An urban 100-meter-level nitrogen dioxide concentration distribution super-resolution analysis method

Abstract

The present application belongs to the field of environmental remote sensing atmospheric chemistry field downscaling technology in the field of atmospheric pollution monitoring technology, specifically a kind of urban hundred-meter level nitrogen dioxide concentration distribution super-resolution analysis method.The present application uses neural network technology, with the preprocessed nitrogen dioxide concentration data obtained by hyperspectral satellite remote sensing, real-time traffic data, POI data, WRF meteorological field data as the training set, using mean square error loss, regional mean square error loss and correlation loss three losses as the loss function in training, nitrogen dioxide near-surface hourly average concentration of national control site as label data set for training, to obtain hundred-meter resolution nitrogen dioxide concentration distribution.The present application needs a variety of high-resolution auxiliary data, the time resolution and spatial resolution of the obtained data set are very high, and the consistency with the verification data is very high.

Description

Technical Field

[0001] This invention belongs to the field of atmospheric pollution monitoring technology, specifically involving a super-resolution analysis method for the distribution of nitrogen dioxide concentration at the 100-meter level in urban areas. Background Technology

[0002] Nitrogen oxides, as precursors to ozone formation, are closely related to the degree of ozone pollution. Nitrogen dioxide, as the most significant type of nitrogen oxide, is a typical air pollutant. Currently, my country is vigorously promoting the reduction and control of nitrogen dioxide emissions. Industrial and transportation sources, as the main sources of nitrogen dioxide pollution, have been receiving much attention. To accurately control factory and vehicle emissions and pinpoint the location of emission sources, higher resolution nitrogen dioxide concentration distribution data is needed.

[0003] Currently, common methods for detecting nitrogen dioxide include point measurements using chemical methods and remote sensing instruments using optical absorption methods. Point instruments are commonly found in national automatic environmental monitoring stations, offering high detection accuracy but limited spatial range. Remote sensing instruments are frequently used in satellite, airborne, and ground-based remote sensing, providing a large detection range but with lower resolution and accuracy. In the field of atmospheric chemical field downscaling, deep learning methods can generally improve the scale to 1 kilometer, but due to the limitations of traditional measurement methods, it is difficult to improve to the scale of hundreds of meters. Summary of the Invention

[0004] The purpose of this invention is to provide a super-resolution method for analyzing nitrogen dioxide concentration at the 100-meter level in any city, which has a large detection range and high detection accuracy.

[0005] The present invention provides a super-resolution method for analyzing nitrogen dioxide concentration at a resolution of 100 meters in any city. This method employs neural network technology, using pre-processed hyperspectral satellite remote sensing data of nitrogen dioxide concentration, real-time traffic data, POI data, and WRF meteorological field data as the training set. It uses three loss functions during training: mean squared error loss, regional mean mean squared error loss, and correlation loss. The hourly average near-surface nitrogen dioxide concentration from national monitoring stations is used as the label dataset for training, resulting in a nitrogen dioxide concentration distribution at a resolution of 100 meters. The specific steps are as follows:

[0006] Step 1: Preprocess the four types of data obtained from hyperspectral satellites: nitrogen dioxide concentration data, real-time traffic data, POI data (point of interest), and some meteorological data from the WRF meteorological field. Match them in terms of resolution and use them as the input dataset for the neural network, denoted as dataset I.

[0007] (a) Preprocessing of nitrogen dioxide data: Before training, perform bilinear interpolation on the data to interpolate it to the required resolution (hundred meters) so that it can match the subsequent high-resolution data.

[0008] (II) Preprocessing of real-time traffic data, including: extracting road network data as the base value, extracting real-time traffic data of roads, extracting latitude and longitude information of each road node, coordinate system transformation, and converting road network and real-time traffic data into grid data:

[0009] (1) Extract urban road information based on road network data. Use the extracted urban road information to extract congestion information for each road (the extracted data is in JSON format), which includes data such as the time of extraction of road congestion information, road name, road congestion level, and road congestion range; the road congestion range is defined by dividing the road into several nodes, and defining the congestion range in a way such as "near a certain node" or "from a certain node to a certain node";

[0010] (2) Use road network data to automatically obtain the latitude and longitude information of road nodes (intersection-type road nodes); the road network data contains a large amount of road vector data. Using this data, the latitude and longitude of each road intersection can be obtained. For two intersecting roads, each segment of their vector data is traversed and it is determined whether they intersect. If they intersect, the latitude and longitude of the intersection point is extracted as the latitude and longitude information of the intersection of the two roads.

