Method for carbon emission zoning of districts and counties based on two-dimensional classification and interpretable machine learning

Abstract

The application discloses a district county carbon emission zoning method based on double-dimensional classification and interpretable machine learning, and belongs to the technical field of urban low-carbon planning and land space governance. It comprises the following steps: obtaining multi-source data of a target region to construct a unified spatial database; extracting urban entity regions (UPA) based on land cover data as a unified spatial base; identifying social and economic development and spatial growth pattern types to construct double-dimensional composite types; calculating district county carbon emission density (CED) based on UPA to construct a gray-green space index system, and identifying the contribution structure of the gray-green space index to the carbon emission density under different composite types by using the XGBoost-SHAP method; and identifying the nonlinear response interval of key indicators by LOWESS fitting to form zoning results covering double-dimensional composite types, dominant indicators and the like. The application improves the spatial consistency of carbon emission density measurement and the interpretability of driving identification, and provides technical support for low-carbon land governance and spatial structure optimization at the district county scale.

Description

Technical Field

[0001] This invention belongs to the technical fields of urban low-carbon planning, land space governance and carbon emission spatial analysis, specifically involving a county-level carbon emission zoning method based on two-dimensional classification and interpretable machine learning. Background Technology

[0002] Against the backdrop of regional green transformation, carbon emission control is gradually shifting from total emission control to refined spatial governance. Districts and counties, as crucial units connecting regional development strategies with specific spatial implementation, not only bear the brunt of construction land expansion, population agglomeration, industrial activities, and public service allocation, but also serve as the basic scale for the implementation of low-carbon governance measures. Therefore, accurately identifying differences in carbon emission density at the district and county level and further proposing zoning control methods has become a significant issue in low-carbon land governance, urban spatial optimization, and national spatial planning.

[0003] Existing spatial analysis methods for carbon emissions mostly use complete administrative regions as statistical and comparative units, judging regional differences through total carbon emissions, emission intensity per unit administrative area, or per capita emission intensity. However, the administrative boundaries of districts and counties often include farmland, forest land, water bodies, and other non-built spaces. If the complete administrative area is directly used as the denominator, it is easy to dilute the carbon emission pressure in the actual built space and reduce the consistency between carbon emission density measurement and urban physical space. At the same time, existing studies often explain carbon emission differences from a single socio-economic development level, a single urban expansion type, or a single landscape indicator, which makes it difficult to simultaneously reflect the combined effects of development pressure sources and spatial growth organization methods.

[0004] Furthermore, gray construction space and green ecological space are important spatial structural elements influencing the carbon emission density of districts and counties. The size and morphological complexity of construction patches, the connectivity of green spaces, the density of green patches, and the ratio of gray to green spaces can all alter the allocation of population, industry, and ecological regulation functions within a unit of space. While existing methods can use machine learning models to identify the importance of variables, they typically remain at the level of ranking global contributions or judging average effects. They struggle to further explain which gray and green indicators are more critical in different types of districts and counties, within which value ranges are key indicators more sensitive, and how to translate the model's interpretation results into differentiated spatial governance strategies.

[0005] Therefore, it is necessary to propose a district-level carbon emission density zoning control method that combines urban physical geographical base, two-dimensional composite classification framework, gray-green space index system and interpretable machine learning method, so as to improve the accuracy of carbon emission density measurement, enhance the contextual explanation ability of district-level carbon emission differences, and provide operable technical support for low-carbon land governance and gray-green space structure optimization. Summary of the Invention

[0006] To address the aforementioned problems in the existing technology, the purpose of this invention is to provide a county-level carbon emission zoning method based on two-dimensional classification and interpretable machine learning.

[0007] This invention provides the following technical solution: a county-level carbon emission zoning method based on two-dimensional classification and interpretable machine learning, comprising the following steps:

[0008] S1. Acquire socioeconomic statistics, multi-source spatial auxiliary data, high-resolution land cover data, and carbon emission raster data of the target area, and preprocess the acquired data;

[0009] S2. Urban Entity Areas (UPAs) are extracted based on preprocessed high-resolution land cover data, and UPAs are used as a unified spatial basis for the measurement of carbon emission density (CED) in districts and counties, identification of population growth patterns, and calculation of gray-green space indicators.

