Reconstruction method of three-dimensional distribution of regional atmospheric pollutants with kilometer resolution
By combining hyperspectral remote sensing networks and deep learning technology, a convolutional neural network was constructed, and the three-dimensional distribution of regional atmospheric pollutants was reconstructed using multiple data sources. This solved the problem of achieving high-resolution three-dimensional reconstruction in existing technologies and enabled high-precision monitoring of pollutant distribution.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-12-10
- Publication Date
- 2026-03-27
AI Technical Summary
Existing environmental monitoring instruments are insufficient for monitoring the three-dimensional distribution of regional air pollutants, especially in complex and variable environments, and cannot meet the requirements for high-resolution three-dimensional reconstruction.
By combining hyperspectral remote sensing networks and deep learning technologies, and constructing a convolutional neural network, we can reconstruct the kilometer-level resolution three-dimensional distribution of regional atmospheric pollutants using MAX-DOAS horizontal remote sensing, UAV remote sensing, and ground monitoring station data, combined with high-resolution satellite data and meteorological information.
It has achieved high-precision three-dimensional reconstruction of the regional atmospheric pollutant distribution, providing a scientific basis and data support for the analysis of the causes of atmospheric pollutants and their reduction and control.
Smart Images
Figure CN116188705B_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The application belongs to the technical field of optical measurement, and particularly relates to a reconstruction method for regional atmospheric pollutant three-dimensional distribution with kilometer resolution. BACKGROUND
[0002] With the progress of society and the improvement of people's living standards, people pay more and more attention to the environment for survival. To avoid the continuous deterioration of the environment and solve the environmental problems has gradually become the consensus of people.
[0003] At present, most of the environmental monitoring instruments are point sampling, and it is not realistic to arrange a large range of three-dimensional distribution due to the value of atmospheric environmental monitoring instruments, and it is difficult to meet the needs of complex and variable environmental monitoring. Some optical monitoring instruments break through this limitation. For example, laser radar can observe the distribution of some pollutants along the observation direction at a long distance; Fourier infrared spectrometer can detect the relative concentration of atmospheric pollutants according to the influence of pollutants on the interference pattern; and hyperspectral instrument can obtain the concentration of pollutants according to the specific absorption of light of different wavelengths. However, these instruments cannot monitor the three-dimensional distribution of regional pollutants, and the existing deep learning results can only reconstruct the horizontal distribution of atmospheric pollutants near the ground. Due to the emission source and complex atmospheric chemical reaction, the distribution of atmospheric pollutants changes very rapidly with time and space, and it is of great research value to obtain the three-dimensional distribution of regional pollutants. SUMMARY
[0004] The purpose of the present application is to provide a reconstruction method for regional atmospheric pollutant three-dimensional distribution with kilometer resolution, which can monitor the three-dimensional distribution of regional pollutants, so as to realize the reconstruction of the three-dimensional spatial distribution of atmospheric pollutants in a certain area, and provide a scientific basis for the cause analysis and control reduction of atmospheric pollutants.
[0005] The reconstruction method for regional atmospheric pollutant three-dimensional distribution with kilometer resolution provided by the present application combines the vertical distribution profile of atmospheric pollutants observed by hyperspectral remote sensing network with deep learning technology, and reconstructs the three-dimensional distribution of atmospheric pollutants in the observation area with kilometer resolution according to the hyperspectral observation data, meteorological simulation data and geographic information. The specific steps are as follows:
[0006] Step 1: combine MAX-DOAS horizontal remote sensing, unmanned aerial vehicle remote sensing and ground monitoring station in-situ monitoring into a data with a length of 3, each data allows 1-2 missing data, construct a data set, and mark it as label L1.
[0007] Step 2, constructing a convolutional neural network N1; constructing a data set D1 from the vertical column concentration of atmospheric pollutants in the target area continuously obtained by high-resolution satellites, meteorological information obtained by atmospheric chemical models, and geographic information such as population, road network, and land cover type; taking D1 as the input data of the convolutional neural network N1 and L1 as the label of the convolutional neural network N1, obtaining a model M1 for high-precision near-surface atmospheric pollutant data through training, and the loss function is the weighted average of the average root mean square error of the non-missing values of the three data in the label L1;
[0008] The specific structure of the constructed convolutional neural network N1 (see Figure 3 ) includes:
[0009] According to the time selection of the label L1 data, the adjacent time D1 data is generated to obtain its time scale data; and according to the latitude and longitude selection of the label L1 data, a three-dimensional matrix (longitude, latitude, height) is generated from the adjacent position D1 data to obtain its spatial scale data.
