynamic inversion, operational status monitoring and anomaly self-repair, and sliding window incremental updates; S6, generating standardized results supporting long-term applications in multiple fields and continuous system optimization. Using this system, high-precision and high-efficiency multi-scale dynamic monitoring of soil organic carbon is achieved.

Description

Technical Field

[0001] This invention relates to the field of soil monitoring and ecological environment information technology, and in particular to a dynamic monitoring system for soil organic carbon based on model fusion. Background Technology

[0002] Soil organic carbon, as a core component of the global terrestrial ecosystem carbon pool, is crucial for maintaining soil fertility, supporting sustainable agricultural development, and advancing the implementation of "dual carbon" goals.

[0003] However, while traditional manual sampling and laboratory chemical analysis are highly accurate, they are time-consuming, labor-intensive, costly, and have limited spatial coverage. Estimation methods based on deep learning or machine learning, while possessing certain inversion capabilities, generally suffer from weak interpretability and insufficient cross-regional generalization. On the other hand, while process mechanism-based models can characterize key processes of carbon input, transformation, and decomposition, they rely on high-quality multi-parameter inputs and have low computational efficiency, making it difficult to support dynamic monitoring and widespread application at regional and larger scales. Summary of the Invention

[0004] The purpose of this invention is to provide a model fusion-based dynamic monitoring system for soil organic carbon, in order to construct a unified monitoring framework that combines mechanistic constraints, prediction accuracy, and operational efficiency, and to solve the technical problem that existing soil organic carbon monitoring methods cannot simultaneously meet the requirements of large scale, dynamic monitoring, and high reliability.

[0005] To achieve the above objectives, this invention provides a soil organic carbon dynamic monitoring system based on model fusion, comprising the following steps: S1. Obtain multi-source data related to soil organic carbon monitoring, and perform standardized preprocessing such as unified formatting, spatial resampling, and missing data imputation on the multi-source data to obtain a standardized dataset. S2, based on the standardized dataset output by S1, quantifies the physical-ecological processes of soil carbon input, decomposition and transformation through a carbon cycle process model, generating mechanistic features that combine mechanistic and spatiotemporal correlation. S3. Based on the preprocessed multi-source data in step S1 and the mechanistic characteristics in step S2, construct an artificial intelligence prediction model to learn the mapping relationship between soil organic carbon and multi-source inputs, so as to achieve rapid inversion of soil organic carbon. S4. Employ one or more strategies among residual coupling, feature coupling, loss function constraints, and Bayesian fusion to fuse the carbon cycle process model with the artificial intelligence prediction model, so that the fused model has both mechanistic interpretability, high accuracy, and high efficiency. S5. Based on time-series meteorological and remote sensing data, through dynamic inversion, operational status monitoring and anomaly self-repair, and sliding window incremental update, long-term continuous monitoring of soil organic carbon and continuous optimization of model parameters are achieved. S6. Based on the dataset generated by the time-series monitoring in step S5, through multi-scale aggregation, customized adaptation for application scenarios, multi-format output and API calls, data archiving and feedback iteration, standardized results supporting long-term applications in multiple fields and continuous system optimization are generated.

[0006] Preferably, the multi-source data in step S1 includes satellite remote sensing data. Meteorological driving data Soil basic attribute data Topographic data and land use type data The information from multivariate data is analyzed to obtain the spatial coordinate system, spatial resolution, temporal resolution, and effective range of data values.

[0007] Preferably, the uniform formatting process in the standardization preprocessing of step S1 is as follows: All raw data are uniformly converted into a standard raster format, and the coordinate systems of different data sources are uniformly converted into the target projected coordinate system; Bilinear interpolation is used to resample multi-source raster data with unified coordinates to the target spatial resolution, thereby achieving resolution normalization. For missing values ​​in the resampled data, spatiotemporal co-imputation is performed by combining spatial correlation and temporal continuity, as shown in the following formula: ; in, The interpolated complete data value; These are the spatiotemporal interpolation weighting coefficients, and , For the spatial correlation strength of data, The strength of the temporal correlation of the data; This is a spatial interpolation estimate; These are time interpolation estimates; Output standardized dataset .

[0008] Preferably, the specific steps of step S2 are as follows: S21. Standardized dataset based on step S1 The parameters required by the model are screened and extracted to form a set of model input parameters, including vegetation-related parameters, meteorological driving parameters, soil basic parameters, land use parameters, and spatial auxiliary parameters. ; S22. Based on the vegetation-related parameters and meteorological driving parameters input in step S21, simulate the vegetation residue carbon input process: Based on the vegetation growth status and meteorological conditions, first estimate the aboveground biomass of vegetation, then calculate the vegetation residue carbon input. The formula for calculating aboveground biomass of vegetation is: ; in, Aboveground biomass of vegetation; Biomass conversion coefficient; Leaf area index; Normalized Difference Vegetation Index; The average daily temperature; This is the optimal temperature for vegetation growth; Daily precipitation; The optimal rainfall for vegetation growth; It is the hyperbolic tangent function; The formula for calculating the carbon input of vegetation residue is: ; in, Carbon input from vegetation residue; The proportion of vegetation residue; Harvest coefficient; The vegetation carbon content coefficient; S23. Based on the soil basic parameters and land use parameters input in step S21, simulate the soil carbon pool decomposition process: divide soil organic carbon into active carbon pool, slow-release carbon pool, and inert carbon pool; first calculate the environmental regulation factor, then determine the decomposition rate of each carbon pool, and finally calculate the decomposition flux of each carbon pool. The formula for calculating the temperature regulation factor is as follows: ; in, Temperature regulation factor; Temperature sensitivity coefficient; Soil temperature; This is the optimal temperature for carbon decomposition. The formula for calculating the humidity regulation factor is: ; in, Humidity regulation factor; This refers to the soil volumetric water content. The optimal soil moisture content for carbon decomposition; The formulas for calculating the decomposition rate of each carbon pool are as follows: ; in, , , The actual decomposition rates of the active, slow-release, and inert carbon libraries are listed in order. , , The basic decomposition rates of the active, slow-release, and inert carbon libraries are listed in order. Soil texture regulator; The formulas for calculating the decomposition flux of each carbon pool are as follows: ; in, , , The decomposition fluxes of the active, slow-release, and inert carbon libraries are listed in order. , , The initial reserves of the active, slow-release, and inert carbon libraries are listed in that order. S24. Based on the initial carbon pool storage and the inter-pool transformation coefficient, calculate the inter-pool transformation flux to simulate the soil carbon pool transformation process. The calculation formula is as follows: ; in, To obtain carbon from the pool arrive The conversion flux; For carbon pool arrive The conversion coefficient; For source carbon pool Reserves; , Corresponding to active, slow-release, and inert carbon libraries respectively. ; S25. Integrate the simulation results of carbon input, decomposition flux, conversion flux, and carbon pool storage to form a set of mechanistic characteristics. .

