Machine Learning-Based Method, Apparatus, and Equipment for Calculating Regional NO2 Spatial Distribution

Abstract

This invention discloses a method, apparatus, and device for calculating the spatial distribution of NO2 in a region based on machine learning. The method includes: collecting remote sensing datasets and auxiliary data; the remote sensing dataset includes an AOD dataset, NO2 column concentration data, and related data; the auxiliary data includes the characteristics of a specific day of the year, the location ID of each pixel, and a 1-kilometer resolution grid in a UTM coordinate system covering the study area; preprocessing the remote sensing dataset and auxiliary data; labeling the location of all pixels with independent IDs, shuffling all IDs using a random algorithm, and incorporating the random IDs into a random forest regression method; and using the random forest regression method to train and predict the spatial distribution of NO2 concentration. Compared with existing RF models, this embodiment shows significant improvements in both performance and display quality.

Description

Technical Field

[0001] This invention relates to the field of environmental monitoring technology, and in particular to a method, apparatus, and equipment for calculating the spatial distribution of regional NO2 based on machine learning. Background Technology

[0002] Nitrogen dioxide (NO2) is one of the major air pollutants. Excessive NO2 concentrations can lead to various environmental disasters (acid rain, vegetation destruction, fine particulate matter, and ozone precursors, etc.) and harm human health. To better address the hazards caused by NO2 and effectively manage air pollutants, it is necessary to refine the spatial distribution of NO2 concentrations through simulation.

[0003] With the development of machine learning, more and more models are playing an increasingly important role in the field of ecology, and Random Forest (RF) is one of the key models among them. Currently, the application of using large amounts of remote sensing data, ground monitoring data, and statistical data, combined with random forest models to simulate the spatial distribution of NO2, and further to conduct key ecological risk assessments is booming.

[0004] The current approach involves collecting NO2 ground-level concentration monitoring data downloaded from the Ministry of Ecology and Environment's air quality forecasting and information release system, collecting remote sensing data from NASA, and collecting regional statistical data. This data and parameters are then input into the RF model, ultimately outputting a daily simulation of the spatial distribution of NO2 concentration across a 1km grid.

[0005] The drawback of this approach is that directly importing the data into the RF model can easily lead to severe "striping" and "patchwork" phenomena in the simulation results, such as... Figure 1 As shown in the diagram, the main reason is the specific spatial distribution of NO2. During its transport in the atmosphere, it is affected by factors such as meteorology and topography, resulting in various phenomena such as spatial aggregation and transformation into nitrate particles. To better simulate the spatial distribution of NO2, more parameters need to be introduced. However, many parameters, as remote sensing data, are often limited by the observation level, leading to differences in resolution between parameters. Machine learning models, during the fitting process, find it difficult to directly eliminate the influence of spatial resolution differences, ultimately causing "banding" and "patchwork" phenomena. Since the spatial distribution of atmospheric pollutants is the result of highly correlated and heterogeneous complex combinations, excessively regular "banding" and "patchwork" severely violate natural laws. Therefore, this phenomenon will directly lead to a decrease in the reliability of simulation results and affect the results of subsequent ecological analyses. Summary of the Invention

[0006] This invention proposes a method, apparatus, and device for calculating the spatial distribution of NO2 in a region based on machine learning, in order to improve the above-mentioned problems.

[0007] This invention provides a machine learning-based method for calculating the spatial distribution of NO2 in a region, comprising:

[0008] Remote sensing datasets and auxiliary data were collected; the remote sensing datasets included AOD datasets, NO2 column concentrations and related data; the auxiliary data included the characteristics of doy on a certain day of the year, the location ID of each pixel and a 1-kilometer resolution grid of the UTM coordinate system covering the study area.

[0009] The remote sensing dataset and auxiliary data are preprocessed;

[0010] All pixel locations are marked with independent IDs, all IDs are shuffled using a random algorithm, and the random IDs are then introduced into the random forest regression method;

[0011] The spatial distribution of NO2 concentration was trained and predicted using a random forest regression method.

