All-sky high-frequency near-surface ozone estimation method

Abstract

The invention discloses an all-sky high-frequency near-surface ozone estimation method. The method comprises the following steps: S1, carrying out data acquisition and preprocessing on influence factors of near-surface ozone; s2, setting a near-surface ozone estimation logic; s3, obtaining estimation target time; s4, estimating a near-surface ozone estimation result according to the near-surface ozone estimation logic; s5, carrying out training optimization on the near-surface ozone estimation logic; and S6, obtaining an all-sky high-frequency near-surface ozone estimation method and applying the method. According to the invention, multi-source information is fused, defects of ground monitoring are made up, and errors caused by weather are eliminated; the near-ground ozone monitoring coverage degree is improved, the defects of site monitoring are overcome, and large-range monitoring is achieved; high-precision estimation of near-surface ozone is realized, and data support is provided for business service and academic research.

Description

Technical Field

[0001] The present invention relates to the field of remote sensing technology, in particular to a method for estimating all-sky high-frequency near-ground ozone. Background Art

[0002] In recent years, ground-level ozone has become the primary summer atmospheric pollutant. In 2013, China's Ministry of Ecology and Environment established a national air quality observation network to monitor near-surface ozone in real time. The construction and maintenance of this ground-based network requires significant human and material resources. Satellite remote sensing is often used to complement ground-based monitoring and enable large-scale monitoring.

[0003] Currently, the application of remote sensing data to estimating ground-level ozone primarily involves traditional statistical models and machine learning algorithms. Machine learning algorithms, with their powerful data mining capabilities, are increasingly widely used. While simple to operate and computationally fast, traditional statistical models are sensitive to outliers and susceptible to collinearity, which can affect estimation results. Currently, most machine learning algorithms applied to ground-level ozone are based on fixed models constructed and estimated based on historical data, resulting in low temporal extrapolation accuracy in operational ground-level ozone estimation.

[0004] Existing technologies estimate ozone by inverting tropospheric ozone profiles. This method is greatly affected by weather and is limited to obtaining information only under clear sky conditions, making it inapplicable to full-sky calculations. Existing technologies typically use several specific parameters to estimate near-ground ozone. Once geophysical research discovers new influencing factors of near-ground ozone, existing technologies will not be able to introduce new variables in a short period of time, resulting in a lag. In addition, existing technologies collect data in a relatively discrete manner, so the time interval for estimating near-ground ozone increases accordingly, making it impossible to accurately reflect the changing trend of near-ground ozone. More existing technologies rely on programmed calculation methods, with fixed software programs that are difficult to modify and iterate, resulting in weak generalization and accuracy of the models.

[0005] The information disclosed in this background technology section is only intended to enhance understanding of the overall background of the invention and should not be regarded as an admission or any form of suggestion that the information constitutes the prior art already known to a person skilled in the art. Summary of the Invention

[0006] The present invention aims to provide a method for estimating all-sky high-frequency near-ground ozone, which can solve the problems existing in the background technology.

[0007] To achieve the above object, the present invention provides a method for estimating high-frequency near-ground ozone in the entire sky, comprising the following steps:

[0008] S1: Data collection and preprocessing of factors affecting near-ground ozone;

[0009] S2: Set the near-surface ozone estimation logic;

[0010] S3: Get estimated target time;

[0011] S4: estimating a near-ground ozone estimation result according to the near-ground ozone estimation logic;

[0012] S5: training and optimizing the near-surface ozone estimation logic;

[0013] S6: Obtain and apply the all-sky high-frequency near-surface ozone estimation method.

[0014] In one or more embodiments, step S1 includes:

[0015] S11: Obtain observation data from authoritative data sources;

[0016] S12: Verify the observation data to ensure that the data contains time coordinates and geographic longitude and latitude coordinates, and corresponding to the time coordinates and geographic longitude and latitude coordinates, associate the corresponding ground environmental monitoring station data, 4 km hourly FY4B Advanced Geostationary Radiation Imager (AGRI) payload zenith radiation data, 3 km CMA-GFS forecast data, 0.25° monthly scale Tsinghua University emission inventory, 90-meter ground elevation data (DEM), 5-kilometer population density distribution data (POP), historical 0.25° daily ozone monitoring instrument (OMI) L3 level NO2 column concentration and O3 column concentration data.

