Machine Learning-Based Method for Predicting NEE in China's Terrestrial Ecosystems

Abstract

This invention discloses a machine learning-based method for predicting the Negative Energy Emission (NEE) of China's terrestrial ecosystems. Belonging to the interdisciplinary field of ecological remote sensing and machine learning, the method includes data quality control, feature engineering, model training, spatial prediction, and interpretability analysis. Based on the XGBoost model, this method conducts SHAP interpretability analysis, quantifying the contribution of each feature to the NEE prediction results using SHAP values, and visualizing the analysis using SHAP feature importance bar charts and bee colony diagrams. This invention achieves superior prediction accuracy with less than half the feature variables of existing technologies, overcoming the limitations of traditional black-box prediction, balancing resolution and computational efficiency, and providing methodological support for terrestrial ecosystem carbon flux simulation.

Description

Technical Field

[0001] This invention relates to the interdisciplinary field of ecological remote sensing and machine learning, specifically to a machine learning-based method for predicting the NEE of China's terrestrial ecosystem. Background Technology

[0002] Net Ecosystem Exchange (NEE) is a core indicator for quantitatively characterizing the intensity of carbon sources and sinks in ecosystems and revealing the processes and mechanisms of land-atmosphere carbon exchange: NEE < 0 indicates that the ecosystem is net carbon absorbed, acting as a carbon sink; NEE > 0 indicates that the ecosystem is net carbon released, acting as a carbon source. Accurately identifying the spatial and temporal variability and seasonal dynamics of NEE can provide data support for the spatial layout and effectiveness evaluation of ecological engineering projects in forestry, agriculture, and grassland management.

[0003] Currently, NEE prediction mainly relies on the Eddy Covariance technique. The FLUXNET2015 and ChinaFLUX flux observation networks have accumulated valuable measured data, but the uneven spatial distribution of stations and limited representativeness make it difficult to directly depict the continuous spatial pattern of NEE across the country. While traditional ecological process models can achieve spatial extrapolation, they suffer from complex mechanisms, numerous parameters, difficult calibration, and limited regional applicability. In recent years, machine learning, with its powerful nonlinear fitting capabilities, multi-source data fusion capabilities, and generalization abilities, has demonstrated significant advantages in ecosystem carbon flux simulation.

[0004] Wu Meiling et al. proposed a machine learning-based method for predicting carbon flux in China's terrestrial ecosystems and analyzing its response to land use change. Based on data from 28 ChinaFLUX stations, ERA-5 meteorological reanalysis data, and remote sensing data on vegetation cover and normalized difference vegetation index (NDE), terrestrial evapotranspiration and gross primary productivity (GPLP), and land use, they constructed a high-dimensional feature set containing 24 characteristic variables. They employed four machine learning models—random forest, XGBoost, LightGBM, and support vector machine—for carbon flux prediction. After model comparison, random forest was selected as the optimal model. The spatiotemporal distribution pattern and seasonal variation characteristics of carbon sinks were analyzed using this model. However, in practical applications, this model employs a high-dimensional feature-driven strategy, incorporating 24 characteristic variables covering boundary layer height, cloud cover, total ozone, various components of evapotranspiration, and leaf area index of both high and low vegetation. Variable calculations are complex, data sources are scattered, and the acquisition threshold is high. Furthermore, redundancy exists among the features, increasing the workload of data preprocessing and limiting the rapid deployment and promotion of the model in practical applications. In addition, the model cannot quantify the contribution intensity, direction of action, nonlinear effects and threshold characteristics of each driving factor to the NEE prediction results, making it difficult to support in-depth analysis of carbon cycle mechanisms and scientific decision-making in ecological management. Summary of the Invention

[0005] To address the technical problems of high-dimensional feature-driven strategies, such as high data acquisition thresholds, computational complexity, feature redundancy, and lack of quantitative analysis of the driving factor mechanism, this invention provides a machine learning-based method for predicting the NEE of China's terrestrial ecosystem.