[0010] Note: Each data of L1 has its latitude and longitude (spatial information, assuming that the data position of L1 is (x, y)) and collection time (time information, assuming that the data collection time of L1 is t);
[0011] Compared with the satellite data and meteorological data in D1, it can be regarded as point sampling, while the data in D1 can be regarded as three-dimensional data (longitude, latitude, height) in space and continuous in time. In the time scale, the data at (x, y) at time t and a period of time before time t is taken, and the data of the surrounding points at (x, y) at time t is taken in the spatial scale.
[0012] The time scale data and spatial scale data of the satellite observed pollutant column concentration, near-surface meteorological information, and geographic information are respectively spliced through two residual structures, and then spliced again through a residual structure to obtain data, which are respectively denoted as D Resnet1 , D Resnet2 , and D Resnet3 .
[0013] After splicing D Resnet1 , D Resnet3 and D Resnet2 through a residual structure, the obtained data is spliced with D Resnet3 again through a residual structure, and then the data is output through a fully connected layer, and the result is denoted as L1'.
[0014] The loss value of L1 and L1' is calculated using the loss function F1, and the network parameters are updated using the RMSProp optimizer to optimize the loss function result.
[0015] The residual structure is specifically (seeFigure 3 The specific structure of the constructed convolutional neural network N2 (see FIG. 2) includes: an average pooling layer, six convolutional layers, and a standardization layer between the convolutional layers, an activation function;
[0016] (1) The data after the input data passes through the average pooling layer is denoted as X; the data obtained by passing X through a convolutional layer and using a batch standardization layer is denoted as F1(X); the data obtained by passing X through a convolutional layer and using a batch standardization and a relu activation function is denoted as F2(X); the data obtained by passing X through three convolutional layers, using a batch standardization and a relu activation function between the layers, and then using a batch standardization is denoted as F3(X);
[0017] (2) The data obtained by adding F1(X) and F3(X) and using a relu activation function is denoted as H1(X), and the data obtained by passing H1(X) through a convolutional layer and using a batch standardization and a relu activation function is denoted as G2(X);
[0018] (3) The data obtained by adding F2(X) and G2(X) is output.
[0019] The loss function F1 used for training the convolutional neural network N1 is:
[0020]
[0021] In the formula, the subscripts 0, 1, and 2 of L1 and L1' respectively represent three kinds of data of the labels L1 and L1', and A, B, and C respectively represent the number of non-missing values of the three kinds of data of the labels L1 and L1'.
[0022] Step 3: Constructing a convolutional neural network N2; constructing a data set D2 using the kilometer-level meteorological data obtained from the atmospheric chemical model, the tropospheric atmospheric pollutant data obtained from satellite observation, the MAX-DOAS background station vertical remote sensing observation results, and the high-precision near-surface atmospheric pollutant data obtained in step 2; taking D2 as the input data of the convolutional neural network N2, and taking the layered standardized MAX-DOAS vertical remote sensing observation results as the label L2; through training, obtaining a model M2 of the atmospheric pollutant stereoscopic distribution results, and using the logarithm of the hyperbolic cosine (Log-Cosh) of the error as the loss function.
[0023] The specific structure of the constructed convolutional neural network N2 (see FIG. 2) includes: Figure 4
[0024] The satellite observation tropospheric atmospheric pollutant data, the MAX-DOAS background station vertical remote sensing observation results, the high-precision near-surface atmospheric pollutant data obtained in step 2, and various kilometer-level meteorological data in D2 are respectively input into the residual structure; the obtained data is spliced and then passed through the convolutional structure to obtain the output results, denoted as L2';
[0025] The loss values of L2 and L2' are calculated using the loss function F2, and the network parameters are updated using the RMSProp optimizer to optimize the loss function results.
[0026] Specifically, the residual structure is described in (see...). Figure 4 (As shown) includes: four convolutional layers, wherein a batch normalization layer and an activation function are placed between each two convolutional layers; and a batch normalization layer and an activation function are placed at the end;
[0027] (1) Let the input data be X. Let F1(X) be the data obtained by passing X through a convolutional layer and then using batch normalization and ReLU activation function. Let F2(X) be the data obtained by passing F1(X) through a convolutional layer and then using batch normalization.
