Regional scale forest carbon sink estimation method based on fusion of remote sensing data and vorticity covariance data

Abstract

The invention belongs to the technical field of forest ecological system carbon cycle monitoring, and relates to a regional scale forest net ecological system carbon exchange estimation method fusing ground vortex covariance observation and multi-source remote sensing data. The problems of insufficient representativeness of flux monitoring forest types and limited precision of pure remote sensing estimation during regional scale estimation of forest carbon sink in a traditional method are solved. A membership function matrix is innovatively constructed by using fuzzy mathematics, EC measured data and multi-source remote sensing data are matched, and effective fusion of point and surface data is realized. According to the method, a regional forest carbon sink measurement model based on remote sensing data and a flux observation network is constructed, high-precision regional carbon sink estimation is realized, and single-point limitation is overcome; the spatial pattern (presenting the gradient characteristic of high southeast and low northwest) of the temperate forest carbon sink and the seasonal change are disclosed; it is determined that carbon sink annual fluctuation is mainly influenced by meteorological conditions and vegetation state changes; and an effective tool is provided for accurate evaluation of regional forest carbon sink and carbon management decision.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of ecological environment monitoring, in particular to a regional scale forest ecosystem carbon sink estimation method fusing eddy covariance observation data and remote sensing information and combining machine learning algorithm, which can be used for temperate forest ecosystem carbon flux monitoring and climate change research. BACKGROUND

[0002] Carbon sink is the ability of an ecosystem to absorb and store carbon dioxide (CO2) from the atmosphere, and is an important part of global carbon cycle research. Eddy covariance (EC) technology is a widely used carbon flux observation technology internationally, which can directly measure the net carbon exchange (NEE) between the site-scale ecosystem and the atmosphere. However, the spatial distribution of EC sites is sparse, which cannot meet the demand of large-scale regional carbon sink estimation. In recent years, the development of remote sensing technology and machine learning methods has provided new possibilities for carbon flux modeling at regional and even global scales. Remote sensing data can provide a wide range of vegetation, biophysical and meteorological parameter information, and machine learning algorithms (such as random forest, gradient boosting machine, etc.) can be used to establish a nonlinear mapping relationship between flux data and environmental factors, and to realize the spatial extrapolation of carbon flux. Although previous studies have preliminarily explored the feasibility of flux data and remote sensing data fusion modeling, there is still a significant lack of a systematic and interpretable regional carbon sink estimation framework, especially in the evaluation of flux tower observation meteorological data and remote sensing inversion meteorological data substitution relationship, variable importance analysis, model uncertainty explanation, etc. SUMMARY

[0003] The present application aims to provide a regional estimation method for forest ecosystem carbon sink based on the fusion of eddy covariance data and remote sensing information. EC technology can provide high-precision and continuous NEE observation data, but its representative range is limited. The forest types observed by the flux network are less than the forest distribution types in the region, making it difficult to be directly extrapolated to the regional scale. Simply relying on remote sensing data has wide coverage, but it has shortcomings in the accuracy and mechanism of carbon flux estimation. The existing methods fail to fully exploit and effectively fuse the complementary advantages of high-precision site-scale EC data and full-coverage remote sensing data, resulting in limited accuracy and reliability of the carbon flux estimation results. The method introduces a membership function matrix of fuzzy mathematics to fuse EC observation data and remote sensing data, and realizes high-precision spatio-temporal simulation of regional-scale forest ecosystem carbon flux (NEE) through random forest modeling and SHAP explanation analysis, providing support for carbon cycle scientific research and carbon neutral policy making.

[0004] The present application is achieved by the following technical means:

[0005] A regional-scale forest carbon sink estimation method based on remote sensing data and vorticity covariance data fusion, comprising the following steps:

[0006] S1: Region selection, determine the temperate forest in Northeast China (33°-54°N, 102°-135°E) as the main range;

[0007] S2: Data preparation, obtain and fuse ChinaFlux flux data, MODIS remote sensing data, TerraClimate meteorological data, unify the spatial resolution to 1km, and the time resolution to monthly scale;

[0008] S3: Construct a multi-source input variable set, including normalized difference vegetation index (NDVI), enhanced vegetation index (EVI), leaf area index (LAI), photosynthetically active radiation absorption ratio (FPAR), terrain factor, meteorological factor and spatial and temporal variables;

[0009] S4: Use the random forest algorithm to establish a multi-source data-driven CO2 flux prediction model, set the number of decision trees n_estimators of the model to 1000 and the maximum depth of the decision tree max_depth to 20, and use 5-fold cross-validation to evaluate the performance;

[0010] S5: Use SHAP (SHapley Additive exPlanations) value to explain the model results, determine the main control variables of carbon flux and their influence mechanism;

[0011] S6: Estimate the monthly forest carbon sink of the whole study area, output the distribution map of regional net ecosystem carbon exchange (NEE) from 2003 to 2022;

[0012] S7: Analyze the interannual variation trend, judge the regional forest carbon source / sink pattern, and evaluate the error and uncertainty.