[0017] S13: Write the observation data into the storage unit.

[0018] In one or more embodiments, step S2 includes.

[0019] S21: Divide the land surface area;

[0020] S22: Calculate the multi-year mean of daily and monthly column concentrations;

[0021] S23: Select two time periods and perform linear interpolation to obtain hourly forecast data.

[0022] In one or more embodiments, step S21 includes:

[0023] S211: Radiation extraction in the FY4B AGRI band;

[0024] S212: 4km×4km matrix network cutting of the national land surface area;

[0025] S213: Obtain 9 hourly FY4B AGRI bands 7 to 15 and satellite zenith angle VZA, satellite azimuth angle VAA, solar zenith angle SZA, and solar azimuth angle SAA data respectively.

[0026] In one or more embodiments, step S22 includes:

[0027] S221: Retrieve historical 0.25° Daily Ozone Monitor (OMI) L3 level NO2 column concentration and O3 column concentration data;

[0028] S222: Calculate the daily multi-year mean data of OMI L3 NO2 column concentration (dOMINO2); the daily multi-year mean data of OMI L3 O3 column concentration (dOMIO3), the monthly multi-year mean data of OMI L3 NO2 column concentration (mOMINO2), and the multi-year mean data of OMI L3 O3 column concentration (mOMIO3);

[0029] The calculation formula of step S222 includes:

[0030]

[0031] Where Nr represents the number of the rth day, Dm represents the number of the mth month, n represents the grid number, d represents the number of the day, and m represents the number of the month.

[0032] In one or more embodiments, step S23 includes:

[0033] S231: Perform linear interpolation on the two middle time periods to obtain hourly forecast data;

[0034] The calculation formula for linear interpolation of the two middle time periods includes:

[0035]

[0036] Where ti represents the number of hours, and i is a non-negative integer multiple of 3.

[0037] In one or more embodiments, step S5 includes:

[0038] S51: Construct 37-dimensional grid data;

[0039] S52: Construct sample set;

[0040] S53: Train and verify the near-surface ozone estimation logic and optimize the estimation logic.

[0041] In one or more embodiments, step S51 includes:

[0042] S511: Select 17 near-surface ozone impact factors, cut out the center longitude and latitude of a 4km×4km rectangular grid of equal latitude and longitude across the entire land surface area of ​​China, search for valid pixels within 2km of the center, and take the average value as the pixel value of the feature;

[0043] S512: Calculate the spherical distance from the pixel to the center point, using the grid center longitude and latitude coordinates (φ n ,λ n ) extract the mean value of the effective pixels within the range of 2km as the grid element value, and calculate the longitude and latitude coordinates (φ) of each pixel to the grid center n ,λ n )’s spherical distance;

[0044] S513: Filter pixels within a 2km range and calculate the regional mean;

[0045] S514: Calculate the spherical distance from each pixel to the four vertices and the center point of a 4 km × 4 km rectangular grid of equal latitude and longitude that cuts out the entire land surface area;

[0046] S515: Obtain a 37-dimensional grid dataset.

[0047] In one or more embodiments, step S52 includes:

[0048] S521: Using the latitude and longitude information of the site observation data, matching the latitude and longitude of the site observation data with the average of the results within the neighborhood of the grid data center with a distance less than 4 km, combining the 37-dimensional grid data with the site observation data to extract the sample set;

[0049] S522: Extracting the sample set hour by hour for the previous 16 days;

[0050] In one or more embodiments, step S53 includes:

[0051] S531: Select LightGBM as the learner to supervise the learning model training;

[0052] S532: Evaluate the model using a 10-fold cross validation (10-CV) method, dividing the sample set into 10 groups;

[0053] S533: Select any one of the 10 sample sets as a test set, and use the remaining 9 sets as training sets to perform model training;

[0054] S534: Traverse all the 10 groups of sample sets as test sets, calculate and save the evaluation index of each model;

[0055] S535: Calculate the average of the 10 groups of test results as the model accuracy estimate, and use it as the accuracy indicator of the model under the current 10-fold cross-validation. The evaluation indicator of the model is the root mean square error (RMSE).

[0056] S536: The model outputs the importance index I(xj) of each feature xj, sorts the feature importance indexes, removes the minimum contributing features in turn, and retrains the model to find the optimal feature subset corresponding to the optimal model accuracy index.

