Near-surface nitrogen dioxide concentration inversion method, system, equipment and medium

Abstract

The invention discloses a near-surface nitrogen dioxide concentration inversion method, system and device and a medium. The method comprises the following steps: acquiring a data set including European mid-term weather forecast center reanalysis data, troposphere nitrogen dioxide column concentration data, MODIS satellite sensing data, global terrain elevation and population distribution; and correcting the troposphere nitrogen dioxide data by using European mid-term weather forecast center data to obtain near-surface concentration data. And the data set is spatially aligned with the near-surface nitrogen dioxide data to form a modeling data set. A lightweight gradient elevator and an extreme random forest model are adopted to construct a nitrogen dioxide concentration inversion model, and the model is trained through the modeling data set; and inverting near-surface nitrogen dioxide concentration data of the to-be-detected area under different spatial and temporal scales by using the trained model. According to the method, the precision of an inversion result can be improved, near-surface nitrogen dioxide concentration products on different time scales can be obtained, and the method has relatively high applicability and authenticity.

Description

Technical Field

[0001] The present invention belongs to the technical field of nitrogen dioxide concentration inversion, and in particular relates to a near-ground nitrogen dioxide concentration inversion method, system, equipment and medium. Background Art

[0002] Nitrogen dioxide (NO2), as one of the main pollutants in air pollution, is a toxic brown-red irritating gas. NO2 is very active in chemical properties and is one of the important precursors of ozone (O3) and aerosol particles. It can trigger photochemical reactions and photochemical smog, and generate nitrate aerosols with strong radiative forcing. At the same time, nitric acid and nitrates formed by NO2 oxidation are the main causes of nitric acid rain. At present, with the rapid development of economy and population, nitrogen dioxide emissions have increased sharply, resulting in a continuous increase in the concentration of nitrogen dioxide in the atmosphere, causing serious harm to human health and the natural environment. Therefore, studying the concentration of NO2 in the atmosphere and its changing trend is of great significance and urgent need for strengthening environmental governance and ecological environmental protection. NO2 in the atmosphere mainly comes from natural emissions and human emissions. Among them, the natural emission sources are mainly lightning oxidation, soil microbial decomposition and biomass combustion, and the human emission sources are mainly fossil fuel combustion and automobile exhaust emissions.

[0003] Nowadays, the inversion of atmospheric NO2 concentration is mainly based on ground monitoring stations and satellite remote sensing. Ground stations use gas sensors or gas sampling methods to invert NO2 concentration. One common method is based on chemical analysis, which converts NO2 into measurable products such as nitrates, and then uses spectrometers to measure the concentration of the products to determine the concentration of NO2. However, ground monitoring stations have the disadvantages of high cost, complex maintenance, fixed monitoring stations and limited coverage. Compared with ground station monitoring, satellite remote sensing has the advantage of long-term, large-scale real-time dynamic continuous monitoring, which is different from ground stations. It is increasingly becoming the main means of inverting NO2 concentration. Among them, the Tropospheric Monitoring Instrument (TROPOMI) is the spectrometer with the highest temporal and spatial resolution and the most advanced technology after the Ozone Monitoring Instrument (OMI), providing better data support. Although satellite remote sensing can monitor the distribution and changes of global NO2, it can only provide the total nitrogen dioxide column in the atmosphere or the nitrogen dioxide column in the troposphere. Although the tropospheric nitrogen dioxide column is of great significance to meteorological and environmental research at the global and regional scales, the ground-level nitrogen dioxide concentration (GNO2) is more directly related to the ecological environment and human life. It can directly reflect the air quality conditions in cities and local areas, so as to timely detect and respond to air pollution problems and take effective control measures. Therefore, GNO2 has more practical research value and application prospects. At present, attempts have been made to convert satellite tropospheric NO2 column concentrations into GNO2 using different chemical transport, statistical, and machine learning models.