[0009] Preferably, the specific steps of step S3 are as follows: S31. Standardize the dataset from step S1. The set of mechanistic features of step S25 Perform feature concatenation to construct the initial input feature set; S32. The mutual information entropy method is used to remove redundancy from the initial input feature set, and the effective feature set is obtained by screening. Then, the feature scale is unified through standardization to obtain the processed feature set. ; S33, Standardized feature set based on step S32 A physical information-enhanced artificial intelligence prediction model is constructed to enable it to learn the mapping relationship between soil organic carbon and input features. The model expression is as follows: ; in, The soil organic carbon content predicted by the model; For artificial intelligence prediction models; For the set of model parameters; S34, Standardized feature set based on step S32 and the corresponding measured values ​​of soil organic carbon Train the model and optimize the parameters by minimizing the loss function. The mixture loss function is defined and calculated using the following formula: ; in, This is a mixed loss value; The fitting deviation between the predicted and measured values. ; For carbon conservation constraint loss, ; To constrain the loss weighting coefficients; For the first SOC prediction value for each sample; For the first Measured SOC values ​​for each sample; The time-series rate of change of the SOC predicted value; This represents the sum of the decomposition fluxes of all carbon pools. The formula for calculating the optimized model parameters is: ; in, This is the optimized set of model parameters. Operator for parameter optimization; The Adam optimizer is used to solve for the optimal parameters. The optimization formula is as follows: ; in, For the first Model parameters for the next iteration This represents the number of iterations. The learning rate; For the first In the next iteration, the loss function affects the parameters. The gradient; S35. The optimized model is validated using metrics such as root mean square error.

[0010] Preferably, the specific steps of step S4 are as follows: S41. Extract the core outputs and auxiliary parameters of the carbon cycle process model and the artificial intelligence prediction model to form a fusion basic dataset; S42. By fusing feature dimensions and correcting residuals, the mechanism-data synergy between the carbon cycle process model and the artificial intelligence prediction model is enhanced. The feature coupling formula is as follows: ; in, For coupling feature set; For feature concatenation function; Output features for the intermediate layer of the AI ​​model; The residual coupling formula is: ; ; in, For the residuals of the mechanism model; This is the residual correction amount; This is a residual correction AI model used to learn the mapping relationship between coupled features and residuals; These are the parameters of the residual sub-model; S43. Embedding multiple constraints balances fitting accuracy, mechanistic interpretability, and residual correction effect to construct a fusion loss function, calculated as follows: ; in, This represents the total loss value of the fusion model; To fuse the fitting loss, and ; For residual constraint loss, and ; These are the residual constraint weights; , , , They are respectively the number Mechanism estimate, residual correction, measured value, and mechanism residual for each sample; S44. Based on dynamic weight allocation due to model uncertainty, the Bayesian weight calculation formula is as follows: ; ; in, Bayesian weights for the mechanistic model; For the Bayesian weights of the AI ​​model; Uncertainty in the mechanism model; For the uncertainty of AI models; The fusion calculation formula is as follows: ; in, This is the estimated SOC value after final fusion; This is the output of the mechanism model after residual correction.

[0011] Preferably, step S4 includes step S45, which specifically involves: using root mean square error and coefficient of determination to verify the performance of the fusion model.

[0012] Preferably, the specific steps of step S5 are as follows: S51. Based on time-series meteorological data and remote sensing data, and combined with the standardized process in step S1, construct a spatiotemporally continuous time-series input dataset, and obtain the effective time-series feature set through a time-series feature processing function. And initialize the system's long-term operating status parameters, using the following formula: ; in, This serves as a status indicator (1 = normal, 0 = abnormal). This is the initial runtime; The current time; This is the threshold for abnormal repair. S52. Based on the time-series effective feature set obtained in step S51 and the fusion model obtained in step S4, the SOC dynamic inversion is realized time-by-time, and the formula is as follows: ; in, for The dynamic inversion value of SOC after fusion at each moment; , for Bayesian weights at each time step; for The estimated SOC value of the mechanism model at time step; for The residual correction amount at time step; for AI model SOC prediction value at time; S53. Monitor the system's operating status in real time and implement self-repair for anomalies, and verify the stability of its operation. The formula is: ; in, for Real-time system operational stability indicators; To verify the window length; , These are the SOC inversion values ​​after fusion at the corresponding time points; The triggering conditions for anomaly repair are: ; The repair formula is: ; in, This is the stability threshold; for SOC inversion value after time-lapse repair; for Normal inversion value at any time; The repair coefficient; To verify the average rate of change of SOC within the verification window; Reset after repair To ensure continuous system operation; S54. Extract incremental time series samples based on a sliding time window. The formula for extracting samples using a sliding time window is: ; ; ; in, for A sliding time window for each moment; To adjust the sliding window size; To update the total number of samples; The number of spatial samples per time step; This represents the effective temporal feature set for each time step within the window. For incremental sample sets; This is the incremental filtering threshold; The formula for constructing the online update loss function is as follows: ; in, To continuously update the total loss value; This refers to the in-window blending loss; Weights for incremental sample loss; For regularization weights; For parameter regularization terms; , These are the model parameters to be updated; The fitting bias is calculated only for incremental samples within the sliding window; The formula for parameter optimization and update is: ; ; in, , The updated optimal parameter set; To update the learning rate online, , The initial learning rate, The attenuation coefficient; , This represents the gradient of the loss function with respect to the corresponding parameters; S55. Calculate the SOC time-series change trend and anomaly detection indicators to generate dynamic monitoring results. The formula for calculating the SOC time-series change rate is as follows: ; in, for The rate of change of SOC over time at any given moment; The time interval used to calculate the rate of change; for SOC inversion value after fusion at any given time; Anomaly detection is performed on SOC, and the calculation formula is as follows: ; ; in, for Time-based anomaly markers; This is the threshold for anomaly detection; The standard deviation of the historical SOC rate of change; S56, Integration , , and the updated model parameters Generate a time-series monitoring dataset: .

[0013] Preferably, the specific steps of step S6 are as follows: S61, Time-series monitoring dataset based on step S56 The SOC core results are aggregated according to a preset scale type to generate different levels. The formula is as follows: ; in, for Scale of time SOC value after aggregation; For scale type identification; It is a multi-scale aggregation function; for Time, coordinates SOC inversion value at the location; For scale The corresponding set of spatial ranges; It is an aggregation method; S62. Calculate the intensity of SOC change trend and integrate the uncertainty of the fusion model to provide core indicators for visualization results. The formula for calculating the intensity of SOC time-series change trend is as follows: ; in, For scale The intensity of the SOC time series change trend; The formula for calculating the overall uncertainty of the fusion model is: ; in, for Time, coordinates The uncertainty after fusion is used to generate an uncertainty layer; , for Time, coordinates Uncertainty in the mechanism model and uncertainty in the AI ​​model; S63. Based on different application scenarios, construct a scenario adaptation model, and calculate the formula as follows: ; in, Identify the application scenario; For customized results; Adapt weights to different scenarios; S64. Convert the visualization results and core data into a standard format through a format conversion function, and build a standardized API interface to support external systems in calling monitoring results and core parameters according to specified scale, time range, data format and data type; S65. Based on multi-format output data and corresponding version model parameters, establish a long-term data archiving and version management mechanism to obtain archived data with version identifiers and associated storage results of model parameters. S66. Based on the measured data from the application scenario feedback, the feedback loss is calculated and the model parameters are optimized by gradient to obtain the optimal parameters of the model after iterative updates, thereby achieving continuous system optimization. S67. Based on the output data, API response results, and archived data, standardized results are obtained.