[0006] To achieve the above objectives, the present invention adopts the following technical solution:

[0007] A method for predicting net CO2 exchange in terrestrial ecosystems includes the following steps:

[0008] S1, Data quality control steps: Obtain monthly statistical data of FLUXNET2015 global flux station observations. Based on the latitude and longitude information of the stations, only the stations in the Northern Hemisphere are retained. All stations are merged into one table. The isolated forest algorithm is used to perform joint anomaly detection and removal on the NEE observations of each vegetation type and the corresponding GPP and RECO feature variables to obtain a high-quality training sample set.

[0009] S2, Feature Engineering Steps: Combining ecological mechanisms, data availability, and computational efficiency, a feature set of driving factors with clear physical meaning is constructed. This feature set contains 10 core features, which are composed of the following four categories:

[0010] 1) Spatial and geographical factors: longitude, altitude and soil organic carbon. Longitude and altitude reflect the control of macro-geographical patterns on the redistribution of water and heat conditions. Soil organic carbon is closely related to soil respiration and is an important variable characterizing the carbon flux of the ecosystem.

[0011] 2) Meteorological factors: month, temperature, shortwave downwave radiation, longwave downwave radiation and total precipitation. Month is used as a periodic time variable to capture the annual seasonal rhythm of carbon flux. Temperature and shortwave radiation directly drive the intensity of photosynthesis and respiration. Longwave radiation reflects the heat exchange process between the surface and the atmosphere. Precipitation determines the availability of water supply.

[0012] 3) Vegetation structure factors: leaf area index and the normalized average carbon flux of each vegetation type in FLUXNET2015. The leaf area index represents the ratio of the total leaf area to the land area per unit land area, and is a key indicator for quantifying canopy photosynthetic capacity and vegetation productivity.

[0013] 4) Interaction features: To enhance the model's ability to capture nonlinear relationships and synergistic effects of factors, six sets of interaction terms were constructed, including the product of air temperature and leaf area index, the product of shortwave radiation and leaf area index, the product of precipitation and leaf area index, the product of month and air temperature, the product of air temperature and precipitation, and the product of altitude and standardized vegetation type.

[0014] The StandardScaler algorithm is used to standardize the features, scaling the data to a standard distribution with a mean of 0 and a variance of 1, thus eliminating dimensional differences.

[0015] S3, Model Training Steps: Using XGBoost as the prediction model, under the constraints of uniform random seed, uniform training / test set partitioning strategy, uniform number of iterations, uniform maximum depth, and uniform learning rate, 10-fold cross-validation is used to evaluate the model performance. The dataset is randomly divided into 10 equal parts, with 9 parts used as the training set and 1 part as the test set, and the cycle is repeated 10 times. The average result is then taken. The three classic regression evaluation indicators, namely the coefficient of determination, root mean square error, and mean absolute error, are selected. The optimal model parameters are determined by automatic Bayesian parameter tuning.

[0016] S4, Spatial Prediction Steps: Input the gridded driving factor data covering the target area into the trained XGBoost model. The model calculates the NEE prediction value grid by grid to generate a monthly NEE spatial distribution dataset for the target area. Based on the monthly NEE prediction results, calculate the multi-year average NEE for each season and draw a spatial distribution map of the seasonal average NEE.

[0017] S5, Interpretability Analysis Steps: Conduct SHAP interpretability analysis based on the XGBoost model, quantify the contribution of each feature to the NEE prediction results using SHAP values, and perform visualization analysis using SHAP feature importance bar charts and bee colony diagrams.

[0018] Furthermore, the FLUXNET2015 dataset in S1 includes 11 vegetation types: evergreen coniferous forest, grassland, deciduous broad-leaved forest, farmland, permanent wetland, sparse shrubland, mixed forest, evergreen broad-leaved forest, dense shrubland, savanna, forested steppe, and snow. The proportion of isolated forest anomalies removed ranges from 4.6% to 8.0% across all ecosystem types.

