Multi-model coupled regional terrestrial ecosystem carbon reserve estimation method

Abstract

The invention discloses a multi-model coupled regional terrestrial ecosystem carbon reserve estimation method, which is suitable for the field of carbon cycle and carbon reserve estimation. The method comprises the following steps: firstly, collecting remote sensing data, and then estimating overground, underground and soil annual carbon density mean values of a regional terrestrial ecosystem by adopting an ARIMA model; the method comprises the following steps: clustering regions with different carbon densities by adopting a CatBoost model, dividing the carbon densities of a regional ecological system into low, medium and high levels, and avoiding the influence of spatial heterogeneity among different density levels on subsequent model training; an RNN model is constructed to evaluate the carbon density conditions of different levels of unit grids, and interannual dynamic carbon density parameters of different types of ecosystems are obtained based on multi-source data; and finally, calculating the carbon reserves of the unit grids, analyzing the annual average change trend and the spatial distribution condition of the carbon reserves, and realizing accurate evaluation of the regional terrestrial ecosystem carbon sink. The method is simple in step and good in using effect, and the change condition of the carbon sink capacity in the corresponding time can be obtained according to the sampling time interval of the input data.

Description

Technical Field

[0001] The multi-model coupling regional terrestrial ecosystem carbon storage estimation method and carbon sink capacity driving mechanism analysis method involved in the present invention are particularly applicable to the field of carbon cycle. Background Art

[0002] At present, due to the rapidly increasing CO2 concentration, the greenhouse effect of the earth has been exacerbated, resulting in rising temperatures, rising sea levels, and damaged ecosystems. As an important part of the global carbon storage and carbon sink, the terrestrial ecosystem plays an important role in alleviating the greenhouse effect. In the prior art, to estimate the regional carbon storage, there are mainly two methods: the "bottom-up" estimation based on point-source measured or grid simulation data, and the "top-down" estimation based on satellite remote sensing or atmospheric inversion data. The "bottom-up" assessment methods mainly include the inventory method, the eddy covariance method, the ecosystem model simulation method, etc.; the "top-down" assessment method mainly relies on remote sensing technology to evaluate the ecosystem carbon sink capacity based on the greenhouse gas concentration distribution retrieved from the atmosphere or the changes in surface vegetation and land use types.

[0003] Fang J et al. evaluated the changes in forest biomass carbon storage in the mid-latitudes of the Northern Hemisphere of China based on the inventory method and using 50-year forest resources inventory data with an improved biomass expansion coefficient. Although the evaluation results of the inventory method have high accuracy, they require a large amount of measured data support, with high evaluation costs and it is difficult to achieve large-scale, high-spatial and long-time resolution evaluations; Rodda SR et al. evaluated the carbon sink capacity of tropical dry deciduous forests in central India based on multi-year eddy covariance data. Although the eddy covariance has the advantages of a wide evaluation range, a fine time scale, and being able to reduce the influence of factors such as climate on the evaluation, the evaluation accuracy is easily affected by complex underlying surface conditions, and it is difficult to further distinguish the carbon budgets of soil, above-ground and below-ground biomass. Various types of ecosystem carbon budget simulation models such as the CASA model, the global dynamic vegetation model, the coupled land use model and the inVEST model are based on the assumption that the carbon density is fixed and cannot distinguish the differences in carbon sink capacity of the same type of land. The present invention uses easily accessible remote sensing data to train the ARIMA-CatBOOST-RNN model, and uses a multi-model coupling method to identify the carbon density of different ecosystem types, breaking through the limitations of traditional methods, improving the carbon storage estimation module of the inVEST model, and realizing large-scale, high-precision, long-time series regional terrestrial ecosystem carbon storage evaluation.

[0004] Patent CN112836610B discloses a method for quantitatively estimating carbon storage based on remote sensing data and land use change data. Based on land use classification data and ground survey data, the carbon density of ground objects is calculated. The correlation between the carbon storage and each characteristic value in each sample plot is analyzed, and the characteristic values with significant correlation are selected to construct a multi-scale convolutional neural network, so as to realize the quantitative estimation of carbon storage in the research area. This method focuses on constructing the non-linear relationship between the characteristic variables of remote sensing data and carbon storage, ignoring the spatial heterogeneity between different research areas. For research areas with a large range, the influencing factors such as temperature and precipitation have different effects on carbon storage. If represented by a non-linear relationship, it will inevitably affect the evaluation accuracy of carbon storage. At the same time, the carbon density of ground objects is not fixed. Using this method to estimate carbon storage based on fixed carbon density values of ground objects will also cause certain errors. Patent CN110750904B discloses a regional carbon storage spatial pattern monitoring system and method based on remote sensing data, including a remote sensing data processing module, a measured data processing module, and a data analysis and output module. Using satellite remote sensing data, a method of establishing a model with precipitation, vegetation coverage, surface temperature, land use data, lighting data, DEM data, and meteorological data is used to monitor the spatio-temporal changes of regional carbon storage, overcoming the disadvantage of the long research cycle of carbon storage in large-scale regions. The method has the advantages of easy access to source data, high computer operation efficiency, low cost, and strong timeliness. However, this method only establishes a relatively simple linear regression model, which is difficult to accurately fit the effects of various influencing factors on carbon storage, and does not consider the impact of spatial heterogeneity on carbon storage evaluation. Summary of the Invention

[0005] Technical Problem: Aiming at the deficiencies of the prior art, a method for estimating carbon storage in regional terrestrial ecosystems by coupling multiple models is provided. It can take into account the influence of spatial heterogeneity, distinguish the carbon storage differences of the same land type, and has higher evaluation accuracy. It can obtain the interaction effects of various influencing factors on the carbon sequestration capacity of different sub-ecosystems, and can fully understand the complex mapping relationship between the carbon sequestration capacity and the driving factors. Aiming at the problem that only fixed final estimation results can be obtained in the prior art, the present invention can obtain estimation results that change according to the sampling time interval of the input sample data.