[0004] However, most previous studies simply selected relevant meteorological factors and terrain factors as training data for machine learning, and rarely considered the impact of the vertical distribution of NO2 on the process of converting the NO2 vertical column (XNO2 obtained from satellites) into the horizontal process of GNO2. As a black box algorithm, the internal implementation details of these traditional machine learning models cannot be understood. Therefore, if there is no high correlation between the training data and GNO2 in a practical sense, the training results obtained will not have a high practical value. Therefore, adding the vertical hierarchical structure of NO2 to the retrieval process is an improvement on the model physics. In addition, it is difficult for traditional machine learning models to fully express the complex spatiotemporal relationship between GNO2 and XNO2, which is also the reason why the current model results are quite different from the field measurement results.

[0005] In general, the inversion method of ground stations uses gas sensors or chemical analysis instruments to obtain real-time data to measure NO2 concentration, which has high timeliness and accuracy, and can promptly detect and respond to air pollution events. However, since ground stations require a large amount of equipment and human resources to maintain and operate, and their spatial layout is limited by various objective factors such as geographical conditions, the distribution of stations is uneven, so it is impossible to provide concentration data over a wide range. Although satellite remote sensing can provide global NO2 concentration data, it can only provide total atmospheric or tropospheric NO2 column concentration data, and cannot directly provide GNO2 concentration data. At the same time, it is affected by weather conditions, clouds and algorithm selection, and there are also problems such as missing data and inaccurate measurement accuracy. Summary of the invention

[0006] The purpose of the present invention is to provide a method, system, equipment and medium for inverting near-ground nitrogen dioxide concentration to solve the problems existing in the above-mentioned prior art.

[0007] To achieve the above object, the present invention provides a method for inverting near-ground nitrogen dioxide concentration, comprising:

[0008] Acquire a data set, wherein the data set includes reanalysis data of the European Centre for Medium-Range Weather Forecasts, tropospheric nitrogen dioxide column concentration data, MODIS satellite sensing data, global terrain elevation data, and global population distribution data;

[0009] Based on the reanalysis data of the European Centre for Medium-Range Weather Forecasts, the tropospheric nitrogen dioxide column concentration data is corrected to obtain near-ground nitrogen dioxide concentration data; the data set and the near-ground nitrogen dioxide concentration data are spatially aligned to obtain a modeling data set;

[0010] A nitrogen dioxide concentration inversion model is constructed based on a lightweight gradient boosting machine model and an extreme random forest model, the nitrogen dioxide concentration inversion model is trained based on the modeling data set, and the nitrogen dioxide concentration inversion of the area to be inverted is performed based on the trained nitrogen dioxide concentration inversion model to obtain nitrogen dioxide concentration data in the area to be inverted at different time and space scales.

[0011] Optionally, the correcting the tropospheric nitrogen dioxide column concentration data specifically includes:

[0012] Simulating vertical variation data of tropospheric nitrogen dioxide concentration based on reanalysis data of the European Centre for Medium-Range Weather Forecasts, and calculating data on the ratio of nitrogen dioxide concentration below the boundary layer to the overall tropospheric concentration based on the vertical variation data;

[0013] The tropospheric nitrogen dioxide column concentration data is resampled, and the resampled tropospheric nitrogen dioxide column concentration data is height-corrected based on the ratio data to obtain near-ground nitrogen dioxide concentration data.

[0014] Optionally, the resampled tropospheric nitrogen dioxide column concentration data is height-corrected based on the ratio data, and the specific calculation formula is:

[0015]

[0016] In the formula, is the height correction value of the satellite-observed tropospheric NO2 column concentration (XNO2), For satellite observation of XNO2 data, is the NO2 concentration at different layers / heights (r), and BLH is the boundary layer height obtained from the ERA5 reanalysis data.

[0017] Optionally, the process of constructing the modeling dataset specifically includes:

[0018] The data set and the near-ground nitrogen dioxide concentration data are spatially aligned, and based on a characteristic correlation analysis method, data having a strong correlation with the nitrogen dioxide concentration is selected, and the modeling data set is constructed based on the data selected by the characteristic correlation analysis method.

[0019] Optionally, the training of the nitrogen dioxide concentration inversion model based on the modeling data set specifically includes:

[0020] The modeling data set is input into the nitrogen dioxide concentration inversion model for concentration inversion, and training is performed according to the physical change mechanism of near-ground nitrogen dioxide to obtain a trained nitrogen dioxide concentration inversion model.