[0019] Furthermore, the soil organic carbon data mentioned in S2 is from the World Soil Database, and the leaf area index is from MODIS.

[0020] Furthermore, the unified constraints described in S3 are: random seed set to 42, maximum depth to 8, learning rate to 0.1, number of iterations to 200, and subsampling rate to 0.8.

[0021] Furthermore, the feature is that the optimal parameters of XGBoost after automatic tuning of the Bayesian parameters in S3 are: number of iterations 181, maximum depth 9, learning rate 0.0874, and subsampling rate 1.0.

[0022] Furthermore, the feature is that the XGBoost model described in S3 has an average coefficient of determination of 0.7404 and a root mean square error of 1.0463 gC·m during 10-fold cross-validation. – ²·d – ¹, the mean absolute error is 0.6460 gC·m – ²·d – ¹.

[0023] Furthermore, the spatial resolution of the gridded driving factor data described in S4 is 0.05° × 0.05°.

[0024] Compared with the prior art, the present invention has the following advantages:

[0025] By adopting a "simplified interaction-driven" feature engineering strategy, only 10 core features are selected and 6 sets of interaction items with clear physical meaning are constructed. The number of feature variables is less than half of that in existing technologies, the data acquisition threshold is significantly reduced, the computational complexity is reduced, and better accuracy is achieved with less than half the number of feature variables.

[0026] Using XGBoost as the prediction model, and after 10-fold cross-validation and Bayesian parameter tuning under uniform constraints, the model achieved an average coefficient of determination of 0.7404 and a root mean square error of 1.0463 gC·m⁻¹ during 10-fold cross-validation. – ²·d – ¹, the mean absolute error is 0.6460 gC·m – ²·d – ¹, its accuracy and stability are significantly better than other models;

[0027] We introduce the SHAP interpretability model, identify the dominant factors through SHAP global importance ranking, and reveal the direction, intensity and distribution of each factor through SHAP bee colony graph. We quantitatively verify that the contribution of interaction terms is higher than that of single factors, and can reveal the threshold effect and saturation characteristics of each driving factor. Attached Figure Description

[0028] Figure 1 A comparison chart of NEE simulation results for six machine learning models.

[0029] Figure 2 Rank SHAPs by global importance.

[0030] Figure 3 This is a beehive diagram showing the direction, intensity, and distribution of the effects of each factor on NEE. Detailed Implementation

[0031] Example 1: A machine learning-based method for predicting the NEE of China's terrestrial ecosystems. The specific implementation steps are as follows:

[0032] S1, Data Quality Control: Monthly statistical data from the FLUXNET 2015 global flux stations were obtained. Based on the station latitude and longitude information, only stations in the Northern Hemisphere were retained. All stations were merged into one table. The isolated forest algorithm was used to perform joint anomaly detection and removal on the NEE observations and corresponding GPP and RECO feature variables for each vegetation type, resulting in a high-quality training sample set. Of the 177 stations, 48 ​​were evergreen coniferous forests, 31 were grasslands, 24 were deciduous broad-leaved forests, 20 were farmland, 19 were permanent wetlands, 13 were sparse shrublands, 8 were mixed forests, 7 were evergreen broad-leaved forests, 2 were dense shrublands, 2 were savannahs, 2 were forested steppes, and 1 was snow and ice. After processing, the number of stations for each vegetation type remained unchanged. The total number of records removed was 5.2%, with snow and ice and savannahs being removed at 0%, forested steppes having the highest removal rate at 8.0%, dense shrublands at 6.7%, and mixed forests at 4.6%. A total of 14,861 records were ultimately used as the training sample.

