Wetland carbon reserve inversion and contribution analysis method considering vegetation type

Abstract

The invention discloses a vegetation type-considered wetland carbon reserve inversion and contribution analysis method, which comprises the following steps of: (1) obtaining a spatial mode of a wetland dominant vegetation type according to multi-source remote sensing data and ground observation plant community sample data; (2) constructing a multi-model collaborative learning framework on the spatial mode of the wetland advantageous vegetation type obtained in the step (1); (3) analyzing the carbon reserve difference of different vegetation types on the basis of the spatial pattern of the carbon reserve of each advantageous vegetation type of the wetland predicted in the step (2) so as to evaluate the influence value of the vegetation types on the soil carbon reserve; and (4) preferentially selecting characteristic factors from the key environment factors identified in the step (3) to obtain direct and indirect influence of each characteristic on carbon reserve change. The method has the following beneficial effects that a multi-model collaborative learning framework integrated with the advantages of various machine learning is combined with a data-driven regression model, and scientific support is provided for fine management and ecological protection strategies of the wetland.

Description

Technical Field

[0001] The present invention relates to a wetland carbon storage inversion and an analysis method of its influencing factors, in particular to a technical solution based on multi-model collaborative learning, which is used for wetland vegetation type classification, regression prediction and SHAP contribution analysis, and belongs to the field of geographic information system development technology. Background Art

[0002] Wetlands have a high carbon storage capacity and are one of the types of terrestrial ecosystems with the richest carbon storage per unit area. Their anoxic soil environment effectively reduces the decomposition rate of organic matter and promotes the long-term accumulation of organic carbon. However, the degradation and loss of wetlands worldwide is posing a serious threat to ecosystem stability, leading to increased carbon emissions, reduced biodiversity, and weakening the carbon storage function of wetlands. At present, research on wetland carbon storage estimation is still in its early development stage. Affected by multiple factors such as vegetation type, hydrological conditions, climate change, and soil physical and chemical properties, the wetland carbon cycle process is complex. The accurate measurement of carbon storage and the analysis of the contribution of its key influencing factors still face many challenges.

[0003] Current research on wetland ecosystems focuses on the classification of wetland vegetation types. Machine learning models are trained by inputting relevant characteristic variables such as spectral bands, vegetation indices, and texture features. The focus is on classification accuracy, specifically, the comparison of the accuracy of different models. In the context of the "Dual Carbon Strategy," while wetland vegetation classification is a crucial foundation for wetland ecological monitoring, estimating wetland carbon storage based on this foundation offers a more intuitive way to quantify wetland carbon sequestration. Some researchers, in their study of wetland carbon storage, tend to treat wetlands as a whole, estimating carbon storage using a series of climate and biological process simulation models. While this approach infers carbon storage and offers clear theoretical insights, it is only applicable to large-scale wetland carbon storage estimates and suffers from data scarcity. Other researchers construct linear or multivariate regression models based on vegetation indices, ignoring the differences in carbon sequestration capacity among different vegetation types and the complex influence of multiple factors such as hydrology, soils, and climate. This makes it difficult to accurately characterize the spatial heterogeneity of wetland carbon storage. Traditional regression models have limited adaptability and generalization, making them difficult to account for the nonlinear interactions and dynamic changes in wetland ecosystems. In contrast, machine learning methods can explore the complex relationships between carbon storage and multiple ecological factors, improving estimation accuracy and interpretability. Wetland carbon is composed of soil carbon and vegetation carbon, which complement each other. Its carbon storage is influenced by factors such as vegetation type, canopy carbon dioxide concentration, flooding frequency, and microtopography. From a systemic perspective, factors influencing wetland carbon storage tend to focus on describing independent factors and lack consideration of the interactions and correlations between factors.