[0021] Optionally, inputting the modeling data set into the nitrogen dioxide concentration inversion model to perform concentration inversion specifically includes:

[0022] The modeling data set is input into the lightweight gradient boosting machine model for concentration prediction to obtain a first concentration inversion result, the residual between the first concentration inversion result and the corresponding measured value is input into the extreme random forest model for optimization to obtain a second concentration inversion result, and the total nitrogen dioxide concentration inversion result is constructed based on the first concentration inversion result and the second concentration inversion result to complete the concentration inversion.

[0023] Optionally, the total nitrogen dioxide concentration inversion result is constructed based on the first concentration inversion result and the second concentration inversion result, and the specific calculation formula is:

[0024] GNO2(i,j)=stableGNO2(i,j)+randomGNO2(i,j)

[0025] Where GNO2(i,j) is the near-ground nitrogen dioxide level of the specified pixel; stableGNO2(i,j) and randomGNO2(i,j) are the stable term and random term of the specified pixel.

[0026] A near-ground nitrogen dioxide concentration inversion system, comprising:

[0027] A data acquisition module is used to obtain a data set, wherein the data set includes reanalysis data of the European Center for Medium-Term Weather Forecasts, tropospheric nitrogen dioxide column concentration data, MODIS satellite sensor data, global terrain elevation data and global population distribution data; based on the reanalysis data of the European Center for Medium-Term Weather Forecasts, the tropospheric nitrogen dioxide column concentration data is corrected, and the data set and the corrected tropospheric nitrogen dioxide column concentration data are spatially aligned to obtain a modeling data set;

[0028] The nitrogen dioxide concentration inversion module is used to construct a nitrogen dioxide concentration inversion model based on a lightweight gradient boosting machine model and an extreme random forest model, train the nitrogen dioxide concentration inversion model based on the modeling data set, perform nitrogen dioxide concentration inversion in the area to be inverted based on the trained nitrogen dioxide concentration inversion model, and obtain nitrogen dioxide concentration data in the area to be inverted at different time and space scales.

[0029] An electronic device comprises a memory and a processor, wherein the memory is used to store a computer program, and the processor runs the computer program to enable the electronic device to perform a method for inverting near-ground nitrogen dioxide concentration.

[0030] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements a method for inverting near-ground nitrogen dioxide concentration.

[0031] The technical effects of the present invention are:

[0032] The present invention integrates various factors that affect the near-ground nitrogen dioxide concentration (GNO2), such as meteorological data, vegetation index (such as NDVI), terrain data (such as DEM) and population data, so that the inversion results are more comprehensive and reliable. This embodiment constructs a two-layer machine learning model by combining LGBM (Lightweight Gradient Boosting Machine Model) and ERF (Extreme Random Forest), which can more accurately capture the stability effect and random effect in the time series of GNO2 concentration, thereby improving the accuracy of the inversion results. This solution provides an efficient, accurate and reliable near-ground nitrogen dioxide concentration inversion method through an advanced two-layer machine learning model, which can obtain GNO2 products on different time scales, more accurately reflect abnormal events, have strong applicability and authenticity, and have important application value for environmental monitoring and air quality management. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0034] The drawings constituting a part of the present application are used to provide a further understanding of the present application. The illustrative embodiments and descriptions of the present application are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0035] Figure 1 It is a flow chart of a method for inverting near-ground nitrogen dioxide concentration based on a double-layer machine learning model in an embodiment of the present invention;

[0036] Figure 2 is a correlation coefficient diagram between different influencing factors in an embodiment of the present invention;

[0037] Figure 3 This is a diagram showing the accuracy comparison between the double-layer machine learning model in an embodiment of the present invention and the traditional model;

[0038] Figure 4 The spatial distribution map of the average values ​​of GNO2 concentration data inverted by the model in the embodiment of the present invention in China and typical regions from 2018 to 2023;

[0039] Figure 5 is the annual spatial distribution of GNO2 levels in China and the annual change rate of GNO2 concentration from 2018 to 2023 in the embodiment of the present invention;

[0040] Figure 6It is a spatial distribution map of the seasonal average values ​​of GNO2 concentration data in China and typical regions from 2018 to 2023 inverted by the model in an embodiment of the present invention;

[0041] Figure 7 The GNO2 concentration data inverted by the model in the embodiment of the present invention is a time series distribution diagram of 2018-2023 in China and typical regions;

[0042] Figure 8 A spatial distribution diagram of daily average values ​​of GNO2 concentration data inverted by the model in an embodiment of the present invention in a typical area where anomalies occur;

[0043] Fig. 9 4 is a structural diagram of a near-ground nitrogen dioxide inversion device in an embodiment of the present invention. DETAILED DESCRIPTION

[0044] Various exemplary embodiments of the present invention will now be described in detail. This detailed description should not be considered as limiting the present invention, but should be understood as a more detailed description of certain aspects, features, and embodiments of the present invention.