[0014] Preferably, the standardized deliverables of step S67 include customized application results, multi-format data files, API call documentation, archived data, and iteration logs.

[0015] Therefore, the present invention, by adopting the above-described method, has the following beneficial effects: 1. By employing a refined simulation of the carbon cycle process, constructing a physical information-enhanced AI model, and combining residual coupling with Bayesian fusion, along with a continuously updated loss function and parameter gradient optimization mechanism, this approach addresses the issues of poor interpretability and insufficient generalization ability in traditional pure data-driven AI models, as well as the limitations in accuracy and fixed parameters of single-process models. It provides physical prior constraints through mechanistic features, ensures consistency of physical laws through a hybrid loss function, and integrates the complementary advantages of two models. Further optimization of parameters through application feedback not only significantly improves the accuracy of soil organic carbon inversion but also fundamentally enhances the mechanistic interpretability and long-term adaptability of the model results. This ensures that the prediction results always conform to ecological and physical laws, providing core support for the long-term reliability of monitoring data.

[0016] 2. Leveraging the efficient predictive capabilities of artificial intelligence models, combined with standardized multi-source data integration processes, the rapid computational characteristics of fusion models, and long-term time-series data reuse mechanisms, this approach solves the problems of long processing times, high costs, and low efficiency associated with traditional manual sampling and laboratory analysis, as well as the high input requirements, low operational efficiency, and difficulty in large-scale application of traditional process models. It enables rapid inversion and long-term dynamic reuse of soil organic carbon at regional and even national scales, improving efficiency by more than 100 times compared to traditional monitoring methods. At the same time, it avoids the inefficiencies of process models and eliminates the need for repeated large-scale sampling and analysis, providing technical support for the efficient implementation of large-scale, long-term monitoring tasks.

[0017] 3. A spatiotemporal co-interpolation technique is used to construct a spatiotemporally continuous multi-source dataset. Combined with a time-series dynamic inversion method, a sliding window incremental update strategy, and a system operation anomaly self-repair mechanism, this solves the problems of limited spatial coverage and difficulty in achieving time-series dynamic tracking in traditional monitoring, insufficient time-series adaptability due to fixed model parameters, and the impact of system operation anomalies on data continuity. By dynamically inverting time-by-time to capture the time-series change characteristics of SOC, iteratively optimizing model parameters with the help of a sliding window, and ensuring data integrity through anomaly self-repair, this approach not only ensures the stability and consistency of inversion results at different time points, but also effectively identifies abnormal fluctuations in the carbon pool caused by extreme climate, land use change, etc., truly realizing long-term, dynamic, and uninterrupted continuous monitoring of soil organic carbon.

[0018] 4. By standardizing data preprocessing processes, modular model structure design, standardizing API interface construction, and establishing a long-term data archiving and version management mechanism, the system solves the problems of poor adaptability of traditional monitoring methods and models, difficulty in compatibility with new data sources, difficulty in expanding monitoring areas, and fragmented and untraceable data. The system supports the access of new data sources such as remote sensing, meteorology, and soil properties, and can flexibly replace AI models or process model structures to adapt to the monitoring needs of different regions. It can quickly integrate with external systems through API interfaces, and combined with version management, it ensures data traceability and reuse. It has good versatility, portability, and scalability, providing possibilities for flexible technology expansion and long-term reuse in multiple scenarios.

[0019] 5. Relying on multi-scale result aggregation, comprehensive uncertainty quantification, scenario-customized adaptation models, and multi-format data output and API call capabilities, it solves the problems of poor scenario adaptability and difficulty in supporting long-term decision-making in multiple fields of traditional monitoring results. The generated data such as the spatial distribution, temporal variation trend, and scenario-customized results of soil organic carbon can be directly applied to fields such as agricultural soil fertility assessment and management optimization, ecosystem carbon pool accounting, and regional ecological protection planning. Combined with the application feedback closed loop to continuously optimize the output results, it provides accurate and continuous dynamic monitoring data support for the national carbon peaking and carbon neutrality goals, assists in the formulation of relevant policies, implementation effect evaluation and dynamic adjustment, and has important practical application value and strategic significance.

[0020] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0021] Figure 1 The flowchart of a model fusion-based dynamic monitoring system for soil organic carbon provided by the present invention. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the embodiments of the present invention will be further described in detail below with reference to the accompanying drawings and examples. It should be understood that the specific embodiments described herein are merely illustrative of the embodiments of the present invention and are not intended to limit the embodiments of the present invention. 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 this application. Examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout.

[0023] It should be noted that the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion, such as a process, method, system, product, or server that includes a series of steps or units, not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such process, method, product, or device.

[0024] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0025] See appendix Figure 1 A model fusion-based dynamic monitoring system for soil organic carbon includes the following steps: S1. Obtain multi-source data related to soil organic carbon monitoring, and perform standardized preprocessing such as unified formatting, spatial resampling, and missing data imputation on the multi-source data to obtain a standardized dataset. The multi-source data in step S1 includes satellite remote sensing data. Meteorological driving data Soil basic attribute data Topographic data and land use type data The information from multivariate data is analyzed to obtain the spatial coordinate system, spatial resolution, temporal resolution, and effective range of data values. The uniform formatting process in the standardization preprocessing of step S1 is as follows: The raw data is converted to a standard raster format to ensure data structure consistency. Simultaneously, the coordinate systems of different data sources are converted to the target projected coordinate system. The projection transformation formula is as follows: ; in, , The plane coordinates of the original data; , These are the transformed target projection coordinates; , , for Orientation projection transformation coefficient; , , for Orientation projection transformation coefficient; The spatial resampling process is as follows: Multi-source raster data with unified coordinates is resampled to the target spatial resolution. The resampled pixel values ​​are then calculated using bilinear interpolation, with the following formula: ; in, This represents the value of the target pixel after resampling; For the first Interpolation weights for each original neighboring cell; For the first The values ​​of the original neighboring pixels; This is the index of the original adjacent pixels; The spatiotemporal collaborative missing data imputation process is as follows: for missing values ​​in the resampled data, imputation is performed by combining spatial correlation and temporal continuity, using the following formula: ; in, The interpolated complete data value; These are the spatiotemporal interpolation weighting coefficients, and , For the spatial correlation strength of data, The strength of the temporal correlation of the data; This is a spatial interpolation estimate; These are time interpolation estimates; Output standardized dataset It contains multi-source data with a unified format, unified coordinates, unified resolution, and no missing data, which can be directly used for carbon cycle process simulation in the subsequent mechanism model simulation module and feature input and model training in the artificial intelligence prediction module.