[0012] Preferably, the AOD dataset includes MAIACAOD retrieved by the multi-angle atmospheric correction algorithm and AHI AOD retrieved by the Himawari-8 satellite; each AOD includes 470nm AOD and 550nm AOD; NO2 column concentration data is collected using an ozone detector with a time resolution of 1 day and a spatial resolution of 0.25°.

[0013] Preferably, the relevant data includes normalized vegetation index, topography, population distribution, roads, and land use.

[0014] Preferably, preprocessing the remote sensing dataset and auxiliary data includes:

[0015] The multi-source AOD dataset is converted to a 1 km / d spatiotemporal resolution using a 1 km resolution grid of the UTM coordinate system in the auxiliary data;

[0016] Elevation ELE and slope SL are generated by SRTM calculation. ELE, SL, land use, road data, 470nm MAIAC AOD, NDVI, population, 470nm AHI AOD and OMI NO2 column concentration are superimposed on a 1 km resolution grid of the UTM coordinate system and reproduced.

[0017] The reconstruction process is calculated based on the weighted average of the data pixel values ​​and the coverage ratio of pixels on a 1-kilometer resolution grid;

[0018] Fill in data with a time resolution exceeding one day;

[0019] Meteorological parameters were interpolated using the co-kriging method, with ELE as the covariate, and the output was UTM 1km / d spatiotemporal resolution.

[0020] A two-step method was used to fill in the missing data for AOD and NO2 concentrations on the 470nm MAIAC column.

[0021] Preferably, the two-step method for filling in the missing data of AOD and NO2 concentrations at the 470nm MAIAC column specifically includes:

[0022] The RF model is used to simulate the overlapping portion between AOD data, and the formula is as follows:

[0023] AOD maiacpre =RF(AHI 470nm ,RID,ELE,SL,POP,NDVI,RL,LU,DOY)

[0024] Where RF represents the RF model, AHI 470nm RID, ELE, SL, POP, NDVI, RL, LU, DOY represent 470nmAHI AOD, random ID, altitude, slope, population, NDVI, road length, land use, and doy, respectively.

[0025] The appropriate interpolation method is selected based on the pixels covered within the moving window to perform interpolation calculations and fill in the missing data;

[0026] The expected value of the pixels covered by the moving window is used as a buffer factor to correct the deviation; the formula is as follows:

[0027]

[0028]

[0029]

[0030]

[0031]

[0032] Where MoranI represents the Global Moran's I method; n represents the number of valid pixels; p i and p j This represents the values ​​of pixels i and j; G represents the average pixel value; dis(i, j) represents the distance between pixels i and j; i，j Represents the inverse distance weight; Scope Window represents the maximum value of MoranI across multiple moving windows; Scope Window is a square, where w represents its length and width. The left arrow represents the iterative operation to search the window; the right arrow represents the process of obtaining w; the left arrow represents the process of obtaining w. iRepresents the value within the Scope Window; S tk Represents the value of the tk-day Scope Window; E w The expected value of the Scope Window, i.e., the buffer factor; This represents the correlation coefficient between tk and t2.

[0033] Preferably, the missing data is filled by selecting the appropriate interpolation method based on the number of pixels covered within the moving window, specifically including:

[0034] Assuming the moving window size is 3*3, when the moving window lacks a center pixel, the inverse distance weighted interpolation method is used; when the moving window is missing five or fewer pixels, the RC-Kriging method is used; when the number of missing units on the second day is greater than or equal to 5 and the number of valid pixels on the first or third day is greater than or equal to 5, spatiotemporal weighted interpolation is used; when the number of pixels in the moving window is too small for three consecutive days, this part of the calculation is ignored; when pixel recovery causes overlap, the missing values ​​are filled according to their priority; the priority is set as inverse distance weighted interpolation > RC Kriging > spatiotemporal weighted interpolation.