[0011] (3) Use geocoding service to automatically obtain road node information, including obtaining the latitude and longitude information of each road node (landmark road node); geocoding service can return the latitude and longitude information of multiple location points according to given information, such as the name of geographic coordinates, extract the latitude and longitude of the first location point, convert the latitude and longitude information from bd09 coordinate system to WG84 coordinate system, and determine the distance to the corresponding road. If the distance is less than 300m, the latitude and longitude information is determined to be the latitude and longitude information of the node.

[0012] (4) Manually filter the node information obtained automatically above, and manually obtain node information whose latitude and longitude cannot be obtained by the above method through the coordinate picking service;

[0013] (5) The basic values ​​based on road network information are gridded. Specifically, the vector information is converted into road density values ​​within each grid point, and different weights are set for different road density levels. See Table 2 for details of the specific road levels, where the different weights are set as follows:

[0014] highway 1 Secondary arterial roads 0.7 Main roads 0.9 Third-level arterial road 0.5 Main roads 0.8 Unnumbered roads 0.25

[0015] Based on the above road grade and weight settings, the road network information is converted into grid data;

[0016] (6) Real-time traffic data is gridded. Specifically, the road congestion level and congestion range in the data are converted into road density values ​​within each grid point. There are four road congestion levels, represented by 1 to 4, and the road congestion range is determined by road node information and road congestion length. The specific formula is as follows:

[0017] RD bi =∑ k l bk , Formula 1

[0018] RD ri =∑ k l rk , Formula 2

[0019] RD i =RD bi +RD ri , Formula 3

[0020] In formula 1, RD bi Let l represent the road base density value of the i-th grid point, k represent the k-th road within the grid point, and l represent the road base density value of the i-th grid point. bk This represents the base density value of the k-th road within the grid; in Formula 2, RD ri Let l represent the road congestion density value at the i-th grid point, k represent the k-th congested road within the grid point, and l represent the road congestion density value at the i-th grid point. rk This represents the road congestion density value of the k-th congested road within the grid point; in Formula 3, RD i This represents the total road density value within the i-th grid point.

[0021] (III) Preprocessing of POI Data: POI data contains information on various locations within a city, including restaurants, factories, and tourist attractions. During processing, POI data is categorized into industrial, energy, and residential emission sources according to the emission inventory. It is then converted into gridded data based on POI density and processed and merged using the following formula:

[0022] POI=η(ind)*ind / sum(ind)+η(pow)*pow / sum(pow)+η(res)*res / sum(res), formula 4

[0023] In Formula 4, ind represents industrial density, pow represents energy density, res represents residential density, and η represents the emission proportions of several emission sources, based on the annual average emissions of the emission inventory.

[0024] (iv) Preprocessing of WRF meteorological field data: The WRF meteorological field data uses the latitude and longitude wind fields at 10m near the ground. To ensure consistent resolution, the relative humidity and temperature data at 2m near the ground are interpolated using the nearest neighbor interpolation method to make them consistent with the bilinear interpolated satellite remote sensing data.

[0025] The above data are normalized to [0,1] and merged into a data set of length 7, which is used as the input dataset for the neural network and denoted as dataset I.

[0026] Step 2: Extract data from nationally monitored sites as a label dataset, denoted as dataset L. Convolutional neural networks are built using dataset I as the network input dataset and dataset L as the network's label dataset. Mean squared error loss, regional mean mean squared error loss, and correlation loss are used as three loss functions for training to obtain the nitrogen dioxide distribution at the 100-meter level. The three loss functions are denoted as Loss... M Loss A Loss P .

[0027] Building a convolutional neural network mainly involves feature extraction networks and feature fusion networks, and their network structures are as follows: Figure 1 As shown.

[0028] The feature extraction network is mainly divided into three parts, which extract features from satellite remote sensing data of nitrogen dioxide, POI data, and real-time road conditions, respectively:

[0029] (1) The satellite remote sensing nitrogen dioxide data is denoted as X1, and the data after passing X1 through a convolutional layer is denoted as F(X1);

[0030] (2) Merge the POI density data with the WRF meteorological data and denote it as X2. Denote the data after passing X2 through a convolutional layer as F(X2);

[0031] (3) Combine the real-time road condition density data with the WRF meteorological data and denote it as X3. Then, denote the data after passing X3 through a convolutional layer as F(X3).

