Pollution and carbon reduction space function partitioning method based on multi-modal deep learning

Abstract

The invention discloses a pollution and carbon reduction space function partitioning method based on multi-modal deep learning. The method comprises the following steps: firstly, constructing a multi-modal spatial index system covering environmental bearing capacity, environmental pressure and pollution reduction and carbon reduction potential, including an air diffusion index, an ecological carbon sink index, and a traffic emission pressure index and pollution reduction and carbon reduction potential considering a road network, population, lamplight and function mixing degree; secondly, uniformly resampling, standardizing and synthesizing the multi-source spatial indexes into a multi-band raster data set; then a multi-branch convolutional neural network model is constructed, the multi-branch convolutional neural network model comprises three independent convolution branches corresponding to bearing capacity, pressure and potential index groups respectively, and after branch output features are fused, the probability that each space unit belongs to different function partitions is calculated in a low-dimensional embedding space through a deep embedding clustering algorithm; and finally, generating a pollution-reducing and carbon-reducing space function partition grid map according to the maximum probability. According to the method, deep fusion, automation and refined partitioning of the multi-source heterogeneous spatial data are realized.

Description

Technical Field

[0001] This invention relates to the field of environmental protection and spatial planning technology, and in particular to a method for functional zoning of pollution reduction and carbon reduction spaces that integrates geospatial analysis, environmental science and artificial intelligence technologies. Background Technology

[0002] Synergistic effects in pollution and carbon reduction are a key path to promoting a comprehensive green transformation of economic and social development. To achieve precise policy implementation, it is necessary to scientifically zon the region's spatial functions to identify the pollution emission pressure, environmental carrying capacity, and pollution and carbon reduction potential of different areas, thereby formulating differentiated management and control strategies.

[0003] Currently, many spatial zoning methods suffer from the following limitations: First, they rely on single or insufficiently integrated indicators. Existing methods typically select a limited number of environmental or economic indicators for simple overlay or weighted scoring, failing to systematically construct a multi-dimensional, multi-modal indicator system encompassing "carrying capacity-pressure-potential," especially lacking refined and comprehensive characterization of key factors such as atmospheric diffusion conditions, ecosystem carbon sequestration endowment, and complex traffic emission pressures. Second, they depend on expert experience and subjective assignment. The Delphi method and analytic hierarchy process are often used to determine indicator weights and zoning thresholds, which are highly subjective, have poor repeatability, and struggle to capture complex nonlinear relationships between multi-source data. Third, their spatial analysis depth is limited. Traditional methods often remain at the statistical and overlay level, unable to automatically extract deep, high-dimensional features and patterns from massive spatial data like deep learning, resulting in coarse zoning results with rigid boundaries that fail to reflect true geographic spatial continuity and heterogeneity.

[0004] Therefore, there is an urgent need for a new method that can automatically and intelligently process multi-source heterogeneous spatial data and achieve scientific, precise, and interpretable functional zoning of pollution reduction and carbon reduction spaces. Summary of the Invention

[0005] In view of this, the purpose of this invention is to provide a spatial functional zoning method for pollution reduction and carbon reduction based on multimodal deep learning, so as to overcome the problems of one-sided zoning indicators, strong subjectivity and insufficient analysis depth in the existing technology, and realize automated, refined and scientific spatial zoning.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: A spatial functional zoning method for pollution and carbon reduction based on multimodal deep learning includes the following steps: Step S1: Construct a multimodal indicator system for spatial functional zoning. This system includes three main categories of indicators: a) Environmental carrying capacity indicators: These are used to characterize the regional natural environment's ability to dilute and disperse pollutants and the ecosystem's carbon sequestration capacity. Core indicators include: daily air diffusion index under "stable" meteorological scenarios (combining wind speed, topography, and urban morphology), and comprehensive ecological carbon sink endowment index (combining net primary productivity, vegetation index, and topographic stability).

[0007] b) Environmental stress indicators: These are used to characterize the pressure of pollutants and carbon emissions caused by human activities. Core indicators include: industrial POI density index, rural residential area winter nighttime surface temperature index, and traffic emission stress index (combining road network, population, nighttime light, and functional mixing).

[0008] c) Pollution and carbon reduction potential indicators: These indicators characterize a region's potential to reduce emissions through economic and technological means. Core indicators include: GDP per capita index and technological innovation potential index (density of high-tech enterprises, research institutes, and universities).

[0009] Step S2: Data Integration and Preprocessing. All kilometer-scale raster index data generated in Step S1 are resampled using a reference raster (e.g., the air diffusion index raster) to ensure complete consistency in spatial range, resolution, and coordinate system. Subsequently, Z-score normalization is applied to the raster data for each band to eliminate dimensional differences. Finally, all processed index raster data are strictly combined in a predetermined order into a multi-band raster dataset, which serves as the unified input for subsequent deep learning models.

[0010] Step S3: Construct and train a multimodal deep embedding clustering model. This model is the core of this invention, and its structure includes: Multi-branch feature extraction network: Based on the indicator modalities (bearing capacity, pressure, potential), the input multi-band data is divided into three groups, and each group is input into three structurally identical but parameter-independent convolutional neural network branches. Each branch extracts the deep spatial features of that modality through convolution, batch normalization, activation, and pooling operations.

[0011] Feature fusion and compression module: The feature maps extracted from the three branches are concatenated along the channel dimension to form a fused feature tensor. Then, an additional convolutional layer is used to perform deep interaction and abstraction on the fused features, and finally, global average pooling is used to compress the spatial features of the entire study area into a fixed-length one-dimensional feature vector.

[0012] Deep Embedding and Soft Clustering Layer: The one-dimensional feature vector is mapped to a low-dimensional embedding space through a fully connected layer. Within this embedding space, a clustering layer containing K learnable cluster centers is defined. By calculating the cosine similarity between the embedding vector of each spatial unit and all cluster centers, and normalizing it using a softmax function, the probability distribution (soft assignment) of the unit belonging to K different functional partitions is obtained.

[0013] Step S4: Model Training and Spatial Partition Generation. A deep embedding clustering algorithm is employed, with the goal of minimizing the KL divergence between the soft-assignment distribution of the model output and an auxiliary target distribution as the loss function. The entire network model (including CNN parameters and cluster center locations) is trained end-to-end. After model training converges, for each spatial pixel in the input multi-band dataset, the model outputs a K-dimensional probability vector. The category with the highest probability value is selected as the final partition label for that pixel, thus dividing the entire study area into K spatial partitions with different pollution reduction and carbon reduction functional orientations, and outputting them as a raster map.