[0033] S2, Feature Engineering: Constructing a streamlined feature set containing 10 core features:

[0034] 1) Spatial and geographical factors: longitude, altitude, and soil organic carbon;

[0035] 2) Meteorological factors: month, temperature, shortwave radiation, longwave radiation, precipitation;

[0036] 3) Vegetation structure factors: leaf area index, standardized value of vegetation type;

[0037] 4) Interactive features: product of air temperature and leaf area index, product of shortwave radiation and leaf area index, product of precipitation and leaf area index, product of month and air temperature, product of air temperature and precipitation, and product of altitude and standardized vegetation type value.

[0038] In the aforementioned driving factor data, the standardized values ​​of longitude, altitude, month, temperature, shortwave radiation, longwave radiation, precipitation, and vegetation type were obtained from the FLUXNET 2015 global flux observation monthly data set; soil organic carbon data were obtained from the World Soil Database; and leaf area index was obtained from MODIS. All data were uniformly processed to the same time scale (monthly mean) and spatial benchmark as NEE observations. StandardScaler was used to standardize the features, scaling the data to a standard distribution with a mean of 0 and a variance of 1.

[0039] S3, Model Training: XGBoost was used as the prediction model, trained under uniform constraints: random seed set to 42, maximum depth to 8, learning rate to 0.1, number of iterations to 200, and subsampling rate to 0.8. Ten-fold cross-validation was used to evaluate model performance. The dataset was randomly divided into 10 equal parts, with 9 parts used as the training set and 1 part as the test set, repeated 10 times, and the average result was taken. The coefficient of determination, root mean square error (RMSE), and mean absolute error (MAO) were selected as the three classic regression evaluation metrics. After automatic Bayesian parameter tuning, the optimal parameters were determined to be: number of iterations to 181, maximum depth to 9, learning rate to 0.0874, and subsampling rate to 1.0. At this point, the model's 10-fold average coefficient of determination reached 0.746. The coefficient of determination on the training set was 0.959, the 10-fold cross-validation average coefficient of determination was 0.7404, and the root mean square error was 1.0463 gC·m. – ²·d – ¹, the mean absolute error is 0.6460 gC·m – ²·d – ¹, the standard deviation of the coefficient of determination is 0.0161, the standard deviation of the root mean square error is 0.0429, and the standard deviation of the mean absolute error is 0.0163.

[0040] S4, Spatial Prediction: A 0.05°×0.05° gridded driving factor data covering the entire territory of China is input into a trained XGBoost model. NEE predictions are calculated grid-by-grid, generating a monthly NEE spatial distribution dataset for China from 2011 to 2020. Meteorological seasonal divisions are used: March to May is spring, June to August is summer, September to November is autumn, and December to February of the following year is winter. Based on the monthly NEE prediction results, the multi-year average NEE for each season (10-year average from 2011 to 2020) is calculated, and multi-year average and seasonal NEE spatial distribution maps for spring, summer, autumn, and winter are drawn. The multi-year average NEE spatial distribution in China from 2011 to 2020 exhibits a pattern of "strong carbon sink in the southeast, weak carbon sink / carbon neutral in the northwest."

[0041] Strong carbon sink region (NEE < -2 gC·m – ²·d – ¹): It is mainly distributed in the humid and semi-humid southeastern region, including the Greater and Lesser Khingan Mountains and Changbai Mountain forest areas in the northeast, the subtropical and tropical forest areas south of the Qinling-Huaihe line, and the southwestern forest areas (southern Yunnan and western Guangxi).

[0042] Medium carbon sink region (-2 to -1 gC·m – ²·d – ¹): Mainly distributed in concentrated farmland areas such as the North China Plain and the middle and lower reaches of the Yangtze River Plain, as well as ecological transition zones such as the Loess Plateau and the Yunnan-Guizhou Plateau;

[0043] Weak carbon sink / carbon neutral zone (-1~0 gC·m)– ²·d – ¹): Covering the arid and semi-arid regions of Northwest China, most of the Qinghai-Tibet Plateau, and the grasslands of central and eastern Inner Mongolia;

[0044] Carbon source region (NEE > 0 gC·m) – ²·d – ¹): Scattered in extremely arid desert areas, urban built-up areas, and some degraded grassland areas, accounting for less than 5% of the total area.