[0004] In summary, it remains difficult to precisely quantify the specific contributions of different vegetation types to wetland carbon storage, and existing carbon storage estimation methods are limited in their ability to account for the combined effects of multiple factors. Questions remain, such as: "Does the succession of vegetation types affect wetland carbon storage?", "To what extent does this influence occur?", "How much carbon storage does each vegetation type contribute?", "Are there significant differences in soil carbon storage under different vegetation types?", and "How do ecological conditions such as climate and hydrology influence wetland carbon storage and its spatiotemporal variations?" These questions require further research to improve the accuracy of wetland carbon storage inversion and optimize carbon sequestration management strategies. Summary of the Invention

[0005] In response to the above-mentioned technical problems existing in the existing technology, the present invention proposes a technical solution based on a multi-model collaborative learning framework to predict wetland carbon storage and analyze the contribution of ecological factors, which helps to improve the accuracy of carbon storage inversion, construct a relationship network, quantify the contribution of key environmental factors to wetland carbon storage, and provide a scientific basis for wetland ecological protection and refined management.

[0006] In order to solve the above problems, the present invention adopts the following technical solutions:

[0007] A method for inverting wetland carbon storage and analyzing its contribution taking into account vegetation types is characterized by comprising the following steps:

[0008] Step 1: Based on multi-source remote sensing data and ground-based plant community sample data, combined with a random forest classification model, the characteristic differences of plant communities and the discrimination of their spatial distribution were evaluated to obtain the spatial pattern of dominant vegetation types in the wetland;

[0009] Step (2): Based on the spatial pattern of dominant wetland vegetation types obtained in step (1), a multi-model collaborative learning framework is constructed. Using vegetation types, soil properties, hydrological conditions, and climate variables as inputs, the spatial pattern of carbon storage in each dominant wetland vegetation type and soil is predicted, and the regional distribution and cumulative characteristics of vegetation and soil carbon storage are derived.

[0010] Step (3): Based on the spatial pattern of carbon storage of each dominant vegetation type in the wetland predicted in step (2), analyze the differences in carbon storage among different vegetation types to assess the impact of vegetation types on soil carbon storage; and use the contribution value quantification method to quantitatively assess the degree of influence of ecological factors on carbon storage changes, in order to identify the key environmental factors that dominate the spatial distribution of carbon storage;

[0011] Step (4): Select characteristic factors from the key environmental factors identified in step (3), establish a relationship network between characteristics and carbon storage, and obtain the direct and indirect effects of each characteristic on carbon storage changes.

[0012] Compared with the prior art, the present invention has the following beneficial effects:

[0013] While existing wetland carbon storage estimation methods can simulate carbon storage distribution at a macroscale, they struggle to accurately quantify the contribution of different vegetation types to carbon storage at a local scale. This results in insufficient representation of spatial heterogeneity and limited estimation accuracy. Furthermore, existing wetland carbon storage inversion methods mostly rely on climate models or simple regression models for prediction, making them difficult to adapt to the needs of wetland carbon storage estimation under different regions and ecological conditions. Furthermore, wetland carbon storage is influenced by multiple ecological factors, and existing studies mostly use single-factor or simple statistical analysis methods, making it difficult to systematically analyze the interactions between these factors and their contributions to carbon storage, limiting the scientific nature and applicability of the predictions.

[0014] The present invention proposes a wetland carbon storage inversion and contribution analysis method that takes into account vegetation types. First, multi-source remote sensing data, ecological factors and measured sample data are input, and a unified data set is formed through preprocessing, feature extraction, data interpolation and registration. Combined with the random forest classification model, the spatial pattern of the dominant vegetation types in the wetland is obtained. Then, the regression algorithms of three models, ridge regression, support vector machine and gradient boosting tree, are used to predict wetland carbon storage, and integrated through a multi-layer perceptron (MLP) as a meta-learner. By constructing a multi-model collaborative learning framework, the advantages of multiple models are combined. For example, ridge regression or support vector machine are strong in processing linear relationships, while gradient boosting tree can improve the fitting ability of complex data patterns by gradually optimizing model errors. Through multi-model collaborative learning, the respective advantages can be fully utilized, thereby more comprehensively exploring the potential patterns in the data and improving the accuracy of prediction.