[0014] In a further embodiment, step S3 involves constructing a multimodal deep embedding clustering model, the model comprising: S31, Data preparation module, used to load the multi-band raster dataset; S32, Multi-branch feature extraction module, contains three independent convolutional neural network branches, which are used to extract features of environmental carrying capacity indicators, environmental pressure indicators and pollution reduction and carbon reduction potential indicators, respectively, and the output features of the three branches are spliced ​​together in the channel dimension; S33, Feature Fusion and Embedding Module, performs convolution and global average pooling on the concatenated features to generate a deep feature vector representing the entire region, and maps the vector to a low-dimensional embedding space. S34. Clustering probability calculation module: calculates the similarity between the embedded vector and several preset cluster center vectors, and normalizes it into a probability distribution to obtain the soft assignment probability of each spatial unit belonging to each functional partition. In a further embodiment, step S1, constructing the environmental carrying capacity index includes: S11. Based on atmospheric boundary layer height data, identify "stable" weather scenarios for the study area; S12. Based on the "stable" meteorological scenario day, combined with wind speed data, terrain location index and sky openness data, the air diffusion index of the "stable" meteorological scenario day is calculated.

[0015] In a further embodiment, step S11 specifically includes: Calculate the 33rd percentile of the daily atmospheric boundary layer height for each spatial location; Dates with daily atmospheric boundary layer height less than or equal to the 33rd percentile of their corresponding location are selected as "stable" weather scenario days for that location.

[0016] In a further embodiment, in step S12, the daily air diffusion index of the "stable" meteorological scenario is calculated using the following formula: ,in, A normalized topographic location index, For the openness of the sky, The wind speed is the intraday normalized wind speed for the "stable" weather scenario.

[0017] In a further embodiment, in step S1, the environmental carrying capacity index further includes the comprehensive ecological carbon sink endowment index, which is constructed in the following way: obtaining net primary productivity data, normalized vegetation index data, and topographic relief data, performing normalization processing on each, and then calculating... ,in , , These are normalized net primary productivity, normalized vegetation index, and topographic relief, respectively.

[0018] In a further embodiment, in step S1, the environmental pressure index includes a traffic emission pressure index, which is constructed by calculating based on weighted road network density data, population density data, nighttime light intensity data, and point-of-interest (POI) functional hybridity data. ,in , , These are the normalized weighted road network density, population density, and nighttime light intensity, respectively. For the range of values ​​within Points of interest feature blending degree.

[0019] In a further embodiment, in step S2, the standardization process is Z-score standardization, which is performed based on the mean and standard deviation of the data for each band.

[0020] In a further embodiment, in step S32, each of the three independent convolutional neural network branches includes at least two convolutional layers, a batch normalization layer, an activation function layer, and a pooling layer. In step S33, the feature fusion and embedding module further includes a fusion convolution submodule, which is used to perform further convolution processing on the concatenated features. The fusion convolution submodule includes at least one convolutional layer, a batch normalization layer, and an activation function layer. In step S34, the similarity is cosine similarity, and the normalization process uses the softmax function.

[0021] In a further embodiment, in step S4, the argmax function is used to operate on the soft allocation probability distribution to determine the partition category with the highest probability for each spatial unit.

[0022] The present invention has the following beneficial effects: Highly scientific: It has constructed a three-in-one multimodal index system of "carrying capacity-pressure-potential", and in particular, it has innovatively proposed a "stable" meteorological scenario daily air diffusion index, an ecological carbon sink "quantity-quality-stability" comprehensive index, and a traffic emission pressure index that includes functional mixing degree, which scientifically depicts the spatial differentiation law of pollution reduction and carbon reduction from multiple dimensions.

[0023] Highly automated and intelligent: It utilizes a multi-branch CNN model to automatically extract deep features from multi-source spatial data, and achieves automated partitioning under unsupervised or weakly supervised conditions through deep embedding clustering, which greatly reduces human intervention and subjective bias, and improves the repeatability and objectivity of the method.

[0024] Refined partitioning results: Deep learning models can capture complex nonlinear spatial relationships, generating partition boundaries that better match the natural transition of geographical entities, with more uniform features within the partitions, resulting in more refined and accurate results.

[0025] Its practical value is outstanding: the generated spatial functional zoning map can directly serve environmental management decisions. For example, it can promote industrial upgrading and green technology application in areas with high carrying capacity, high pressure and high potential; and implement strict emission control in areas with low carrying capacity. It provides a direct spatial blueprint and decision-making basis for formulating differentiated and precise pollution reduction and carbon reduction policies.

[0026] Excellent technological integration: It successfully integrates GIS spatial analysis, environmental simulation and deep learning technologies, providing a novel and powerful technological paradigm for solving complex geospatial problems. Attached Figure Description

[0027] Figure 1 This is a flowchart illustrating the overall method of an embodiment of the present invention.

[0028] Figure 2 This is a schematic diagram of the structure of the multimodal deep embedding clustering model in an embodiment of the present invention.

[0029] Figure 3 This is a schematic diagram of the process for constructing the daily air diffusion index of a "stable" meteorological scenario in an embodiment of the present invention.

[0030] Figure 4 This is a schematic diagram of the process for constructing the traffic emission pressure index in an embodiment of the present invention.

[0031] Figure 5This is a schematic diagram of the functional zoning of the pollution reduction and carbon reduction space generated in the final embodiment of the present invention.

[0032] Figure 6 This is a flowchart illustrating the overall workflow of the present invention. Detailed Implementation

[0033] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0034] In the description of this invention, it should be noted that the terms "center," "upper," "lower," "left," "right," "vertical," "horizontal," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are used only for the convenience of describing the invention and for simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on the invention. Furthermore, the terms "first," "second," and "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0035] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances. Furthermore, the technical features involved in the different embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0036] A spatial functional zoning method for pollution and carbon reduction based on multimodal deep learning, such as... Figure 1 As shown, the specific steps are as follows: 1. Construction of a multimodal indicator system 1.1 Construction of Environmental Carrying Capacity Indicators 1.1.1 Daily Air Diffusion Index under a "Stable" Meteorological Scenario See Figure 3 .

[0037] (1) Obtain the daily atmospheric boundary layer height (PBLH) raster data provided by ERA5 reanalysis data. In ArcGIS Pro, create a PBLH multidimensional raster layer.