[0035] Preferably, the location of all pixels is marked with an independent ID, all IDs are shuffled using a random algorithm, and the random IDs are introduced into the random forest regression method. The random forest regression method is then used to train and predict the spatial distribution of NO2 concentration. Specifically, this includes:

[0036] Randomize the position parameters by shuffling the position ID values ​​using a random algorithm and assigning them to each pixel.

[0037] The position parameters and random position IDs are normalized using a 0-1 normalization algorithm;

[0038] Using the random forest regression method to train and predict the spatial distribution of NO2 concentration:

[0039] RID=normalization(random ID)

[0040]

[0041] pre NO =RF(NO column AOD 470nm ,RID,ELE,SL,POP,NDVI,RL,LU,DOY)

[0042] The number of IDs is equal to the 1 km resolution grid of the UTM; random and normalization represent randomization and normalization algorithms, respectively; RF represents the random forest model; pre NOThis represents the predicted NO2 value; NO column AOD 470nm RID, ELE, SL, POP, NDVI, RL, LU, and DOY represent NO2 column concentration, 470nm MAIAC AOD, random ID, altitude, slope, NDVI, road length, land use, and doy, respectively.

[0043] An embodiment of the present invention provides a regional NO2 spatial distribution computing device based on machine learning, comprising:

[0044] The data acquisition unit is used to acquire remote sensing datasets and auxiliary data. The remote sensing datasets include AOD datasets, NO2 column concentrations, and related data. The auxiliary data include the characteristics of doy on a certain day of the year, the location ID of each pixel, and a 1-kilometer resolution grid of the UTM coordinate system covering the study area.

[0045] The preprocessing unit is used to preprocess the remote sensing dataset and auxiliary data;

[0046] The random ID unit is used to mark the position of all pixels with an independent ID. All IDs are shuffled by a random algorithm and then introduced into the random forest regression method.

[0047] The prediction unit is used to train and predict the spatial distribution of NO2 concentration using the random forest regression method.

[0048] This invention also provides a machine learning-based regional NO2 spatial distribution calculation device, which includes a memory and a processor. The memory stores a computer program that can be executed by the processor to implement the machine learning-based regional NO2 spatial distribution calculation method described above.

[0049] This embodiment optimizes the random forest model without sacrificing cross-validation accuracy. By appropriately introducing auxiliary variables, it minimizes fitting anomalies caused by resolution differences between parameters, ultimately constructing a random forest model adapted to the spatial distribution of NO2. Compared to existing RF models, this embodiment demonstrates significant improvements in both performance and display quality. Attached Figure Description

[0050] Figure 1 This is a schematic diagram of the spatial distribution of NO2 generated by an existing RF model.

[0051] Figure 2 This is a flowchart illustrating the regional NO2 spatial distribution calculation method based on machine learning provided in the first embodiment of the present invention.

[0052] Figure 3This is a schematic diagram illustrating the principle of a machine learning-based regional NO2 spatial distribution calculation method provided in an embodiment of the present invention.

[0053] Figure 4(a) shows the cross-validation results of this embodiment.

[0054] Figure 4(b) shows the cross-validation results of the traditional RF model.

[0055] Figure 5(a) is a display diagram of the simulated spatial distribution of NO2 in this embodiment.

[0056] Figure 5(b) shows the simulated spatial distribution of NO2 using a traditional RF model.

[0057] Figure 6 This is a schematic diagram of the structure of the machine learning-based regional NO2 spatial distribution computing device provided in the second embodiment of the present invention. Detailed Implementation

[0058] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. 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.

[0059] See Figure 2 and Figure 3 This invention provides a method for calculating the spatial distribution of NO2 in a region based on machine learning, comprising:

[0060] S101, Collect remote sensing datasets and auxiliary data; the remote sensing datasets include AOD datasets, NO2 column concentrations and related data; the auxiliary data include the characteristics of doy on a certain day of the year, the location ID of each pixel and a 1-kilometer resolution grid of the UTM coordinate system covering the study area.