[0015] SHAP analysis assigns a numerical value to each input feature, representing its impact on the model output. The contribution of each feature to the model prediction is then calculated, thereby assessing the impact of ecological factors on carbon storage. Incorporating structural equation modeling, the path relationships between latent and observed variables are defined, helping to quantify the direct and indirect effects of various factors. By establishing a causal network between these factors, the interaction mechanisms among different ecological factors are analyzed. Ultimately, this reveals the complex ecological processes underlying carbon storage changes.

[0016] In light of this, this paper focuses on wetlands, encompassing wetland vegetation classification, carbon storage prediction, and contribution analysis. To mitigate the scaling issues inherent in large-scale model-based approaches and the accuracy issues associated with single vegetation index regression, this approach combines a multi-model collaborative learning framework that integrates the strengths of various machine learning approaches with a data-driven regression model, providing scientific support for refined wetland management and ecological conservation strategies. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1The figure is an overall flow chart of the method proposed in the present invention. DETAILED DESCRIPTION

[0018] The present invention is further described in detail below with reference to the accompanying drawings and examples.

[0019] In response to scientific problems such as inaccurate carbon storage inversion caused by the dynamics and complexity of the wetland environment, this study aims to accurately estimate wetland carbon storage and reveal the interactive and complex ecological processes of carbon storage changes from a geographical perspective. Multi-source remote sensing data, ecological factors (vegetation type, soil properties, hydrology, climate, etc.) and field sampling survey data are coupled. A multi-model collaborative learning framework (including ridge regression, support vector machine and gradient boosting tree) is used, combined with a meta-learner to optimize prediction performance, to achieve wetland vegetation classification and carbon storage regression. SHAP analysis is used to evaluate the contribution of each ecological factor to carbon storage, and the interaction between factors is analyzed through structural equation models. This study provides a scientific basis for the accurate prediction of wetland carbon storage and the relationship between ecological factors through comprehensive analysis, providing support for wetland management and ecological protection strategies. Figure 1 As shown, the present invention includes the following specific method steps.

[0020] (1) Wetland vegetation classification and carbon storage prediction under a multi-model collaborative learning framework

[0021] Specific implementation steps:

[0022] Sub-step 1-1: Data Acquisition and Preprocessing: Integrate the measured wetland sample data (spatial location of various wetland vegetation types and laboratory analysis results of vegetation / soil carbon storage). To ensure data timeliness, obtain multi-source remote sensing imagery data, vegetation data (NDVI, GI, etc.), climate data (precipitation, temperature, etc.), soil data (such as soil organic carbon, pH, etc.), and hydrological data (such as flooding frequency, water level, flow rate, etc.) for the same period within the study area. Declouding and denoising the raw data, filling missing values, and using interpolation methods (such as linear interpolation or Kriging interpolation) to fill missing data to ensure data integrity, and spatially align different data sources to ensure consistent resolution and projection coordinates.

[0023] Sub-steps 1-2: Extract vegetation indices from remote sensing images and combine them with climate, soil, and hydrological data to extract relevant features (such as vegetation cover, soil moisture, precipitation, etc.), ensuring that data from different sources are spatially aligned, scale-matched, and time-synchronized. Detailed information on each classification feature is shown in Table 1:

[0024] Table 1 Introduction to characteristic variables

[0025]

[0026]

[0027] Sub-steps 1-3: Select the random forest classification algorithm to classify wetland vegetation, divide the vegetation spatial location sampling data into a training set and a validation set, input characteristic variables such as spectral bands, vegetation characteristics, polarization characteristics, and hydrological characteristics into the basic model, perform preliminary model training, and analyze the impact of different characteristics on the separability of vegetation communities. The optimal feature combination is selected and the model parameters are optimized to finally obtain the spatial distribution of various dominant vegetation types in the wetland.