[0013] The remote sensing data includes MCD15A3H (FPAR / LAI) and MOD13Q1 (NDVI / EVI) in MODIS products, and is processed by spatial resampling and flux tower data registration.

[0014] The random forest model input variables include: air temperature, total precipitation, shortwave radiation, NDVI, EVI, LAI, FPAR, DEM elevation, land cover type, site longitude and latitude, and month code, etc.

[0015] The model explanation analysis uses the SHAP (SHapley Additive exPlanations) method, outputs the global variable importance ranking and local dependence graph, to quantify the influence direction and strength of the variable on carbon sink prediction.

[0016] The predicted object is net ecosystem exchange (NEE) in gC / m 2 • month, the interpolation and prediction range covers missing values, low quality values and spatial expansion areas.

[0017] A regional forest carbon flux estimation system of a regional scale forest carbon sink estimation method based on remote sensing data and eddy covariance data fusion, comprising the following modules:

[0018] 1: Data management module: for downloading, preprocessing and fusing multi-source data;

[0019] 2: Modeling module: for training random forest model and evaluating accuracy;

[0020] 3: Model interpretation module: for calculating SHAP value and outputting variable explanation;

[0021] 4: Regional simulation module: based on the model, the carbon flux of the whole region is predicted;

[0022] 5: Visualization module: for generating carbon sink spatial graph, time trend graph and variable importance graph;

[0023] 6: The system running platform is based on Python language, and is realized by using scikit-learn, xarray, shap and the like.

[0024] A regional forest carbon flux estimation system of a regional scale forest carbon sink estimation method based on remote sensing data and eddy covariance data fusion, the forest carbon flux estimation result output by the system can be used for regional carbon balance research, carbon peak / carbon neutral policy support and ecosystem response analysis.

[0025] A regional forest carbon flux estimation system of a regional scale forest carbon sink estimation method based on remote sensing data and eddy covariance data fusion, the application of the system in the following aspects:

[0026] 1: Estimation and monitoring of regional scale forest carbon flux;

[0027] 2: Forest carbon sink simulation under different climate scenarios;

[0028] 3: Scientific support for carbon neutralization and carbon trading policy;

[0029] 4: Ecosystem management and remote sensing data integration evaluation.

[0030] The beneficial effects of the present application are:

[0031] 1. Using random forest feature importance and SHAP method, the contribution of different variables to NEE prediction and its nonlinear influence path are evaluated, the interpretability of model prediction mechanism is enhanced, and the accuracy and stability of NEE prediction by replacing flux tower meteorological data with remote sensing data are improved. The response mechanism between remote sensing variables and NEE is clarified.

[0032] 2. Combined with cross-validation residual analysis and SHAP global contribution value, the relationship between model prediction error and key variable fluctuation is evaluated, and the scientificity and robustness of carbon sink estimation are improved.

[0033] 3. A carbon sink estimation framework that can be applied to China's temperate forest ecosystem is established; the model has good interpretability and is easy to support policy and ecological management decisions. BRIEF DESCRIPTION OF DRAWINGS

[0034] Figure 1 is a regional distribution map and a flux site location distribution map.

[0035] Figure 2 is a modeling process diagram and a data processing process.

[0036] Figure 3 is a flux and remote sensing variable and NEE correlation analysis result diagram.

[0037] Figure 4 is a regional scale NEE estimation diagram and an annual variation trend diagram. DETAILED DESCRIPTION

[0038] In order to make the purpose, technical scheme and beneficial effects of the present application clearer and more explicit, the following will combine the drawings Figures 1-4 and examples to further illustrate the present application. It should be understood that these examples are only used to explain the present application, and are not used to limit the protection scope of the present application.

[0039] A regional scale forest carbon sink estimation method based on remote sensing data and eddy covariance data fusion, comprising the following steps:

[0040] S1: Region selection, determine the main range of China's temperate forest (33°-54°N, 102°-135°E);

[0041] S2: Data preparation, obtain and fuse ChinaFlux flux data, MODIS remote sensing data, TerraClimate meteorological data, unify the spatial resolution to 1km, and the time resolution to monthly scale;

[0042] S3: Construct a multi-source input variable set, including normalized difference vegetation index (NDVI), enhanced vegetation index (EVI), leaf area index (LAI), photosynthetically active radiation absorption ratio (FPAR), terrain factors, meteorological factors, and spatial and temporal variables;

[0043] S4: Use the random forest algorithm to establish a multi-source data-driven CO2 flux prediction model, set the number of decision trees n_estimators = 1000 and the maximum depth of the decision tree max_depth = 20, and use 5-fold cross-validation to evaluate the performance;

[0044] S5: Use SHAP (SHapley Additive exPlanations) values to explain the model results and determine the main control variables of carbon flux and their influence mechanism;

[0045] S6: Estimate the monthly forest carbon sink of the entire study area, and output the distribution map of regional net ecosystem carbon exchange (NEE) from 2003 to 2022;

[0046] S7: Analyze the interannual variation trend, judge the regional forest carbon source / sink pattern, and evaluate the error and uncertainty.