[0010] S3. Based on the population change rate, per capita GDP change rate, and built-up area change rate, identify the socio-economic development type and obtain the socio-economic development type layer of the target area.

[0011] S4. Identify growth grids based on population grid change results, calculate the proportion of growth grids to the total area of ​​UPA, and use the growth grid proportion as the core indicator for spatial growth pattern identification; make threshold judgments based on the growth grid proportion to identify no-growth and full-growth types; for samples in the intermediate growth range, further combine the maximum patch dominance, average growth probability and patch number, and use the K-means clustering method to subdivide the spatial growth pattern type to obtain a county-level spatial growth pattern type layer;

[0012] S5. Combine the socio-economic development type obtained in S3 with the spatial growth pattern type obtained in S4 to construct a two-dimensional composite type of socio-economic and spatial growth, and use it as the basic analytical unit for subsequent carbon emission density differentiation comparison, gray-green space driving structure identification and zonal regulation.

[0013] S6. Extract the total carbon emissions within the UPA range of each district and county based on carbon emission raster data, and calculate the carbon emission density (CED) of the district and county using the corresponding UPA area as the denominator; perform a logarithmic transformation on the CED to obtain LnCED.

[0014] S7. Based on the gray construction space and green ecological space within the UPA scope, construct a gray-green space indicator system to obtain gray-green space structure indicators at the district and county level;

[0015] S8. Using the gray-green space index obtained in S7 as the input variable and LnCED obtained in S6 as the output variable, construct an XGBoost regression model and use the SHAP method to identify the contribution structure of the gray-green space index to the carbon emission density of districts and counties under different two-dimensional composite types.

[0016] S9. Based on the SHAP contribution ranking of gray-green space indicators in different two-dimensional composite types, the dominant driving group is merged, and the nonlinear response interval of key gray-green space indicators is identified by LOWESS fitting, forming the county-level carbon emission density zoning control results of composite type-dominant indicator-response interval-control direction.

[0017] Furthermore, in S1, socioeconomic statistics include resident population, GDP per capita, and built-up area; multi-source spatial auxiliary data includes topographic data, slope data, climate data, urban center distance data, road network data, nighttime light data, normalized difference vegetation index (NDE) data, and water system data; high-resolution land cover data is used for UPA extraction and gray-green space index calculation; carbon emission raster data is used for district / county CED measurement. The data preprocessing process includes projection unification, spatial registration, resolution resampling, boundary clipping, and attribute matching, ultimately forming a comprehensive database with standardized scale, spatial alignment, and consistent attributes.

[0018] Furthermore, in S2, the extraction process of the urban entity area UPA includes: extracting built-up area pixels from high-resolution land cover data; performing hole filling and boundary aggregation processing on the built-up area pixels to reconstruct continuous built-up space; removing isolated patches with an area smaller than a preset threshold; and spatially matching the processed continuous built-up space with district / county administrative units to form the UPA boundary corresponding to each district / county. The UPA is used to replace the area of ​​a complete administrative region to reduce the dilution effect of non-built-up spaces such as farmland, forest land, and water bodies on carbon emission density measurement.

[0019] Furthermore, in S3, the process of identifying socio-economic development types is as follows: taking prefecture-level cities as the identification object, the change rate of permanent residents, the change rate of per capita GDP, and the change rate of built-up area are calculated between two research time points; the three indicators are standardized; the K-means clustering algorithm is used to identify socio-economic development types; and by combining the clustering results and type interpretability, socio-economic development types such as high-quality development type, scale expansion type, restructuring and transformation type, and economic-driven transformation type are formed.