[0026] S2. Based on the standardized dataset output from S1, the physical-ecological processes of soil carbon input, decomposition, and transformation are quantified using a carbon cycle process model, generating mechanistic features that combine mechanistic and spatiotemporal correlations to provide physical constraints for subsequent AI models; the specific steps of step S2 are as follows: S21. Standardized dataset based on step S1 The parameters required by the model are screened and extracted to form a set of model input parameters, including vegetation-related parameters, meteorological driving parameters, soil basic parameters, land use parameters, and spatial auxiliary parameters. ; Vegetation-related parameters include: Normalized Difference Vegetation Index (NDVI). Leaf area index Meteorological driving parameters: daily average temperature Daily precipitation Soil basic parameters: soil texture, soil bulk density, soil moisture, land use type code; Spatial auxiliary parameters: slope, altitude; S22. Based on the vegetation-related parameters and meteorological driving parameters input in step S21, simulate the vegetation residue carbon input process: Based on the vegetation growth status and meteorological conditions, first estimate the aboveground biomass of vegetation, then calculate the vegetation residue carbon input. The formula for calculating aboveground biomass of vegetation is: ; in, Aboveground biomass of vegetation; The biomass conversion coefficient is determined based on land use type; for example, 2.8 × 10⁻⁶ is used for arable land. 3 5.6 × 10⁻⁶ mm of forest land 3 , used to convert vegetation index signals into actual biomass values; Leaf area index, reflecting the photosynthetic area of ​​vegetation; Normalized Difference Vegetation Index; The average daily temperature; The optimal temperature for vegetation growth is used to correct the inhibitory effect of extreme temperatures on biomass accumulation. Daily precipitation; The optimal precipitation for vegetation growth is determined based on the regional climate type, such as 80 mm / month in the temperate monsoon region, to correct for the impact of water stress on biomass. It is a hyperbolic tangent function, used to limit the impact of precipitation to the 0-1 range, so as to avoid the unreasonable amplification of biomass estimation by excessive precipitation; The formula for calculating the carbon input of vegetation residue is: ; in, Carbon input from vegetation residue; The proportion of vegetation residue is determined according to the vegetation type, such as 0.35 for herbaceous plants and 0.65 for woody plants. The harvest coefficient is 0.6 for cultivated land and 0.05 for forest / grassland. The vegetation carbon content coefficient, and ; S23. Based on the soil basic parameters and land use parameters input in step S21, simulate the soil carbon pool decomposition process: divide soil organic carbon into active carbon pool, slow-release carbon pool, and inert carbon pool; first calculate the environmental regulation factor, then determine the decomposition rate of each carbon pool, and finally calculate the decomposition flux of each carbon pool. The formula for calculating the temperature regulation factor is as follows: ; in, Temperature regulation factor; This is the temperature sensitivity coefficient, an empirical parameter with a value of 0.08, used to adjust the degree of attenuation of the decomposition rate when the temperature deviates from the optimal value; Soil temperature; This is the optimal temperature for carbon decomposition, an empirical parameter, and is set to 20℃. The formula for calculating the humidity regulation factor is: ; in, Humidity regulation factor; This refers to the soil volumetric water content. The optimal soil moisture content for carbon decomposition is an empirical parameter, with a value of 0.3. The formulas for calculating the decomposition rate of each carbon pool are as follows: ; in, , , The actual decomposition rates of the active, slow-release, and inert carbon libraries are listed in order. , , The baseline decomposition rates for the active, slow-release, and inert carbon libraries are 0.8, 0.05, and 0.001, respectively. Soil texture regulator; The formulas for calculating the decomposition flux of each carbon pool are as follows: ; in, , , The decomposition fluxes of the active, slow-release, and inert carbon libraries are listed in order. , , The initial reserves of the active, slow-release, and inert carbon pools, in that order, are obtained through carbon pool allocation coefficient conversion and form the basis for decomposition flux calculation. S24. Based on the initial carbon pool storage and the inter-pool transformation coefficient, calculate the inter-pool transformation flux to simulate the soil carbon pool transformation process. The calculation formula is as follows: ; in, To obtain carbon from the pool arrive The conversion flux; For carbon pool arrive The conversion coefficient is 0.15 when active to slow-acting, 0.02 when slow-acting to active, and 0.005 when slow-acting to inert. For source carbon pool Reserves; , Corresponding to active, slow-release, and inert carbon libraries respectively. ; S25. Input the vegetation residue carbon amount from step S22. Decomposition flux of the active, slow-acting, and inert carbon libraries in step S23 , , The conversion flux in step S24 And the initial reserves of active, slow-release, and inert carbon libraries. , , and temperature regulation factor Humidity regulating factor Integrating to form a set of mechanistic features .

[0027] S3. Based on the preprocessed multi-source data from step S1 and the mechanistic characteristics from step S2, construct an artificial intelligence prediction model to learn the mapping relationship between soil organic carbon and multi-source inputs, thereby achieving rapid inversion of soil organic carbon. The specific steps of step S3 are as follows: S31. Standardize the dataset from step S1. The set of mechanistic features of step S25 The model input feature set is fused and constructed using the following formula: ; in, This is the initial input feature set after fusion; This is a feature concatenation function used to concatenate standardized multi-source data with mechanistic features column-wise, achieving complementary data dimensions. S32. The initial input feature set from step S31 Redundancy removal and scaling are performed to improve model training efficiency and accuracy. The feature selection formula is as follows: ; in, For the first Mutual information entropy between a feature and the measured value of soil organic carbon; For the initial feature set The One feature; This represents the measured value of soil organic carbon. A threshold for mutual information entropy is selected and determined through cross-validation. This is used to eliminate redundant features that are weakly associated with the target variable; The selected effective feature set; The feature standardization formula is: ; in, For the first The standardized values ​​of each effective feature are used to eliminate the impact of feature scale differences on model training. For effective feature set The first in One feature; For the first The mean of each effective feature is used to center the feature data; For the first The standard deviation of each effective feature is used to normalize the feature data; The resulting standardized feature set is ; S33, Standardized feature set based on step S32 A physical information-enhanced artificial intelligence prediction model is constructed to enable it to learn the mapping relationship between soil organic carbon and input features. The model expression is as follows: ; in, The soil organic carbon content predicted by the model; For artificial intelligence prediction models; For the set of model parameters; The model The network structure includes: Input layer: Number of neurons is Hidden layers: 3 layers, each with 64 neurons, using ReLU activation function; Output layer: 1 neuron, output... ; S34, Standardized feature set based on step S32 and the corresponding measured values ​​of soil organic carbon Train the model and optimize the parameters by minimizing the loss function. The mixture loss function is defined and calculated using the following formula: ; in, This is a mixed loss value used to balance model fitting accuracy and physical consistency; To characterize the fitting deviation between the predicted and measured values, ; For carbon conservation constraint loss, ; The constraint loss weight coefficient is determined through cross-validation and is set to 0.3. It is used to balance the contribution ratio of fitting loss and constraint loss. For the first SOC prediction value for each sample; For the first Measured SOC values ​​for each sample; The time-series rate of change of the SOC predicted value; This represents the sum of the decomposition fluxes of all carbon pools. The formula for calculating the optimized model parameters is: ; in, This is the optimized set of model parameters. For parameter optimization operators, used to find the loss function The minimum parameter value; The Adam optimizer is used to solve for the optimal parameters. The optimization formula is as follows: ; in, For the first Model parameters for the next iteration This represents the number of iterations. The learning rate, with a value of 0.001, is used to control the parameter update step size in each iteration. For the first In the next iteration, the loss function affects the parameters. The gradient is used to guide the direction of parameter updates; S35. Perform performance verification on the optimized model, and achieve rapid SOC inversion based on the validation set. The performance verification index formula is: ; in, The root mean square error is used to quantify the accuracy of model predictions; the smaller the value, the higher the accuracy. The number of samples in the validation set; For the verification set SOC prediction value for each sample; For the verification set Measured SOC values ​​for each sample; Standardized multi-source data of the region to be inverted and corresponding mechanism characteristics Input values ​​in steps S1-S2 to obtain Input the optimized model The output is the fast inversion result, which can be substituted into the formula as follows: .