[0038] (2) Calculate the 33rd percentile of PBLH in the time series of each raster cell to obtain the single-value raster PBLH_P33.

[0039] (3) Extract from the original multidimensional PBLH data those rasters whose PBLH values ​​are less than or equal to the corresponding PBLH_P33 value on all dates. These dates are defined as the “stable” weather scenario days for that location, and generate the corresponding date-matching multidimensional rasters.

[0040] (4) Obtain daily wind speed data from CLDAS-V2.0, create a multidimensional grid, and resample to 1 km resolution.

[0041] (5) Based on the date matching relationship obtained in step (3), extract the wind speed data corresponding to the "stable" meteorological scenario day from the wind speed multidimensional data.

[0042] (6) Calculate the Terrain Location Index (TPI) and Sky Openness (SVF), and resample them to 1 km grid.

[0043] (7) Perform Min-Max normalization on the TPI grid and the "stable" diurnal wind speed grid (for each time slice). The SVF value itself is in the [0,1] interval and does not need to be normalized.

[0044] (8) For each "stable" day, perform the following using the raster calculator: The air diffusion index for that day is obtained. Finally, the average of the index results for all "stable" days is taken to generate the final "stable" weather scenario daily air diffusion index kilometer grid (Air_Diffusion_Idx).

[0045] 1.1.2 Comprehensive Ecological Carbon Sink Endowment Index (1) Obtain annual average NPP and annual average NDVI data from MODIS, as well as high-precision DEM data.

[0046] (2) Calculate the terrain relief (TR) based on DEM: First, calculate the maximum and minimum elevation values ​​within the neighborhood (e.g., a radius of 500 meters) through focus statistics, subtract the two to obtain the original resolution TR, and then resample to 1 km.

[0047] (3) Resample the NPP and NDVI data to a 1-kilometer resolution that matches the TR grid space.

[0048] (4) Perform Min-Max normalization on NPP, NDVI, and TR rasters respectively.

[0049] (5) Perform the following using the grid calculator: The Eco Carbon Sink Comprehensive Endowment Index (Eco_Carbon_Idx) is generated.

[0050] 1.2 Construction of Environmental Stress Indicators 1.2.1 Industrial Sector POI Density Index Based on text mining (TF-IDF and LDA topic models), industrial source keywords such as "steel", "coking", and "chemical" were identified from policy documents. Points of corresponding categories were selected from the POI data, and kernel density analysis (search radius 1 km) was performed to generate an industrial sector POI density raster (Ind_POI_Density).

[0051] 1.2.2 Surface temperature index of rural settlements at night in winter (1) Use land use data to extract rural settlements, generate binary raster and resample to 1 km.

[0052] (2) Obtain monthly nighttime surface temperature (LST) data for winter (e.g., from November to March of the following year), and calculate the average value to obtain the average LST raster for winter nights.

[0053] (3) Using a conditional assignment function (such as Con), the average LST value of winter nights is assigned only to the grid where the rural settlement is located, and other areas are set to NoData, thus generating a rural settlement winter night LST index grid (Rural_LST_Idx).

[0054] 1.2.3 Traffic Emission Pressure Index See Figure 4 .

[0055] (1) Weighted road network density: The road vector data is weighted according to the level (highway / expressway=1, national highway / main road=0.8, etc.), and line density analysis is performed to generate a weighted road network density raster (Weighted_Road_Density).

[0056] (2) POI Functional Mixing: POIs are classified according to function (residential, employment, leisure, public services), and kernel density is calculated for each category. For each 1-kilometer grid, the proportion p of POI density for each category is calculated using the information entropy formula. Calculate the feature mix index (POI_MFI). A higher value indicates a more mixed feature mix.

[0057] (3) Obtain 1-kilometer raster data of population density (Pop_Density) and nighttime light intensity (NTL).

[0058] (4) Resample all data to align with the weighted road network density grid space.

[0059] (5) Perform Min-Max normalization on Weighted_Road_Density, Pop_Density, and NTL rasters. POI_MFI rasters are already in [0,1] and do not require processing.

[0060] (6) Perform the following using the raster calculator: .in This reflects the assumption that a high degree of functional hybridization may reduce reliance on motor vehicles, thereby lowering emission pressure. A traffic emission pressure index raster (Traffic_Pressure_Idx) is generated.

[0061] 1.3 Construction of Potential Indicators for Pollution Reduction and Carbon Reduction 1.3.1 GDP per capita index Acquire spatial raster data of GDP and population, resample and align them, then perform GDP / Population to generate a GDP per capita raster (GDP_per_Capita).

[0062] 1.3.2 Technological Innovation Potential Index Locations of high-tech enterprises, research institutes, and research universities (985 / 211) are selected from the POI data, and kernel density rasters are calculated for each. After normalizing the three density rasters, they are summed with equal weights to generate the Tech_Innovation_Idx raster.

[0063] 2. Data Preprocessing and Model Input Preparation (1) Spatial alignment and resampling: Using the Air_Diffusion_Idx raster as the spatial reference, the six rasters Eco_Carbon_Idx, Ind_POI_Density, Rural_LST_Idx, Traffic_Pressure_Idx, GDP_per_Capita, and Tech_Innovation_Idx were resampled to the same spatial range, 1 km resolution and coordinate system using ArcGIS Pro's "Resampling" tool.

[0064] (2) Band Combining: Strictly follow the following order to combine the seven rasters into a 7-band multi-band raster dataset (MultiBand_Input): Band 1: Air_Diffusion_Idx Band 2: Eco_Carbon_Idx Band 3: Ind_POI_Density, Band 4: Rural_LST_Idx Band 5: Traffic_Pressure_Idx Band 6: GDP_per_Capita, Band 7: Tech_Innovation_Idx.

[0065] (3) Z-score standardization: Calculate the mean (μ) and standard deviation (μ) of each band in the MultiBand_Input. For each pixel value in each band. ,implement This generates a standardized multiband raster dataset, Normalized_Input.

[0066] (4) Data format conversion: Write a Python script to use the arcpy.RasterToNumPyArray function to convert Normalized_Input to a format of . A NumPy array (X_input) is used, where H is the number of rows and W is the number of columns. Simultaneously, metadata such as the raster's spatial extent, cell size, and coordinate system is extracted and saved.

[0067] 3. Construction and Training of Multimodal Deep Embedding Clustering Model See Figure 2 The model is implemented using the PyTorch framework.

[0068] (1) Multi-branch feature extraction: Input: X_input [H, W, 7].