[0057] Compared with the prior art, the multiple technical solutions and embodiments provided by the present invention have at least the following technical effects or advantages:

[0058] This solution integrates satellite remote sensing, CMA-GFS weather forecasts, emission inventories, and topographic population information to comprehensively characterize ozone formation conditions, fuse multi-source information, make up for the shortcomings of ground-based monitoring, and eliminate errors caused by weather.

[0059] Outputs continuous hourly O3 concentration estimates for the entire sky at a resolution of 4 km, improving near-surface ozone monitoring coverage, addressing the shortcomings of station monitoring, and enabling large-scale monitoring.

[0060] By sorting and gradually eliminating feature importance indicators, the model's generalization ability is improved, and high-precision estimation of near-ground ozone is achieved, providing data support for business services and academic research. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] The drawings described herein are used to provide a further understanding of the present invention and constitute a part of this application. The illustrative embodiments of the present invention and their descriptions are used to explain the present invention and can be considered as illustrative of various combinations of preferred embodiments and do not constitute an undue limitation of the present invention. In the drawings:

[0062] Figure 1 A schematic diagram of the overall steps of a method for estimating high-frequency near-ground ozone in the entire sky provided by the present invention;

[0063] Figure 2 The present invention provides a flowchart of a method for estimating high-frequency near-ground ozone in the entire sky. DETAILED DESCRIPTION

[0064] Unless otherwise expressly stated, throughout the specification and claims, the term "comprise" or its variations such as "include" or "comprising", "made of", etc. will be understood to include the stated elements or components but not to exclude other elements or other components.

[0065] The purpose of the present invention is to provide a method for estimating high-frequency near-surface ozone in the entire sky, so as to make up for the shortcomings of ground-based monitoring and realize large-scale monitoring; avoid the influence of variable collinearity and improve the time extrapolation accuracy in the operational estimation of near-surface ozone.

[0066] Example 1:

[0067] This embodiment provides a method for estimating high-frequency near-ground ozone in the entire sky, including the following steps:

[0068] S1: Data collection and preprocessing of factors affecting near-ground ozone;

[0069] S2: Set the near-surface ozone estimation logic;

[0070] S3: Get estimated target time;

[0071] S4: estimating a near-ground ozone estimation result according to the near-ground ozone estimation logic;

[0072] S5: training and optimizing the near-surface ozone estimation logic;

[0073] S6: Obtain and apply the all-sky high-frequency near-surface ozone estimation method.

[0074] Specifically, in order to provide a full-sky high-frequency near-ground ozone estimation method that is not affected by weather, can accurately and immediately reflect the changing trend of near-ground ozone, and has weak generalization ability and precision, this application integrates multi-source information, considers multiple factors affecting near-ground ozone, and estimates their impact on the change of near-ground ozone content through intelligent means. After the preliminary estimation, in order to verify the results, this application uses the training optimization method of step S5, which not only verifies the results, but also can feedback and optimize the verification algorithm by repeatedly selecting existing data from the data acquisition library for verification, thereby improving the estimation accuracy and generalization ability of the algorithm.

[0075] As a preferred implementation of this embodiment, step S1 includes:

[0076] S11: Obtain observation data from authoritative data sources;

[0077] S12: verifying the observation data, ensuring that the data contains time coordinates and geographic longitude and latitude coordinates, and corresponding ground environment monitoring station data, 4 km per hour FY4B advanced geostationary orbit radiometric imaging (AGRI) payload zenith radiation data, 3 km CMA-GFS forecast data, 0.25° monthly Tsinghua University emission inventory, 90 meter ground elevation data (DEM), 5 km population density distribution data (POP), historical 0.25° daily ozone detector (OMI) L3 NO2 column concentration and O3 column concentration data corresponding to the time coordinates and geographic longitude and latitude coordinates.

[0078] S13: writing the observation data into a storage unit.

[0079] Specifically, one or more prior art is based on the inversion of the tropospheric ozone profile to estimate the near-surface ozone, and there is a problem that the observation data is greatly affected by the weather. Under severe weather conditions, the prior art cannot obtain observation data, and thus cannot obtain the exact value of the near-surface ozone; and based on the multi-source data obtained from the authoritative data source, the estimation algorithm of the present application includes mutual verification of multiple data. Once individual data is affected, other data can be used for supplementation; as another preferred embodiment, the estimation algorithm of the present embodiment can measure the reliability of the data, and divide the observation data by weight, for example, when individual data is affected by local severe weather, the estimation algorithm is based on the reliability of the observation data. Reduce its weight, and thus reduce the interference of external factors on the estimation result.