[0006] In order to overcome the defects and deficiencies existing in the prior art, the present invention discloses a method for estimating carbon storage in regional terrestrial ecosystems by coupling multiple models, including the following steps:

[0007] Step 1: Multi-source data fusion: Collect remote sensing data across the country to generate six types of impact factor data, carbon density statistical data, and NASA carbon stock remote sensing data to construct a model training and validation dataset. The six types of impact factor data include annual average precipitation (PRS), temperature (TEM), soil moisture (SM), normalized difference vegetation index (NDVI), net primary productivity (NPP), and land use data over consecutive years; the carbon density statistical data includes the statistical results of underground, above-ground, and soil carbon density in the entire study area over many years; the NASA carbon stock remote sensing data includes the raster values of underground, above-ground, soil, and dead organic matter carbon stocks in the study area in a certain year.

[0008] Step 2: Prediction of annual average carbon density parameters: Based on the carbon density survey data of above-ground, underground, and soil biomass in different regions across the country, calculate their annual average carbon density; use the autoregressive integrated moving average model ARIMA to predict the annual average carbon density in the future based on the average carbon density over continuous time. Let d represent the order of differencing, and use SPSS data analysis software to determine that the data differencing order d = 1; based on the Akaike Information Criterion (AIC) and the Bayesian Information Criterion (BIC), traverse the values of the autoregressive term number p and the moving term number q, and select the minimum values of the autoregressive term number p and the moving term number q as the best fitting parameters p = 5, q = 1 to construct the ARIMA model. The formula is as follows:

[0009] C dm = φ1Y t-1 + φ2Y t-2 +…+ φ p Y t-p + e t - θ1e t-1 - θ2e t-2 -…- θ q e t-q

[0010] In the formula: C dm is the prediction result, p is the autoregressive term number, q is the moving term number, e is the natural constant, t is the sample time, φ is the autoregressive coefficient, and θ is the moving coefficient;

[0011] Use the training dataset to train the ARIMA model. During training, the ARIMA model performs training iterations according to the moving term number, calculates the error during the iteration process, and adjusts the parameters to gradually reduce the error, making the predicted value gradually approach the true value. After multiple iterations, fit the changing rule of carbon density over time and estimate the annual average carbon density in the next few consecutive years in the training dataset.

[0012] Step 3: Based on the impact factor data, use the CatBoost model, which is good at handling classification tasks, to achieve the classification of carbon density levels based on impact factors: First, determine the carbon density classification system, and divide the normalized carbon density data into three categories: low, medium, and high; then spatially overlay the impact factor data and the carbon density data to obtain the corresponding tabular data of six types of impact factors and carbon density levels at the same location. Based on the CatBoost model, obtain the spatial characteristic relationship between the impact factor data and the carbon density levels, and achieve the classification of carbon density levels based on the six types of impact factor data;

[0013] Step 4: Use the RNN model to estimate the carbon density value: Use the hidden layer of the RNN model to extract the characteristic information of the impact factor data and the carbon density data, and perform cycling and transmission in the form of hidden states; use the training data set to train the RNN model to fit the functional relationship between the carbon density and the influencing factors, rasterize the research area grid units, and calculate the normalized carbon density value (0 - 1) of each grid unit in the research area; use the average carbon density predicted by the ARIMA model and the carbon density level predicted by the CatBoost model to impose spatio-temporal constraints on the prediction results of the RNN model to obtain the true carbon density value formula of each grid unit in the research area as follows:

[0014] C prei = f(X1X2…X6)

[0015]

[0016] In the formula: C prei is the result of the RNN model predicting the normalized carbon density value of the i-th grid unit, f is the fitting function of the RNN model, X1 - X6 are independent variables, i.e., the impact factor values, C di is the true carbon density value of the i-th grid unit, C dm is the average carbon density of the j-th carbon pool predicted by the ARIMA model, C dj is the average carbon density of the j-th carbon pool in the base year, B max and B min are the maximum and minimum values of the carbon density classification obtained by the CatBoost model;

[0017] Repeat Steps 2 - 4 to calculate the aboveground, underground, and soil biomass carbon densities respectively;

[0018] Step 5: Calculate the carbon storage in the study area: Estimate the carbon density value of dead organic matter based on the aboveground, underground, and soil carbon density values estimated by the RNN model. Use these four types of carbon density values as the carbon storage parameters of the inVSET model, and use the inVSET model to calculate the aboveground, underground, soil, and dead organic matter carbon storage in each grid cell in the study area respectively, so as to realize the estimation of the carbon storage of the terrestrial ecosystem at the grid cell level, make up for the deficiency of the traditional method that overly relies on land use types for carbon storage assessment, fully consider various influencing factors, improve its ability to identify the differences in carbon storage at a fine scale, and realize the estimation of the carbon storage of the regional terrestrial ecosystem. The formula is as follows:

[0019] C = C above + C below + C soil + C dead

[0020]