[0061] In this embodiment, the remote sensing dataset includes (1) the AOD dataset, (2) the NO2 column concentration, and (3) other relevant data.

[0062] The AOD dataset includes AOD retrieved by the multi-angle atmospheric correction algorithm (hereinafter referred to as MAIAC AOD) and AOD retrieved by the Himawari-8 satellite (hereinafter referred to as AHI AOD), and both MAIAC AOD and AHI AOD include 470nm AOD and 550nm AOD.

[0063] In this study, considering the strong absorption line of NO2 between near-ultraviolet and visible light, and the fact that the differential optical absorption spectroscopy (DOAS) algorithm used for NO2 retrieval uses a 405-465 nm spectrum which is closer to the 470 nm AOD, the 470 nm MAIAC AOD and AHI AOD were selected. The MAIAC AOD (earthdata.nasa.gov) has a spatial resolution of 1 km and a temporal resolution of 1 day, while the AHI AOD of the L3 day product Himawari-8 has a spatial resolution of 5 km.

[0064] In this embodiment, the NO2 column concentration can be obtained using NO2 L3 data collected by an ozone detector (OMI), with a time resolution of 1 day and a spatial resolution of 0.25°.

[0065] In this embodiment, other relevant data include the Normalized Difference Vegetation Index (NDVI), topography, population distribution, roads, and land use. NDVI was calculated from MODIS data (earthdata.nasa.gov) with a temporal resolution of 16 days and a spatial resolution of 1 km. Topographic data, including elevation and slope, was extracted from the Space Shuttle Imaging Radar Topographic Mapping Mission (SRTM) data with a spatial resolution of 90 m. Population data was obtained from LandScan with a spatial resolution of approximately 1 km. Road data was provided by OpenStreet and consisted of .shp format road layers. Land use data came from the Copernicus Climate Change Service (C3S) with a spatial resolution of 300 m.

[0066] In this embodiment, the auxiliary data includes characteristics of a day of the year (doy), such as whether it is a weekday or a non-weekday (wdon), the location ID of each pixel, and a 1-kilometer resolution grid of the UTM coordinate system covering the study area.

[0067] S102, preprocess the remote sensing dataset and auxiliary data.

[0068] Specifically, including:

[0069] S1021, the multi-source AOD dataset is converted to a 1 km / d spatiotemporal resolution using a 1 km resolution grid of the UTM coordinate system in the auxiliary data.

[0070] In this embodiment, since the AOD dataset includes two data sources, MAIAC AOD and AHI AOD, alignment and fusion transformation are required. Specifically, this can be achieved by converting the 1 km resolution grid of the UTM coordinate system in the auxiliary data to a 1 km / d spatiotemporal resolution.

[0071] S1022 generates elevation ELE and slope SL through SRTM calculation. ELE, SL, land use, road data, 470nm MAIAC AOD, NDVI, population, 470nm AHI AOD and OMI NO2 column concentration are superimposed on a 1 km resolution grid of the UTM coordinate system and then reproduced.

[0072] S1023, the reconstruction process is calculated based on the weighted average of the data pixel values ​​and the coverage ratio of pixels on a 1-kilometer resolution grid.

[0073] S1024, filling data with a time resolution exceeding one day.

[0074] For example, NDVI fills in gaps in the time range with NDVI values ​​based on a 16-day time resolution.

[0075] S1025 uses co-kriging to interpolate meteorological parameters, with ELE as the covariate, and outputs a UTM spatiotemporal resolution of 1 km / d.

[0076] The meteorological parameters include temperature (TEM), pressure (PR), humidity (HUM), and wind speed (WS).