[0080] Further, the machine learning algorithm model currently applied to the near-surface ozone is mostly based on the construction and estimation of fixed models based on historical data, and lacks a systematic feature importance sorting and simplification mechanism, and has the problem of low extrapolation accuracy in the business estimation of near-surface ozone. To solve this problem, the present patent uses FY4B geostationary satellite zenith radiation data, OMI NO2 and O3 column concentration data, combined with ground environment station data, numerical prediction data, emission inventory data, surface representation data and social production representation data, to construct a LightGBM near-surface ozone estimation model. Based on historical data, add real-time update data to combine samples for optimal feature selection and model iteration training, to improve the estimation accuracy of the model in actual business application.

[0081] As a preferred embodiment of the present embodiment, the step S2 comprises.

[0082] S21: dividing the land surface area;

[0083] S22: calculating the daily and monthly column concentration multi-year average;

[0084] S23: selecting two time periods and performing linear interpolation to obtain hourly forecast data.

[0085] As a preferred implementation of this embodiment, step S21 includes:

[0086] S211: Radiation extraction in the FY4BAGRI band;

[0087] S212: 4km×4km matrix network cutting of the national land surface area;

[0088] S213: Obtain 9 hourly FY4BAGRI bands 7 to 15 and satellite zenith angle VZA, satellite azimuth angle VAA, solar zenith angle SZA, and solar azimuth angle SAA data respectively.

[0089] Specifically, to establish a standardized spatial reference frame and set the resolution for near-surface ozone concentrations, radiation extraction was performed in the FY4B AGRI band. The national land surface was segmented into a 4km×4km rectangular grid of equal latitude and longitude, serving as the smallest unit. Hourly FY4B AGRI data were acquired for bands 7 through 15 (B7, B8…B15) and four angles (satellite zenith angle VZA, satellite azimuth angle VAA, solar zenith angle SZA, and solar azimuth angle SAA). By dividing the longitude and latitude grid, unique coordinates were assigned to any location on the Earth's surface, enabling precise positioning of observation points and sampled data and eliminating location ambiguity.

[0090] As a preferred implementation of this embodiment, step S22 includes:

[0091] S221: Retrieve historical 0.25° Daily Ozone Monitor (OMI) L3 level NO2 column concentration and O3 column concentration data;

[0092] S222: Calculate the daily multi-year mean data of OMI L3 NO2 column concentration (dOMINO2); the daily multi-year mean data of OMI L3 O3 column concentration (+dOMIO3), the monthly multi-year mean data of OMI L3 NO2 column concentration (mOMINO2), and the multi-year mean data of OMI L3 O3 column concentration (mOMIO3);

[0093] The calculation formula of step S222 includes:

[0094]

[0095]

[0096] Where Nr represents the number of the rth day, Dm represents the number of the mth month, n represents the grid number, d represents the number of the day, and m represents the number of the month.

[0097] Specifically, there is a complex nonlinear relationship between ground-level ozone (O3) concentrations and nitrogen dioxide (NO2) column concentrations, a relationship driven primarily by photochemical reaction mechanisms, regional transport, and meteorological conditions. Preferably, daily NO2 and O3 column concentrations from 2005 to 2021 are obtained from the OMI L3 daily product to calculate multi-year averages of daily OMI L3 NO2 column concentrations (dOMINO2), daily OMI L3 O3 column concentrations (dOMIO3), monthly OMI L3 NO2 column concentrations (mOMINO2), and monthly OMI L3 O3 column concentrations (mOMIO3). The authoritative data are further calculated using the multi-year average calculation formulas for daily and monthly column concentrations and used in subsequent calculations.

[0098] As a preferred implementation of this embodiment, step S23 includes:

[0099] S231: Perform linear interpolation on the two middle time periods to obtain hourly forecast data;

[0100] The calculation formula for linear interpolation of the two middle time periods includes:

[0101]

[0102] Where ti represents the number of hours, and i is a non-negative integer multiple of 3.