[0069] Branch splitting: Load-bearing capacity branch: Input (Band 1-2)

[0070] Pressure branch: Input Bands 3-5).

[0071] Potential branch: Input (Band 6-7)

[0072] Each branch has the same structure: The size of the output feature map for each branch is approximately [H / 2, W / 2, 64].

[0073] (2) Feature fusion and compression: Concatenation: The output feature maps of the three branches are concatenated along the channel dimension to obtain the fused feature [H / 2, W / 2, 192].

[0074] Fusion Convolution: After two fusion convolutional layers: Keep the spatial dimensions unchanged.

[0075] Global average pooling: for the fused feature map Perform global average pooling to obtain a 256-dimensional feature vector. .

[0076] (3) Deep embedding and soft clustering: Embedded layer: A fully connected layer will Mapping to the 32-dimensional embedding space yields the embedding vector e.

[0077] Clustering layer: Defines a learnable cluster center matrix , shape ,in The preset number of partitions is (in this embodiment, K=4, representing 4 types of functional partitions).

[0078] Probability calculation: Calculation and The cosine similarity of each cluster center is used to obtain a The similarity vectors are then converted into a probability distribution using the softmax function. , This indicates that the current sample belongs to the first... The probability of each partition.

[0079] (4) Model training: Loss function: Using the Deep Embedding Clustering (DEC) algorithm, an auxiliary target distribution is constructed. and minimize With model output distribution KL divergence between: .

[0080] Optimizer: Use the Adam optimizer.

[0081] Training process: Load the X_input data and iteratively train the model. The loss function simultaneously drives the CNN feature extraction network to learn feature representations that are beneficial for clustering and optimizes the location of cluster centers.

[0082] 4. Spatial Partition Generation and Output (1) After the model training converges, the X_input data of the entire study area is input into the model to obtain the probability matrix P_all of each 1-kilometer raster cell belonging to 4 partitions, with the shape of .

[0083] (2) For each pixel, use the argmax function to find the category index with the highest probability, and obtain the category label matrix Labels, with the shape as follows: .

[0084] (3) Using the spatial metadata saved in the preprocessing stage, the labels matrix is ​​converted back to raster data using the arcpy.NumPyArrayToRaster function, and the spatial reference is completely consistent with the original input data.

[0085] (4) Perform necessary post-processing on the generated preliminary zoning raster (such as small patch removal and boundary smoothing) to finally obtain the pollution reduction and carbon reduction spatial functional zoning map of the study area (e.g., Figure 5 (Illustrative image). The map divides the region into four categories: key pollution reduction and carbon reduction areas (high pressure, low carrying capacity), ecological conservation and carbon sink enhancement areas (high carrying capacity, high carbon sink), green development and potential tapping areas (high potential, medium pressure), and general control areas (all indicators are at a medium level).

[0086] Through the above steps, this invention completes the entire process from multi-source data to intelligent spatial zoning, providing environmental managers with a scientific and detailed "spatial battle map for pollution reduction and carbon reduction." like Figure 6 As shown, the working principle of this invention is as follows: I. Construction of Environmental Carrying Capacity Indicators 1. Identify “stable” weather scenarios for days.

[0087] The first step is to obtain daily atmospheric boundary layer height raster data from the European Centre for Medium-Range Weather Forecasts' Generation 5 Reanalysis (ERA5) dataset.

[0088] The second step is to use the "Create Multidimensional Raster Layer" tool in ArcGIS Pro, taking the data obtained in the first step as input, specifying the time field (usually the date contained in the file name), to generate a multidimensional raster layer, where each raster represents the daily atmospheric boundary layer height at that location and its corresponding date.

[0089] The third step involves using the "Summary Multidimensional Raster" tool, taking the results generated in the second step as input, selecting the "StdTime" dimension, selecting the "Percentile" statistical type and setting the value to 33, to generate a single-value raster. Each raster represents the 33rd percentile value of the daily atmospheric boundary layer height at that location after sorting from smallest to largest.

[0090] The fourth step involves using the "Extract by Attribute" tool, taking the results generated in the second step as input, selecting the "BY_ATTRIBUTE" method and the "StdTime" dimension, and using the results generated in the third step as the standard to set the SQL query statement to generate a new multidimensional raster layer. Each raster represents the atmospheric boundary layer height at that location that is less than or equal to the 33rd percentile and its corresponding date, which can be regarded as a "stable" meteorological scenario day.

[0091] 2. Generate daily wind speed per kilometer grid data for "stable" meteorological scenarios.

[0092] The first step is to obtain daily wind speed raster data from the real-time product dataset of the China Meteorological Administration's Land Surface Data Assimilation System (CLDAS-V2.0) from the National Meteorological Science Data Center.

[0093] The second step is to use the "Create Multidimensional Raster Layer" tool in ArcGIS Pro, taking the data obtained in the first step as input, specifying the time field, and generating a multidimensional raster layer, where each raster represents the daily wind speed at that location and its corresponding date.

[0094] The third step involves using the "resampling" tool, taking the result generated in the second step as input, setting the cell size to 1 kilometer, selecting an appropriate resampling method (such as "bilinear"), and generating a multi-dimensional raster layer of daily wind speed with a spatial resolution of 1 kilometer, while maintaining the integrity of the original time dimension.

[0095] The fourth step involves using the "Extract by Attribute" tool, taking the results generated in the third step as input, selecting the "BY_ATTRIBUTE" method and the "StdTime" dimension, and using the results generated in the fourth step of Part 1 as the standard to set the SQL query statement, generating a new multidimensional raster layer. Each raster represents all wind speeds at that location that perfectly match the date of the aforementioned "stable" weather scenario.

[0096] 3. Generate Topographic Location Index (TPI) kilometer grid data.

[0097] The first step is to acquire high-precision (e.g., 12.5-meter or 30-meter resolution) digital elevation model (DEM) raster data.

[0098] The second step is to use the "Focus Statistics" tool in ArcGIS Pro, with the data obtained in the first step as input, select the "Circular" neighborhood, set an appropriate neighborhood range (such as a radius of 500 meters or 1 kilometer), select the "Average" statistics type, and generate a neighborhood average elevation raster.

[0099] The third step is to use the "Raster Calculator" tool to perform the "DEM-Neighborhood Average Elevation" calculation to generate a TPI raster.

[0100] The fourth step involves using the "resampling" tool, taking the results generated in the third step as input, setting the cell size to 1 kilometer, selecting an appropriate resampling method (such as "nearest neighbor"), and generating a TPI kilometer grid.