[0032] In the feature fusion network, F(X1), F(X2), and F(X3) are merged into a single data point of length 3, denoted as X'. This is first processed through a convolutional layer, then activated by the LReLU activation function, and finally processed through another convolutional layer. The data after convolution is denoted as Y, representing the output of the convolutional neural network, which is the nitrogen dioxide result per 100 meters.

[0033] In this invention: the correlation loss P The formula for calculation is:

[0034]

[0035] Among them, Loss P For correlation loss, X is the output data of the neural network, and Y is the label data of the national control point.

[0036] The mean square error loss Loss M The formula for calculation is:

[0037] Loss M =E(XY) 2 , formula 6

[0038] Where X represents the output data of the neural network, and Y represents the label data of the national control point.

[0039] The mean squared error loss of the region A The formula for calculation is:

[0040]

[0041] in, The mean of the region output data of the neural network. The mean of the input data for the neural network.

[0042] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0043] (1) This invention proposes a method to downscale satellite remote sensing data to a resolution of 100 meters. Real-time traffic data and POI (Point of Interest) data containing high-resolution information are added to the input data as part of the input data, which reduces the scale of satellite remote sensing to the 100-meter level and can effectively identify the specific sources of urban pollution.

[0044] (2) In the training of neural networks, this invention proposes to use the correlation between network output data and label data as a loss function, which makes the network converge faster during training and significantly improves the consistency between the downscaled data and the national control point verification. Attached Figure Description

[0045] Figure 1 This is a framework diagram of the method of the present invention.

[0046] Figure 2 This is an example of real-time traffic conditions in Hefei.

[0047] Figure 3 This is a schematic diagram of Hefei's road network data.

[0048] Figure 4 Road data converted to grid data (this map does not include real-time traffic data).

[0049] Figure 5This is a schematic diagram illustrating the data density of residential buildings as POI (Point of Interest) data.

[0050] Figure 6 To verify the distribution of ground-based national monitoring stations.

[0051] Figure 7 An application case of identifying pollution sources using nitrogen dioxide concentration reconstruction results at a resolution of 100 meters.

[0052] Figure 8 A comparison of time series plots of national control points for nitrogen dioxide reconstructed at the 100-meter level with the original results at the kilometer level. This comparison specifically corresponds to the time series plots of four national control points, from 1270A to 1273A.

[0053] Figure 9 A comparison of time series plots of national control points for reconstructed nitrogen dioxide at the 100-meter level and the original results at the kilometer level. This comparison specifically corresponds to the time series plots of four national control points, 1274A-1277A.

[0054] Figure 10 A comparison of time series plots of national control points for nitrogen dioxide reconstructed at the 100-meter level and the original results at the kilometer level is presented. The comparison specifically corresponds to the time series plots of three national control points: 1278A, 1279A, and 3464A. Detailed Implementation

[0055] The present invention will be further described in detail with reference to the accompanying drawings.

[0056] In the following implementation case, the implementation area of ​​the 100-meter-level nitrogen dioxide super-resolution method is Hefei City. Taking this area as an example, the complete implementation process of the present invention will be explained.

[0057] Step 1 involves preprocessing four types of data: nitrogen dioxide concentration data obtained from hyperspectral satellites, real-time traffic data, POI data (point of interest), and some meteorological data from the WRF meteorological field. These data are matched in terms of resolution and used as the input dataset for the neural network, denoted as dataset I.

[0058] Satellite remote sensing data of nitrogen dioxide was preprocessed and reconstructed to a kilometer resolution using artificial intelligence methods, covering the period from January 2021 to April 2022 with a time resolution of 15 minutes. Then, it was subjected to bilinear interpolation, interpolating 16 times to a 62.5m×62.5m grid to match the subsequent high-resolution data.

[0059] Preprocessing of real-time traffic data mainly includes: extracting road network data as base values, extracting real-time traffic data for roads, extracting latitude and longitude information for each road node, coordinate system transformation, and converting the road network and real-time traffic data into grid data.

[0060] (1) Based on road network data (such as...) Figure 3 As shown in the image, road data for Hefei was extracted. Real-time traffic conditions were extracted for each road. Due to a limited number of extractions, only 62 roads from the first three levels (road classification standards are detailed in Table 2) were selected for extraction, basically covering all major arterial roads in Hefei. Real-time traffic conditions are updated in real-time (minute-level). The congested sections and congestion levels of each road are displayed, providing extremely high spatiotemporal resolution road and vehicle information (e.g., ...). Figure 2 As shown in the figure, to ensure correspondence with the kilometer-resolution nitrogen dioxide data, the extraction time range was from 8:00 AM to 4:00 PM daily, with extraction every 15 minutes. The real-time traffic data was extracted from February 1st to April 30th, 2023. The road congestion range was defined by dividing the road into several nodes, using terms such as "near a certain node" or "from a certain node to a certain node".