[0028] (2) Add F2(X) and X together and then use the ReLU activation function to obtain the number H1(X);
[0029] (3) The data obtained by passing H1(X) through a convolutional layer and then using batch normalization and ReLU activation function is denoted as F3(X), and the data obtained by passing F3(X) through a convolutional layer and then using batch normalization is denoted as F4(X);
[0030] (4) Add F4(X) and H1(X) together, activate them using the ReLU function, and then output the result.
[0031] Specifically, the convolutional structure is described in (see...). Figure 4 (As shown) includes: a convolutional layer, a max pooling layer and a normalized layer between the two, another convolutional layer, another max pooling layer and a normalized layer between the two, an average pooling layer, and two fully connected layers.
[0032] (1) Let the input data be X. The data obtained by passing X through a 2*2 convolutional layer with a stride of 1 and then using batch normalization and ReLU activation function is denoted as F1(X). The data obtained by passing F1(X) through a 2*2 max pooling layer is denoted as F2(X).
[0033] (2) The data obtained by passing F2(X) through a 2*2 convolutional layer with a stride of 1 and then using batch normalization and ReLU activation function is denoted as F3(X). The data obtained by passing F3(X) through a 2*2 max pooling layer is denoted as F4(X).
[0034] (3) Output F4(X) after passing it through one average pooling layer and two fully connected layers;
[0035] The loss function F2 used to train the convolutional neural network N2 is:
[0036]
[0037] wherein M is the layer number of the MAX-DOAS vertical remote sensing observation result, N is the total number of the MAX-DOAS vertical remote sensing observation result, L2 i,j , L2' i,j respectively represent the i-th layer and the j-th result of the MAX-DOAS vertical remote sensing observation result.
[0038] Step 4, input the km-level meteorological data obtained by the atmospheric chemical model, the tropospheric atmospheric pollutant data obtained by satellite observation, the MAX-DOAS background station vertical remote sensing observation result and the high-precision near-surface atmospheric pollutant data obtained in step 2 corresponding to a certain moment of all spatial points in the region into the model M2 to obtain the three-dimensional atmospheric pollutant distribution in the region at the moment.
[0039] Considering that the peak values of pollutant concentrations at different heights are different, the layered standardization is adopted when the MAX-DOAS vertical remote sensing observation result used in the present application is standardized, so as to ensure the consistency of the order of magnitude of the data after standardization; considering that the interaction between meteorological information and pollutant concentration in different regions may be different, the application range of the model is limited, and a plurality of MAX-DOAS observation instruments are densely arranged at different positions in the region, which is more conducive to extracting the corresponding relationship between the meteorological information obtained by the atmospheric chemical model and the pollutant concentration distribution.
[0040] The reconstruction method of the km-level resolution three-dimensional distribution of regional atmospheric pollutants provided by the present application firstly uses one or more MAX-DOAS station observations in the region as input to construct a model, no longer relies on only the meteorological information obtained by the atmospheric chemical model, inputs sufficient vertical information to the model, makes the vertical reconstruction result more reliable, uses satellite observation data and ground observation data to reconstruct high-precision near-surface atmospheric pollutant data, and introduces the tropospheric atmospheric pollutant data obtained by satellite observation to provide the model with rich horizontal information, so that the horizontal reconstruction result is more reasonable, thereby realizing the reconstruction of the km-level resolution three-dimensional distribution of atmospheric pollutants. BRIEF DESCRIPTION OF DRAWINGS
[0041] Figure 1 It is a flowchart of the reconstruction method of the km-level resolution three-dimensional distribution of regional atmospheric pollutants.
[0042] Figure 2 It is a schematic diagram of the selected reconstruction range (116.95°E-117.58°E, 31.65°N
[0043] N-32.05°N) of the three-dimensional distribution of atmospheric pollutants of the embodiment and the observation instrument position.
[0044] Figure 3A flow chart for constructing the convolutional neural network N1 to obtain a high-precision pollutant near-surface horizontal concentration distribution reconstruction model M1.
[0045] Figure 4 A flow chart for constructing the convolutional neural network N1 to obtain an atmospheric pollutant three-dimensional distribution result model M2.
[0046] Figure 5 A three-dimensional reconstruction result of NO2 in a selected area range (116.95°E-117.58°E, 31.65°N-32.05°N) at a certain time obtained in the embodiment.
[0047] Figure 6 A comparison result of the test set label and the reconstruction result at different heights obtained in the embodiment.