[0021] In the formula, C is the total carbon storage of the terrestrial ecosystem, C above is the total aboveground carbon storage, C below is the total underground carbon storage, C soil is the total soil carbon storage, C dead is the total carbon storage of dead organic matter, C i is the carbon storage value of a certain grid, C d is the carbon density value of a certain grid, S k is the area of the grid cell, n is the total number of grid cells in the study area, k is the serial number of the current grid cell;

[0022] Use the inVEST model to evaluate the carbon sink capacity of the study area: Based on the interannual changes in the carbon storage of the study area, calculate the carbon accumulation rate between different years, and use the carbon accumulation rate index to evaluate the interannual carbon sink capacity. The formula is as follows;

[0023]

[0024] In the formula: CAR k is the carbon accumulation rate of the kth grid, ΔC k is the change in the carbon storage of the kth grid within the time period T, ΔT is the time change. Within the time interval, if the carbon accumulation rate is positive, it is a carbon sink area, if it is negative, it is a carbon source area, and the larger the positive value, the stronger the carbon sink capacity of the study area;

[0025] Attribution Analysis of Carbon Sink Changes: Referring to the ecosystem classification standard, the terrestrial ecosystem is divided into seven sub-ecosystems: forest, grassland, cultivated land, settlement, desert, water body, and others. The Geodetector model is used to analyze the impacts of human activities and natural factors on carbon sink capacity, explore the driving mechanism of carbon sink capacity changes, and provide data and theoretical support for formulating low-carbon development policies.

[0026] Furthermore, before the multi-source data fusion process in Step 1, the remote sensing data and verification data used as influencing factors are subjected to operations such as calculating the annual average value, projection, resampling, and clipping using MATLAB and ArcGis software, unifying the coordinate systems of all remote sensing data to WGS1984 and the spatial resolution to 1KM; the carbon density statistical data is successively processed such as unit normalization, extracting carbon density data within the study area, screening out outliers, and filtering noise data; all attributes are mapped to the interval of (0-1) to avoid the influence of data with different value ranges on model training; the remote sensing data is rasterized according to a vector grid of 1KM*1KM using the fishnet tool and zonal statistics tool and converted into tabular data, and a training dataset and a verification dataset are constructed in a ratio of 7:3.

[0027] Furthermore, in the estimation using the inVSET model and the evaluation of carbon sink capacity: The study calculated the carbon density of aboveground, underground, soil, and dead organic matter carbon for each grid cell (1KM) in the study area based on a multi-model coupling method, obtained carbon density parameters at the grid cell level, recommended the estimation of carbon storage at the traditional land use type scale to the grid cell scale, fully considered the impacts of driving factors such as precipitation, temperature, and population density on carbon storage, broke the excessive dependence on land use data by traditional methods, and achieved continuous long-term dynamic carbon storage evaluation, improving the accuracy and efficiency of carbon storage estimation.

[0028] Furthermore, the analysis method of the driving mechanism of carbon sink capacity is as follows:

[0029] The study divided the terrestrial ecosystem into seven categories: farmland ecosystem, forest ecosystem, grassland ecosystem, water body and wetland ecosystem, settlement ecosystem, desert ecosystem, and other ecosystems according to the ecosystem classification system based on land use data; from the perspectives of single-factor detection and interaction detection based on the Geodetector model, the driving mechanism of human activities and the natural environment on carbon sink capacity changes was analyzed. Multiple grid cells in the above seven ecosystem types were randomly selected as samples N respectively, and the Geodetector model was used to conduct driving mechanism analysis by stratifying according to the influencing factor category h = 6. The formula is as follows:

[0030]

[0031] Where: q is the influence degree of the above-mentioned influencing factor on the carbon storage, and the value range of q is [0,1]. The larger the value, the stronger the explanatory power of the influencing factor for the spatio-temporal differentiation of the carbon sequestration capacity, and vice versa. N represents the total number of samples, and N h represents the total number of samples in layer h, and δ 2 represents the variance of the carbon storage value of the entire region of the terrestrial ecosystem carbon storage, represents the variance of the carbon storage value in layer h. SSW represents the sum of the within-layer variances, and SST represents the total variance of the entire region of the terrestrial ecosystem carbon storage.

[0032] A computer device includes a processor and a memory. The processor is electrically connected to the memory. The memory is used to store instructions and data, and the processor is used to execute the method for estimating the carbon storage of the regional terrestrial ecosystem by multi-model coupling.

[0033] Beneficial effects

[0034] In the present invention, the ARIMA-CatBoost-RNN three models are coupled to estimate the mean value of the carbon density of the terrestrial ecosystem. According to specific requirements, vector fishing nets of different spatial scales are generated to extract the raster values of the influencing factors and input them into the model. The raster value input model outputs the estimation results of multiple spatial scales according to the resolution of the input raster values. At the same time, based on the inVEST model and combined with the carbon density parameters estimated from the influencing factor data at different times, the change of the carbon storage distribution in the study area under a long time series can be calculated, realizing the assessment of the carbon storage of the terrestrial ecosystem at multiple spatio-temporal scales.