[0101] 4. Generate sky openness (SVF) kilometer grid data.

[0102] The first step is to obtain surface vector data containing the building's base outline and building height attributes.

[0103] The second step is to use the "Feature to 3D" tool in ArcGIS Pro, taking the data obtained in the first step as input, selecting the building height field as "Height", and generating a 3D polygon feature class with three-dimensional height information.

[0104] The third step involves using the "Layer to Polyhedron" tool, taking the result generated in the second step as input, selecting "MULTIPATCH" as the output geometry type, and generating a 3D architectural model polyhedron feature class.

[0105] The fourth step involves using the "Create Fishing Net" tool, setting both the cell width and height to 1 kilometer, selecting the "Point" geometry type, and generating kilometer-scale grid point features covering the study area as observation points for calculating SVF values.

[0106] The fifth step involves using the "Skyline" tool, with the result generated in the third step as the "Input Polyhedron" and the result generated in the fourth step as the "Input Observation Point," to generate skyline features and skyline point features, which are used to record the outline of obstacles observed from each observation point.

[0107] Step 6: Using the "Skyline Map" tool, with the results generated in step 4 as the "Input Observation Points" and the results generated in step 5 as the "Input Skyline Elements", calculate and generate a statistical table containing the SVF value of each observation point.

[0108] Step 7: Use the "Connect Fields" tool to connect the SVF values ​​in the statistical table from Step 6 back to the kilometer grid point features based on the observation point ID field.

[0109] Step 8: Using the "point to raster" tool, take the result generated in step 7 as input, select SVF value as the value field, set the cell size to 1 kilometer, select "average" as the assignment type, and generate an SVF value kilometer grid.

[0110] 5. Generate "stable" meteorological scenario daily air diffusion index kilometer grid data.

[0111] The first step is to use the "Resampling" tool in ArcGIS Pro, taking any grid cell from the "Stable" meteorological scenario daily wind speed kilometer grid multidimensional grid generated in Part 2 as the capture grid and spatial reference, and resampling the TPI grid generated in Part 3 and the SVF grid generated in Part 4 respectively. Set the cell size to 1 kilometer, select an appropriate resampling method (such as "Bilinear"), and ensure that all input data are spatially perfectly matched.

[0112] The second step involves using a "raster calculator" tool to perform minimum-maximum value normalization on each time slice of the spatially aligned TPI raster and the "stable" meteorological scenario daily wind speed multidimensional raster, uniformly converting the numerical range to the 0-1 interval, thus generating normalized TPI and wind speed multidimensional rasteres. Since the SVF raster has a value range of [0,1], no normalization is required, and it can be directly used for subsequent calculations.

[0113] The third step involves using the "Iterative Multidimensional Raster" tool to automatically traverse each time slice in the normalized wind speed multidimensional raster generated in the second step. Combining the normalized TPI raster and SVF raster, the "Raster Calculator" tool is used to perform the "Normalized TPI × SVF × Normalized Wind Speed" calculation to generate a multidimensional raster layer, where each raster represents the daily air diffusion index at that location.

[0114] The fourth step involves using the "Summary Multidimensional Raster" tool, taking the results generated in the third step as input, selecting the "StdTime" dimension, selecting the "Average" statistical type, and generating a kilometer grid of the daily air diffusion index average value for the "Stable" meteorological scenario, which serves as the core alternative indicator for characterizing the atmospheric environmental capacity of the study area.

[0115] 6. Generate kilometer grid data for the comprehensive ecological carbon sink endowment index.

[0116] The first step is to acquire annual net primary productivity (NPP) raster data, normalized difference vegetation index (NDVI) raster data, and high-precision DEM raster data from MODIS satellite remote sensing products.

[0117] The second step, in ArcGIS Pro, is to first use the "Focus Statistics" tool, taking the DEM raster data obtained in the first step as input, selecting the "Circular Neighborhood," setting an appropriate neighborhood range (e.g., a radius of 500 meters or 1 kilometer), and selecting the "Maximum" statistics type to generate the neighborhood's maximum elevation raster. Then, using the same tools and parameters, selecting the "Minimum" statistics type, generate the neighborhood's minimum elevation raster. Next, using the "Raster Calculator" tool, perform the "Neighborhood Maximum Elevation - Neighborhood Minimum Elevation" calculation to generate the original resolution Topographic Relief (TR) raster. Finally, using the "Resampling" tool, taking the original resolution TR raster as input, setting the cell size to 1 kilometer, and selecting an appropriate resampling method (e.g., "Nearest Neighbor"), generate the TR kilometer grid.

[0118] The third step involves using a "resampling" tool, with the results generated in the second step as the capture raster and spatial reference, to resample the NPP raster and NDVI raster obtained in the first step, setting the cell size to 1 kilometer and selecting an appropriate resampling method (such as "bilinear") to ensure that all input data are spatially perfectly matched.

[0119] The fourth step involves using the "Raster Calculator" tool to perform minimum-maximum value normalization on the spatially aligned NPP, NDVI, and TR rasters, converting the numerical range to the 0-1 interval to generate normalized NPP, NDVI, and TR rasters.

[0120] Fifth step, use the "Grid Calculator" tool to perform the following steps: The calculation generates a kilometer-scale grid of ecological carbon sink comprehensive endowment index, which serves as a core alternative indicator characterizing the carbon sequestration capacity of the study area. This index comprehensively reflects the intensity of the ecosystem's carbon sequestration function, vegetation growth status, and natural stability, achieving a comprehensive assessment that integrates quantity, quality, and stability, moving beyond a single-function measurement.

[0121] II. Construction of Environmental Stress Indicators 7. Identify keywords related to pollution sources.

[0122] The first step involves collecting relevant texts of air pollution prevention and carbon emission reduction implementation plans for the study area (such as the "Implementation Plan for the Action Plan for Continuous Improvement of Air Quality") and saving them as TXT files encoded in UTF-8. Using the Python programming language, the jieba library is employed for Chinese text segmentation. A Chinese stop word list (such as the Harbin Institute of Technology stop word list) is loaded to filter out meaningless words, generating a cleaned word list.

[0123] The second step involves constructing a Term Frequency-Inverse Document Frequency (TF-IDF) model using the `TfidfVectorizer` class from the `sklearn.feature_extraction.text` module. The parameter `max_features=500` is set to retain the 500 most important feature words. The `fit_transform` method is used to fit the model and transform the text data, generating a document-term TF-IDF matrix. For each document, the top 50 feature words with the highest TF-IDF weights in its corresponding vector are extracted, generating a preliminary set of emission source keywords.