[0048] Figure 7 A comparison result of the bottom layer reconstruction result obtained in the embodiment and the NO2 data of the national control point in the selected area range. DETAILED DESCRIPTION
[0049] The application will be further described below in combination with the drawings and embodiments.
[0050] Figure 1 A flow chart of a reconstruction method of a regional atmospheric pollutant three-dimensional distribution with a kilometer resolution, Figure 2 A schematic diagram of the atmospheric pollutant three-dimensional distribution reconstruction range (116.95°E-117.58°E, 31.65°N-32.05°N) selected in the embodiment and the observation instrument position. The reconstruction method of NO2 three-dimensional distribution with a kilometer resolution based on hyperspectral remote sensing and deep learning provided in the embodiment includes the following steps:
[0051] Step 1: MAX-DOAS horizontal remote sensing, unmanned aerial vehicle remote sensing, and ground monitoring station in-situ monitoring are combined into a data with a length of 3, 1-2 missing data are allowed for each data, a data set is constructed, and is denoted as label L1.
[0052] MAX-DOAS horizontal remote sensing slant column concentration data obtained according to the standard of fitting root mean square error rms<0.001; and unmanned aerial vehicle remote sensing slant column concentration data obtained according to the standard of fitting root mean square error rms<0.002, and atmospheric radiation transmission model is used to simulate atmospheric quality factors to correct the unmanned aerial vehicle remote sensing data.
[0053] After standardization, MAX-DOAS horizontal remote sensing, unmanned aerial vehicle remote sensing, and ground monitoring station in-situ monitoring data are combined into a data with a length of 3, 1-2 missing data are allowed, and the missing values are uniformly set as null values, a near-surface NO2 data set label L1 is constructed.
[0054] Step 2: Construct a convolutional neural network N1. A dataset, denoted as D1, is constructed using data continuously obtained from high-resolution satellites, including vertical column concentrations of atmospheric NO2 in the target area, meteorological information from atmospheric chemistry models, and geographic information such as population, road networks, and land cover types. D1 is used as the input data for the convolutional neural network N1, and L1 is used as its label. A model M1 with high-precision near-surface atmospheric NO2 data is obtained through training. Its loss function is the weighted average of the root mean square errors of the three non-missing values of label L1.
[0055] The specific structural components of the constructed convolutional neural network N1 include:
[0056] Based on the time of the label L1 data, a sequence of neighboring time D1 data is generated to obtain its time-scale data; based on the latitude and longitude of the label L1 data, a three-dimensional matrix of neighboring location D1 data is generated: longitude, latitude, and height to obtain its spatial-scale data.
[0057] The time-scale and spatial-scale data of satellite-observed NO2 column concentration, near-surface meteorological information, and geographic information are respectively processed through two residual structures and then stitched together. The data is then processed through a residual structure again to obtain the data, which is denoted as D. Resnet1 D Resnet2 D Resnet3 ;
[0058] D Resnet1 D Resnet3 respectively with D Resnet2 After splicing and passing through the residual structure, the obtained data is the same as D. Resnet3 The data is then concatenated again through the residual structure, and the resulting data is passed through a fully connected layer and output as L1'.
[0059] The loss values of L1 and L1' are calculated using the loss function F1, and the network parameters are updated using the RMSProp optimizer to optimize the loss function results; the initial learning rate is set to 0.0001 and gradually decreased as the number of iterations increases;
[0060] Specifically, the residual structure is described in (see...). Figure 3 (As shown) includes: an average pooling layer, six convolutional layers, and normalization layers and activation functions between the convolutional layers;
[0061] (1) The input data after passing through the average pooling layer is denoted as X; the data obtained by passing X through a convolutional layer and then using a batch normalization layer is denoted as F1(X); the data obtained by passing X through a convolutional layer and then using batch normalization and the ReLU activation function is denoted as F2(X); the data obtained by passing X through three convolutional layers, using batch normalization and the ReLU activation function between layers, and then using batch normalization is denoted as F3(X);
[0062] (2) The data obtained by adding F1(X) and F3(X) and using the relu activation function is denoted as H1(X), and the data obtained by passing H1(X) through a convolutional layer and using batch normalization and the relu activation function is denoted as G2(X);
[0063] (3) The output after adding F2(X) and G2(X).