[0035] The technical advantages of the present invention include:

[0036] (1) Promote the innovation of the regional carbon density estimation method and improve the carbon storage accounting ability of traditional models:

[0037] The spatio-temporal distribution of carbon density and carbon storage is the determining factor of the carbon sequestration capacity and also the intuitive manifestation of the interannual fluctuation of the urban terrestrial ecosystem. The spatio-temporal distribution and evolution of carbon density and carbon storage are affected by various factors such as temperature, precipitation, population density, and vegetation coverage. The traditional coefficient method evaluation method that only considers the change of land use type cannot accurately depict the spatio-temporal evolution process of carbon density and carbon storage. Therefore, through the ARIMA-CatBoost-RNN coupling model, the present invention explores the mapping relationship between multiple influencing factors and carbon density, realizing high-precision carbon storage accounting at the multi-temporal raster cell level;

[0038] (2) Clarify the spatio-temporal distribution and evolution of the carbon pool in the regional terrestrial ecosystem:

[0039] Terrestrial ecosystem carbon pools mainly include four types: aboveground biomass carbon pool, belowground biomass carbon pool, soil carbon pool, and dead organic matter carbon pool. Based on the spatial classification method of terrestrial ecosystems, the urban land area is divided into seven categories: farmland ecosystem, forest ecosystem, grassland ecosystem, water body and wetland ecosystem, settlement ecosystem, desert ecosystem, and other ecosystems. Based on the carbon storage accounting results of the coupling model and the improved inVEST model, the annual changes of carbon pools are analyzed from four perspectives: underground, aboveground, soil, and dead organic matter biomass, and the distribution and evolution of carbon storage in terrestrial ecosystems are studied in detail. (3) Explore the mechanism of changes in carbon sink capacity and assist in formulating low-carbon policies:

[0040] The carbon sink capacity of regional terrestrial ecosystems is mainly reflected in the rate of carbon accumulation; the carbon sink capacity is evaluated based on the carbon accumulation rate, and the geographical detector model is used to explore the influence mechanism of human activities and natural conditions on the carbon sink capacity, revealing the spatio-temporal laws of carbon sequestration and increasing sinks in regional terrestrial ecosystems. Provide data and theoretical support for formulating low-carbon development strategies.

[0041] Explanation of attached figures and tables

[0042] Figure 1 It is a schematic flow chart of the method for estimating carbon storage in regional terrestrial ecosystems by multi-model coupling of the present invention.

[0043] Figure 2 It is a schematic diagram of the construction process of the geographically weighted regression model in the present invention.

[0044] Figure 3 It is the evaluation result of carbon density and carbon storage by the coupling model in the embodiment of the present invention. Specific implementation method

[0046] The method of the present invention will be further described below in conjunction with the accompanying drawings and examples:

[0047] Such as Figure 1As shown in the figure, the present invention discloses a multi-model coupled regional terrestrial ecosystem carbon sink capacity assessment model that can identify the differences in carbon sink capacities of the same land use type. The ARIMA model is used to estimate the annual mean values of above-ground, below-ground, and soil carbon densities of the regional terrestrial ecosystem based on the carbon density statistical data from 2000 to 2014. Secondly, the CatBoost model is trained to cluster the regions with different carbon densities in the Xinjiang Uygur Autonomous Region to avoid the influence of spatial heterogeneity between different density levels on subsequent model training. An RNN model is constructed to evaluate the carbon density values of grid cells at different levels based on the prediction results of the ARIMA model and the CatBoost model. The carbon density parameters in the inVEST model are optimized based on the evaluation results of the multi-model coupling, the carbon storage of the grid cells is calculated, and the annual change trend and spatial distribution of the carbon storage are analyzed to achieve an accurate assessment of the regional terrestrial ecosystem carbon sink capacity. Finally, based on the geographical detector model, the spatial driving mechanism of the influence of natural conditions and human activities on the carbon sink capacity is explored to promote the innovation of the regional terrestrial ecosystem carbon sink capacity assessment method. The steps are as follows:

[0048] Step 1: Data preprocessing and construction of the training dataset. The collected basic data is divided into impact factor data for training the model and validation data for verifying the model accuracy. The impact factor data includes annual mean precipitation (PRS), annual mean temperature (TEM), annual mean soil moisture (SM), normalized difference vegetation index (NDVI), net primary productivity (NPP), and land use data. The validation data includes below-ground, above-ground, soil, and dead organic matter carbon densities. Based on MATLAB and ArcGis software, operations such as calculating the annual mean value, projection, resampling, and clipping are performed to unify the data coordinate system to WGS1984 and the spatial resolution to 1KM. All data is processed by normalization, extraction, screening, and division, filtering out noise data, and a training dataset and a validation dataset are constructed in a 7:3 ratio, with a total of 1,635,396 training data extracted;

[0049] Multi-source data fusion: Comprehensive consideration is given to multi-source data such as annual mean precipitation, annual mean temperature, NDVI, soil carbon density, and carbon density for carbon storage assessment. In order to improve the training efficiency of the model, reduce the difficulty of parameter tuning, and avoid the risk of overfitting, enabling the model to better understand the mapping relationship between impact factors and various carbon density values. First, the multi-source data is unified under the WGS1984 coordinate system with a resolution of 1km, and the normalization of the dataset is achieved based on the deviation normalization method, mapping all attributes to the interval (0 - 1) to avoid the influence of data with different value ranges on model training. A python script is written to uniformly extract the training data that meets the conditions from the overall dataset to achieve the purpose of improving training efficiency and model accuracy.