[0020] Furthermore, in S4, the spatial growth pattern identification process is as follows: a standard grid is established using the UPA as the unified analysis boundary; a population grid model is constructed based on multi-source spatial auxiliary variables and county-level resident population statistics, generating population density grids for two research time points; the difference between the two periods' population density grids is calculated, and grids with population changes greater than zero are defined as growing grids; the proportion of growing grids to the total area of ​​the UPA is calculated, and this proportion is used as the core indicator for spatial growth pattern identification; a threshold judgment is made based on the proportion of growing grids to identify no-growth and fully-growing types; for samples in the intermediate growth range, the maximum patch dominance, average growth probability, and number of patches are further calculated, and K-means clustering is used to further subdivide them into edge-expanding growth, strong-core growth, multi-center growth, scattered growth, and fragmented growth. The spatial growth pattern types include fully-growing, no-growth, edge-expanding growth, strong-core growth, multi-center growth, scattered growth, and fragmented growth.

[0021] Furthermore, in S5, the socio-economic-spatial growth dual-dimensional composite type is encoded using a socio-economic development type code—spatial growth form type code. This dual-dimensional composite type simultaneously characterizes the development background and spatial growth organization of a district or county, where the socio-economic development type reflects the sources of development pressure, and the spatial growth form type reflects the way population growth activities are carried out in the urban physical space.

[0022] Furthermore, in S6, the calculation process of the district / county CED is as follows: the carbon emission raster data is spatially overlaid with the boundary of each district / county UPA to extract the carbon emission raster values ​​within the UPA range; the sum of all carbon emission raster values ​​within the UPA range is obtained to get the total carbon emissions within the district / county UPA range; the total carbon emissions are divided by the corresponding district / county UPA projected area to obtain the district / county CED; extreme value processing and natural logarithm transformation are performed on the CED to obtain LnCED, which is used as the dependent variable of the subsequent machine learning model.

[0023] Furthermore, in S7, the gray-green space indicator system includes gray space morphology indicators, green space configuration indicators, and gray-green space relationship indicators. Among these, the gray space morphology indicators include the gray space landscape shape index (LSI_Grey) and the average gray space patch area (AREA_MN_Grey); the green space configuration indicators include green space patch density (PD_Green) and green space patch connectivity (PCGreen); and the gray-green space relationship indicators include the gray-green space area ratio (GGR) and landscape sprawl (CONTAG). These indicators are used to characterize the gray construction space, green ecological space, and their combination relationships within the district / county UPA.

[0024] Furthermore, in S8, the XGBoost regression model is trained and validated using a training set and a test set partitioning method. In the training set, Bayesian optimization and cross-validation are combined to optimize hyperparameters. After model training, the SHAP method is used to calculate the contribution of each gray-green space index to the LnCED prediction result. The average absolute SHAP value of each index is then summarized according to the two-dimensional composite type to obtain the relative contribution ranking and driving structure of the gray-green space indices under different two-dimensional composite types.

[0025] Furthermore, in S9, the process of merging the dominant driving groups is as follows: Dominant and secondary dominant indicators are identified by ranking the gray-green space indicators in each two-dimensional composite type according to their average absolute SHAP values; two-dimensional composite types with similar combinations of dominant indicators are merged into the same dominant driving group; LOWESS fitting is performed on the key indicators of each dominant driving group to identify the stage changes, local peaks, local valleys, and sensitive response intervals between the key indicator values ​​and SHAP values; and corresponding low-carbon land governance control directions are formed based on the identification results.

[0026] The method of this invention constructs a technical system of: spatial basis construction - two-dimensional composite classification - carbon emission density measurement - gray-green driving identification - response interval output. Under a unified spatial basis, it realizes the identification of county-level carbon emission density, typified scenario determination, gray-green spatial driving structure analysis and differentiated control result output.

[0027] At the spatial base construction level, urban entity areas (UPAs) are extracted from high-resolution land cover data, and UPAs are used as a common base for carbon emission density measurement, population growth pattern identification, and gray-green space index calculation, which mainly corresponds to steps S1, S2, and S6.

[0028] At the level of two-dimensional composite classification, the development background and spatial organization of growth activities of districts and counties are characterized by the identification of socio-economic development type and spatial growth pattern type, respectively. Through cross-combination, a two-dimensional composite type of socio-economic-spatial growth is formed, which mainly corresponds to steps S3, S4 and S5.