[0028] in, SOC prediction for the region to be inverted; This is the standardized effective feature set of the region to be inverted.

[0029] S4. Employ one or more strategies among residual coupling, feature coupling, loss function constraints, and Bayesian fusion to fuse the carbon cycle process model and the artificial intelligence prediction model, so that the fused model possesses mechanistic interpretability, high accuracy, and high efficiency; the specific steps of step S4 are as follows: S41. Extract the core outputs and auxiliary parameters of the carbon cycle process model and the artificial intelligence prediction model to form a fused basic dataset; specifically, this includes: the SOC estimate from the carbon cycle process model. Standardized feature set in step S32 Step S33 AI model SOC prediction value Optimal model parameters in step S34 Carbon conservation constraint parameters Uncertainty in mechanism models and the uncertainty of AI models .

[0030] S42. By fusing feature dimensions and correcting residuals, the mechanism-data synergy between the carbon cycle process model and the artificial intelligence prediction model is enhanced. The feature coupling formula is as follows: ; in, For coupling feature set; For feature concatenation function; Output features for the intermediate layer of the AI ​​model; The residual coupling formula is: ; ; in, For the residuals of the mechanism model; This is the residual correction amount, used to compensate for biases in the mechanism model; This is a residual correction AI model used to learn the mapping relationship between coupled features and residuals; These are the parameters of the residual sub-model; S43. Embedding multiple constraints balances fitting accuracy, mechanistic interpretability, and residual correction effect to construct a fusion loss function, calculated as follows: ; in, The total loss value of the fusion model is used for optimization. and This ensures that the fusion model has both accuracy and consistency in mechanism; To fuse the fitting loss, and ; For residual constraint loss, and ; This is the residual constraint weight, with a value of 0.2, used to balance the contribution of residual correction. , , , They are respectively the number Mechanism estimate, residual correction, measured value, and mechanism residual for each sample; S44. Based on dynamic weight allocation due to model uncertainty, the Bayesian weight calculation formula is as follows: ; ; in, Bayesian weights for the mechanistic model; For the Bayesian weights of the AI ​​model; Uncertainty in the mechanism model; For the uncertainty of AI models; The fusion calculation formula is as follows: ; in, This is the estimated SOC value after final fusion; This is the output of the mechanistic model after residual correction, used to improve the accuracy of the mechanistic model.

[0031] S45. The root mean square error and coefficient of determination are used to verify the performance of the fusion model, as shown in the following formula: ; ; in, The root mean square error of the fusion model is used to quantify the prediction accuracy; the smaller the value, the higher the accuracy. The coefficient of determination for the fusion model; For the verification set Estimated SOC of each sample fusion; For the verification set Measured SOC values ​​of fusion for each sample; To verify the mean of the measured values ​​of the SOC set.

[0032] S5. Based on time-series meteorological and remote sensing data, through dynamic inversion, operational status monitoring and anomaly self-repair, and sliding window incremental updates, long-term continuous monitoring of soil organic carbon and continuous optimization of model parameters are achieved; the specific steps of step S5 are as follows: S51. Based on time-series meteorological data and remote sensing data, and combined with the preceding standardization process, construct a spatiotemporally continuous time-series input dataset, as shown in the formula: ; ; in, This is a time-series standardized dataset; This is a time-series effective feature set; This is a time-series feature processing function; for A set of mechanistic characteristics at any given moment; The formula for initializing the long-term operating status parameters of the system is: ; in, This serves as a status indicator (1 = normal, 0 = abnormal). This is the initial runtime; The current time; This is the threshold for abnormal repair. S52. Based on the time-series effective feature set obtained in step S51 and the fusion model obtained in step S4, the SOC dynamic inversion is realized time-by-time, and the formula is as follows: ; in, for The dynamic inversion value of SOC after fusion at each moment; , for Bayesian weights at each time step; for The estimated SOC value of the mechanism model at time step; for The residual correction amount at time step; for AI model SOC prediction value at time; S53. Monitor the system's operating status in real time and implement self-repair for anomalies, and verify the stability of its operation. The formula is: ; in, for The system's operational stability index reflects the degree of fluctuation in recent SOC inversion values; the smaller the value, the more stable the system. To verify the window length; , These are the fused SOC inversion values ​​at the corresponding time points; The triggering conditions for anomaly repair are: ; The repair formula is: ; in, The stability threshold is defined as 0.05 to 0.15. for SOC inversion value after time-lapse repair; for Normal inversion value at any time; This is the repair coefficient, with a value ranging from 0.6 to 0.9; To verify the average rate of change of SOC within the verification window, and ; When stability index Exceeding the threshold When an abnormal system operation is detected, the current inversion value is corrected using a repair formula based on the previous normal value and the historical average rate of change. The system is then updated after the correction. Reset the stability metric count to ensure continuous and normal system operation; reset after repair. To ensure continuous system operation; S54. Extract incremental time series samples based on a sliding time window. The formula for extracting samples using a sliding time window is: ; ; ; in, for A sliding time window for each moment; To adjust the sliding window size; To update the total number of samples; The number of spatial samples per time step; This represents the effective temporal feature set for each time step within the window. For incremental sample sets; This is the incremental filtering threshold; The formula for constructing the online update loss function is as follows: ; in, To continuously update the total loss value; For in-window blending loss, and , The fitting loss is within the window. For carbon conservation constraint loss, This is the residual constraint loss; Weights for incremental sample loss; For regularization weights; For parameter regularization terms; , These are the model parameters to be updated; To calculate the fitting bias only for incremental samples within the sliding window, and , This represents the incremental sample size. , The order is number 1 The fusion inversion value and measured value of each incremental sample; The formula for parameter optimization and update is: ; ; in, , The updated optimal parameter set; To update the learning rate online, , The initial learning rate, The attenuation coefficient is 0.95. , This represents the gradient of the loss function with respect to the corresponding parameters; S55. Calculate the SOC time-series change trend and anomaly detection indicators to generate dynamic monitoring results. The formula for calculating the SOC time-series change rate is as follows: ; in, for The rate of change of SOC over time at any given moment; The time interval used to calculate the rate of change; for SOC inversion value after fusion at any given time; Anomaly detection is performed on SOC, and the calculation formula is as follows: ; ; in, for Anomaly markers at specific times are used to identify SOC abrupt events, such as anomalous changes caused by extreme weather or abrupt land use changes. This is the threshold for anomaly detection; The standard deviation of the historical SOC rate of change; S56, Integration , , and the updated model parameters Generate a time-series monitoring dataset: .