[0047] The remote sensing data includes MCD15A3H (FPAR / LAI) and MOD13Q1 (NDVI / EVI) in MODIS products, and is processed by spatial resampling and flux tower data registration.

[0048] The random forest model input variables include: air temperature, total precipitation, shortwave radiation, NDVI, EVI, LAI, FPAR, DEM elevation, land cover type, site latitude and longitude, and month code, etc.

[0049] The model explanation analysis uses the SHAP (SHapley Additive exPlanations) method, which outputs global variable importance ranking and local dependence graph to quantify the influence direction and strength of variables on carbon sink prediction.

[0050] The prediction object is net ecosystem exchange (NEE), with the unit of gC / m 2 ·month, the interpolation and prediction range covers missing values, low quality values and spatial expansion areas.

[0051] A regional forest carbon flux estimation system based on a regional scale forest carbon sink estimation method based on remote sensing data and eddy covariance data fusion, comprising the following modules:

[0052] 1: Data management module: used for downloading, preprocessing and fusing multi-source data;

[0053] 2: Modeling module: for training random forest model and evaluating accuracy;

[0054] 3: Model interpretation module: for calculating SHAP values and outputting variable explanations;

[0055] 4: Regional simulation module: based on the model to predict carbon flux in the whole region;

[0056] 5: Visualization module: for generating carbon sink spatial map, time trend chart and variable importance chart;

[0057] 6: The system running platform is based on Python language, using scikit-learn, xarray, shap and other libraries to realize.

[0058] A regional forest carbon flux estimation system of a regional scale forest carbon sink estimation method based on remote sensing data and vorticity covariance data fusion, the forest carbon flux estimation result output by the system can be used for regional carbon balance research, carbon peak / carbon neutral policy support and ecosystem response analysis.

[0059] A regional forest carbon flux estimation system of a regional scale forest carbon sink estimation method based on remote sensing data and vorticity covariance data fusion, the application of the system in the following aspects:

[0060] 1: Estimation and monitoring of regional scale forest carbon flux;

[0061] 2: Forest carbon sink simulation under different climate scenarios;

[0062] 3: Scientific support for carbon neutralization and carbon trading policy;

[0063] 4: Ecosystem management and remote sensing data integration evaluation.

[0064] Embodiment, data preparation and processing method as follows

[0065] 1: Obtain the monthly CO2 flux (NEE) and supporting meteorological data (temperature, precipitation, shortwave radiation, wind speed) of flux tower sites from 2003 to 2022 from ChinaFlux and AsiaFlux;

[0066] 2: Obtain MODIS remote sensing products (MCD15A3H and MOD13Q1), extract vegetation indices such as FPAR, LAI, NDVI, EVI, and obtain monthly meteorological data provided by TerraClimate;

[0067] 3: All variables are resampled to 1km resolution, monthly time series, and spatially registered and time aligned;

[0068] 4: Remove outliers and missing values, construct a complete sample dataset, including NEE, meteorological factors, remote sensing variables, terrain data, spatial information, etc.

[0069] The model construction and training method is as follows:

[0070] 1: Respectively using flux tower meteorological data and remote sensing inversion meteorological data as input, constructing two basic random forest models (RF Model(Tower), RF Model(RS)) for NEE prediction;

[0071] 2: Further integrate multi-source variables, including remote sensing vegetation factors, DEM, forest canopy height, land cover type and spatial and temporal variables, to build RF Model 1 and RF Model 2;

[0072] 3: The model adopts 5-fold cross-validation strategy for training and evaluation, and the average performance indicators (R 2 , RMSE) are calculated repeatedly;

[0073] The model interpretability analysis process is as follows:

[0074] 1: Use the built-in feature importance function of the random forest model to get the contribution ranking of each variable to the model prediction results;

[0075] 2: Introduce SHAP interpretation method to make local and global interpretation of model output, analyze the nonlinear relationship between variables such as shortwave radiation, precipitation, NDVI and NEE;

[0076] 3: Combine SHAP value distribution chart and dependence chart to further clarify the threshold interval and positive and negative effect mechanism of key variables, and provide reference basis for ecological management and model optimization.