[0061] (2) Use road network data to automatically obtain the latitude and longitude information of road nodes (intersection-type road nodes). The road network data contains a large amount of road vector data. Use this data to obtain the latitude and longitude of each road intersection. For two intersecting roads, traverse each segment of their vector data and determine whether they intersect. If they intersect, extract the latitude and longitude of the intersection point as the latitude and longitude information of the intersection of the two roads.

[0062] (3) Use geocoding service to automatically obtain the latitude and longitude information of each road node (landmark road node). The geocoding service can return the latitude and longitude information of multiple location points according to the given information, such as the name of the geographic coordinates. Extract the latitude and longitude of the first location point and convert the latitude and longitude information from the bd09 coordinate system to the WG84 coordinate system. Determine the distance to the corresponding road. If the distance is less than 300m, determine that the latitude and longitude is the latitude and longitude information of the node.

[0063] (4) Manually filter the node information obtained automatically above, and manually obtain node information whose latitude and longitude cannot be obtained by the above method through the coordinate picking service;

[0064] (5) The basic values ​​based on road network information are gridded. Specifically, the vector information is converted into road density values ​​within each grid point, and different weights are set for different road density levels. See Table 2 for details of the specific road levels, where the different weights are set as follows:

[0065] highway 1 Secondary arterial roads 0.7 Main roads 0.9 Third-level arterial road 0.5 Main roads 0.8 Unnumbered roads 0.25

[0066] Based on the above road classification and weight settings, the road network information is converted into grid data (such as...). Figure 4 (as shown);

[0067] (6) Real-time traffic data is gridded. Specifically, the road congestion level and congestion range in the data are converted into road density values ​​within each grid point. There are four road congestion levels, represented by 1 to 4, and the road congestion range is determined by road node information and road congestion length. The specific formula is as follows:

[0068] RD bi =∑ k l bk , Formula 1,

[0069] RD ri =∑ k l rk , Formula 2,

[0070] RD i =RD bi +RD ri , Formula 3,

[0071] In Formula 1, RD bi Let l represent the road base density value of the i-th grid point, k represent the k-th road within the grid point, and l represent the road base density value of the i-th grid point. bk This represents the base density value of the k-th road within the grid; RD in Formula 2 ri Let l represent the road congestion density value at the i-th grid point, k represent the k-th congested road within the grid point, and l represent the road congestion density value at the i-th grid point. rk This represents the road congestion density value of the k-th congested road within the grid point; RD in Formula 3 i This represents the total road density value within the i-th grid point.

[0072] The POI data underwent preprocessing. POI data contains information on various locations within a city, including restaurants, factories, and tourist attractions. Hefei City has over 600,000 POI records. When processing this data, the proportion of various nitrogen dioxide emission sources excluding transportation sources was referenced from the emissions inventory. Industrial sources accounted for 86.64%, energy sources for 10.17%, and residential sources for 3.2%. The POI data was categorized into these three types based on keywords, and then manually filtered. The resulting data included over 4,000 industrial records, 995 energy records, and over 500,000 residential records. This data was then converted into grid data according to POI density and processed and merged using the following formula:

[0073] POI=86.64%*ind / sum(ind)+10.17%*pow / sum(pow)+3.2%*res / sum(res), formula 4,

[0074] In the formula, ind represents industrial density, pow represents energy density, and res represents residential density.

[0075] The WRF meteorological field data was preprocessed. The WRF meteorological field data used latitudinal and longitude wind fields at 10m near the ground. To ensure consistent resolution, the relative humidity and temperature data at 2m near the ground were interpolated using the nearest neighbor interpolation method to match the bilinearly interpolated satellite remote sensing data.

[0076] The above data are normalized to [0,1] and merged into a data set of length 7, which is used as the input dataset for the neural network and denoted as dataset I.

[0077] Step 2: Extract data from nationally monitored sites as a label dataset, denoted as dataset L. Convolutional neural networks are built using dataset I as the network input dataset and L as the network's label dataset. Mean squared error loss, regional mean mean squared error loss, and correlation loss are used as three loss functions for training to obtain the nitrogen dioxide distribution at the 100-meter level. The three loss functions are denoted as Loss... M Loss A Loss P .