[0077] S1026 uses a two-step method to fill in the missing data for AOD and NO2 concentrations on the 470nm MAIAC column.

[0078] Specifically:

[0079] First, the RF model is used to simulate the overlapping portion of AOD data, as shown in the following formula:

[0080] AOD maiacpre =RF(AHI 470nm ,RID,ELE,SL,POP,NDVI,RL,LU,DOY)

[0081] Where RF represents the RF model, AHI 470nm RID, ELE, SL, POP, NDVI, RL, LU, and DOY represent 470nmAHI AOD, random ID, altitude, slope, population, NDVI, road length, land use, and doy, respectively.

[0082] Then, based on the pixels covered within the moving window, the corresponding interpolation method is selected to perform interpolation calculations to fill in the missing data.

[0083] The moving window is set to 3*3. When the center pixel is missing in the moving window, the inverse distance weighted interpolation method is used; when five or fewer pixels are missing in the moving window, the RC-Kriging method is used; when the number of missing units on the second day is greater than or equal to 5 and the number of effective pixels on the first or third day is greater than or equal to 5, the spatiotemporal weighted interpolation method is used; when the number of pixels in the moving window is too small for three consecutive days, this part of the calculation is ignored; when pixel recovery causes overlap, the missing values ​​are filled according to their priority; the priority is set as inverse distance weighted interpolation > RC Kriging > spatiotemporal weighted interpolation.

[0084] The formulas for spatiotemporal interpolation IDW and constrained kriging (RC kriging) are as follows:

[0085]

[0086]

[0087] Where Z1 and Z2 represent the IDW and RC Kriging interpolation results; G i，j Represents the inverse distance weight; s i，j Represents the values ​​of i and j; μ represents the Lagrange multiplier; λ² i，j Represents weight; Cov(s) i，j ) and Cov(s j，i ) represents s i，j and s j，i covariance; E w Represents the mathematical expectation of the Scope Window.

[0088] The formula for spatiotemporal weighted interpolation STW is as follows:

[0089]

[0090]

[0091]

[0092]

[0093] in, The STW represents the estimation of AOD and NO2 concentrations on a 470nm MAIAC column; T represents time; t1, t2, and t3 represent 3 days of the estimation time axis; S t Represents the effective pixels; E is the mathematical expectation; E tc It is the global mathematical expectation; λ represents the correlation coefficient R between t1 and tn. tnThe time weights for n days represent the distance (n∈(1,2,3)); N is the number of pixels in the moving window; distance(tn) i ,tn j ) represents tn i and tn j Spatial distance.

[0094] Next, the expected value of the pixels covered by the moving window is used as a buffer factor to correct the deviation; as follows:

[0095]

[0096]

[0097]

[0098]

[0099]

[0100] Where MoranI represents the Global Moran's I method; n represents the number of valid pixels; p i and p j This represents the values ​​of pixels i and j; G represents the average pixel value; dis(i, j) represents the distance between pixels i and j; i，j Represents the inverse distance weight; Scope Window represents the maximum value of MoranI across multiple moving windows; Scope Window is a square, where w represents its length and width. The left arrow represents the iterative operation to search the window; the right arrow represents the process of obtaining w; the left arrow represents the process of obtaining w. i Represents the value within the Scope Window; S tk Represents the value of the tk-day Scope Window; E w The mathematical expectation of the Scope Window, i.e., the buffer factor; This represents the correlation coefficient between tk and t2.

[0101] S103: Mark the position of all pixels with an independent ID, shuffle all IDs using a random algorithm, and introduce the random IDs into the random forest regression method.

[0102] S104 uses a random forest regression method to train and predict the spatial distribution of NO2 concentration.

[0103] Specifically,

[0104] First, the position parameters are randomized by shuffling the position ID values ​​using a random algorithm and assigning them to each pixel.