[0064] The loss function F1 used for training the convolutional neural network N1 is:
[0065]
[0066] In the formula, the subscripts 0, 1, and 2 of L1 and L1' represent three kinds of data of the labels L1 and L1', respectively, and A, B, and C represent the number of non-missing values of the three kinds of data of the labels L1 and L1'.
[0067] Step 3, constructing a convolutional neural network N2; constructing a data set D2 using the kilometer-level meteorological data obtained from the atmospheric chemical model, the tropospheric atmospheric NO2 data obtained from satellite observation, the MAX-DOAS background station vertical remote sensing observation results, and the high-precision near-surface atmospheric NO2 data obtained in step 2, taking D2 as the input data of the convolutional neural network N2, taking the layered standardized MAX-DOAS vertical remote sensing observation results L2 as the label, and obtaining the model M2 of the atmospheric NO2 stereoscopic distribution results through training, and the loss function adopts the logarithm of the hyperbolic cosine (Log-Cosh) of the error.
[0068] The tropospheric atmospheric NO2 data obtained from satellite observation comes from the paper Zhang, C., Liu, C., Chan, K. L. et al. First observation of tropospheric nitrogen dioxide from the Environmental Trace Gases Monitoring Instrument onboard the GaoFen-5 satellite. Light Sci Appl 9, 66 (2020);
[0069] https: / / doi.org / 10.1038 / s41377-020-0306-z.
[0070] The specific structure and composition of the constructed convolutional neural network N2 include:
[0071] The tropospheric atmospheric NO2 data obtained by satellite observation in D2, the MAX-DOAS background station vertical remote sensing observation results, the high-precision near-surface atmospheric NO2 data obtained in step 2, and various kilometer-level meteorological data are respectively input into the residual structure; the obtained data is spliced and then passed through the convolution structure to obtain the output result, denoted as L2';
[0072] The loss value of L2 and L2' is calculated using the loss function F2, the network parameter is updated using the RMSProp optimizer, and the loss function result is optimized. The initial value of the learning rate is set to 0.00008, which gradually decreases as the number of iterations increases;
[0073] The residual structure specifically (see Figure 4 ) includes four convolutional layers, wherein batch normalization layers and activation functions are arranged between the front and rear two convolutional layers; and finally batch normalization layers and activation functions are arranged;
[0074] (1) Denote the input data as X. The data obtained by passing X through a convolutional layer and then using batch normalization and a relu activation function is denoted as F1(X), and the data obtained by passing F1(X) through a convolutional layer and then using batch normalization is denoted as F2(X);
[0075] (2) The data obtained by adding F2(X) and X and then using a relu activation function is denoted as H1(X);
[0076] (3) The data obtained by passing H1(X) through a convolutional layer and then using batch normalization and a relu activation function is denoted as F3(X), and the data obtained by passing F3(X) through a convolutional layer and then using batch normalization is denoted as F4(X);
[0077] (4) The output is obtained by adding F4(X) and H1(X) and then using a relu activation function.
[0078] The convolutional structure specifically (see Figure 4 ) includes a convolutional layer, a max pooling layer, and a normalization layer between the two, another convolutional layer, another max pooling layer, and a normalization layer between the two, an average pooling layer, and 2 fully connected layers;
[0079] (1) Denote the input data as X. The data obtained by passing X through a 2*2 convolutional layer with a step of 1 and then using batch normalization and a relu activation function is denoted as F1(X), and the data obtained by passing F1(X) through a 2*2 max pooling layer is denoted as F2(X);
[0080] (2) The data obtained by passing F2(X) through a 2*2 convolution layer with a step size of 1, using batch normalization and a relu activation function, is denoted as F3(X), and the data obtained by passing F3(X) through a 2*2 max-pooling layer is denoted as F4(X);
[0081] (3) The output obtained by passing F4(X) through an average pooling layer and two fully connected layers.
[0082] The loss function F2 used to train the convolutional neural network N2 is:
[0083]
[0084] In the formula, M is the number of layers of MAX-DOAS vertical remote sensing observation results, N is the total number of MAX-DOAS vertical remote sensing observation results, L2 i,j , and L2' i,j respectively represent the i-th layer and the j-th result of the MAX-DOAS vertical remote sensing observation results.
[0085] Step 4: Input the km-level meteorological data obtained from the atmospheric chemical model, the tropospheric atmospheric NO2 data obtained from satellite observations, the MAX-DOAS background station vertical remote sensing observation results, and the high-precision near-surface atmospheric NO2 data obtained in step 2 into the model M2 to obtain the three-dimensional atmospheric NO2 distribution in the region at that time.