[0045] Seasonal NEE Dynamics: China's NEE exhibits a significant seasonal differentiation pattern, with overall carbon sequestration intensity showing the order of summer > spring > autumn > winter. In spring, as vegetation enters its growing season, carbon sequestration intensity gradually recovers in the southeast region, while the northeast and North China regions transition from carbon neutral to weak carbon sequestration. The Qinghai-Tibet Plateau remains predominantly carbon neutral. During summer, most areas of the country are carbon sequestration zones, with the southeastern humid region reaching its highest annual carbon sequestration intensity (NEE < -3 gC / m²). – ² d – ¹), carbon sequestration capacity significantly improved in the arid Northwest and the Qinghai-Tibet Plateau, with only the extreme desert region remaining carbon neutral. As the autumn vegetation growing season ended, overall carbon sequestration intensity declined, with the Southeast region maintaining moderate carbon sequestration, the Northeast and North China regions transitioning to weak carbon sequestration, and the Northwest region approaching carbon neutrality. Winter saw the weakest carbon sequestration capacity nationwide, with most areas in the Northeast, North China, and Northwest becoming carbon sources or carbon neutral, except for the tropical and subtropical forest regions of South China maintaining weak carbon sequestration.

[0046] S5 Interpretability Analysis: SHAP interpretability analysis was conducted based on the XGBoost model, using SHAP values ​​to quantify the contribution of each feature to NEE prediction results. The global importance ranking of SHAP showed that the product of shortwave radiation and leaf area index, shortwave radiation, standardized vegetation type value, longitude, and the product of temperature and leaf area index were the five dominant factors influencing NEE in China, with a cumulative contribution exceeding 60%.

[0047] Comparative Example 1: Compared with the prediction method disclosed in the paper by Wu Meiling et al., the existing technology uses 24 feature variables, including boundary layer height, low cloud cover, total cloud cover, total atmospheric ozone column, total column water, vegetation cover, normalized difference vegetation index, total primary productivity, vegetation transpiration, soil evaporation, canopy interception evaporation, water body ice and snow evaporation, 2-meter dew point temperature, downward surface solar radiation, downward surface thermal radiation, predicted albedo, high vegetation leaf area index, low vegetation leaf area index, evaporation, 10-meter zonal wind, 10-meter radial wind, surface pressure, and total precipitation. This invention uses only 10 core features and 6 sets of interaction terms, with less than half the number of feature variables as Wu Meiling et al. Wu Meiling et al. selected random forest as the optimal model, with a 10-fold cross-validation determination coefficient of 0.73 and a root mean square error of 35.09 g·(m²·month). –¹. This invention employs the XGBoost model, with a 10-fold cross-validation coefficient of determination of 0.7404 and a root mean square error of 1.0463 gC·m. – ²·d – ¹, achieving better prediction accuracy with less than half of the feature variables.

[0048] Comparative Example 2: Comparison of prediction performance with different machine learning models. S1-S2 are the same as in Example 1; S3 uses Artificial Neural Network (ANN), Convolutional Neural Network (CNN), Random Forest (RF), XGBoost, LightGBM, and Support Vector Machine (SVM) as prediction models, with the following parameter settings for each model:

[0049]

[0050] The models were trained under uniform constraints and evaluated using 10-fold cross-validation. Specific parameters and evaluation results for each model are shown in Table A; S4-S5 are the same as in Example 1. Comparison results are shown in the following table:

[0051]

[0052] It can be seen that gradient boosting models (XGBoost and LightGBM) significantly outperform other types of models in all evaluation metrics, demonstrating a strong ability to fit the complex ecological variable NEE. Among them, the XGBoost model after hyperparameter optimization performed best: its average R² reached 0.7404, the highest among all models; RMSE and MAE were 1.0463 gC·m⁻¹, respectively. – ²·d – ¹ and 0.6460 gC·m – ²·d – ¹, all of which are the minimum values ​​among all models. More importantly, the standard deviation of the XGBoost model is the smallest across all three metrics (R² standard deviation 0.0161, RMSE standard deviation 0.0429, MAE standard deviation 0.0163), indicating that the model is insensitive to the training / test set splitting method and has excellent stability and generalization ability.