[0105] Then, the position parameters and random position IDs are normalized using a 0-1 normalization algorithm;

[0106] Finally, the random forest regression method was used to train and predict the spatial distribution of NO2 concentration:

[0107] RID=normalization(random ID)

[0108]

[0109] pre NO =RF(NO column AOD 470nm ,RID,ELE,SL,POP,NDVI,RL,LU,DOY)

[0110] The number of IDs is equal to the 1 km resolution grid of the UTM; random and normalization represent randomization and normalization algorithms, respectively; RF represents the random forest model; pre NO This represents the predicted NO2 value; NO column AOD 470nm RID, ELE, SL, POP, NDVI, RL, LU, and DOY represent NO2 column concentration, 470nm MAIAC AOD, random ID, altitude, slope, NDVI, road length, land use, and doy, respectively.

[0111] To facilitate understanding of the present invention, a comparison diagram of this embodiment and traditional RF prediction of NO2 spatial distribution is provided below.

[0112] As shown in Figure 4, comparing the two embodiments with the traditional model using the commonly used comparison parameters R and RMSE in cross-validation, it can be found that the R of this embodiment is improved by about 6% compared with the traditional model, while the RMSE is reduced by about 1.4. This indicates that this embodiment has better performance.

[0113] As shown in Figure 5, the spatial distribution of NO2 simulated by the traditional RF model exhibits obvious "banded" and "patchwork" phenomena. However, this embodiment does not show these characteristics.

[0114] In summary, this embodiment optimizes the random forest model without sacrificing cross-validation accuracy. By appropriately introducing auxiliary variables, it minimizes fitting anomalies caused by resolution differences between parameters, ultimately constructing a random forest model adapted to the spatial distribution of NO2. Compared to existing RF models, this embodiment demonstrates significant improvements in both performance and display quality.

[0115] Please see Figure 6 The second embodiment of the present invention provides a regional NO2 spatial distribution computing device based on machine learning, comprising:

[0116] Data acquisition unit 210 is used to acquire remote sensing datasets and auxiliary data; the remote sensing datasets include AOD datasets, NO2 column concentrations and related data; the auxiliary data include the characteristics of doy on a certain day of the year, the location ID of each pixel and a 1-kilometer resolution grid of the UTM coordinate system covering the study area.

[0117] Preprocessing unit 220 is used to preprocess the remote sensing dataset and auxiliary data;

[0118] Random ID unit 230 is used to mark the position of all pixels with an independent ID. All IDs are shuffled by a random algorithm and the random IDs are introduced into the random forest regression method.

[0119] Prediction unit 240 is used to train and predict the spatial distribution of NO2 concentration using a random forest regression method.

[0120] The third embodiment of the present invention also provides a machine learning-based regional NO2 spatial distribution calculation device, which includes a memory and a processor. The memory stores a computer program, which can be executed by the processor to realize the machine learning-based regional NO2 spatial distribution calculation method described above.

[0121] Exemplarily, the computer program of the present invention can be divided into one or more modules, which are stored in the memory and executed by the processor to complete the present invention. The one or more modules may be a series of computer program instruction segments capable of performing a specific function, which describe the execution process of the computer program in the device.

[0122] The processor referred to can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor or any conventional processor. The processor is the control center of the printing method, connecting various parts of the method via various interfaces and lines.

[0123] The memory can be used to store the computer program and / or modules. The processor implements various functions of the printing method by running or executing the computer program and / or modules stored in the memory and by calling the data stored in the memory. The memory may mainly include a program storage area and a data storage area. The program storage area may store the operating system, at least one application program required for a function (such as sound playback function, text conversion function, etc.), etc.; the data storage area may store data created according to the use of the user terminal (such as audio data, text message data, etc.). In addition, the memory may include high-speed random access memory, and may also include non-volatile memory, such as hard disk, memory, plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, at least one disk storage device, flash memory device, or other volatile solid-state storage device.