[0028] (1) The random forest classification model comes from the "RandomForestClassifier" class of the "scikit-learn" library in Python, and is imported using "from sklearn.ensemble import RandomForestClassifier";

[0029] (2) The optimal feature combination selection refers to selecting by ranking the feature variables of the random forest model. The importance score of each feature is calculated through the "feature_importances_" attribute of the "RandomForestClassifier" class, so as to select the most discriminative feature combination and improve the model performance;

[0030] (3) The optimal feature combination selection refers to selecting by ranking the feature variables of the random forest model. The importance score of each feature is calculated through the "feature_importances_" attribute of the "RandomForestClassifier" class, so as to select the most discriminative feature combination and improve the model performance;

[0031] (4) Model parameter optimization is to improve the model accuracy by optimizing core parameters (such as the number of trees n_estimators, the maximum depth of the tree max_depth, the minimum number of samples for node splitting min_samples_split, etc.), and search for the optimal value of each core parameter for vegetation classification through the grid search method. The implementation code is “RandomizedSearchCV(RandomForestClassifier(),rf_params)”.

[0032] (II) Analysis of the contribution of ecological factors to changes in wetland carbon storage

[0033] Specific implementation steps:

[0034] Sub-step 2-1: Convert the collected soil and vegetation sample data, the obtained soil organic carbon content, and the measured results of the above-ground and below-ground biomass and organic carbon content of various types of vegetation into carbon storage. Vegetation carbon storage includes the sum of the carbon storage of each component in the system. Vegetation carbon storage is divided into carbon storage in the above-ground part of vegetation and carbon storage in the underground root system. Soil carbon storage is calculated by multiplying the soil organic carbon density and area, where soil organic carbon density is related to soil bulk density, soil organic carbon content, soil layer thickness, and gravel content. The specific formula is as follows:

[0035]

[0036] Where C D is the carbon storage of wetland vegetation (g C / km 2 ),∑ K C i is the sum of carbon storage of K vegetation in wetlands (g C / km 2 ).

[0037] AGC i =p×A i ×Q i

[0038] Where, AGC i is the aboveground carbon storage of the i-th type of vegetation (g), p is the carbon conversion coefficient (g / kg), which refers to the organic carbon fixed by 1g of organic dry matter formed by vegetation during photosynthesis, obtained from field sampling data, and A i is the aboveground biomass of vegetation (kg / m 2 ), Q i is the area (m 2 ).

[0039] BGC i =AGC i ×r

[0040] Where, BGC i is the underground carbon storage of the i-th type of vegetation (g), and r is the root-to-stem ratio of vegetation. i =AGC i +BGC i .

[0041] C S =SOCD×Q i

[0042] Where C S is soil carbon storage (g), SOCD is soil organic carbon density (g C / cm 2 ), Q i is the area (m 2 ).

[0043] SOCD=ρ i ×C i ×T i ×(1-g)

[0044] Where SOCD is soil organic carbon density (g C / cm 2 ), ρ i Soil bulk density (g / cm 3 ), C i is the soil organic carbon content at a depth of 0.3 m (g / kg), T i is the thickness of the soil layer (30 cm), and g is the volume ratio (%) of gravel larger than 2 mm.

[0045] ρ i =M / V

[0046] Where, ρ i Soil bulk density (g / cm 3 ), M is the dry weight of soil (g), V is the original volume at the time of sampling (cm 3 ).

[0047] Sub-step 2-2: Train multiple models for collaborative learning, select a base model for preliminary carbon storage prediction, divide the original data set into a training set and a validation set, input characteristic variables such as vegetation characteristics, soil characteristics, hydrology, and climate into the base model, train each model and optimize the model parameters, input the preliminary results of the base model training into the meta-learner for collaborative learning, and generate the final prediction results:

[0048] The original data set includes the data in sub-steps 1-1 and 1-2, the carbon storage data includes vegetation carbon storage data and soil carbon storage data; the basic model is ridge regression, support vector machine and gradient boosting tree;

[0049] (1) The ridge regression model comes from the "Ridge" class of the "scikit-learn" library in Python, and is imported using "fromsklearn.linear_model import Ridge";

[0050] (2) The support vector machine model comes from the "SVR" class of the "scikit-learn" library in Python, and is imported using "from sklearn.svm import SVR";

[0051] (3) The gradient boosting tree model comes from the “XGBRegressor” class of the “xgboost” library in Python, and is imported using “from xgboost import XGBRegressor”.