[0029] At the level of identifying gray-green driving forces, by constructing an index system of gray space morphology, green space configuration and gray-green space relationship, and combining it with the XGBoost-SHAP interpretable machine learning model, we can identify the differentiated contributions of gray-green space indicators to carbon emission density under different two-dimensional composite types, which mainly correspond to steps S7 and S8.

[0030] At the level of zonal regulation output, the dominant driving groups are merged by SHAP contribution ranking, and the nonlinear response intervals of key indicators are identified by LOWESS fitting, generating a low-carbon land governance result of composite type—dominant indicator—response interval—regulation direction, which mainly corresponds to step S9.

[0031] The above levels logically form a progressive mechanism of unified base, composite classification, driving identification, and interval regulation: unified base improves the spatial consistency of carbon emission density measurement, composite classification provides contextual units for explaining carbon emission differences, driving identification reveals the mechanism by which gray-green spatial structure affects carbon emission density, and interval regulation transforms the model interpretation results into actionable directions for low-carbon land governance.

[0032] By employing the above-described technology, the beneficial effects of the present invention compared to the prior art are as follows:

[0033] 1) This invention extracts urban physical area UPA from high-resolution land cover data as a unified spatial base to replace traditional administrative boundaries, effectively eliminating interference from non-built space and improving the matching degree between district / county carbon emission density (CED) and the actual built space distribution.

[0034] 2) This invention constructs a two-dimensional composite classification framework of socio-economic and spatial growth, which couples development pressure and spatial organization characteristics, overcoming the limitations of single-dimensional analysis and realizing the analysis of carbon emission density differences between districts and counties.

[0035] 3) This invention integrates the gray-green space index system with the XGBoost-SHAP interpretable model to identify the differentiated contributions of gray-green space indicators under different composite types, and clarifies the driving structure of construction form and ecological space on carbon emissions.

[0036] 4) This invention is based on SHAP value merging of dominant driving groups and identifies the nonlinear response intervals of key indicators through LOWESS fitting. The model results are transformed into a zoning scheme with composite type, dominant indicator, response interval, and regulation direction, forming a regulation scheme for differentiated governance of national land space. Attached Figure Description

[0037] Figure 1 This is the overall flowchart of the county-level carbon emission density zoning control method of the present invention;

[0038] Figure 2 This is a clustering result diagram of socio-economic development types in this invention;

[0039] Figure 3 This is a clustering result diagram of spatial growth morphology types in this invention;

[0040] Figure 4 This is a schematic diagram illustrating the rules for determining spatial growth morphology types and their specific meanings in this invention;

[0041] Figure 5 This is a schematic diagram illustrating the construction of the socio-economic-spatial growth dual-dimensional composite type in this invention;

[0042] Figure 6This is an example diagram illustrating the differences in the distribution of CEDs (Central Economic Development Departments) across districts and counties under different socioeconomic development types in this invention, taking the HQD (High Quality District) type as an example.

[0043] Figure 7 This is an example diagram illustrating the differences in the distribution of CEDs in counties and districts under the dual-dimensional composite type in this invention, taking the combination of various spatial growth patterns under the HQD type as an example;

[0044] Figure 8 This is an example diagram of CED differentiation under the same socio-economic development type in this invention, taking the HQD type as an example;

[0045] Figure 9 This is an example diagram illustrating the CED differentiation of the same spatial growth pattern in different socio-economic development types in this invention, taking the CG type as an example;

[0046] Figure 10 This is a typical comparison diagram of the contribution structure of gray-green spatial indicators to SHAP under the dual-dimensional composite type in this invention. (a) takes the counties and districts of various spatial growth patterns under the socio-economic development type HQD as an example, and (b) takes the spatial growth pattern type CG under different socio-economic development types as an example.