[0053] LightGBM performed just behind XGBoost, with an R² of 0.7331 and an RMSE of 1.0606 gC·m. – ²·d – ¹, MAE is 0.6636 gC·m – ²·d –¹. Although both error metrics are slightly higher than XGBoost, they are still at a very high level. The performance difference between the two models may be due to XGBoost's more refined hyperparameter tuning and its regularization strategy being better adapted to this dataset.

[0054] The ANN model performed at a moderate level (R²=0.7013). While its prediction accuracy was better than traditional machine learning methods RF and SVM, it lagged behind gradient boosting models. Considering that the input feature dimension in this study was only 10 dimensions, the "depth" advantage of ANNs (adept at handling high-dimensional data such as images and sequences) could not be fully utilized. The performance of CNN (R²=0.6326) was slightly worse than that of ANN. This may be because 1D-CNN assumes local correlation between features, while the feature ranking in this study, although containing some physical meaning, does not have strong spatial local correlation, and the convolution operation failed to extract meaningful local patterns.

[0055] The Random Forest model performed moderately well (R²=0.6591), but its advantages lay in its robustness to noisy data and training efficiency. SVM performed the worst on this task (R² only 0.3422), and its error index was also significantly higher than other models. This result is not surprising, as SVM has inherent limitations when dealing with large-sample (the effective sample size in this study was relatively large), highly nonlinear, and noisy ecological data: ① It is sensitive to kernel function selection; linear kernels are difficult to fit nonlinear relationships, and nonlinear kernels (such as RBF) are difficult to tune and computationally expensive; ② It is sensitive to outliers; ③ It takes a long time to train on large-scale datasets.

[0056] Figure 1 Scatter plots comparing the predicted and observed values ​​of six models on the test set are presented. It is clear that the scatter plots of XGBoost and LightGBM are closely distributed along a 1:1 diagonal, indicating a high degree of consistency between their predictions and observations. The scatter plots of ANN and RF are relatively dispersed, indicating some systematic bias. The scatter plots of CNN and SVM are even more discrete, especially showing larger prediction biases in low-value regions (strong carbon sink regions) and high-value regions (strong carbon source regions). Based on comprehensive analysis, this study selected XGBoost as the optimal model for subsequent NEE spatiotemporal prediction and SHAP interpretability analysis.