[0045] It should be understood that the terms described in the present invention are only for describing special embodiments and are not intended to limit the present invention. In addition, for the numerical range in the present invention, it should be understood that each intermediate value between the upper and lower limits of the scope is also specifically disclosed. Each smaller range between the intermediate value in any stated value or stated range and any other stated value or intermediate value in the described range is also included in the present invention. The upper and lower limits of these smaller ranges can be independently included or excluded in the scope.

[0046] The words “include,” “including,” “have,” “contain,” etc. used in this article are open-ended terms, meaning including but not limited to.

[0047] It should be noted that, in the absence of conflict, the embodiments and features in the embodiments of the present application can be combined with each other. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0048] Embodiment 1

[0049] like Figure 1 - Fig. 9As shown, in this embodiment, a method for inverting near-ground nitrogen dioxide concentration is provided, including: obtaining a data set, the data set including European Centre for Medium-Term Weather Forecasts reanalysis data, tropospheric nitrogen dioxide column concentration data, MODIS satellite sensor data, global terrain elevation data and global population distribution data; correcting the tropospheric nitrogen dioxide column concentration data based on the European Centre for Medium-Term Weather Forecasts reanalysis data to obtain near-ground nitrogen dioxide concentration data; spatially aligning the data set and the near-ground nitrogen dioxide concentration data to obtain a modeling data set; constructing a nitrogen dioxide concentration inversion model based on a lightweight gradient boosting machine model and an extreme random forest model, training the nitrogen dioxide concentration inversion model based on the modeling data set, performing nitrogen dioxide concentration inversion in the area to be inverted based on the trained nitrogen dioxide concentration inversion model, and obtaining nitrogen dioxide concentration data in the area to be inverted at different time and space scales.

[0050] This embodiment discloses a method, device, system and storage medium for inverting near-ground nitrogen dioxide concentration (GNO2) based on a double-layer machine learning model. The method takes into account the impact of the vertical layered structure of NO2 on GNO2 and the problem that the traditional machine learning model is difficult to fully express the complex spatiotemporal relationship between GNO2 and XNO2. By combining the advanced lightweight gradient boosting machine (LGBM) and extreme random forest (ERF) model, a double-layer machine learning (DLML) model is established to invert the GNO2 concentration in China from 2018 to 2023. The annual average GNO2 data obtained in this embodiment is basically consistent with the measured data of the monitoring site, and GNO2 products on different time scales can be obtained at the same time, which can more accurately reflect abnormal events and has strong applicability and authenticity.

[0051] The implementation steps of this embodiment include:

[0052] 1) Step 1: Analyze the sources and influencing factors of GNO2 concentration and perform height correction on the tropospheric NO2 column concentration data provided by TROPOMI;

[0053] 2) Step 2: Preprocess the data products such as the highly corrected NO2 data, meteorological data, topographic data, and population data and unify them to the same resolution for feature correlation analysis;

[0054] 3) Step 3: The processed variable data are combined into a modeling data set, and a two-layer machine learning model is used to model and adjust the matching data set to achieve the best model verification result;

[0055] 4) Step 4: Use the adjusted optimal prediction model to perform inversion to obtain the GNO2 concentration data set at different spatial and temporal scales in the region.

[0056] Perform height correction on the existing tropospheric NO2 column concentration data products: obtain the CAMS data set and boundary layer height (BLH) data provided by the European Centre for Medium-Range Weather Forecasts (ECMWF) to perform height correction on the tropospheric NO2 column concentration data. The spatial resolution of the CAMS data set product is 40km and the temporal resolution is 1 day. The spatial resolution of the boundary layer height (BLH) data is 0.25°×0.25°

[0057] The existing tropospheric NO2 column concentration data product is obtained. The spatial resolution of this product is 5.5×3.5km and the temporal resolution is 1 day. However, there are missing values ​​in the daily NO2 column concentration data. Therefore, the bilinear interpolation method is used to resample the TROPOMI data to a 0.05°×0.05° grid.