[0052] The model parameter optimization requires tuning the core parameters of each base model to ensure that each model has been tuned to a better state.

[0053] (1) The core parameters of the ridge regression model include alpha (typical value range is [1e-5, 1e5]) and solver (typical value range is {auto, svd, sag}). The grid search method is used to search for the optimal value of each core parameter for carbon storage prediction. The implementation code is “RandomizedSearchCV(Ridge(),ridge_params)”;

[0054] (2) The core parameters of the support vector machine model include C (typical value range is [0.1, 1000]), kernel (typical value range is {linear, rbf, poly}), and gamma (typical value range is [1e-5, 10]). The grid search method is used to search for the optimal value of each core parameter for carbon storage prediction. The implementation code is “GridSearchCV(SVR(kernel='rbf'),svm_params)”;

[0055] (3) The core parameters of the gradient boosting tree model include max_depth (typical value range is [0.1, 1000]), learning_rate (typical value range is [0.01, 0.1]), and n_estimators (typical value range is [50, 1000]). The grid search method is used to search for the optimal value of each core parameter for carbon storage prediction. The implementation code is “GridSearchCV(xgb_model, param_grid, scoring = 'neg_mean_squared_error')”.

[0056] The multi-model collaborative learning is to input the output results of the basic model into the meta-learner for weighted fusion, including meta-feature generation, meta-learner configuration and weight allocation, and integrated prediction and evaluation to generate the final prediction result;

[0057] (1) Meta-feature generation refers to using the "cross_val_predict" method in Python to generate cross-validation prediction values ​​of each basic model on the training set, and splicing the prediction values ​​of each basic model by column to form a new meta-feature matrix;

[0058] (2) Meta-learner configuration and weight assignment refers to selecting ridge regression to train the meta-learner and inputting the meta-feature matrix to calculate the meta-learner weights. In order to dynamically adjust the weights, it is necessary to evaluate the prediction performance of each basic model on the validation set (e.g., by calculating R2 ), and the initial weights of the validation set are assigned by heuristic weighting. Finally, the weighted average of the meta-learner weight and the initial weight of the validation set is taken (optimal weight assignment value = 0.7 * meta-learner weight + 0.3 * initial weight of the validation set) to obtain the optimal weight assignment for each base model;

[0059]

[0060] The above is the formula for calculating the weight of the meta-learner, y train is the actual value of the training set, y train,i is the predicted value from each base model, ω i is the weight of the base model, λ‖w‖ 2 is a regularization term used in the ridge regression model to combine the best results.

[0061]

[0062] The above is the calculation formula for the initial weight of the validation set. is the R of the base model i on the validation set 2 value, is the initial weight of base model i.

[0063] Sub-step 2-3: Based on the integrated prediction and evaluation described in sub-step 2-2, the integrated prediction part is achieved by weighted fusion of the prediction results of the basic models according to the optimal weight distribution value of each basic model, and finally the model is evaluated according to the performance indicators:

[0064] Validation indicators include coefficient of determination (R 2 ), Root Mean Squared Error (RMSE) and Mean Absolute Error (MAE). The calculation method is as follows:

[0065]

[0066] Where, T i is the measured value of the i-th sample, is the mean of the measured data, P i is the simulated value of the i-th sample.

[0067]

[0068] Where n is the number of samples, T i is the measured value of the i-th sample, P i is the simulated value of the i-th sample.