[0078] Building a convolutional neural network mainly involves feature extraction networks and feature fusion networks, and their network structures are as follows: Figure 1 As shown.

[0079] The feature extraction network is mainly divided into three parts to extract features from satellite remote sensing nitrogen dioxide data, POI data, and real-time road condition data respectively:

[0080] (1) The satellite remote sensing nitrogen dioxide data is denoted as X1, and the data after passing X1 through a convolutional layer is denoted as F(X1);

[0081] (2) Merge the POI density data with the WRF meteorological data and denote it as X2. Denote the data after passing X2 through a convolutional layer as F(X2);

[0082] (3) Combine the real-time road condition density data with the WRF meteorological data and denote it as X3. Then, denote the data after passing X3 through a convolutional layer as F(X3).

[0083] In the feature fusion network, F(X1), F(X2), and F(X3) are merged into a single data point of length 3, denoted as X'. This data is then passed through a convolutional layer, processed using the LReLU activation function, and then through another convolutional layer. The resulting convolutional data is denoted as Y, representing the output of the convolutional neural network, which is the nitrogen dioxide result over 100 meters. The proportions of the various losses during training are as follows: Loss M Loss A Loss PThe ratio was 8:1:1, resulting in a final correlation of 0.8126. Specific correlation validation results are shown in Table 3. The 11-fold cross-validation (using a model trained with data from the target site removed to predict and validate the data from that site) is shown in Table 4. The time series plot is shown below. Figure 8 As shown in Table 3, the correlation between the near-surface nitrogen dioxide concentration distribution at the 100-meter level after neural network super-resolution and the national monitoring stations, whether for individual national monitoring stations or for overall verification, is significantly improved compared to the kilometer-level data. The excellent 11-fold cross-validation results in Table 4 demonstrate the spatial extrapolation capability of the neural network model.

[0084] Table 1. List of Loss Functions for Convolutional Neural Networks

[0085]

[0086] Table 2 Road Classification

[0087]

[0088] Table 3. Correlation verification changes of Hefei national monitoring stations from January 2021 to April 2022 at a 15-minute time resolution.

[0089]

[0090] Table 4. 11-fold cross-validation table of Hefei national monitoring stations from January 2021 to April 2022 at 15-minute time resolution.

[0091]

[0092] References:

[0093] [1] Kurokawa, J. and Ohara, T.: Long-term historical trends in airpollutant emissions in Asia: Regional Emission inventory in ASia(REAS) version 3, Atmos. Chem. Phys., 20, 12761–12793, https: / / doi.org / 10.5194 / acp-20-12761-2020, 2020.

Claims

1. A super-resolution analysis method for urban nitrogen dioxide concentration distribution at the 100-meter level, characterized in that, Neural network technology was employed, using preprocessed hyperspectral satellite remote sensing data of nitrogen dioxide concentration, real-time traffic data, point-of-interest (POI) data, and meteorological field data from the WRF numerical model for meteorological research and forecasting as the training set. Three loss functions were used in training: mean squared error loss, regional mean mean squared error loss, and correlation loss. The near-surface hourly mean concentration of nitrogen dioxide from national monitoring stations was used as the label dataset for training, resulting in a nitrogen dioxide concentration distribution with a 100-meter resolution. The specific steps are as follows: Step 1 involves preprocessing the nitrogen dioxide concentration data, real-time traffic data, POI data, and WRF weather field data obtained from hyperspectral satellites, matching them in terms of resolution, and using these as the input dataset for the neural network; where: (a) Preprocessing of nitrogen dioxide data: Bilinear interpolation is performed to interpolate the data to the required 100-meter resolution so that it can be matched with subsequent high-resolution data; (ii) Preprocessing of real-time traffic data: including: extracting road network data as the base value, extracting real-time traffic data of roads, extracting latitude and longitude information of each road node, and converting road network and real-time traffic data into grid data through coordinate system transformation; (III) Preprocessing of POI Data: POI data contains information on various locations within a city, including restaurants, factories, and tourist attractions. During processing, POI data is categorized into industrial, energy, and residential emission sources according to the emission inventory. It is then converted into grid data based on POI density and processed and merged using the following formula: POI=η(ind)*ind / sum(ind)+η(pow)*pow / sum(pow)+η(res)*res / sum(res), In the formula, ind is the industrial density, pow is the energy density, res is the residential density, and η is the proportion of emissions from several emission sources, which is based on the annual average emissions of the emission inventory. (iv) Preprocessing of WRF meteorological field data: The WRF meteorological field data uses near-surface latitudinal wind field and longitude wind field, near-surface relative humidity and temperature data. In order to ensure consistent resolution, the WRF data is interpolated using the nearest neighbor interpolation method to be consistent with the bilinear interpolated satellite remote sensing data. The above data are normalized to [0,1] and merged into a data set of length 7, which is used as the input dataset for the neural network and denoted as dataset I. Step 2: Extract data from national monitoring stations as a label dataset, denoted as dataset L. Convolutional neural networks are built using dataset I as the network input dataset and dataset L as the network's label dataset. Mean squared error loss, regional mean mean squared error loss, and correlation loss are used as loss functions during training to obtain the nitrogen dioxide distribution at the 100-meter level. The three loss functions are denoted as Loss... M Loss A Loss P ; Building a convolutional neural network involves two main components: a feature extraction network and a feature fusion network. The feature extraction network consists of three parts, which extract features from satellite remote sensing data on nitrogen dioxide, POI data, and real-time road conditions, respectively: (1) The satellite remote sensing nitrogen dioxide data is denoted as X1, and the data after passing X1 through a convolutional layer is denoted as F(X1); (2) Merge the POI density data with the WRF meteorological data and denote it as X2. Denote the data after passing X2 through a convolutional layer as F(X2); (3) Combine the real-time road condition density data with the WRF meteorological data and denote it as X3. Denote the data after passing X3 through a convolutional layer as F(X3); The feature fusion network combines F(X1), F(X2), and F(X3) into a single data point of length 3, denoted as X. ’ First, the data is processed through a convolutional layer, then through an LReLU activation function, and then through another convolutional layer. The data after convolution is denoted as Y, which is the output of the convolutional neural network, i.e., the nitrogen dioxide result per 100 meters.