[0058] The vertical variation of NO2 concentration was simulated using the NO2 concentration profile data provided by CAMS, and the ratio of NO2 concentration below the boundary layer to the overall concentration in the troposphere was calculated. Based on the calculated ratio data, the NO2 column concentration product of TRO POMI was corrected, that is, the NO2 concentration of TROPOM in the troposphere was corrected to the surface concentration to obtain GNO2. The specific calculation formula is as follows:

[0059]

[0060] in, is the height correction value of the satellite-observed tropospheric NO2 column concentration (XNO2), For satellite observation of XNO2 data, is the NO2 concentration at different layers / heights (r), and BLH is the boundary layer height obtained from the ERA5 reanalysis data.

[0061] Data products such as height-corrected NO2 data, meteorological data, topographic data, and population data were preprocessed and unified to the same resolution, specifically: hourly meteorological data from ECMWF were obtained, including 2m temperature (TEMP), relative humidity (RELH), surface pressure (PRES), 10m meridional wind (u10), and 10m longitudinal wind (v10), with a spatial resolution of 0.25°×0.25°.

[0062] Obtain NDVI provided by MODIS, DEM data provided by SRTM, and LandScan population data set provided by EVC. The temporal and spatial resolutions of NDVI data are 16 days and 250M respectively, the resolution of DEM data is 30m, and the resolution of LandScan is 1km.

[0063] The highly corrected NO2 data, meteorological data, terrain data and other data were spatially unified into a 0.05°×0.05° grid to obtain the modeling data set. By resampling all data to a uniform spatial resolution (such as 0.05°×0.05°), the compatibility and comparability between different data sources were ensured, and the consistency of the inversion data set was improved.

[0064] The processed variable data are combined into a modeling data set, and the matching data set is modeled using a two-layer machine learning model, specifically: using the data set, a DLML model is established by combining advanced LGBM and ERF models to describe the stable effect and random effect of GNO2 concentration in the time series, respectively. The retrieval residuals of the first-layer LGBM model are regarded as random terms. Then, the ERF model in the second layer can estimate the random terms to correct the inversion level of GNO2. This embodiment decomposes the change in GNO2 concentration into stable terms and random terms, allowing the model to handle these two different changes separately, thereby improving the model's ability to identify and respond to different influencing factors. By combining with a two-layer model, the advantages of different models can be better utilized, thereby improving the overall retrieval performance.

[0065] It is assumed that the true GNO2 level consists of two parts, namely the stable term and the random term. The stable term refers to the cross-year stability of GNO2. Specifically, the changes in GNO2 concentration at the same time in different years can be ignored. The numerical deviations are usually not large and show similar trends. The random term refers to the random fluctuations of GNO2. GNO2 is affected by meteorological conditions or human activities and will change unpredictably at the same time in different years. It may manifest as a rapid change in the GNO2 concentration of a certain pixel every day. Therefore, according to this mechanism, the true GNO2 level can be decomposed into:

[0066] GNO2(i,j)=stableGNO2(i,j)+randomGNO2(i,j)

[0067] Among them, GNO2(i,j) refers to the true GNO2 level of the specified pixel (i,j); stableGNO2(i,j) and randomGNO2(i,j) are the stable term and random term of the specified pixel (i,j).

[0068] The stable term represents the long-term pattern of GNO2 levels, and the LGBM model inversion with more efficient training capacity and more stable training results is chosen to simulate the stable term of GNO2. The LGBM model is retrieved as the first layer:

[0069]

[0070] in, is the stable term for the specified pixel (i, j); the explanatory variables of the LGBM model include CNO2, NDVI, DEM, POP and other meteorological factors. The predictor variable of the LGBM model is the monthly average of the ground station data.