[0069] Sub-step 3-1: After predicting carbon storage using multi-model collaborative learning in step 2, the contribution of each ecological factor to carbon storage is calculated using a contribution value quantification method. A bee swarm plot is used (red dots indicate a positive impact on carbon storage when the factor value is high, and blue dots indicate a negative impact) to visualize the positive or negative impact of each factor on carbon storage. The SHAP value measures the contribution of each feature to the prediction result, showing the magnitude and direction of the influence of each influencing factor, and has good visualization performance:

[0070]

[0071] Where N is the set of carbon storage influencing factors. Assuming there are p subsets, S is the set that does not include the influencing factor x. i The subset of |S| is the size of the subset, f(S) is the model output result corresponding to the subset S, and f(S∪{i}) is the influence factor x i Output after joining.

[0072] Step (4): Select characteristic factors from the key environmental factors identified in step (3), establish a relationship network between the characteristics and carbon storage, and obtain the direct and indirect effects of each characteristic on carbon storage changes. The specific steps include the following:

[0073] Sub-step 4-1: Based on the quantitative contribution value results obtained in step (3), select the top 5-10 variables as input data to establish a relationship network between characteristics and carbon storage;

[0074] Sub-step 4-2: Select a structural equation model to demonstrate the quantitative relationships between the selected variables and to identify the direct and indirect effects between the variables;

[0075] Sub-step 4-3: Use a path diagram to display the interaction network between factors and show the direct or indirect influence values ​​between the selected variables.

[0076] In other words, the use of structural equation models to define the path relationship between latent variables and observed variables helps quantify the direct and indirect effects between factors, and by establishing a causal network between factors, analyzes the interaction mechanism between different ecological factors.

[0077] The structural equation model, derived from the "lavaan" package in R, defines latent variables and path relationships. For example, "Climate = ~Temperature + Precipitation" indicates that climate is measured by these two indicators. Path relationships, such as "Carbon_Storage ~ Climate + Soil + Vegetation," represent the direct effects of these latent variables on carbon storage. Finally, a path diagram (derived from the "semPaths" package in R) was used to visualize the interaction network between factors, showing the direct and indirect effects of the selected variables.

[0078] Finally, it should be noted that the above implementation examples are only used to illustrate the technical solutions of the present invention and are not limiting. Although the present invention has been described in detail with reference to preferred examples, those skilled in the art should understand that the technical solutions of the present invention can be modified or replaced by equivalents without departing from the purpose and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A wetland carbon storage inversion and contribution analysis method taking into account vegetation types, characterized by: It includes the following steps: Step 1: Based on multi-source remote sensing data and ground-based plant community sample data, combined with a random forest classification model, the characteristic differences of plant communities and the discrimination of their spatial distribution were evaluated to obtain the spatial pattern of dominant vegetation types in the wetland; Step (2): Based on the spatial pattern of dominant wetland vegetation types obtained in step (1), a multi-model collaborative learning framework is constructed. Using vegetation types, soil properties, hydrological conditions, and climate variables as inputs, the spatial pattern of carbon storage in each dominant wetland vegetation type and soil is predicted, and the regional distribution and cumulative characteristics of vegetation and soil carbon storage are derived. Step (3): Based on the spatial pattern of carbon storage of each dominant vegetation type in the wetland predicted in step (2), analyze the differences in carbon storage of different vegetation types to evaluate the impact of vegetation type on soil carbon storage; and use the contribution value quantification method to quantitatively evaluate the degree of influence of ecological factors on carbon storage changes, so as to identify the key environmental factors that dominate the spatial distribution of carbon storage; Step (4): Select characteristic factors from the key environmental factors identified in step (3), establish a relationship network between characteristics and carbon storage, and obtain the direct and indirect effects of each characteristic on carbon storage changes.