[0047] Figure 11 This is a SHAP ranking chart of the representative gray-green space index of the dominant driving group in this invention, taking group G1 as an example;

[0048] Figure 12 This is a nonlinear response interval diagram of the key gray-green indicators in group G1 in this invention;

[0049] Figure 13 This is a nonlinear response interval diagram of the key gray-green indicators in group G2 in this invention;

[0050] Figure 14 This is a nonlinear response interval diagram of the key gray-green indicators in group G3 of this invention;

[0051] Figure 15 This is a nonlinear response interval diagram of the key gray-green indicators in group G4 of this invention;

[0052] Figure 16 This is a nonlinear response interval diagram of the key gray-green indicators in group G5 of this invention. Detailed Implementation

[0053] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0054] Conversely, this invention encompasses any substitutions, modifications, equivalent methods, and solutions made within the spirit and scope of the invention as defined in the claims. Furthermore, to provide a better understanding of the invention, certain specific details are described in detail below. However, those skilled in the art will fully understand the invention even without these detailed descriptions.

[0055] This embodiment takes region A as the research object and provides a method for zoning and regulating county-level carbon emission density based on a two-dimensional composite classification framework and interpretable machine learning. This method is used to identify differences in county-level carbon emission density, analyze the driving structure of gray-green space, and generate zoning and regulating results for low-carbon land governance. (Refer to...) Figure 1 The method flow shown in this embodiment specifically includes the following steps.

[0056] S1. Data Collection and Preprocessing: Acquire socioeconomic statistics, multi-source spatial auxiliary data, high-resolution land cover data, and carbon emission raster data for region A. As shown in Table 1, socioeconomic statistics include resident population, GDP per capita, and built-up area; multi-source spatial auxiliary data includes topography, slope, climate, distance from city center, road network, nighttime light, NDVI, and water system data; high-resolution land cover data is used for urban entity area UPA extraction and gray-green space index calculation; carbon emission raster data is used for district / county carbon emission density (CED) measurement. The above data undergoes projection unification, spatial registration, resampling, boundary clipping, and attribute matching to form a spatially aligned and attribute-consistent comprehensive database. Table 1 contains detailed information on the data used.

[0057] S2. Urban Entity Area (UPA) Extraction: Built-up area pixels are extracted from high-resolution land cover data, and then subjected to hole filling, boundary aggregation, and small patch removal to obtain continuous urban entity area (UPA). The UPA is then overlaid with district / county administrative boundaries to form the corresponding UPA boundaries for each district / county. The UPA serves as a unified spatial basis for carbon emission density (CED) measurement, population growth pattern identification, and gray-green space index calculation.

[0058] S3. Socioeconomic Development Type Identification: Taking prefecture-level cities as the research object, the rates of change in permanent residents, per capita GDP, and built-up area were calculated between two research time points. After standardizing the three indicators, K-means clustering was used to identify socioeconomic development types, yielding development types reflecting differences in population change, economic growth, and spatial expansion of construction. The clustering results of socioeconomic development types are as follows: Figure 2 As shown, where Figure 2 (a) shows the spatial distribution results of the four types of socio-economic development. Figure 2(b) represents the distribution of indicator characteristics for the four types of socio-economic development.

[0059] S4. Spatial Growth Pattern Identification: A standard grid was established using UPA as the unified analysis boundary. A population gridded model was constructed based on multi-source spatial auxiliary variables and county-level resident population statistics, generating population density grids for two research time points. The population grid results were corrected using total control, as shown in the following formula: ;

[0060] In the formula, To calculate the population of the i-th grid within the j-th district after correction; These are the initial population weights predicted by the model; The total population of the j-th district / county is calculated. It is the sum of the initial population weights of all grid cells within the j-th district / county.

[0061] The difference between population grids at two study time points was calculated, and grids with population changes greater than 0 were defined as growing grids. The proportion of growing grids to the total area of ​​the UPA was further calculated, and this proportion was used as the core indicator for identifying spatial growth patterns. Threshold determination was performed based on the growing grid proportion to identify non-growing and fully growing patterns; for samples in the intermediate growth range, the maximum patch dominance (R0) was further considered. max ), average growth probability (R) mean ) and number of plaques (N) patch The spatial growth patterns were further subdivided into spatial growth morphology types using K-means clustering. These spatial growth morphology types include fully growing (CG), no-growth (NG), edge-expanding growth (EE), strong-core growth (SC), multi-center growth (SG), fragmented growth (PG), and scattered growth (FG). The clustering results for spatial growth morphology types are shown below. Figure 3 As shown, Figure 3 (a) represents the distribution of growth types in the two-dimensional feature space. Figure 3 (b) Distribution of growth types in orthogonal three-dimensional feature space; see the determination rules. Figure 4 .