2. The super-resolution analysis method for urban nitrogen dioxide concentration distribution at the 100-meter level according to claim 1, characterized in that, The preprocessing of real-time traffic data mentioned in step 1 is as follows: (1) Extract urban road information based on road network data; use the extracted urban road information to extract congestion information for each road, including the time of extraction of road congestion information, road name, road congestion level, and road congestion range data; the road congestion range is defined by dividing the road into several nodes and delineating the congestion range. (2) Use road network data to automatically obtain the latitude and longitude information of road nodes at intersections; the road network data contains a large amount of road vector data. Using this data, the latitude and longitude of each road intersection can be obtained. For two intersecting roads, each segment of their vector data is traversed and it is determined whether they intersect. If they intersect, the latitude and longitude of the intersection point is extracted as the latitude and longitude information of the intersection of the two roads. (3) Use geocoding service to automatically obtain road node information, including obtaining the latitude and longitude information of each landmark road node; geocoding service returns the latitude and longitude information of multiple location points according to the given information, including the name of the geographic coordinates, extracts the latitude and longitude of the first location point, converts the latitude and longitude information from the bd09 coordinate system to the WG84 coordinate system, and judges the distance to the corresponding road. If the distance is less than 300m, the latitude and longitude is determined to be the latitude and longitude information of the node. (4) Manually filter the node information obtained automatically above, and manually obtain the node information whose latitude and longitude cannot be obtained by the above method; (5) The basic values ​​based on road network information are gridded. Specifically, the vector information is converted into road density values ​​within each grid point, and different weights are set for different levels of road density. (6) Real-time traffic data is gridded. Specifically, the road congestion level and road congestion range in the data are converted into road density values ​​within each grid point. There are four levels of road congestion, represented by 1 to 4. The road congestion range is determined by road node information and road congestion length. The specific calculation formula is as follows: Official 1 Official 2 Official 3 In formula 1, This represents the road base density value at the i-th grid point, and k represents the k-th road within that grid point. This represents the base density value of the k-th road within the grid; in Formula 2, This represents the road congestion density value at the i-th grid point, where k represents the k-th congested road within that grid point. This represents the road congestion density value of the k-th congested road within the grid point; in Formula 3, This represents the total road density value within the i-th grid point.

3. The super-resolution analysis method for urban nitrogen dioxide concentration distribution at the 100-meter level according to claim 1, characterized in that, In step 2, the correlation loss The formula for calculation is: ， in, For correlation loss, Output data for the neural network. National control point label data; The mean square error loss Loss M The formula for calculation is: Loss M , in, Output data for the neural network. National control point label data; The mean squared error loss of the region A The formula for calculation is: Loss A , in, The mean of the region output data of the neural network. The mean of the input data for the neural network.

Images