Claims

1. A machine learning-based method for predicting the Neighbor-Earth Ecosystem (NEE) of China's terrestrial ecosystems, characterized by... Includes the following steps: S1, Data quality control steps: Obtain monthly statistical data of FLUXNET2015 global flux station observations. Based on the latitude and longitude information of the stations, only the stations in the Northern Hemisphere are retained. All stations are merged into one table. The isolated forest algorithm is used to perform joint anomaly detection and removal on the NEE observations of each vegetation type and the corresponding GPP and RECO feature variables to obtain a high-quality training sample set. S2, Feature Engineering Steps: Combining ecological mechanisms, data availability, and computational efficiency, a feature set of driving factors with clear physical meaning is constructed. This feature set contains 10 core features, which are composed of the following four categories: 1) Spatial and geographical factors: longitude, altitude and soil organic carbon. Longitude and altitude reflect the control of macro-geographical patterns on the redistribution of water and heat conditions. Soil organic carbon is closely related to soil respiration and is an important variable characterizing the carbon flux of the ecosystem. 2) Meteorological factors: month, temperature, shortwave downwave radiation, longwave downwave radiation and total precipitation. Month is used as a periodic time variable to capture the annual seasonal rhythm of carbon flux. Temperature and shortwave radiation directly drive the intensity of photosynthesis and respiration. Longwave radiation reflects the heat exchange process between the surface and the atmosphere. Precipitation determines the availability of water supply. 3) Vegetation structure factors: leaf area index and the normalized average carbon flux of each vegetation type in FLUXNET2015. The leaf area index represents the ratio of the total leaf area to the land area per unit land area, and is a key indicator for quantifying canopy photosynthetic capacity and vegetation productivity. 4) Interaction features: To enhance the model's ability to capture nonlinear relationships and synergistic effects of factors, six sets of interaction terms were constructed, including the product of air temperature and leaf area index, the product of shortwave radiation and leaf area index, the product of precipitation and leaf area index, the product of month and air temperature, the product of air temperature and precipitation, and the product of altitude and standardized vegetation type. The StandardScaler algorithm is used to standardize the features, scaling the data to a standard distribution with a mean of 0 and a variance of 1, thus eliminating dimensional differences. S3, Model Training Steps: Using XGBoost as the prediction model, under the constraints of uniform random seed, uniform training / test set partitioning strategy, uniform number of iterations, uniform maximum depth, and uniform learning rate, 10-fold cross-validation is used to evaluate the model performance. The dataset is randomly divided into 10 equal parts, with 9 parts used as the training set and 1 part as the test set, and the cycle is repeated 10 times. The average result is then taken. The three classic regression evaluation indicators, namely the coefficient of determination, root mean square error, and mean absolute error, are selected. The optimal model parameters are determined by automatic Bayesian parameter tuning. S4, Spatial Prediction Steps: Input the gridded driving factor data covering the target area into the trained XGBoost model. The model calculates the NEE prediction value grid by grid to generate a monthly NEE spatial distribution dataset for the target area. Based on the monthly NEE prediction results, calculate the multi-year average NEE for each season and draw a spatial distribution map of the seasonal average NEE. S5, Interpretability Analysis Steps: Conduct SHAP interpretability analysis based on the XGBoost model, quantify the contribution of each feature to the NEE prediction results using SHAP values, and perform visualization analysis using SHAP feature importance bar charts and bee colony diagrams.

2. The prediction method according to claim 1, characterized in that, The FLUXNET2015 dataset in S1 includes 11 vegetation types: evergreen coniferous forest, grassland, deciduous broad-leaved forest, farmland, permanent wetland, sparse shrubland, mixed forest, evergreen broad-leaved forest, dense shrubland, savanna, forested steppe, and snow. The proportion of isolated forest anomalies removed ranges from 4.6% to 8.0% across all ecosystem types.

3. The prediction method according to claim 1, characterized in that, The soil organic carbon data mentioned in S2 are from the World Soil Database, and the leaf area index is from MODIS.

4. The prediction method according to claim 1, characterized in that, The unified constraints described in S3 are: random seed set to 42, maximum depth to 8, learning rate to 0.1, number of iterations to 200, and subsampling rate to 0.

8.

5. The prediction method according to claim 4, characterized in that, The optimal parameters for XGBoost after automatic tuning of the Bayesian parameters described in S3 are: number of iterations 181, maximum depth 9, learning rate 0.0874, and subsampling rate 1.

0.

6. The prediction method according to claim 1, characterized in that, The XGBoost model described in S3 has an average coefficient of determination of 0.7404 and a root mean square error of 1.0463 gC·m⁻¹ after 10-fold cross-validation. – ²·d – ¹, the mean absolute error is 0.6460 gC·m – ²·d – ¹.

7. The prediction method according to claim 1, characterized in that, The spatial resolution of the gridded driving factor data described in S4 is 0.05° × 0.05°.

Images