[0062] S5. Construction of a Two-Dimensional Composite Type for Socioeconomic-Spatial Growth: The socioeconomic development type obtained in S3 is cross-combined with the spatial growth pattern type obtained in S4 to form a two-dimensional composite type for socioeconomic-spatial growth, encoded using a socioeconomic development type code followed by a spatial growth pattern code. This two-dimensional composite type simultaneously represents the development background and spatial growth organization methods, serving as the basic analytical unit for subsequent CED differentiation comparison, gray-green spatial driving force identification, and zonal regulation. The construction logic of the two-dimensional composite type is as follows: Figure 5 As shown, Figure 5(a) represents the clustering process along the socio-economic development dimension. Figure 5 (b) represents the spatial growth morphological dimension clustering process. Figure 5 (c) in the figure represents a two-dimensional composite framework.

[0063] S6. Measurement of Carbon Emission Density in Districts and Counties: Based on carbon emission grid data, extract the carbon emission grid values ​​within the UPA range of each district and county and sum them to obtain the total carbon emissions within the UPA range; then calculate the district / county carbon emission density (CED) using the UPA area as the denominator, as shown in the following formula: ;

[0064] In the formula, Let be the carbon emission density of the i-th district / county based on UPA; This is the sum of carbon emission grid values ​​within the UPA area of ​​that district / county; This represents the projected area of ​​the UPA in that district / county.

[0065] To reduce the influence of skewed distribution and extreme values, the CED is transformed by the natural logarithm to obtain LnCED, as shown in the following formula: ;

[0066] In the formula, This represents the logarithmically transformed carbon emission density. Differences in CED distribution across districts and counties under different socioeconomic development types are shown below. Figure 6 As shown, the differences in the distribution of CEDs in districts and counties under the two-dimensional composite type are as follows: Figure 7 As shown; the CED differentiation of different spatial growth patterns under the same socio-economic development type is as follows: Figure 8 As shown, the CED differentiation of the same spatial growth pattern under different socio-economic development types is as follows: Figure 9 As shown.

[0067] S7. Construction of the Gray-Green Space Indicator System: Based on high-resolution land cover data within the UPA area, gray construction space and green ecological space are extracted, and six indicators are constructed from three dimensions: gray space morphology, green space configuration, and gray-green space relationship. The gray space morphology indicators include the Gray Space Landscape Shape Index (LSI_Grey) and the Gray Space Average Patch Area (AREA_MN_Grey); the green space configuration indicators include the Green Space Patch Density (PD_Green) and the Green Space Patch Connectivity (PCGreen); and the gray-green space relationship indicators include the Gray-Green Space Area Ratio (GGR) and the Landscape Spread (CONTAG). These six gray-green space indicators are used to characterize gray construction space, green ecological space, and their combined relationships within the UPA.

[0068] S8. XGBoost-SHAP-Driven Structure Identification: An XGBoost regression model is constructed using six gray-green space indicators as input variables and LnCED as the output variable. The model is validated using training and testing sets, and parameters are determined through Bayesian optimization and cross-validation. After model training, the SHAP method is used to calculate the contribution of each gray-green space indicator to LnCED, and the average absolute SHAP value is summarized according to the two-dimensional composite type. Typical comparisons of the gray-green space indicator contribution structure relative to SHAP under the two-dimensional composite type are shown below. Figure 10 As shown, Figure 10 (a) shows the time-series changes in the contribution rates of various spatial growth types under the HQD background. Figure 10 (b) represents the temporal variation of the contribution rates of various socio-economic development types under the CG background.