[0033] S6. Based on the dataset generated by the time-series monitoring in step S5, through multi-scale aggregation, customized adaptation to application scenarios, multi-format output and API calls, data archiving and feedback iteration, standardized results supporting long-term applications in multiple fields and continuous system optimization are generated; the specific steps of step S6 are as follows: S61, Time-series monitoring dataset based on step S56 The SOC core results are aggregated according to a preset scale type to generate different levels. The formula is as follows: ; in, for Scale of time SOC value after aggregation; These are scale type identifiers, namely grid scale, administrative unit scale, and ecological zone scale; It is a multi-scale aggregation function used to summarize and calculate the SOC inversion values ​​within a specified spatial range; for Time, coordinates SOC inversion value at the location; For scale The corresponding set of spatial ranges; The aggregation methods include aggregation of mean, median, maximum, and minimum values; S62. Calculate the intensity of SOC change trend and integrate the uncertainty of the fusion model to provide core indicators for visualization results. The formula for calculating the intensity of SOC time-series change trend is as follows: ; in, For scale The intensity of the SOC time series change trend under this scale represents the annual / monthly average change of SOC, which is used to generate a change trend map; The formula for calculating the overall uncertainty of the fusion model is: ; in, for Time, coordinates The uncertainty after fusion represents the reliability of the SOC inversion result and is used to generate the uncertainty layer; , for Time, coordinates Uncertainty in the mechanism model and uncertainty in the AI ​​model; S63. Based on different application scenarios, construct a scenario adaptation model, and calculate the formula as follows: ; in, Identify application scenarios, including agricultural management, ecological protection, and carbon inventory accounting; For customized results; Adapt weights to different scenarios, and , For scale With Scene The degree of matching; The total number of scales participating in the adaptation; S64. Convert the visualization results and core data to a standard format using a format conversion function. The conversion formula is as follows: ; ; in, The converted data is in a standard format and supports data applications in multiple scenarios, such as GIS analysis, academic research, and decision-making system access. This is a data format conversion function that adapts the core result to the target format. The target data format types cover the export requirements for spatial data, time-series data, and vector data. Build a standardized API interface to support external systems in calling monitoring results and core parameters. The formula is as follows: ; in, To enable external sharing and integration of monitoring results through API interface response; The API response function parses the request parameters and returns the corresponding data; Define the API request parameter set and clarify the core requirements of external calls; For time range parameters; For data format type; Specify the data request type, including SOC spatial distribution, changing trends, uncertainty, and full data, and clarify the specific data type to be called; S65. Establish a long-term data archiving and version management mechanism for multi-format output data and corresponding version model parameters to obtain archived data with version identifiers and associated storage results of model parameters; wherein, the archiving formula is: ; in, for Version, scale , Archived data at any given time; For scale , time, Formatted output data; for The set of model parameters corresponding to the version; This is a data and parameter association operator that enables the binding and storage of archived data with corresponding model parameters; The version update formula is: ; in, This is the latest historical version number; archiving is triggered when the archiving period is the initial value. S66. Based on the measured data from application scenario feedback, the optimal parameters of the model are obtained by calculating the feedback loss and performing gradient optimization on the model parameters to achieve continuous system optimization. The formula for calculating the feedback loss is as follows: ; The parameter optimization formula is: ; in, The number of feedback samples for the application; For the first Customized output values ​​for each feedback sample in a given scenario; For the first The actual measured values ​​of the feedback sample; This serves as the optimal set of parameters for the optimized AI model. The learning rate is used for feedback, and its value ranges from 0.0001 to 0.001. For the feedback loss function pair The gradient; S67. Perform integrity and consistency checks on the output data, API response results, and archived data, and finally output standardized results including customized application results, multi-format data files, API call documents, archived data, and iteration logs.

[0034] Example: Taking an agricultural area in the Yellow River Basin as the monitoring area, focusing on three core ecosystems: agriculture, grassland, and forest, the monitoring period is divided into two parts: historical dynamic monitoring (1990-2024) and future scenario prediction (2025-2050). The specific implementation is as follows: S1: Obtain remote sensing data, meteorological data, soil survey data, and land use data for this region from 1990 to 2024. After coordinate unification, 1km resolution resampling, and spatiotemporal collaborative interpolation, generate a standardized multi-source dataset. S2: Using a carbon cycle process model, parameters were customized for different ecosystems to obtain the annual vegetation residue carbon input for three types of ecosystems from 1990 to 2024, with an average annual input of 320 gC / hm² in agricultural areas. 2 Grassland averages 280 gC / hm² per year 2 Forests average 410 gC / hm² per year 2 The three carbon pools decompose fluxes and integrate to generate a set of mechanistic features that combine mechanistic and ecosystem-specific characteristics; S3: Combine the standardized data from S1 with the mechanistic features from S2 to construct the input feature set, and then train it to obtain the AI ​​prediction model; S4: Model fusion is achieved through feature coupling, residual correction, and Bayesian weights; S5: Historical Dynamic Monitoring: The input dataset is constructed based on annual time-series data, and dynamic inversion is performed annually. A sliding window with τ=5 years is used to update the model parameters. Data continuity is ensured through operational status monitoring and anomaly self-repair mechanism. Key trend anomalies are identified: from 1990 to 2000, excessive fertilization in agricultural areas led to a slow accumulation of SOC (average annual increase of 0.35%), and from 2000 to 2010, overgrazing in grasslands led to a slow decline in SOC (average annual decrease of 0.21%). S6: ① Historical Results Output: Generate SOC spatial distribution and long-term trend maps from 1990 to 2024 at three scales: 1km grid, municipal administrative units, and ecological zones, quantifying and integrating uncertainty; ② Future Scenario Output: For climate scenarios from 2026 to 2050 and three types of intervention schemes, generate SOC change trend prediction maps and carbon sequestration potential assessment reports, and customize three types of scenario-based results: "Agricultural Carbon Sequestration Optimization Scheme," "Grassland Ecological Restoration Suggestions," and "Forest Carbon Sequestration Enhancement Strategy"; ③ Data and Interfaces: Output spatial data, time-series data, and scenario prediction data in multiple formats, build a standardized API interface, and support calls from the National Carbon Accounting Platform and regional ecological management systems; ④ Archiving and Iteration: Establish a historical data version library from 1990 to 2024, and combine newly added measured data from 2024 to 2025 to iterate model parameters and improve future prediction accuracy.

[0035] In summary, this invention systematically addresses the shortcomings of traditional soil organic carbon monitoring by integrating core technologies such as spatiotemporal collaborative preprocessing of multi-source data, refined simulation of carbon cycle mechanisms, physical information-enhanced AI modeling, residual coupling-Bayesian fusion, time-series dynamic inversion, sliding window incremental updates, system operation anomaly self-repair, application feedback closed-loop iteration, long-term data archiving and version management, and multi-scale-multi-scenario customized output. These include low efficiency and limited coverage of traditional soil organic carbon monitoring, poor interpretability and insufficient generalization of pure AI models, inefficient operation of process models and difficulty in adapting to large-scale applications, difficulty in multi-source data fusion and lack of dynamic monitoring capabilities, and system operation instability. Key issues such as insufficient continuity, lack of data traceability, and weak long-term decision support across multiple scenarios were addressed. Ultimately, soil organic carbon monitoring achieved a synergistic improvement in accuracy, efficiency, mechanistic interpretability, dynamic continuity, and scenario adaptability. It also possesses strong scalability, large-scale adaptability, and multi-ecosystem coverage capabilities. It can provide accurate and reliable historical dynamic monitoring data for agricultural soil management, ecological protection, and carbon pool accounting, and can also provide scientific support for the formulation of differentiated carbon sequestration strategies through future scenario prediction. It injects practical momentum into the advancement of regional and national carbon peaking and carbon neutrality goals, and has both significant technological innovation value and far-reaching practical application significance.