[0050] Step 2: Based on the carbon density statistical data from 2000 to 2014, use the ARIMA model, which is good at time series prediction, to estimate the average annual carbon density of above-ground, underground, and soil biomass in the region from 2015 to 2020 as the time constraint parameter for the prediction results of the RNN model;

[0051] Estimation of the mean carbon density: The ARIMA model includes an autoregressive (AR) part, which represents the relationship between the current observation and past observations. Usually, the number of autoregressive terms is p; the integration (I) part represents differencing the time series. The number of differencing times is called the integration order, usually denoted as d. The differenced series becomes a stationary series, which can reduce the uncertainty introduced by non-stationarity; the moving average (MA) part represents the impact of past noise errors on the current observation. Usually, the number of moving terms is q. The model formula is as follows:

[0052] C dm =φ1Y t-1 +φ2Y t-2 +…+φ p Y t-p +e t -θ1e t-1 -θ2e t-2 -…-θ q e t-q

[0053] In the formula: C dm is the prediction result, p is the number of autoregressive terms, q is the number of moving terms, e is the natural constant, t is the sample time, θ is the autoregressive coefficient, and θ is the moving coefficient.

[0054] The research is based on the average annual statistical data of above-ground, underground, and soil carbon densities from 2000 to 2014, and analyzes the changing trends of various carbon densities from 2000 to 2020. Use the SPSS data analysis software to determine the data differencing order d; based on the Akaike Information Criterion (AIC) and the Bayesian Information Criterion (BIC), traverse the values of p and q, and select the minimum of the two as the best fitting parameter to realize the construction of the ARIMA model, and evaluate the prediction accuracy of the model based on the mean square error index.

[0055] Step 3: Based on the above impact factor data and carbon density data, use the CatBoost model, which is good at handling classification tasks, to realize the classification of carbon density levels based on impact factors as the spatial constraint parameter for the prediction results of the RNN model, reducing the impact of spatial heterogeneity on subsequent model training;

[0056] Carbon density level classification: The CatBoost model is a classification algorithm based on gradient-boosted decision trees. It constructs a symmetric tree structure that enables the model to more reasonably analyze categorical feature data when dealing with classification tasks and minimize overfitting to the greatest extent. The CatBoost model performs well when dealing with data with a large number of categorical features. Therefore, in this study, the CatBoost model was trained to extract the relationships between various influencing factors and carbon density, achieve aboveground, underground, and soil carbon density level classification, and verify the classification accuracy of various carbon densities based on existing data.

[0057] Step 4: Based on the constructed training set and validation set, train the RNN model that is good at feature mining to fit the implicit relationship between the raster carbon density value and the influencing factors, and impose spatio-temporal constraints on the RNN by combining the prediction results of the ARIMA model and the CatBoost model to achieve high-precision carbon density value evaluation.

[0058] Carbon density estimation: The RNN model usually consists of five parts: an input layer, a hidden layer, a recurrent structure, a hidden state, and an output layer. Among them, the hidden layer is the key part of the model, responsible for extracting the feature information of the data and circulating and transmitting it in the form of a hidden state. Based on the training dataset, including aboveground, underground, and soil carbon density data and six types of influencing factor data, the RNN model was trained to fit the functional relationship between carbon density and influencing factors, impose spatio-temporal constraints on the RNN model with the carbon density mean predicted by the ARIMA model and the carbon density level predicted by the CatBoost model, and achieve carbon density estimation from 2000 to 2020 based on the influencing factors from 2000 to 2020. The formula is as follows:

[0059] C prei =f(X1X2…X n )

[0060]

[0061] In the formula: C prei is the result of the RNN model predicting the normalized value of the carbon density of the i-th raster cell, f is the fitting function of the RNN model, X i is the independent variable, that is, the value of the influencing factor, C i is the true carbon density value of the i-th raster cell, C dm is the mean carbon density of the j-th carbon pool predicted by the ARIMA model, C dj is the mean carbon density of the j-th carbon pool in the base year, B max and B min are the maximum and minimum values of the carbon density classification obtained by the CatBoost model.

[0062] Step 5: Improve the estimation parameters of the inVSET model according to the carbon density values, make up for its deficiency of relying too much on land use types for carbon storage assessment, fully consider various influencing factors, improve its ability to identify carbon storage differences at fine scales, realize the estimation of regional terrestrial ecosystem carbon storage, and evaluate the annual carbon sink capacity based on the carbon accumulation rate index;

[0063] Estimation and carbon sink capacity evaluation using the inVSET model: The carbon storage of regional terrestrial ecosystems mainly includes four categories: aboveground, underground, soil, and dead organic matter carbon storage. The traditional InVEST model's carbon storage accounting module mainly calculates regional carbon storage using the carbon density coefficients of land use type data. This method overly relies on the classification accuracy of land use data, cannot distinguish the carbon density differences of the same land use type, and is difficult to specifically analyze the changes in ecosystem carbon storage of different subtypes, seriously affecting the accuracy of regional carbon storage estimation. In this study, based on the multi-model coupling method, the carbon density values of the four carbon pools of various sub-ecosystems are estimated, and the parameters of the carbon storage accounting module of the InVEST model are optimized to the grid cell level, enabling it to identify the carbon density value of each grid, no longer restricted by the land use carbon coefficient. The accounting formula is as follows:

[0064] C = C above + C below + C soil + C dead

[0065]

[0066] In the formula, C is the total carbon storage of the terrestrial ecosystem, C above is the total aboveground carbon storage, C below is the total underground carbon storage, C soil is the total soil carbon storage, C dead is the total carbon storage of dead organic matter, C i is the carbon storage value of a certain grid, C d is the carbon density value of a certain grid, S k is the area of the grid cell, n is the total number of grids in the study area, and k is the serial number of the current grid.

[0067] Step 6: Refer to the ecosystem classification standard, divide the terrestrial ecosystem into seven sub-ecosystems: forest, grassland, cultivated land, settlement, desert, water body, and others. Use the Geodetector model to analyze the impacts of human activities and natural factors on the carbon sink capacity, explore the driving mechanism of carbon sink capacity changes, and provide data and theoretical support for formulating low-carbon development policies.