[0124] The third step involves converting the segmented and stop-word-removed text into a bag-of-words model, and using the `gensim.models.LdaModel` module to model potential Dirichlet distribution topics. Based on the conventional classification framework of my country's environmental policy system (e.g., industry, energy, transportation, construction, agriculture), the number of topics (`num_topics`) is set to 5. Using common empirical values ​​in natural language processing, the number of iterations (`passes`) is set to 10, and the random state (`random_state`) to 42 to complete model training. Subsequently, the `show_topic` method is used to extract the top 15 keywords with a probability greater than 0.01 for each topic. Cross-validation is performed using TF-IDF results, retaining keywords significant in both methods, ultimately determining the core terms for the three major topics of industry, energy, and transportation as the identified primary pollution sources.

[0125] 8. Generate the density kilometer grid data of the first industrial sector POI.

[0126] The first step is to obtain POI data. Based on the main polluting industry keywords (such as steel, coking, etc.) identified in Part 7, the corresponding types of POI data are selected.

[0127] The second step involves using the "kernel density analysis" tool in ArcGIS Pro, taking the data selected in the first step as input, setting the search radius and cell size to 1 kilometer, and generating a density kilometer grid of industrial POIs as a core alternative indicator to represent the emission pressure of major industrial pollution sources.

[0128] 9. Generate kilometer grid data of winter nighttime surface temperature (LST) for rural settlements.

[0129] The first step is to acquire high-precision (e.g., 30-meter resolution) land use type raster data and monthly nighttime LST kilometer grid data (e.g., from November to March of the following year) to characterize winter conditions.

[0130] The second step involves using the "Reclassification" tool in ArcGIS Pro, taking the land use type data obtained in the first step as input, assigning a value of 1 to the rural settlement type and 0 to other types, and generating a binary raster of rural settlements at the original resolution.

[0131] The third step involves using the "resampling" tool, taking the results generated in the second step as input, setting the pixel size to 1 kilometer, and selecting an appropriate resampling method (such as "bilinear") to generate a 1-kilometer resolution binary raster of rural settlements.

[0132] The fourth step involves using the "Cell Statistics" tool, taking the LST raster data obtained in the first step as input, selecting the "Average" statistics type, and generating a kilometer-scale raster of the average LST value for winter nights in the study area.

[0133] The fifth step involves using the "Raster Calculator" tool to perform the conditional assignment operation "Con("1 km rural settlement binary grid" == 1, "winter night LST average kilometer grid")" to generate an initial result grid that retains only the LST value on the rural settlement grid.

[0134] The sixth step involves using a "resampling" tool, with the results generated in the third step as the capture grid and spatial reference, to resample the results generated in the fifth step. The pixel size is set to 1 kilometer, and an appropriate resampling method (such as "bilinear") is selected to generate a rural residential area LST kilometer grid for winter nights, which serves as a core alternative indicator to characterize the emission pressure of major energy pollution sources.

[0135] 10. Generate traffic emission pressure kilometer grid data The first step is to acquire road vector data, population density kilometer grid data, nighttime light intensity kilometer grid data, and POI data.

[0136] The second step is to add a "Weight" field to the road vector data obtained in the first step in ArcGIS Pro. Based on the road level, assign a weight of 1 to "Expressway / Urban Expressway", 0.8 to "National Highway / Urban Arterial Road", 0.6 to "Provincial Highway / Urban Secondary Arterial Road", 0.4 to "County Road", and 0.2 to "Urban Branch Road".

[0137] The third step involves using the "Line Density" tool, taking the weighted road data generated in the second step as input, selecting the weight field as the "Population" field, setting the search radius and cell size to 1 kilometer, and generating a weighted road network density kilometer grid.

[0138] The fourth step involves using the "kernel density analysis" tool, taking the POI data obtained in the first step as input, to perform multi-category kernel density calculations according to residence, employment (such as companies and enterprises), leisure (such as catering and shopping), and public services (such as education and medical care), setting the search radius and cell size to 1 kilometer, and generating POI density kilometer grids for each category.

[0139] The fifth step, to quantify the degree of land function mixing, uses an ArcPy script to convert the POI density raster of each category generated in the fourth step into a three-dimensional NumPy array. The sum of the density values ​​of each category within each kilometer grid is calculated, thus obtaining the proportion value p of each POI category. The information entropy formula is then used. A grid-by-grid calculation is performed to generate a kilometer grid of POI functional mix. The higher the value, the more diverse the parcel's functions and the greater its pedestrian and cycling friendliness.

[0140] The sixth step involves using a "resampling" tool, with the results generated in the third step as the capture raster and spatial reference, to resample the population density raster, nighttime light intensity raster, and POI functional mixing degree raster obtained in the first step, respectively. The cell size is set to 1 kilometer, and an appropriate resampling method (such as "bilinear") is selected to ensure that all input data are spatially perfectly matched.

[0141] Step 7: Using the "Raster Calculator" tool, perform minimum-maximum value normalization on the spatially aligned weighted road network density, population density, and nighttime light intensity rasters respectively, converting the numerical range to the 0-1 interval to generate normalized weighted road network density, population density, and nighttime light intensity rasters. Since the POI function's hybridity raster has a value range of [0,1], no normalization is required, and it can be directly used for subsequent calculations.

[0142] The eighth step involves using the "grid calculator" tool to perform the calculation of "normalized weighted road network density × normalized population density × normalized nighttime light intensity × (1 - POI functional mixing degree)" to generate a traffic emission pressure kilometer grid, which serves as a core alternative indicator representing the emission pressure of major traffic pollution sources.

[0143] III. Construction of Potential Indicators for Pollution Reduction and Carbon Reduction 11. Generate per capita GDP per kilometer grid data.

[0144] The first step is to obtain GDP and population grid data.

[0145] The second step involves using a "resampling" tool. The GDP raster obtained in the first step is used as the capture raster and spatial reference. The population raster obtained in the first step is resampled separately. The cell size is set to 1 kilometer, and an appropriate resampling method (such as "bilinear") is selected to ensure that all input data are spatially perfectly matched.

[0146] The third step involves using the "grid calculator" tool to perform a "GDP / population" calculation on the spatially aligned GDP and population grids, generating a per capita GDP per kilometer grid.

[0147] 12. Generate grid data of kilometers with potential for technological innovation.