[0124] If the module implementing the user terminal is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the above embodiments of the present invention can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include: any entity or device capable of carrying the computer program code, recording media, USB flash drives, portable hard drives, magnetic disks, optical disks, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc. It should be noted that the content included in the computer-readable medium can be appropriately added or removed according to the requirements of legislation and patent practice in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, computer-readable media do not include electrical carrier signals and telecommunication signals.

[0125] It should be noted that the device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Furthermore, in the accompanying drawings of the device embodiments provided by this invention, the connection relationships between modules indicate that they have communication connections, which can be specifically implemented as one or more communication buses or signal lines. Those skilled in the art can understand and implement this without any creative effort.

[0126] The above description is merely a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

[0127] The above description is merely a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for calculating the spatial distribution of NO2 in a region based on machine learning, characterized in that, include: Remote sensing datasets and auxiliary data were collected; the remote sensing datasets included AOD datasets, NO2 column concentrations and related data; the auxiliary data included the characteristics of doy on a certain day of the year, the location ID of each pixel and a 1-kilometer resolution grid of the UTM coordinate system covering the study area. The relevant data includes normalized vegetation index, topography, population distribution, roads, and land use; The remote sensing dataset and auxiliary data are preprocessed, including: converting the multi-source AOD dataset to a 1km / d spatiotemporal resolution using a 1km resolution grid in the UTM coordinate system of the auxiliary data; generating elevation (ELE) and slope (SL) using SRTM; overlaying ELE, SL, land use, road data, 470nm MAIAC AOD, NDVI, population, 470nm AHIAOD, and OMI NO2 column concentration onto a 1km resolution grid in the UTM coordinate system and reloading the data; calculating the reconstruction process based on the weighted average of the data pixel values ​​and the coverage ratio of the pixels on the 1km resolution grid; filling in data with a temporal resolution exceeding one day; interpolating meteorological parameters using co-kriging with ELE as the covariate, and outputting the UTM 1km / d spatiotemporal resolution; and using a two-step method to fill in the missing data for 470nm MAIAC AOD and NO2 column concentration. All pixel locations are marked with independent IDs, all IDs are shuffled using a random algorithm, and the random IDs are then introduced into the random forest regression method; The spatial distribution of NO2 concentration was trained and predicted using a random forest regression method; specifically, this included: The location parameters are randomized by shuffling the location ID values ​​using a random algorithm and assigning them to each pixel; the location parameters and random location IDs are normalized using a 0-1 normalization algorithm; and a random forest regression method is used to train and predict the spatial distribution of NO2 concentration. in The number is equal to the number of grids in a 1-kilometer resolution UTM; and These represent randomization and normalization algorithms, respectively. Represents the random forest model; This represents the predicted NO2 value; The numerical value representing the ID cell; These represent NO2 column concentration, 470 nm MAIAC AOD, random ID, altitude, slope, NDVI, road length, land use, and doy, respectively.

2. The method for calculating the spatial distribution of NO2 in a region based on machine learning according to claim 1, characterized in that, The AOD dataset includes MAIAC AOD retrieved by the multi-angle atmospheric correction algorithm and AHI AOD retrieved by the Himawari-8 satellite; each AOD includes 470nm AOD and 550nm AOD; NO2 column concentration data were collected using an ozone detector with a time resolution of 1 day and a spatial resolution of 0.25°.

3. The method for calculating the spatial distribution of NO2 in a region based on machine learning according to claim 1, characterized in that, The two-step method used to fill in the missing data for AOD and NO2 concentrations at the 470nm MAIAC column specifically includes: The RF model is used to simulate the overlapping portion between AOD data, and the formula is as follows: Where RF represents the RF model. These represent 470nm AHI AOD, random ID, altitude, slope, population, NDVI, road length, land use, and doy, respectively. The appropriate interpolation method is selected based on the pixels covered within the moving window to perform interpolation calculations and fill in the missing data; The expected value of the pixels covered by the moving window is used as a buffer factor to correct the deviation; the formula is as follows: in, This represents Global Moran's I method; here, n represents the number of valid pixels. represent The values ​​of two pixels; Represents the average pixel value; represent The distance between two pixels; Represents inverse distance weights; Represents the maximum value of MoranI across multiple moving windows; It is a square. This represents its length and width values; This represents an iterative search of the window; Representative to obtain The process; represent The value inside; represent day The value; represent The mathematical expectation, i.e., the buffer factor; represent and The correlation coefficient between them.