[0068] Analysis of the Driving Mechanism of Carbon Sink Capacity: According to land use data, the research classifies terrestrial ecosystems into seven categories, namely farmland ecosystem, forest ecosystem, grassland ecosystem, water body and wetland ecosystem, settlement ecosystem, desert ecosystem, and other ecosystems, using the ecosystem classification system; evaluates the carbon sink capacity of regions based on the carbon accumulation rate; and analyzes the driving mechanisms of human activities and natural environment on the change of carbon sink capacity based on the Geodetector model. The formula is as follows.

[0069]

[0070] In the formula: CAR k is the carbon accumulation rate of the k-th grid, ΔC k is the change in carbon storage of the k-th grid within time T, and ΔT is the time change. Within a certain period, a positive carbon accumulation rate indicates a carbon sink area, a negative value indicates a carbon source area, and the larger the positive value, the stronger the carbon sink capacity.

[0071] The Geodetector model is a method for geographical data mining and spatial analysis. It uses single-factor detection and interaction detection in the Geodetector to analyze the relationship between the carbon sink capacity (Y) and its driving factors (X). The larger the resulting value q, the greater the explanatory strength of the driving factor for the spatial variability of the carbon sink capacity. The calculation formula of the Geodetector is as follows:

[0072]

[0073] In the formula: q is the influence degree of the factor, N, N h are the total number of samples and the total number of samples in the h-th layer respectively, δ 2 , are the variances of Y values in the whole region and the h-th layer respectively, and SSW and SST are the sum of variances within layers and the total variance of the whole region respectively. The value range of q is [0, 1]. The larger the value, the stronger the explanatory power of the influencing factor for the spatio-temporal differentiation of the carbon sink capacity, and vice versa. Example:

[0074] A method for evaluating the carbon sink capacity of regional terrestrial ecosystems; First, six main influencing factors are selected by comprehensively considering existing research results. Based on the influencing factor data, verification data, and statistical data, an ARIMA-CatBoost-RNN coupled model is trained to estimate the average carbon density of four types of carbon pool grid cells. Then, according to the carbon storage classification system of the inVEST model, the terrestrial ecosystem is divided into four parts: aboveground biomass carbon pool, underground biomass carbon pool, soil carbon density carbon pool, and dead organic matter carbon pool. The aboveground biomass carbon pool refers to the total carbon stored in the biomass of aboveground animals and plants in the ecosystem. The underground biomass carbon pool refers to the carbon stored in the biomass of the underground part of the ecosystem. The soil carbon density carbon pool refers to the total amount of organic carbon and inorganic carbon in the soil. The dead organic matter carbon pool includes the total amount of withered and decayed biological carbon above and below the ground. Secondly, based on multiple grid cell-level carbon density parameters, the carbon storage assessment module of the inVEST model is optimized to achieve high-precision estimation of regional terrestrial ecosystem carbon storage and assessment of carbon sink capacity. Finally, based on geographical detectors, the impacts of human activities and natural factors on carbon sink capacity are analyzed, and the driving mechanism of carbon sink capacity change is explored. The spatial resolution variability of the estimation results of the present invention can generate grid values of influencing factors extracted from vector fishing nets at different spatial scales according to specific requirements and input them into the model. The model can input estimation results at multiple spatial scales according to the resolution of the input grid values. By extracting data based on a 100m*100m fishing net, the carbon density distribution result with a spatial resolution of 100m can be estimated. By extracting data based on a 10KM*10KM fishing net, the carbon density distribution result with a spatial resolution of 10KM can be estimated. Based on the inVEST model and combined with carbon density parameters estimated from influencing factor data at different times, the carbon storage distribution of long time series can be calculated to achieve the assessment of terrestrial ecosystem carbon storage at multiple spatio-temporal scales.

[0075] As Figure 2 shown, it is the construction process of multi-model coupling in the present invention. The invention is based on the statistical data of aboveground, underground, and soil carbon density in the region from 2000 to 2014. The ARIMA model, which is good at time series prediction, is used to estimate the average annual carbon density of aboveground, underground, and soil biomass from 2015 to 2020 respectively. Based on multiple influencing factor data and carbon density data, the CatBoost model, which is good at handling classification tasks, is used to realize the classification of carbon density levels based on influencing factors, and the research area is divided into three types of areas with different carbon densities to reduce the impact of spatial heterogeneity on subsequent model training. The RNN model, which is good at feature mining, is used to fit the implicit relationship between grid point carbon density values and influencing factors, and combined with the time series results of the ARIMA model and the classification results of the CatBoost model, the carbon density values of different types and different area grid cells are estimated to achieve high-precision long-time series estimation of carbon density.