[0069] S9. Merging of Dominant Driving Groups and Output of Partitioned Regulation Results: Based on the ranking of the average absolute SHAP values ​​of gray-green space indicators in different two-dimensional composite types, dominant and secondary dominant indicators are identified, and two-dimensional composite types with similar dominant structures are merged into dominant driving groups. The SHAP ranking of representative gray-green space indicators in the dominant driving group is as follows: Figure 11 As shown, the dominant modes of each dominant driving group and the two-dimensional composite types they contain are listed in Table 2: Table 2. Dominant Driving Groups and Their Subsidiary Two-Dimensional Composite Types

[0070] Further LOWESS fitting was performed on key indicators of each dominant driving group to identify the nonlinear response relationship, stage changes, and sensitive intervals between indicator values ​​and SHAP values. Based on this, a zonal control result was formed, consisting of composite type, dominant indicator, response interval, and control direction. For different dominant driving groups, targeted control directions were adopted, including improving green space connectivity, controlling the scale of gray construction patches, optimizing the overall gray-green space pattern, coordinating the gray-green ratio, and restoring existing space. The nonlinear response intervals of key gray-green indicators for groups G1-G5 are as follows: Figure 12 — Figure 16 As shown; where: Figure 12 (a) represents the nonlinear distribution of PCGreen's SHAP values. Figure 12 (b) is the segmentation feature of the SHAP value of AREA_MN_Grey.

[0071] Figure 13 (a) shows the segmented distribution of PCGreen's SHAP values. Figure 13 (b) represents the segmentation feature of the SHAP value of CONTAG.

[0072] Figure 15(a) shows the three-segment distribution of the SHAP value of CONTAG. Figure 15 (b) represents the four-segment characteristic of GGR's SHAP value.

[0073] Through the above steps, this embodiment establishes a complete technical process from data acquisition, UPA spatial basis construction, two-dimensional composite type identification, CED measurement, gray-green space driving structure analysis, to the output of zoning regulation results. This process enables refined identification of carbon emission density at the district / county level and transforms machine learning interpretation results into regulatory directions applicable to low-carbon land governance and gray-green space optimization.

[0074] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A county-level carbon emission zoning method based on two-dimensional classification and interpretable machine learning, characterized in that, Includes the following steps: S1. Obtain carbon emission raster data, land cover data, socio-economic statistics data and multi-source spatial auxiliary data for the target area. Extract urban entity area UPA based on land cover data, and use UPA as a unified spatial basis for carbon emission density measurement, population growth pattern identification and gray-green space index calculation. S2. Identify socioeconomic development types based on socioeconomic statistics, identify spatial growth patterns based on population grid change characteristics, and combine the two to form a socioeconomic-spatial growth dual-dimensional composite type. S3. Construct a gray-green space index system based on UPA, calculate and transform the carbon emission density of districts and counties, and use the XGBoost-SHAP model to identify the contribution structure of gray-green space indices to carbon emission density (CED) under different two-dimensional composite types; where: The XGBoost regression model is trained and validated by dividing the training set and the test set. In the training set, Bayesian optimization and cross-validation are combined to optimize the hyperparameters. After the model is trained, the SHAP method is used to calculate the contribution of each gray-green space index to the LnCED prediction result. The average absolute SHAP value of each index is summarized according to the two-dimensional composite type to obtain the relative contribution ranking and driving structure of the gray-green space index under different two-dimensional composite types. S4. Based on the SHAP contribution value merging of the dominant driving group, the nonlinear response intervals of key gray-green space indicators are identified through LOWESS fitting, and the output includes the two-dimensional composite type, the dominant gray-green indicator, the indicator response interval, and the partitioned control results of the adapted control strategy; where: The process of merging the dominant driving groups is as follows: Based on the average absolute SHAP values ​​of gray-green space indicators in each two-dimensional composite type, the dominant and secondary dominant indicators are identified, and two-dimensional composite types with similar combinations of dominant indicators are grouped into the same dominant driving group. The dominant indicators corresponding to each dominant driving group are fitted with LOWESS to identify the stage changes, local peaks, local valleys and sensitive response intervals between the key indicator values ​​and SHAP values, so as to form the corresponding low-carbon land governance regulation direction.