[0103] Specifically, since the CMA-GFS forecast is made every three hours, in order to address the problem of missing information during the long forecast interval, linear interpolation is performed on the two intermediate time intervals to obtain hourly forecast data. The following formula applies to all CMA-GFS meteorological elements (2-meter temperature TEM_103, relative humidity RHU_103, 10-meter zonal wind WIU_103, 10-meter meridional wind WIV_103, air pressure PRS_1, precipitation PRE_10, cloud cover TCC_1, boundary layer height PBLH_1):

[0104]

[0105] t i Where i is the i-th multiple of 3, and the unit is hours. For example, i3 is the 9th hour, t i+1 t i 3 hours after the last time.

[0106] As a preferred implementation of this embodiment, step S5 includes:

[0107] S51: Construct 37-dimensional grid data;

[0108] S52: Construct sample set;

[0109] S53: Train and verify the near-surface ozone estimation logic and optimize the estimation logic.

[0110] As a preferred implementation of this embodiment, step S51 includes:

[0111] S511: Select 17 near-surface ozone impact factors, cut out the center longitude and latitude of a 4km×4km rectangular grid of equal latitude and longitude across the entire land surface area of ​​China, search for valid pixels within 2km of the center, and take the average value as the pixel value of the feature;

[0112] S512: Calculate the spherical distance from the pixel to the center point, using the grid center longitude and latitude coordinates (φ n ,λ n ) extract the mean value of the effective pixels within the range of 2km as the grid element value, and calculate the longitude and latitude coordinates (φ) of each pixel to the grid center n ,λ n )’s spherical distance;

[0113] S513: Filter pixels within a 2km range and calculate the regional mean;

[0114] S514: Calculate the spherical distance from each pixel to the four vertices and the center point of a 4 km × 4 km rectangular grid of equal latitude and longitude that cuts out the entire land surface area;

[0115] S515: Obtain a 37-dimensional grid dataset.

[0116] Specifically, in step S51, the core goal of each step is to integrate the original data of different resolutions into a 4kmx4km grid through spatial smoothing while retaining local features. This includes:

[0117] 1. Data Standardization: Step S511 divides the entire land area into a 4km x 4km checkerboard grid, with each grid center serving as the reference coordinate. For feature data from 17 different sources, regardless of the original resolution, the average value of the valid pixels within 2km of the grid center is used to represent the value of that grid. By standardizing the minimum pixel, differences in original data resolution are eliminated, resulting in a spatially continuous 37-dimensional dataset.

[0118] 2. Local feature extraction: Step S512 can provide smooth noise while retaining local details, for example, urban pollution hotspots will not be over-averaged. Specifically, the following formula is used to take the grid center longitude and latitude coordinates (φ n ,λ n ) within a range of 2km, the mean value of the effective pixels is extracted as the grid element value to calculate each pixel (φ i ,λ i) to the grid center longitude and latitude coordinates (φ n ,λ n ) of the spherical distance D ni

[0119]

[0120] Among them, R = 6347km.

[0121] As a preferred embodiment of this embodiment, the spherical distance value from the valid pixel to the longitude and latitude coordinates of the grid center is retained and included in the calculation of the mean, and the weight of the disturbed pixel is reduced to avoid interference from outliers.

[0122] 3. Step S513 is used to filter pixels within a 2km range and calculate the regional mean: Specifically, the filter set S is defined. n ={i|D ni ≤2km, v i ≠null} implementation, v i Represents the effective value of the i-th pixel. For the grid center point (φ n ,λ n ), the corresponding elements are assigned as follows:

[0123]

[0124] 4. Calculate each pixel (φ i ,λ i ) to the spherical distance from the four vertices and the center point of a 4km×4km rectangular grid of equal latitude and longitude that cuts out the entire land surface of the country. Calculation formula:

[0125]

[0126] Among them, R = 6347 km, the value of k represents the position, and k = 1, 2, 3, 4, and 5 represent the upper left, upper right, lower left, lower right, and center point respectively.