[0036] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A soil organic carbon dynamic monitoring system based on model fusion, characterized in that: Includes the following steps: S1. Obtain multi-source data related to soil organic carbon monitoring, and perform standardized preprocessing such as unified formatting, spatial resampling, and missing data imputation on the multi-source data to obtain a standardized dataset. S2, based on the standardized dataset output by S1, quantifies the physical-ecological processes of soil carbon input, decomposition and transformation through a carbon cycle process model, generating mechanistic features that combine mechanistic and spatiotemporal correlation. S3. Based on the preprocessed multi-source data in step S1 and the mechanistic characteristics in step S2, construct an artificial intelligence prediction model to learn the mapping relationship between soil organic carbon and multi-source inputs, so as to achieve rapid inversion of soil organic carbon. S4. The carbon cycle process model and the artificial intelligence prediction model in step S3 are fused using a residual coupling and Bayesian fusion strategy to obtain a fusion model that combines mechanistic interpretability, high accuracy and high efficiency. S5. Based on time-series meteorological and remote sensing data, through dynamic inversion, operational status monitoring and anomaly self-repair, and sliding window incremental update, long-term continuous monitoring of soil organic carbon and continuous optimization of model parameters are achieved. S6. Based on the dataset generated by the time-series monitoring in step S5, through multi-scale aggregation, customized adaptation for application scenarios, multi-format output and API calls, data archiving and feedback iteration, standardized results supporting long-term applications in multiple fields and continuous system optimization are generated.

2. The soil organic carbon dynamic monitoring system based on model fusion according to claim 1, characterized in that: The multi-source data in step S1 includes satellite remote sensing data. Meteorological driving data Soil basic attribute data Topographic data and land use type data The information from multivariate data is analyzed to obtain the spatial coordinate system, spatial resolution, temporal resolution, and effective range of data values.

3. The soil organic carbon dynamic monitoring system based on model fusion according to claim 2, characterized in that: The uniform formatting process in the standardization preprocessing of step S1 is as follows: All raw data are uniformly converted into a standard raster format, and the coordinate systems of different data sources are uniformly converted into the target projected coordinate system; Bilinear interpolation is used to resample multi-source raster data with unified coordinates to the target spatial resolution, thereby achieving resolution normalization. For missing values ​​in the resampled data, spatiotemporal co-imputation is performed by combining spatial correlation and temporal continuity, as shown in the following formula: ； in, This is the complete data value after interpolation; These are the spatiotemporal interpolation weighting coefficients, and , For the spatial correlation strength of data, The strength of the temporal correlation of the data; This is a spatial interpolation estimate; These are time interpolation estimates; Output standardized dataset .

4. The soil organic carbon dynamic monitoring system based on model fusion according to claim 3, characterized in that: The specific steps of step S2 are as follows: S21. Standardized dataset based on step S1 The parameters required by the model are screened and extracted to form a set of model input parameters, including vegetation-related parameters, meteorological driving parameters, soil basic parameters, land use parameters, and spatial auxiliary parameters. ; S22. Based on the vegetation-related parameters and meteorological driving parameters input in step S21, simulate the vegetation residue carbon input process: Based on the vegetation growth status and meteorological conditions, first estimate the aboveground biomass of vegetation, then calculate the vegetation residue carbon input. The formula for calculating aboveground biomass of vegetation is: ； in, Aboveground biomass of vegetation; Biomass conversion coefficient; Leaf area index; Normalized Difference Vegetation Index; The average daily temperature; This is the optimal temperature for vegetation growth; Daily precipitation; The optimal rainfall for vegetation growth; It is the hyperbolic tangent function; The formula for calculating the carbon input of vegetation residue is: ； in, Carbon input from vegetation residue; The proportion of vegetation residue; Harvest coefficient; The vegetation carbon content coefficient; S23. Based on the soil basic parameters and land use parameters input in step S21, simulate the soil carbon pool decomposition process: divide soil organic carbon into active carbon pool, slow-release carbon pool, and inert carbon pool; first calculate the environmental regulation factor, then determine the decomposition rate of each carbon pool, and finally calculate the decomposition flux of each carbon pool. The formula for calculating the temperature regulation factor is as follows: ； in, Temperature regulation factor; This refers to the temperature sensitivity coefficient. Soil temperature; This is the optimal temperature for carbon decomposition. The formula for calculating the humidity regulation factor is: ； in, Humidity regulation factor; This refers to the soil volumetric water content. The optimal soil moisture content for carbon decomposition; The formulas for calculating the decomposition rate of each carbon pool are as follows: ； in, , , The actual decomposition rates of the active, slow-release, and inert carbon libraries are listed in order. , , The basic decomposition rates of the active, slow-release, and inert carbon libraries are listed in order. Soil texture regulator; The formulas for calculating the decomposition flux of each carbon pool are as follows: ； in, , , The decomposition fluxes of the active, slow-release, and inert carbon libraries are listed in order. , , The initial reserves of the active, slow-release, and inert carbon libraries are listed in that order. S24. Based on the initial carbon pool storage and the inter-pool transformation coefficient, calculate the inter-pool transformation flux to simulate the soil carbon pool transformation process. The calculation formula is as follows: ； in, To obtain carbon from the pool arrive The conversion flux; For carbon pool arrive The conversion coefficient; For source carbon pool Reserves; , Corresponding to active, slow-release, and inert carbon libraries respectively. ; S25. Integrate the simulation results of carbon input, decomposition flux, conversion flux, and carbon pool storage to form a set of mechanistic characteristics. .

5. The soil organic carbon dynamic monitoring system based on model fusion according to claim 4, characterized in that: The specific steps of step S3 are as follows: S31. Standardize the dataset from step S1. The set of mechanistic features of step S25 Perform feature concatenation to construct the initial input feature set; S32. The mutual information entropy method is used to remove redundancy from the initial input feature set, and the effective feature set is obtained by screening. Then, the feature scale is unified through standardization to obtain the processed feature set. ; S33, Standardized feature set based on step S32 A physical information-enhanced artificial intelligence prediction model is constructed to enable it to learn the mapping relationship between soil organic carbon and input features. The model expression is as follows: ； in, The soil organic carbon content predicted by the model; For artificial intelligence prediction models; For the set of model parameters; S34, Standardized feature set based on step S32 and the corresponding measured values ​​of soil organic carbon Train the model and optimize the parameters by minimizing the loss function. The mixture loss function is defined and calculated using the following formula: ； in, This is a mixed loss value; The fitting deviation between the predicted and measured values. ; For carbon conservation constraint loss, ; To constrain the loss weighting coefficients; For the first SOC prediction value for each sample; For the first Measured SOC values ​​for each sample; The time-series rate of change of the SOC predicted value; This represents the sum of the decomposition fluxes of all carbon pools. The formula for calculating the optimized model parameters is: ； in, This is the optimized set of model parameters. Operator for parameter optimization; The Adam optimizer is used to solve for the optimal parameters. The optimization formula is as follows: ； in, For the first Model parameters for the next iteration This represents the number of iterations. The learning rate; For the first In the next iteration, the loss function affects the parameters. The gradient; S35. The optimized model is validated using metrics such as root mean square error.