[0076] AsFigure 3 As shown, the estimation results of the coupling model indicate that the carbon density distribution in Xinjiang region generally shows a state of high in the north and low in the south. Most of the carbon in northern Xinjiang is concentrated in areas with high forest vegetation coverage such as Yili, Tacheng, and Altay. The carbon density of soil, aboveground, underground, and dead organic matter in these areas is significantly higher than that in other areas. However, the desertification in Junggar and Tarim Basins is relatively severe, resulting in a lower carbon density. Generally speaking, the carbon density from 2000 to 2020 shows a trend of first decreasing and then increasing. The areas with the most obvious fluctuations in carbon density values are concentrated in the northeast of Xinjiang with high carbon density. The spatial distribution change of carbon density on an annual basis is not obvious. Most areas with high carbon density will maintain a high carbon level for a long time, and it is also very difficult for areas with low carbon density to exceed the high-density areas during the evolution process. Among them, a is the evaluation result map of the spatial distribution of carbon storage in Xinjiang region in 2000, and the size of carbon storage values gradually increases from green to red; b is the evaluation result map of the spatial distribution of carbon storage in Xinjiang region in 2005, and the size of carbon storage values gradually increases from green to red; c is the evaluation result map of the spatial distribution of carbon storage in Xinjiang region in 2010, and the size of carbon storage values gradually increases from green to red; d is the evaluation result map of the spatial distribution of carbon storage in Xinjiang region in 2015, and the size of carbon storage values gradually increases from green to red; e is the evaluation result map of the spatial distribution of carbon storage in Xinjiang region in 2020, and the size of carbon storage values gradually increases from green to red.

[0077] In terms of the proportion of various carbon densities, soil is the main carbon pool in Xinjiang region, accounting for about 55%-61% of the overall carbon storage; followed by underground biomass carbon storage, accounting for 27%-33%, the third is aboveground biomass carbon storage, accounting for 8%-11%, and finally, the carbon storage of dead organic matter only accounts for 3%. The estimation results of this study are highly consistent with the spatio-temporal distribution of the carbon storage results in Xinjiang region released by NASA in 2010, indicating that the model estimation results have high accuracy. The research results comprehensively consider the influence of multiple factors on the spatio-temporal evolution of carbon storage, breaking through the limitation of the traditional model's over-reliance on carbon density parameters of land use. Compared with the traditional model, it has the advantages of easy data acquisition and high evaluation accuracy, and has practical promotion value.

Claims

1. A method for estimating the carbon storage of regional terrestrial ecosystems with multi-model coupling, characterized in that, Including the following steps: Step 1: Multi-source data fusion: Collect remote sensing data across the country to generate six types of impact factor data, carbon density statistical data, and NASA carbon stock remote sensing data to construct a model training and validation dataset. The six types of impact factor data include annual average precipitation (PRS), temperature (TEM), soil moisture (SM), normalized difference vegetation index (NDVI), net primary productivity (NPP), and land use data over consecutive years; the carbon density statistical data includes the statistical data of the carbon density of the whole region, above and below ground, and in soil in the study area over many years; the NASA carbon stock remote sensing data includes the raster values of the carbon stocks of the below ground, above ground, soil, and dead organic matter in the study area in a certain year. Step 2: Prediction of annual average carbon density parameters: Based on the carbon density survey data of above and below ground and soil biomass in different regions of the country collected, calculate its annual average carbon density; use the autoregressive integrated moving average model ARIMA to predict the annual average carbon density in the future based on the carbon density mean over continuous time; use d to represent the order of differencing, and use SPSS data analysis software to determine that the order of differencing of the data d = 1; based on the Akaike Information Criterion (AIC) and the Bayesian Information Criterion (BIC), traverse the values of the number of autoregressive terms p and the number of moving terms q, and select the minimum values of the number of autoregressive terms p and the number of moving terms q as the best fitting parameters p = 5, q = 1 to construct an ARIMA model. The formula is as follows: C dm = φ1Y t-1 + φ2Y t-2 + … + φ p Y t-p + e t - θ1e t-1 - θ2e t-2 - … - θ q e t-q Where: C dm is the prediction result, p is the number of autoregressive terms, q is the number of moving terms, e is the natural constant, t is the sample time, φ is the autoregressive coefficient, and θ is the moving coefficient; Use the training dataset to train the ARIMA model. During training, the ARIMA model is trained iteratively according to the number of moving terms. During the iterative process, calculate the error and adjust the parameters to gradually reduce the error, making the predicted value gradually approach the true value. After multiple iterations, fit the variation law of carbon density over time and estimate the annual average carbon density in the next few consecutive years in the training dataset. Step 3: Based on the impact factor data, use the CatBoost model, which is good at handling classification tasks, to achieve carbon density level classification based on impact factors: First, determine the carbon density classification system and divide the normalized carbon density data into three levels: low, medium, and high; then perform a spatial overlay of the impact factor data and the carbon density data to obtain the corresponding tabular data of the six types of impact factors and carbon density levels at the same location. Based on the CatBoost model, obtain the spatial feature relationship between the impact factor data and the carbon density level, and achieve carbon density level classification based on the six types of impact factor data. Step 4: Estimate the carbon density value using the RNN model: Extract the characteristic information of the influencing factor data and carbon density data using the hidden layer of the RNN model, and perform cycling and transmission in the form of hidden states; Use the training dataset to train the RNN model to fit the functional relationship between carbon density and influencing factors, rasterize the grid units in the study area, and calculate the normalized carbon density value (0 - 1) of each grid unit in the study area; Use the average carbon density predicted by the ARIMA model and the carbon density level predicted by the CatBoost model to impose spatio-temporal constraints on the prediction results of the RNN model to obtain the true carbon density value formula of each grid unit in the study area as follows: C prei = f(X1X2…X6) Where: C prei is the result of the RNN model predicting the normalized value of the carbon density of the i-th grid cell, f is the fitting function of the RNN model, X1 - X6 are the independent variables, i.e., the values of the influencing factors, C di is the true carbon density value of the i-th grid cell, C dm is the average carbon density of the j-th carbon pool predicted by the ARIMA model, C dj is the average carbon density of the j-th carbon pool in the base year, B max and B min are the maximum and minimum values of the carbon density classification obtained by the CatBoost model; Repeat Step 2 - Step 4 to calculate the aboveground, underground, and soil biomass carbon densities respectively; Step 5: Calculate the carbon storage in the study area: Deduce the carbon density value of dead organic matter based on the aboveground, underground, and soil carbon density values estimated by the RNN model. Use these four types of carbon density values as the carbon storage parameters of the inVSET model, and use the inVSET model to calculate the aboveground, underground, soil, and dead organic matter carbon storages of each grid unit in the study area respectively, realizing the estimation of the terrestrial ecosystem carbon storage at the grid unit level, making up for the deficiency of the traditional method that overly relies on land use types for carbon storage assessment, fully considering various influencing factors, improving its ability to identify carbon storage differences at fine scales, and realizing the estimation of the regional terrestrial ecosystem carbon storage. The formula is as follows: C=C above +C below +C soil +C dead where C is the total carbon storage of the terrestrial ecosystem, C above is the total aboveground carbon storage, C below is the total belowground carbon storage, C soil is the total soil carbon storage, C dead is the total carbon storage of dead organic matter, C i is the carbon storage value of a certain grid, C d is the carbon density value of a certain grid, S k is the area of the grid cell, n is the total number of grid cells in the study area, and k is the serial number of the current grid cell; Use the inVEST model to evaluate the carbon sink capacity of the study area: Based on the interannual change of the carbon storage in the study area, calculate the carbon accumulation rate between different years, and use the carbon accumulation rate index to evaluate the interannual carbon sink capacity. The formula is as follows; Where: CAR k is the carbon accumulation rate of the k-th grid, ΔC k is the change in the carbon storage of the k-th grid within time T, and ΔT is the change in time. Within the time interval, if the carbon accumulation rate is positive, it is a carbon sink area; if it is negative, it is a carbon source area. The larger the positive value, the stronger the carbon sink capacity of the study area; Attribution analysis of carbon sink changes: Refer to the ecosystem classification standard, divide the terrestrial ecosystem into seven sub-ecosystems: forest, grassland, cultivated land, settlement, desert, water body, and others. Use the Geodetector model to analyze the impacts of human activities and natural factors on the carbon sink capacity, and explore the driving mechanism of carbon sink capacity changes, providing data and theoretical support for formulating low-carbon development policies.