[0127] After step S51, 37-dimensional data is obtained, which is used as material for subsequent training, including:

[0128] 9 hourly FY4B AGRI channel 7 to 15 (B7, B8…B15) and 4 angles (satellite zenith angle VZA, satellite azimuth angle VAA, solar zenith angle SZA, solar azimuth angle SAA) data;

[0129] 4 OMI NO2 daily mean column concentration (dOMINO2), O3 daily mean column concentration element (dOMIO3), NO2 monthly mean column concentration (mOMINO2), O3 monthly mean column concentration element (mOMIO3);

[0130] 8 CMA-GFS forecast elements (2-meter temperature TEM_103, relative humidity RHU_103, 10-meter zonal wind WIU_103, 10-meter meridional wind WIV_103, air pressure PRS_1, precipitation PRE_10, cloud cover TCC_1, boundary layer height PBLH_1);

[0131] Three monthly average data from Tsinghua University's emission inventories (total carbon monoxide emissions MixCO, total volatile organic compound emissions MNMVOC, and total nitrogen oxide emissions MNOx);

[0132] 2 auxiliary data: 90-meter ground elevation data (DEM), 5-kilometer population density distribution data (POP); spatiotemporal distribution characteristic data:

[0133] The geographical longitude and latitude of the two observation sites (LON, LAT) and the five characteristic parameters (PS1, PS2, PS3, PS4, PS5) representing the space calculated based on the existing Haversine method.

[0134] As a preferred implementation of this embodiment, step S52 includes:

[0135] S521: Using the latitude and longitude information of the site observation data, matching the latitude and longitude of the site observation data with the average of the results within the neighborhood of the grid data center with a distance less than 4 km, combining the 37-dimensional grid data with the site observation data to extract the sample set;

[0136] S522: Extracting the sample set hour by hour for the previous 16 days;

[0137] Specifically, step S52 processes the 37-dimensional network data to determine a sample set. Using the latitude and longitude information of the site observation data, the latitude and longitude of the site observation data is matched with the average of the results within the neighborhood of the grid data center with a distance less than 4 km, and the 37-dimensional grid data is combined with the site observation data to extract the sample set. i (s i ∈P,s i represents the i-th observation station, P represents the set of all ground observation stations) as the center of the first 16 days of hourly sample sets extracted

[0138]

[0139] in, Indicates site S i The mean of the features within the 4km neighborhood at time t; Indicates the measured ozone concentration at the corresponding moment; T 16dIt represents 16 days × 24 hours, a total of 384 moments, the number of samples at each site is 384 (assuming all are valid values), the total number of samples is 384 × |P|, and the sample sets of all sites are combined into the initial sample set

[0140] As a preferred implementation of this embodiment, step S53 includes:

[0141] S531: Select LightGBM as the learner to supervise the learning model training;

[0142] S532: Evaluate the model using a 10-fold cross validation (10-CV) method, and divide the sample set into 10 groups;

[0143] S533: Select any one of the 10 sample sets as a test set, and use the remaining 9 sets as training sets to perform model training;

[0144] S534: Traverse all the 10 groups of sample sets as test sets, calculate and save the evaluation index of each model;

[0145] S535: Calculate the average of the 10 groups of test results as the model accuracy estimate, and use it as the accuracy indicator of the model under the current 10-fold cross-validation. The evaluation indicator of the model is the root mean square error (RMSE).

[0146] S536: The model outputs the importance index I(xj) of each feature xj, sorts the feature importance indexes, removes the minimum contributing features in turn, and retrains the model to find the optimal feature subset corresponding to the optimal model accuracy index.

[0147] Specifically, step S52 is to optimize the ozone estimation model using a machine learning method, and its optimization steps are generally divided into two key stages:

[0148] The first is the model training and validation stage: an efficient decision tree model LightGBM is selected as a learner to supervise the learning model training; the 10-fold cross validation (10-CV) method is used to evaluate the model, and the sample set is divided into 10 groups. One of the groups is used as the test set in turn, and the remaining nine groups are used as training sets for model training. After training on each training set, a model is obtained, and the model is tested on the corresponding test set. The evaluation indicators of each model are calculated and saved. The average value of the 10 groups of test results is calculated as the model accuracy estimate, and is used as the accuracy indicator of the model under the current 10-fold cross validation to ensure the sample size requirements of the training process and the final accuracy. Among them, the model evaluation indicator uses the root mean square error (RMSE), which refers to the square root of the error between the estimated value and the true value. The smaller the error, the more accurate the estimate. Use the cross-validation method to ensure model stability and avoid overfitting.