6. The soil organic carbon dynamic monitoring system based on model fusion according to claim 5, characterized in that: The specific steps of step S4 are as follows: S41. Extract the core outputs and auxiliary parameters of the carbon cycle process model and the artificial intelligence prediction model to form a fusion basic dataset; S42. By fusing feature dimensions and correcting residuals, the mechanism-data synergy between the carbon cycle process model and the artificial intelligence prediction model is enhanced. The feature coupling formula is as follows: ； in, For coupling feature set; For feature concatenation function; Output features for the intermediate layer of the AI ​​model; The residual coupling formula is: ； ； in, For the residuals of the mechanism model; This is the residual correction amount; This is a residual correction AI model used to learn the mapping relationship between coupled features and residuals; These are the parameters of the residual sub-model; S43. Embedding multiple constraints balances fitting accuracy, mechanistic interpretability, and residual correction effect to construct a fusion loss function, calculated as follows: ； in, This represents the total loss value of the fusion model; To fuse the fitting loss, and ; For residual constraint loss, and ; These are the residual constraint weights; , , , They are respectively the number Mechanism estimate, residual correction, measured value, and mechanism residual for each sample; S44. Based on dynamic weight allocation due to model uncertainty, the Bayesian weight calculation formula is as follows: ； ； in, Bayesian weights for the mechanistic model; For the Bayesian weights of the AI ​​model; Uncertainty in the mechanism model; For the uncertainty of AI models; The fusion calculation formula is as follows: ； in, This is the estimated SOC value after final fusion; This is the output of the mechanism model after residual correction.

7. The soil organic carbon dynamic monitoring system based on model fusion according to claim 6, characterized in that: Step S4 includes step S45, which specifically involves verifying the performance of the fusion model using root mean square error and coefficient of determination.

8. The soil organic carbon dynamic monitoring system based on model fusion according to claim 7, characterized in that: The specific steps of step S5 are as follows: S51. Based on time-series meteorological data and remote sensing data, and combined with the standardized process in step S1, construct a spatiotemporally continuous time-series input dataset, and obtain the effective time-series feature set through a time-series feature processing function. And initialize the system's long-term operating status parameters, using the following formula: ； in, This is a running status indicator (1 = normal, 0 = abnormal). This is the initial runtime; The current time; This is the threshold for abnormal repair. S52. Based on the time-series effective feature set obtained in step S51 and the fusion model obtained in step S4, the SOC dynamic inversion is realized time-by-time, and the formula is as follows: ； in, for The dynamic inversion value of SOC after fusion at each moment; , for Bayesian weights at each time step; for The estimated SOC value of the mechanism model at time step; for The residual correction amount at time step; for AI model SOC prediction value at time; S53. Monitor the system's operating status in real time and implement self-repair for anomalies, and verify the stability of its operation. The formula is: ； in, for Real-time system operational stability indicators; To verify the window length; , These are the SOC inversion values ​​after fusion at the corresponding time points; The triggering conditions for anomaly repair are: ; The repair formula is: ; in, This is the stability threshold; for SOC inversion value after time-lapse repair; for Normal inversion value at any time; The repair coefficient; To verify the average rate of change of SOC within the verification window; Reset after repair To ensure continuous system operation; S54. Extract incremental time series samples based on a sliding time window. The formula for extracting samples using a sliding time window is: ； ； ； in, for A sliding time window for each moment; To adjust the sliding window size; To update the total number of samples; The number of spatial samples per time step; This represents the effective temporal feature set for each time step within the window. For incremental sample sets; This is the incremental filtering threshold; The formula for constructing the online update loss function is as follows: ； in, To continuously update the total loss value; This refers to the in-window blending loss; Weights for incremental sample loss; For regularization weights; For parameter regularization terms; , These are the model parameters to be updated; The fitting bias is calculated only for incremental samples within the sliding window; The formula for parameter optimization and update is: ； ； in, , The updated optimal parameter set; To update the learning rate online, , The initial learning rate, The attenuation coefficient; , This represents the gradient of the loss function with respect to the corresponding parameters; S55. Calculate the SOC time-series change trend and anomaly detection indicators to generate dynamic monitoring results. The formula for calculating the SOC time-series change rate is: ； in, for The rate of change of SOC over time at any given moment; The time interval used to calculate the rate of change; for SOC inversion value after fusion at any given time; Anomaly detection is performed on the SOC, and the calculation formula is as follows: ； ； in, for Time-based anomaly markers; This is the threshold for anomaly detection; The standard deviation of the historical SOC rate of change; S56, Integration , , and the updated model parameters Generate a time-series monitoring dataset: 。 9. A soil organic carbon dynamic monitoring system based on model fusion according to claim 8, characterized in that: The specific steps of step S6 are as follows: S61, Time-series monitoring dataset based on step S56 The SOC core results are aggregated according to a preset scale type to generate different levels. The formula is as follows: ； in, for Scale of time SOC value after aggregation; For scale type identification; It is a multi-scale aggregation function; for Time, coordinates SOC inversion value at the location; For scale The corresponding set of spatial ranges; It is an aggregation method; S62. Calculate the intensity of SOC change trend and integrate the uncertainty of the fusion model to provide core indicators for visualization results. The formula for calculating the intensity of SOC time-series change trend is as follows: ； in, For scale The intensity of the SOC time series change trend; The formula for calculating the overall uncertainty of the fusion model is: ； in, for Time, coordinates The uncertainty after fusion is used to generate an uncertainty layer; , for Time, coordinates Uncertainty in the mechanism model and uncertainty in the AI ​​model; S63. Based on different application scenarios, construct a scenario adaptation model, and calculate the formula as follows: ； in, Identify the application scenario; For customized results; Adapt weights to different scenarios; S64. Convert the visualization results and core data into a standard format through a format conversion function, and build a standardized API interface to support external systems in calling monitoring results and core parameters according to specified scale, time range, data format and data type; S65. Based on multi-format output data and corresponding version model parameters, establish a long-term data archiving and version management mechanism to obtain archived data with version identifiers and associated storage results of model parameters. S66. Based on the measured data from the application scenario feedback, the feedback loss is calculated and the model parameters are optimized by gradient to obtain the optimal parameters of the model after iterative updates, thereby achieving continuous system optimization. S67. Based on the output data, API response results, and archived data, standardized results are obtained.

10. A soil organic carbon dynamic monitoring system based on model fusion according to claim 9, characterized in that: The standardized deliverables of step S67 include customized application results, multi-format data files, API call documentation, archived data, and iteration logs.

Images