4. The method for calculating the spatial distribution of NO2 in a region based on machine learning according to claim 3, characterized in that, Based on the pixels covered within the moving window, an appropriate interpolation method is selected to perform interpolation calculations to fill in the missing data. Specifically, this includes: Assuming the moving window size is 3*3, when the moving window lacks a center pixel, the inverse distance weighted interpolation method is used; when the moving window is missing five or fewer pixels, the RC-Kriging method is used; when the number of missing units on the second day is greater than or equal to 5 and the number of valid pixels on the first or third day is greater than or equal to 5, spatiotemporal weighted interpolation is used; when the number of pixels in the moving window is too small for three consecutive days, this part of the calculation is ignored; when pixel recovery causes overlap, missing values ​​are filled according to their priority; the priority is set as inverse distance weighted interpolation > RC Kriging > spatiotemporal weighted interpolation.

5. A machine learning-based regional NO2 spatial distribution computing device, characterized in that, include: The data acquisition unit is used to acquire remote sensing datasets and auxiliary data. The remote sensing datasets include AOD datasets, NO2 column concentrations, and related data. The auxiliary data include the characteristics of doy on a certain day of the year, the location ID of each pixel, and a 1-kilometer resolution grid of the UTM coordinate system covering the study area. The related data include normalized vegetation index, topography, population distribution, roads, and land use. A preprocessing unit is used to preprocess the remote sensing dataset and auxiliary data; this includes: converting the multi-source AOD dataset to a 1km / d spatiotemporal resolution using a 1km resolution grid in the UTM coordinate system of the auxiliary data; generating elevation (ELE) and slope (SL) using SRTM; overlaying ELE, SL, land use, road data, 470nm MAIAC AOD, NDVI, population, 470nm AHI AOD, and OMI NO2 column concentration onto a 1km resolution grid in the UTM coordinate system and reloading the data; calculating the reconstruction process based on the weighted average of the data pixel values ​​and the coverage ratio of the pixels on the 1km resolution grid; filling in data with a temporal resolution exceeding one day; interpolating meteorological parameters using co-kriging with ELE as the covariate, and outputting the UTM 1km / d spatiotemporal resolution; and using a two-step method to fill in missing data for 470nm MAIAC AOD and NO2 column concentration. The random ID unit is used to mark the position of all pixels with an independent ID. All IDs are shuffled by a random algorithm and then introduced into the random forest regression method. A prediction unit is used to train and predict the spatial distribution of NO2 concentration using a random forest regression method; specifically, it includes: The location parameters are randomized by shuffling the location ID values ​​using a random algorithm and assigning them to each pixel; the location parameters and random location IDs are normalized using a 0-1 normalization algorithm; and a random forest regression method is used to train and predict the spatial distribution of NO2 concentration. in The number is equal to the number of grids in a 1-kilometer resolution UTM; and These represent randomization and normalization algorithms, respectively. Represents the random forest model; This represents the predicted NO2 value; The numerical value representing the ID cell; These represent NO2 column concentration, 470 nm MAIAC AOD, random ID, altitude, slope, NDVI, road length, land use, and doy, respectively.

6. A machine learning-based regional NO2 spatial distribution computing device, characterized in that, It includes a memory and a processor, wherein the memory stores a computer program that can be executed by the processor to implement the machine learning-based regional NO2 spatial distribution calculation method as described in any one of claims 1 to 4.

Images