[0148] The first step is to obtain POI data and filter out the POI data of high-tech enterprises, research institutes, and research universities (such as 985 and 211 universities).

[0149] The second step involves using the "kernel density analysis" tool in ArcGIS Pro, taking the data selected in the first step as input, setting the search radius and cell size to 1 kilometer, and generating POI density kilometer grids for each category.

[0150] The third step involves using the "grid calculator" tool to perform minimum-maximum value normalization on the results generated in the second step, converting the numerical range to the 0-1 interval, and generating normalized POI density grids for high-tech enterprises, research institutes, and research universities.

[0151] The fourth step involves using a "weighted summation" tool to overlay the three normalized density grids generated in the third step, setting the weight coefficients to equal weights of 1:1:1, to generate a kilometer grid representing the technological innovation potential, which serves as a comprehensive indicator of the level of regional technological innovation resource agglomeration.

[0152] IV. Construction of Multimodal Deep Embedding Clustering Model This section aims to construct a multimodal deep learning model based on the seven kilometer grid indicators in three categories—environmental carrying capacity, environmental pressure, and pollution reduction and carbon reduction potential—constructed in the first to third sections above, in order to achieve refined spatial zoning of the pollution reduction and carbon reduction functions of the study area.

[0153] 13. Data preparation and preprocessing.

[0154] The first step, in ArcGIS Pro, is to use the "Resampling" tool, with the generated "Stable" meteorological scenario daily air diffusion index kilometer grid as the capture grid and spatial reference, to resample the generated ecological carbon sink comprehensive endowment index, industrial POI density, rural residential winter night surface temperature, traffic emission pressure, per capita GDP, and technological innovation potential grids, respectively. The cell size is set to 1 kilometer, and an appropriate resampling method (such as "Bilinear") is selected to ensure that all input data are spatially perfectly matched.

[0155] The second step involves using the "band synthesis" tool to add the seven raster files processed in the first step one by one in a strictly predetermined order: daily air diffusion index for a "stable" meteorological scenario, comprehensive ecological carbon sink endowment index, industrial POI density, winter nighttime surface temperature of rural settlements, traffic emission pressure, GDP per capita, and technological innovation potential. The "Maintain band order" option is checked to generate a multiband raster dataset containing seven bands, which fully preserves the spatial location and attribute values ​​of all pixels.

[0156] The third step involves using the "Band Ensemble Statistics" tool, taking the multi-band raster generated in the second step as input, to calculate the mean and standard deviation of each band. Then, using the "Raster Calculator" tool, based on the statistical results from the previous step, the "(original value - mean) / standard deviation" calculation is performed on each band (this process can be done in batches via loops) to generate a multi-band raster dataset that has undergone Z-score standardization, eliminating dimensional differences and providing standardized input for model training.

[0157] The fourth step involves writing a separate Python script in a text editor and running it in the ArcGIS Pro Python window or command line. This script imports the arcpy, numpy, and json libraries. The `arcpy.RasterToNumPyArray` function is used to convert the standardized multiband raster generated in the third step into a three-dimensional NumPy array with a shape of [number of rows, number of columns, 7], where the band order strictly matches the composition order. The `arcpy.Raster` function is used to create raster objects, accessing their attributes such as extent, spatialReference, meanCellWidth, meanCellHeight, width, height, and bandCount to extract key spatial metadata, which is then saved as a JSON file using the `json.dump` function. Finally, the `numpy.save` function saves the NumPy array as a .npy format file, completing the data input preparation for the deep learning model.

[0158] 14. Construct a multi-branch convolutional neural network model.

[0159] The first step, using the NumPy library, is to divide the preprocessed input tensor of shape [number of rows, number of columns, 7] into three independent branches based on the channel index, including: Environmental carrying capacity branch: Channels 0 and 1 of the slice input tensor ([:,:,0:2]), corresponding to the daily air diffusion index and ecological carbon sink comprehensive endowment index of the "stable" meteorological scenario.

[0160] Environmental pressure branch: Channels 2, 3, and 4 of the slice input tensor ([:,:,2:5]), corresponding to industrial POI density, rural residential area winter night surface temperature, and traffic emission pressure.

[0161] The potential branch for pollution reduction and carbon reduction: channels 5 and 6 of the slice input tensor ([:,:,5:7]), corresponding to per capita GDP and technological innovation potential.

[0162] The second step involves using the torch.nn library to build an independent convolutional neural network (CNN) model for each branch. Each CNN model includes a layer structure in the following order: Convolutional layer 1: Uses 32 convolutional kernels of size 3x3 with a stride of 1, and sets padding=1 to achieve "Same" padding to keep the spatial size of the output feature map unchanged; followed by batch normalization layer and ReLU activation function.

[0163] Convolutional layer 2: Uses 64 3x3 convolutional kernels with a stride of 1, and also sets padding=1 to achieve "Same" padding; followed by batch normalization layer and ReLU activation function.

[0164] Max pooling layer: Uses a 2x2 pooling window with a stride of 2 to downsample the feature map, reducing its spatial dimensions (width and height) to about half of the input.

[0165] The third step involves using the `torch.cat` function to concatenate the feature maps output by the three CNN models along the channel dimension, generating a unified, fused feature tensor. A fusion convolutional module is then constructed to perform deep interaction and abstraction on the concatenated features. This module includes the following layer structure in sequence: Fusion convolutional layer 1: Uses 128 convolutional kernels of size 3x3 with a stride of 1 and padding=1 to maintain spatial dimensions; followed by a batch normalization layer and a ReLU activation function.

[0166] Fusion convolutional layer 2: Uses 256 convolutional kernels of size 3x3 with a stride of 1 and padding=1; followed by a batch normalization layer and ReLU activation function.

[0167] The fourth step involves using the `torch.nn.AdaptiveAvgPool2d` function to perform global average pooling on the final feature map processed by the fusion convolution module. This operation aggregates all spatial information from each channel into a scalar value, thereby converting the three-dimensional feature tensor into a fixed-length one-dimensional feature vector, which serves as the deep embedding representation of the entire study area.

[0168] 15. Embedded spatial mapping and calculation of cluster probability.

[0169] The first step is to construct a fully connected layer using the torch.nn.Linear function, which maps the one-dimensional feature vector generated in the previous steps to a preset low-dimensional embedding space (e.g., 32-dimensional) as the deep feature representation of the region; then a batch normalization layer is used to stabilize the embedding vector.