2. The method for estimating the regional terrestrial ecosystem carbon storage by multi-model coupling according to claim 1, wherein Before the multi-source data fusion process in Step 1, use MATLAB and ArcGis software to perform operations such as calculating the annual average value, projection, resampling, and clipping on the remote sensing data and verification data used as influencing factors, and unify the coordinate systems of all remote sensing data to WGS1984 and the spatial resolution to 1KM; Perform processing on the carbon density statistical data such as unit normalization, extracting the carbon density data within the study area range, screening out outliers, and filtering noise data; Map all attributes to the interval of (0 - 1) to avoid the influence of data with different value ranges on model training; Use the fishnet tool and zonal statistics tool to rasterize the remote sensing data according to the 1KM * 1KM vector grid and convert it into tabular data, and construct the training dataset and verification dataset in a ratio of 7:

3.

3. The method for estimating regional terrestrial ecosystem carbon storage by multi-model coupling according to claim 2, wherein In the process of estimating and evaluating the carbon sequestration capacity using the inVSET model: The study calculated the carbon densities of aboveground, underground, soil, and dead organic matter carbon in each grid cell of the study area based on a multi-model coupling method, obtained the carbon density parameters at the grid cell level, recommended the estimation of carbon storage at the traditional land use type scale to the grid cell scale, fully considered the impacts of driving factors such as precipitation, temperature, and population density on carbon storage, broke the over-reliance of traditional methods on land use data, and achieved continuous long-term dynamic carbon storage assessment, improving the accuracy and efficiency of carbon storage estimation.

4. The method for estimating the carbon storage of regional terrestrial ecosystems with multi-model coupling according to claim 1, wherein, The analysis method of the driving mechanism of carbon sequestration capacity is as follows: The study classified terrestrial ecosystems into seven categories, namely farmland ecosystem, forest ecosystem, grassland ecosystem, water body and wetland ecosystem, settlement ecosystem, desert ecosystem, and other ecosystems according to the ecosystem classification system based on land use data; analyzed the driving mechanisms of human activities and natural environment on the change of carbon sequestration capacity from two perspectives of single-factor detection and interaction detection based on the Geodetector model. Multiple grid cells in the above seven ecosystem types were randomly selected as samples N respectively, and the Geodetector model was used to conduct driving mechanism analysis by stratifying according to the impact factor category h = 6. The formula is as follows: where: q is the degree of influence of the above influencing factor on carbon storage, and the value range of q is [0, 1]. The larger the value, the stronger the explanatory power of the influencing factor for the spatio-temporal differentiation of carbon sink capacity, and vice versa. N represents the total number of samples, N h represents the total number of samples in layer h, δ 2 represents the variance of the carbon storage value of the entire region of the terrestrial ecosystem carbon storage, represents the variance of the carbon storage value in layer h, SSW represents the sum of the within-layer variances, and SST represents the total variance of the entire region of the terrestrial ecosystem carbon storage.

5. A computer device, characterized in that, It includes a processor and a memory. The processor is electrically connected to the memory. The memory is used to store instructions and data. The processor is used to execute the multi-model coupling regional terrestrial ecosystem carbon storage estimation method according to any one of claims 1-4.

Images