[0149] The second is the feature screening stage, step S536 specifically includes: the model outputs the importance index I (x j ), sort the feature importance indicators, remove the features with the smallest contribution in turn, and retrain the model to find the optimal feature subset corresponding to the model accuracy indicator (represented by M).

[0150] The first step is to set the initial feature set (37 dimensions) as F0, corresponding to the initial sample set F0={x1,x2,...,x 37};

[0151] In the second step, the initial model training target is set as:

[0152] The third step is to define the iterative process after sorting the feature importance (the following expression is expressed in the kth iteration, from the set F k Remove an indicator I(x j )The smallest element, get the new set F k+1 Calculation formula):

[0153] The fourth step is to retrain the first three steps, corresponding to the accuracy index M(k), and repeat the model training goal:

[0154]

[0155] The fifth step is to find the optimal model accuracy index. The final optimal feature subset is defined as:

[0156]

[0157] Final feature set example:

[0158] F * ={B8, B9, B10, B12, B13, B14, B15, dOMINO2, dOMINO2, dOMIO3, DEM, POP, MixCO, MNMVOC, MNOx,

[0159] TEM_103, RHU_103, WIU_103, WIV_103, PRS_1, PRE_10, TCC_1, PBLH_1}

[0160] Estimate ozone products using the final model corresponding to the final feature set:

[0161]

[0162] Example 2:

[0163] This embodiment provides a full-sky near-ground ozone estimation system. Based on the same concept, this embodiment is used to implement the full-sky high-frequency near-ground ozone estimation method described in Example 1, which includes:

[0164] Acquisition module, used for data collection and preprocessing of factors affecting near-ground ozone;

[0165] A setting module is used to set the near-surface ozone estimation logic;

[0166] An acquisition module is used to obtain the estimated target time;

[0167] an estimation module, configured to estimate a near-ground ozone estimation result according to the near-ground ozone estimation logic;

[0168] An optimization module, configured to train and optimize the near-surface ozone estimation logic;

[0169] The application module is used to obtain and apply the all-sky high-frequency near-ground ozone estimation method.

[0170] Example 3:

[0171] This embodiment provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the all-sky high-frequency near-ground ozone estimation method described in Embodiment 1 is implemented.

[0172] Example 4:

[0173] A non-transitory computer-readable storage medium having a computer program stored thereon, wherein when the computer program is executed by a processor, the method for estimating high-frequency near-ground ozone in the entire sky as described in embodiment 1 is implemented.

[0174] The foregoing descriptions of specific exemplary embodiments of the present invention are for purposes of illustration and description. These descriptions are not intended to limit the invention to the precise forms disclosed, and it is apparent that many variations and modifications are possible in light of the foregoing teachings. The exemplary embodiments have been selected and described for the purpose of explaining the specific principles of the invention and their practical application, thereby enabling those skilled in the art to realize and utilize a variety of exemplary embodiments of the invention and various options and modifications. The scope of the invention is intended to be defined by the claims and their equivalents.

Claims

1. A method for estimating high-frequency near-surface ozone in the entire sky, characterized in that: The steps include: S1: Data collection and preprocessing of factors affecting near-ground ozone; S2: Set the near-surface ozone estimation logic; S3: Get estimated target time; S4: estimating a near-ground ozone estimation result according to the near-ground ozone estimation logic; S5: training and optimizing the near-surface ozone estimation logic; S6: Obtain and apply the all-sky high-frequency near-surface ozone estimation method.

2. The all-sky high-frequency near-ground ozone estimation method according to claim 1, characterized in that: The step S1 comprises: S11: Obtain observation data from authoritative data sources; S12: Verify the observation data to ensure that the data contains time coordinates and geographic longitude and latitude coordinates, and corresponding to the time coordinates and geographic longitude and latitude coordinates, associate the corresponding ground environmental monitoring station data, 4 km hourly FY4B Advanced Geostationary Radiation Imager (AGRI) payload zenith radiation data, 3 km CMA-GFS forecast data, 0.25° monthly scale Tsinghua University emission inventory, 90-meter ground elevation data (DEM), 5-kilometer population density distribution data (POP), historical 0.25° daily ozone monitoring instrument (OMI) L3 level NO2 column concentration and O3 column concentration data. S13: Write the observation data into the storage unit.