2. A wetland carbon storage inversion and contribution analysis method taking vegetation types into account according to claim 1, characterized in that: The step (1) includes the following sub-steps: Sub-step 1-1: Acquire multi-source remote sensing image data, vegetation data, climate data, soil data, and hydrological data for the study area over the same period; decloud and de-noise the original data, fill in missing values, use interpolation methods to fill missing data to ensure data integrity, and spatially align different data sources to ensure consistent resolution and projection coordinates; Sub-step 1-2: Extracting vegetation indices from the multi-source remote sensing image data and combining them with climate, soil, and hydrological data to extract relevant features, including vegetation cover, soil moisture, and precipitation, ensuring spatial alignment of data from different sources, and performing scale matching and temporal synchronization; Sub-steps 1-3: Construct a random forest classification model to input relevant features to achieve wetland vegetation classification.

3. The wetland carbon storage inversion and contribution analysis method taking vegetation types into account according to claim 2 is characterized in that: The sub-steps 1-3 include: A random forest classification algorithm was selected for wetland vegetation classification. The vegetation spatial location sampling data was divided into a training set and a validation set. Feature variables such as spectral bands, vegetation characteristics, polarization characteristics, and hydrological characteristics were input into the basic model, and the model was preliminarily trained. By analyzing the impact of different features on the separability of vegetation communities, the optimal feature combination was selected, and the model parameters were optimized. Finally, the spatial distribution of various dominant vegetation types in the wetland was obtained. The basic models were: ridge regression, support vector machine, and gradient boosting tree.

4. The wetland carbon storage inversion and contribution analysis method taking vegetation types into consideration according to claim 2, characterized in that: The step (2) includes the following sub-steps: Sub-step 2-1: Convert the collected soil and vegetation sample data, the measured results of soil organic carbon content, the measured results of aboveground biomass of various vegetation types, the measured results of belowground biomass of various vegetation types, and the measured results of organic carbon content of various vegetation types into carbon storage; the carbon storage includes vegetation carbon storage and soil carbon storage; Sub-step 2-2: Based on the wetland vegetation type classification results obtained in step (1), configure the collaborative learning model group to predict carbon storage and perform accuracy verification.

5. A wetland carbon storage inversion and contribution analysis method taking vegetation types into account according to claim 4, characterized in that: The sub-step 2-2 includes: Train multiple models for collaborative learning and select a base model for preliminary carbon storage prediction: Divide the original data set into a training set and a validation set. Input characteristic variables such as vegetation characteristics, soil characteristics, hydrology, and climate into the base model. Train each model and optimize the model parameters. Input the preliminary results of the base model training into the meta-learner for collaborative learning to generate the final prediction results: The original data set includes: the carbon storage data in sub-steps 1-4, the carbon storage data includes vegetation carbon storage data and soil carbon storage data; the basic model is: ridge regression, support vector machine and gradient boosting tree; The model parameter optimization requires tuning the core parameters of each base model to ensure that each model has been tuned to a better state. Among them, the multi-model collaborative learning is to input the output results of the basic model into the meta-learner for weighted fusion, including meta-feature generation, meta-learner configuration and weight allocation, and integrated prediction and evaluation to generate the final prediction results.

6. The wetland carbon storage inversion and contribution analysis method taking vegetation types into account according to claim 1 is characterized in that: The step (three) includes the following sub-steps: Sub-step 3-1: Use the variance analysis statistical method to determine the significance of carbon storage differences between different vegetation types, calculate the P value and determine the statistical significance of carbon storage changes; Sub-step 3-2: While predicting carbon storage through multi-model collaborative learning in step (2), the contribution value of each ecological factor to carbon storage is calculated through the contribution value quantification method, and the positive or negative impact of each factor on carbon storage is visualized using a beeswax diagram.

7. The wetland carbon storage inversion and contribution analysis method taking vegetation types into consideration according to claim 1, characterized in that: The step (four) includes the following sub-steps: Sub-step 4-1: Based on the quantitative contribution value results obtained in step (3), select the top 5-10 variables as input data to establish a relationship network between characteristics and carbon storage; Sub-step 4-2: Select a structural equation model to demonstrate the quantitative relationships between the selected variables and to identify the direct and indirect effects between the variables; Sub-step 4-3: Use a path diagram to display the interaction network between factors and show the direct or indirect influence values ​​between the selected variables.

Images