[0170] The second step involves defining a trainable clustering layer using the `torch.nn.Parameter` function. Its core is a learnable parameter matrix of shape [K, embedding dimension], where K is the preset number of partitions (e.g., 4), representing the K cluster center vectors. The `torch.nn.functional.cosine_similarity` function is then used to calculate the cosine similarity between the embedding vector output from the first step and all cluster center vectors in the clustering layer, generating a K-dimensional original similarity score vector.

[0171] The third step is to use the torch.nn.functional.softmax function to normalize the original similarity score vector generated in the second step, convert it into a probability distribution, and output the soft assignment probability of each spatial unit belonging to K different partitions, with the sum of all probabilities being 1.

[0172] 16. Generate and output partitioned rasters.

[0173] The KL divergence is defined as the loss function using the `torch.nn.KLDivLoss` function. A deep embedding clustering algorithm is employed, aiming to minimize the KL divergence between the soft-assignment probability distribution output by the model and an auxiliary target distribution calculated from the current clustering results. The positional parameters of the cluster centers in the convolutional neural network, fully connected layers, and clustering layers are jointly optimized using backpropagation. After model training convergence, the `torch.argmax` function is used to operate on the soft-assignment probability distribution output from the previous steps, determining the partition category label with the highest probability for each spatial pixel, resulting in an integer partition category matrix. The `arcpy.NumPyArrayToRaster` function is used to reconstruct the partition category matrix into a pollution reduction and carbon reduction spatial partition raster map that is completely consistent with the spatial baseline of the original input data, based on the metadata (including the number of rows and columns, spatial range, pixel size, and coordinate system) saved during data preparation and preprocessing. This completes the construction of the entire model and its spatial application output.

[0174] Obviously, the above embodiments are merely illustrative examples for clear explanation and are not intended to limit the implementation. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively list all possible implementations here. However, obvious variations or modifications derived therefrom are still within the scope of protection of this invention.

Claims

1. A spatial functional zoning method for pollution reduction and carbon reduction based on multimodal deep learning, characterized in that, include: S1. Construct a multimodal indicator system for spatial functional zoning, wherein the multimodal indicator system includes environmental carrying capacity indicators, environmental pressure indicators, and pollution reduction and carbon reduction potential indicators; S2. Spatial alignment and resampling to a uniform resolution are performed on all the indicator data constructed in step S1, and then the data is standardized and merged into a multi-band raster dataset. S3. Construct a multimodal deep embedding clustering model, the model including: S31, Data preparation module, used to load the multi-band raster dataset; S32, Multi-branch feature extraction module, contains three independent convolutional neural network branches, which are used to extract features of environmental carrying capacity indicators, environmental pressure indicators and pollution reduction and carbon reduction potential indicators, respectively, and the output features of the three branches are spliced ​​together in the channel dimension; S33, Feature Fusion and Embedding Module, performs convolution and global average pooling on the concatenated features to generate a deep feature vector representing the entire region, and maps the vector to a low-dimensional embedding space. S34. Clustering probability calculation module: calculates the similarity between the embedded vector and several preset cluster center vectors, and normalizes it into a probability distribution to obtain the soft assignment probability of each spatial unit belonging to each functional partition. S4. The trained multimodal deep embedding clustering model is used to process the multiband raster dataset, and the partition category label with the highest probability is assigned to each spatial unit to generate a pollution reduction and carbon reduction spatial functional partition raster map.

2. The spatial functional zoning method for pollution reduction and carbon reduction based on multimodal deep learning according to claim 1, characterized in that, In step S1, constructing the environmental carrying capacity index includes: S11. Based on atmospheric boundary layer height data, identify "stable" weather scenarios for the study area; S12. Based on the "stable" meteorological scenario day, combined with wind speed data, terrain location index and sky openness data, the air diffusion index of the "stable" meteorological scenario day is calculated.

3. The spatial functional zoning method for pollution reduction and carbon reduction based on multimodal deep learning according to claim 2, characterized in that, Step S11 specifically includes: Calculate the 33rd percentile of the daily atmospheric boundary layer height for each spatial location; Dates with daily atmospheric boundary layer height less than or equal to the 33rd percentile of their corresponding location are selected as "stable" weather scenario days for that location.

4. The spatial functional zoning method for pollution reduction and carbon reduction based on multimodal deep learning according to claim 2, characterized in that, In step S12, the daily air diffusion index for the "stable" meteorological scenario is calculated using the following formula: ,in, A normalized topographic location index, For the openness of the sky, The wind speed is the intraday normalized wind speed for the "stable" weather scenario.

5. The spatial functional zoning method for pollution reduction and carbon reduction based on multimodal deep learning according to claim 1, characterized in that, In step S1, the environmental carrying capacity index also includes the comprehensive ecological carbon sink endowment index, which is constructed as follows: net primary productivity data, normalized vegetation index data, and topographic relief data are obtained, normalized respectively, and then calculated. ,in , , These are normalized net primary productivity, normalized vegetation index, and topographic relief, respectively.

6. The spatial functional zoning method for pollution reduction and carbon reduction based on multimodal deep learning according to claim 1, characterized in that, In step S1, the environmental pressure index includes the traffic emission pressure index, which is constructed as follows: based on weighted road network density data, population density data, nighttime light intensity data, and point-of-interest (POI) functional hybridity data, it is calculated... ,in , , These are the normalized weighted road network density, population density, and nighttime light intensity, respectively. For the range of values ​​within Points of interest feature blending degree.

7. The spatial functional zoning method for pollution reduction and carbon reduction based on multimodal deep learning according to claim 1, characterized in that, In step S2, the standardization process is Z-score standardization, which is performed based on the mean and standard deviation of the data for each band.

8. The spatial functional zoning method for pollution reduction and carbon reduction based on multimodal deep learning according to claim 1, characterized in that, In step S32, each of the three independent convolutional neural network branches contains at least two convolutional layers, a batch normalization layer, an activation function layer, and a pooling layer. In step S33, the feature fusion and embedding module further includes a fusion convolution submodule, which is used to perform further convolution processing on the concatenated features. The fusion convolution submodule includes at least one convolutional layer, a batch normalization layer, and an activation function layer. In step S34, the similarity is cosine similarity, and the normalization process uses the softmax function.

9. The method for spatial functional zoning for pollution reduction and carbon reduction based on multimodal deep learning according to claim 1, characterized in that, In step S4, the argmax function is used to operate on the soft allocation probability distribution to determine the partition category with the highest probability for each spatial unit.

Images