2. The county-level carbon emission zoning method based on two-dimensional classification and interpretable machine learning according to claim 1, characterized in that, In S1, socioeconomic statistics include resident population, GDP per capita, and built-up area; multi-source spatial auxiliary data include topographic data, slope data, climate data, distance from city center data, road network data, nighttime light data, normalized difference vegetation index data, and water system data; land cover data uses a preset resolution land cover product for UPA extraction and gray-green space index calculation; carbon emission raster data uses ODIAC fossil fuel CO2 emission data for county-level carbon emission density (CED) measurement.

3. The county-level carbon emission zoning method based on two-dimensional classification and interpretable machine learning according to claim 1, characterized in that, In S1, the UPA extraction process is as follows: Built-up area pixels are extracted from land cover data. Hole filling and boundary aggregation are performed on the built-up area pixels. Isolated patches with an area smaller than a preset threshold are removed. The continuous built-up space is superimposed with the administrative boundaries of districts and counties to obtain the UPA boundaries of each district and county.

4. The county-level carbon emission zoning method based on two-dimensional classification and interpretable machine learning according to claim 1, characterized in that, In S2, the process for identifying socioeconomic development types is as follows: Using the target region as the identification object, the change rate of permanent resident population, the change rate of per capita GDP, and the change rate of built-up area between two research time points are calculated. After standardizing the three indicators, the K-means clustering algorithm is used to identify the socio-economic development type, which includes high-quality development type (HQD), scale expansion type (SE), restructuring and transformation type (RT), and economic-driven transformation type (EDT).

5. The county-level carbon emission zoning method based on two-dimensional classification and interpretable machine learning according to claim 1, characterized in that, In S2, the spatial growth pattern type is identified in the following way: A standard grid was established using the UPA (Universal Population Area) as the unified analytical boundary. A population gridded model was constructed based on multi-source spatial auxiliary variables and county-level resident population statistics, generating population density grids for two research time points. The difference between the two population density grids was calculated, and grids with population changes greater than zero were defined as growing grids. The proportion of growing grids to the total area of ​​the UPA was calculated, and a threshold was used to identify non-growing and fully growing types based on this proportion. For samples in the growth interval between the non-growing (NG) and fully growing (CG) types, the maximum patch dominance R was combined with the results. max Average growth probability R mean And the number of patches N patch The K-means clustering method was used to classify the following types into edge-expanding growth type (EE), strong-core growth type (SC), multi-center growth type (PG), scattered growth type (SG), and fragmented growth type (FG).

6. The county-level carbon emission zoning method based on two-dimensional classification and interpretable machine learning according to claim 1, characterized in that, In S2, the socio-economic-spatial growth dual-dimensional composite type adopts a splicing coding method, which combines the socio-economic development type code and the spatial growth form type code to generate the identifier of the socio-economic-spatial growth dual-dimensional composite type, and serves as the basic analysis unit for subsequent CED differentiation comparison, gray-green spatial driving structure identification and zoning regulation.

7. The county-level carbon emission zoning method based on two-dimensional classification and interpretable machine learning according to claim 1, characterized in that, In S3, the gray-green space index system includes gray space morphology index, green space configuration index, and gray-green space relationship index; the gray space morphology index includes gray space landscape shape index LSI_Grey and gray space average patch area AREA_MN_Grey; the green space configuration index includes green space patch density PD_Green and green space patch connectivity PCGreen; the gray-green space relationship index includes gray-green space area ratio GGR and landscape spread CONTAG.

8. The county-level carbon emission zoning method based on two-dimensional classification and interpretable machine learning according to claim 1, characterized in that, In S3, the calculation process for the district / county CED is as follows: Carbon emission raster data is spatially overlaid with the boundaries of each district / county UPA, carbon emission raster values ​​within the UPA range are extracted and summed to obtain the total carbon emissions, and the total carbon emissions are divided by the corresponding district / county UPA projected area to obtain the CED; extreme value processing and natural logarithmic transformation are performed on the CED to obtain LnCED, which is used as the dependent variable for subsequent machine learning models.

Images