[0077] The regional simulation and result analysis process is as follows:

[0078] 1: Use the trained best model (RF Model 2) to simulate the monthly spatial distribution of the entire study area (Northeast China forest region);

[0079] 2: Generate NEE spatial distribution maps for 2005, 2010, 2015 and 2020, and extract the average NEE values under different years and different ecological types;

[0080] 3: Analyze the interannual variation trend of NEE and calculate the regional carbon source-sink balance status;

[0081] 4: Verify the consistency of the model results with existing literature and national carbon flux statistical data to ensure the scientificity and reliability of regional estimation.

[0082] The system implementation is as follows:

[0083] 1: Data management module: processing multi-source data such as raw flux, remote sensing, and weather data;

[0084] 2: Modeling and training module: modeling based on random forest algorithm;

[0085] 3: Variable interpretation module: introducing SHAP toolkit for model transparency analysis;

[0086] 4: Regional simulation module: realizing the application of the model at regional scale;

[0087] 5: Visualization output module: generating spatial images, trend charts, importance rankings, and other results for display.

[0088] 6: The system is developed in Python language, and the core libraries include scikit-learn, xarray, shap, numpy, matplotlib, etc., supporting automatic processing and model deployment.

Claims

1. A regional scale forest carbon sink estimation method based on remote sensing data and vorticity covariance data fusion, characterized in that, Comprising the following steps: S1: Region selection, determine the main range of temperate forest in Northeast China (33°-54°N, 102°-135°E); S2: Data preparation, obtain and fuse ChinaFlux flux data, MODIS remote sensing data, TerraClimate meteorological data, unify the spatial resolution to 1km and the time resolution to monthly scale; S3: Construct a set of multi-source input variables, including normalized difference vegetation index (NDVI), enhanced vegetation index (EVI), leaf area index (LAI), photosynthetically active radiation absorption ratio (FPAR), terrain factor, meteorological factor and spatial and temporal variables; S4: Use the random forest algorithm to establish a multi-source data-driven CO2 flux prediction model, set the number of decision trees n_estimators = 1000 and the maximum depth of the decision tree max_depth = 20, and use 5-fold cross-validation to evaluate the performance; S5: Use SHAP (SHapley Additive exPlanations) values to explain the model results and determine the main control variables of carbon flux and their influence mechanism; S6: Estimate the monthly forest carbon sink of the entire study area, output the distribution map of regional net ecosystem carbon exchange (NEE) from 2003 to 2022; S7: Analyze the interannual variation trend, judge the regional forest carbon source / sink pattern, and evaluate the error and uncertainty.

2. The method of claim 1, wherein, The remote sensing data includes MCD15A3H (FPAR / LAI) and MOD13Q1 (NDVI / EVI) in MODIS products, and is processed by spatial resampling and flux tower data.

3. The method of claim 1, wherein, The input variables of the random forest model include air temperature, total precipitation, shortwave radiation, NDVI, EVI, LAI, FPAR, DEM elevation, land cover type, site latitude and longitude, and month code.

4. The method of claim 1, wherein, The model explanation analysis uses the SHAP (SHapley Additive exPlanations) method to output global variable importance ranking and local dependence graph to quantify the influence direction and strength of variables on carbon sink prediction.

5. The method of claim 1, wherein, The predicted object is the net ecosystem exchange (NEE) in gC / m 2 • month, the imputation and prediction range covers missing values, low quality values and spatially extended areas.

6. A regional forest carbon flux estimation system based on a regional scale forest carbon sink estimation method of fusing remote sensing data and vorticity covariance data, characterized in that, Comprising the following modules: 1: Data management module: used for downloading, preprocessing and fusing multi-source data; 2: Modeling module: used for training random forest model and evaluating accuracy; 3: Model explanation module: used for calculating SHAP value and outputting variable explanation; 4: Regional simulation module: based on the model to predict carbon flux of the whole region; 5: Visualization module: used to generate carbon sink spatial graph, time trend graph and variable importance graph; 6: The system running platform is based on Python language, which uses scikit-learn, xarray, shap and other libraries to realize.

7. The regional forest carbon flux estimation system of the regional scale forest carbon sink estimation method based on remote sensing data and eddy covariance data fusion according to claim 6, characterized in that: The forest carbon flux estimation results output by the system can be used for regional carbon balance research, carbon peak / carbon neutral policy support and ecosystem response analysis. 8.The regional forest carbon flux estimation system based on the remote sensing data and vorticity covariance data fusion according to the regional scale forest carbon sink estimation method of claim 6, wherein, Applications in the following aspects: 1: Estimation and monitoring of forest carbon flux at regional scale; 2: Simulation of forest carbon sink under different climate scenarios; 3: Scientific support for carbon neutralization and carbon trading policies; 4: Assessment of ecosystem management and remote sensing data integration.

Images