[0086] In the embodiment, the NO2 three-dimensional distribution of Hefei city (116.95°E-117.58°E, 31.65°N-32.05°N) at a certain time is obtained, as shown in Figure 5 The comparison results of the test set labels and the reconstruction results at different altitudes are obtained in the embodiment, as shown in Figure 6 The correlation coefficients at 0m, 100m, 200m, 300m, 400m, 500m, 600m, 700m, 800m, 900m are 0.8455, 0.8582, 06624, 0.6861, 0.6825, 0.7080, 0.6672, 0.5508, 0.6186, 0.6184, respectively. The data of 10 national control monitoring stations (Pearl Plaza Station, Sanli Street Station, Amber Villa Station, Dongpu Reservoir Station, Changjiang Road Station, Lu Yang District Station, Yaohe District Station, Baohe District Station, Binhu New District Station, and Gaoxin District Station) in Hefei city are compared with the bottom layer data of the reconstruction results at their locations, as shown in Figure 7 The correlation coefficients are 0.6945, 0.6286, 0.7139, 0.7169, 0.6624, 0.6651, 0.6845, 0.7007, 0.6474, and 0.7419, respectively.
[0087] The above detailed description of the specific embodiments has described the technical solutions and beneficial effects of the present application, and it should be understood that the types of atmospheric pollutants selected and the regions selected for stereoscopic reconstruction are only used to explain the present application, and the above description is only the most preferred embodiment of the present application and is not intended to limit the present application. Any modification, supplement and equivalent replacement, etc. made within the principle range of the present application shall be included in the protection scope of the present application.
Claims
1. A method for reconstructing the three-dimensional distribution of regional atmospheric pollutants with kilometer-level resolution, characterized in that, The vertical distribution profiles of atmospheric pollutants observed by hyperspectral remote sensing networks are combined with deep learning technology to reconstruct the kilometer-level resolution three-dimensional distribution of atmospheric pollutants in areas not covered by observations based on hyperspectral observation data, meteorological simulation data, and geographic information. The specific steps are as follows: Step 1: Combine the data obtained from MAX-DOAS horizontal remote sensing, UAV remote sensing, and in-situ monitoring from ground monitoring stations into a dataset of length 3. Each dataset is allowed to have 1 to 2 missing data types. Construct a dataset and denote it as label L1. Step 2: Construct a convolutional neural network N1. A dataset, denoted as D1, is constructed using the vertical column concentrations of atmospheric pollutants in the target area continuously obtained from high-resolution satellites, meteorological information obtained from atmospheric chemical models, and geographic information on population, road networks, and land cover types. D1 is used as the input data for the convolutional neural network N1, and L1 is used as the label for N1. Through training, a high-precision near-surface atmospheric pollutant data model M1 is obtained. Its loss function is the weighted average of the root mean square errors of the three non-missing data values of label L1. The specific structural components of the constructed convolutional neural network N1 include: Based on the time of the label L1 data, a sequence of neighboring time D1 data is generated to obtain its time-scale data; based on the latitude and longitude of the label L1 data, a three-dimensional matrix of neighboring location D1 data is generated: longitude, latitude, and height to obtain its spatial-scale data. The time-scale and spatial-scale data of satellite-observed pollutant column concentrations, near-surface meteorological information, and geographic information are respectively processed through two residual structures and then stitched together. The data is then processed through a residual structure again to obtain the final data, denoted as D. Resnet1 D Resnet2 D Resnet3 ; D Resnet1 D Resnet3 respectively with D Resnet2 After splicing and passing through the residual structure, the obtained data is the same as D. Resnet3 The data is then concatenated again through the residual structure, and the resulting data is passed through a fully connected layer and output as L1'. The loss values of L1 and L1' are calculated using the loss function F1, and the network parameters are updated using the RMSProp optimizer to optimize the loss function results. Step 3: Construct a convolutional neural network N2. Using kilometer-level meteorological data obtained from atmospheric chemistry models, tropospheric atmospheric pollutant data obtained from satellite observations, vertical remote sensing observation results from the MAX-DOAS background station, and high-precision near-surface atmospheric pollutant data obtained in Step 2, construct a dataset denoted as D2. Use D2 as the input data for the convolutional neural network N2, and use the hierarchically standardized MAX-DOAS vertical remote sensing observation results as the label L2. Through training, obtain a model M2 for the three-dimensional distribution of atmospheric pollutants. The loss function is the logarithm of the hyperbolic cosine of the error. The specific structural components of the constructed convolutional neural network N2 include: The tropospheric atmospheric pollutant data obtained from satellite observations in D2, the vertical remote sensing observation results from the MAX-DOAS background station, the high-precision near-surface atmospheric pollutant data obtained in step 2, and various kilometer-level meteorological data are respectively input into the residual structure; the obtained data are then stitched together and passed through a convolution structure to obtain the output result, denoted as L2'. The loss values of L2 and L2' are calculated using the loss function F2, and the network parameters are updated using the RMSProp optimizer to optimize the loss function results. Step 4: Input the kilometer-level meteorological data obtained by atmospheric chemistry model for all spatial points in the region at a certain time, the tropospheric atmospheric pollutant data obtained by satellite observation, the vertical remote sensing observation results of the MAX-DOAS background station, and the high-precision near-surface atmospheric pollutant data obtained in Step 2 into model M2 to obtain the three-dimensional atmospheric pollutant distribution in the region at that time.