3. The all-sky high-frequency near-ground ozone estimation method according to claim 1, characterized in that: The step S2 includes: S21: Divide the land surface area; S22: Calculate the multi-year mean of daily and monthly column concentrations; S23: Select two time periods and perform linear interpolation to obtain hourly forecast data.

4. The all-sky high-frequency near-ground ozone estimation method according to claim 3, characterized in that: The step S21 includes: S211: Radiation extraction in the FY4BAGRI band; S212: 4km×4km matrix network cutting of the national land surface area; S213: Obtain 9 hourly FY4B AGRI bands 7 to 15 and satellite zenith angle VZA, satellite azimuth angle VAA, solar zenith angle SZA, and solar azimuth angle SAA data respectively.

5. The all-sky high-frequency near-ground ozone estimation method according to claim 4, characterized in that: The step S22 includes: S221: Retrieve historical 0.25° Daily Ozone Monitor (OMI) L3 level NO2 column concentration and O3 column concentration data; S222: Calculate the daily multi-year mean data of OMI L3 NO2 column concentration (dOMINO2); the daily multi-year mean data of OMI L3 O3 column concentration (dOMIO3), the monthly multi-year mean data of OMI L3 NO2 column concentration (mOMINO2), and the multi-year mean data of OMI L3 O3 column concentration (mOMIO3); The calculation formula of step S222 includes: Among them, N r represents the number of the rth day, D m Indicates the day of the mth month, n is the grid number, d is the day, and m is the month.

6. The all-sky high-frequency near-ground ozone estimation method according to claim 5, characterized in that: The step S23 includes: S231: Perform linear interpolation on the two middle time periods to obtain hourly forecast data; The calculation formula for linear interpolation of the two middle time periods includes: Among them, t i Represents the number of hours, i is a non-negative multiple of 3.

7. The all-sky high-frequency near-ground ozone estimation method according to claim 6, characterized in that: The step S5 comprises: S51: Construct 37-dimensional grid data; S52: Construct sample set; S53: Train and verify the near-surface ozone estimation logic and optimize the estimation logic.

8. The all-sky high-frequency near-ground ozone estimation method according to claim 7, characterized in that: The step S51 includes: S511: Select 17 near-surface ozone impact factors, cut out the center longitude and latitude of a 4km×4km rectangular grid of equal latitude and longitude across the entire land surface area of ​​China, search for valid pixels within 2km of the center, and take the average value as the pixel value of the feature; S512: Calculate the spherical distance from the pixel to the center point, using the grid center longitude and latitude coordinates (φ n ,λ n ) extract the mean value of the effective pixels within the range of 2km as the grid element value, and calculate the longitude and latitude coordinates (φ) of each pixel to the grid center n ,λ n )’s spherical distance; S513: Filter pixels within a 2km range and calculate the regional mean; S514: Calculate the spherical distance from each pixel to the four vertices and the center point of a 4 km × 4 km rectangular grid of equal latitude and longitude that cuts out the entire land surface area; S515: Obtain a 37-dimensional grid dataset.

9. The all-sky high-frequency near-ground ozone estimation method according to claim 8, characterized in that: The step S52 includes: S521: Using the latitude and longitude information of the site observation data, matching the latitude and longitude of the site observation data with the average of the results within the neighborhood of the grid data center with a distance less than 4 km, combining the 37-dimensional grid data with the site observation data to extract the sample set; S522: Extract the sample set hour by hour for the previous 16 days.

10. The all-sky high-frequency near-ground ozone estimation method according to claim 9, characterized in that: The step S53 includes: S531: Select LightGBM as the learner to supervise the learning model training; S532: Evaluate the model using a 10-fold cross validation (10-CV) method, dividing the sample set into 10 groups; S533: Select any one of the 10 sample sets as a test set, and use the remaining 9 sets as training sets to perform model training; S534: Traverse all the 10 groups of sample sets as test sets, calculate and save the evaluation index of each model; S535: Calculate the average of the 10 groups of test results as the model accuracy estimate, and use it as the accuracy indicator of the model under the current 10-fold cross-validation. The evaluation indicator of the model is the root mean square error (RMSE). S536: The model outputs the importance index I(xj) of each feature xj, sorts the feature importance indexes, removes the minimum contributing features in turn, and retrains the model to find the optimal feature subset corresponding to the optimal model accuracy index.

Images