2. The method for reconstructing the three-dimensional distribution of regional atmospheric pollutants with kilometer-level resolution according to claim 1, characterized in that, The residual structure described in step (2) includes: an average pooling layer, six convolutional layers, and normalization layers and activation functions between the convolutional layers; wherein: (1) The input data after passing through the average pooling layer is denoted as X; the data obtained by passing X through a convolutional layer and then using a batch normalization layer is denoted as F1(X); the data obtained by passing X through a convolutional layer and then using batch normalization and the ReLU activation function is denoted as F2(X); the data obtained by passing X through three convolutional layers, using batch normalization and the ReLU activation function between layers, and then using batch normalization is denoted as F3(X); (2) The data obtained by adding F1(X) and F3(X) and then using the ReLU activation function is denoted as H1(X). The data obtained by passing H1(X) through a convolutional layer and then using batch normalization and the ReLU activation function is denoted as G2(X). (3) Add F2(X) and G2(X) together and output the result; The loss function F1 used to train the convolutional neural network N1 is: In the formula, the subscripts 0, 1, and 2 of L1 and L1′ represent the three types of data for labels L1 and L1′, respectively, and A, B, and C represent the number of non-missing values for the three types of data for labels L1 and L1′, respectively.
3. The method for reconstructing the three-dimensional distribution of regional atmospheric pollutants with kilometer-level resolution according to claim 2, characterized in that, In step (3): The residual structure includes: four convolutional layers, wherein a batch normalization layer and an activation function are provided between each two consecutive convolutional layers; and a batch normalization layer and an activation function are provided at the end. (1) Let the input data be X, let F1(X) be the data obtained by passing X through a convolutional layer and using batch normalization and ReLU activation function, and let F2(X) be the data obtained by passing F1(X) through a convolutional layer and using batch normalization. (2) Add F2(X) and X together and then use the ReLU activation function to obtain the number H1(X); (3) The data obtained by passing H1(X) through a convolutional layer and then using batch normalization and ReLU activation function is denoted as F3(X), and the data obtained by passing F3(X) through a convolutional layer and then using batch normalization is denoted as F4(X); (4) Add F4(X) and H1(X) together, activate them using the ReLU function, and then output the result. The convolutional structure includes: a convolutional layer, a max-pooling layer and a normalized layer between them, another convolutional layer, another max-pooling layer and a normalized layer between them, an average pooling layer, and two fully connected layers; wherein: (1) Let the input data be X; let F1(X) be the data obtained by passing X through a 2*2 convolutional layer with a stride of 1 and then using batch normalization and ReLU activation function; let F2(X) be the data obtained by passing F1(X) through a 2*2 max pooling layer. (2) The data obtained by passing F2(X) through a 2*2 convolutional layer with a stride of 1 and then using batch normalization and ReLU activation function is denoted as F3(X). The data obtained by passing F3(X) through a 2*2 max pooling layer is denoted as F4(X). (3) Output F4(X) after passing it through one average pooling layer and two fully connected layers; The loss function F2 used to train the convolutional neural network N2 is: In the formula, M represents the number of layers in the MAX-DOAS vertical remote sensing observation results, and N represents the total number of MAX-DOAS vertical remote sensing observation results; L2 i,j L2' i,j These represent the i-th layer and j-th result of the MAX-DOAS vertical remote sensing observation, respectively.