[0071] At the same time, it is difficult to characterize the daily random variation of GNO2 levels using only the traditional LGBM model. Therefore, the ERF model is introduced as the second layer to further optimize the residual results of the LGBM model inversion, thereby reducing the negative impact of random effects on a single traditional machine learning model, and thus inverting the real GNO2 random term. Use the ERF model to describe the random term of GNO2:

[0072]

[0073] in, The random term of GNO2 for a given pixel (i, j); CNO2, NDVI, DE M, POP and other meteorological factors are input into the ERF model as a series of explanatory variables. In addition, the predictor variable of the ERF model is the overall residual of the LGBM model.

[0074] The established GNO2 concentration inversion model is used for prediction to obtain GNO2 concentration data sets at different spatial and temporal scales in the region. Specifically, all modeling data are processed into standard spatial grids, and the established two-layer machine learning model is used to invert the GNO2 value of each pixel in the region to obtain GNO2 concentration data sets at different spatial and temporal scales in the region, and respond to abnormal events.

[0075] This embodiment integrates various factors affecting GNO2 concentration, such as meteorological data, vegetation index (such as NDVI), terrain data (such as DEM) and population data, so that the inversion result is more comprehensive and reliable.

[0076] This embodiment can more accurately capture the stability effect and random effect in the time series of GNO2 concentration by combining the two-layer machine learning model constructed by LGBM and ERF, thereby improving the accuracy of the inversion results. This solution provides an efficient, accurate and reliable near-ground nitrogen dioxide concentration inversion method through an advanced two-layer machine learning model, which can obtain GNO2 products on different time scales, more accurately reflect abnormal events, has strong applicability and authenticity, and has important application value for environmental monitoring and air quality management.

[0077] The specific application process of this embodiment includes:

[0078] 1. Analyze the sources and influencing factors of GNO2 concentration, perform height correction on the tropospheric NO2 column concentration data provided by TROPOMI, pre-process the meteorological data, topographic data, population data and other data products and unify them to the same resolution, and perform feature correlation analysis.

[0079] Specifically, the tropospheric NO2 column concentration data used in this embodiment is obtained by the Tropospheric Observation Instrument (TROPOMI) carried on the Sentinel-5P satellite launched by the European Space Agency (ESA) in 2017. The TROPOMI observation width is about 2600km, which can cover the world every day. The spatial resolution has been increased from 7km×3.5km to 5.5km×3.5km since August 6, 2019. It is the atmospheric monitoring spectrometer with the highest spatial resolution today. The TROPOMI tropospheric NO2 column concentration is obtained by inversion based on the differential absorption spectroscopy (DOAS) algorithm in the spectral range of 405-465nm. At present, there are two levels of tropospheric NO2 column concentration data products that TROPOMI can provide, namely, the data format has two levels of L1B and L2. This embodiment uses L2 data, and the time range is from January 1, 2022 to December 31, 2022. Since there are data gaps in the tropospheric NO2 column concentration data provided by satellites, the TR OPOMI data are resampled to a 0.05° × 0.05° grid using bilinear interpolation.

[0080] This example uses the NO2 stratification data provided by CAMS to simulate the vertical variation of tropospheric NO2 concentration and calculate the ratio of NO2 concentration below the boundary layer to the overall tropospheric concentration, thereby correcting the TROPOMI tropospheric NO2 column concentration product. The specific formula is as follows:

[0081]

[0082] in, is the height correction value of the satellite-observed tropospheric NO2 column concentration (XNO2), For satellite observation of XNO2 data, is the NO2 concentration at different layers / heights (r), and BLH is the boundary layer height obtained from the ERA5 reanalysis data.

[0083] The model input variables were selected according to the source and influencing factors of GNO2 concentration, namely meteorological data, vegetation index data, terrain data and population data, and all data were uniformly resampled to a spatial resolution of 0.05°×0.05° for feature correlation analysis. The correlation between each data set and GNO2 concentration is shown in Figure 2 .

[0084] 2. The processed variable data are combined into a modeling data set, and the matching data set is modeled using a two-layer machine learning model. Specifically, a modeling data set is constructed by analyzing the correlation of variables. The model selected in this embodiment is a two-layer machine learning model. It is difficult for traditional machine learning models to fully express the complex spatiotemporal relationship between GNO2 and tropospheric NO2 column concentrations. To this end, this embodiment divides the real GNO2 into two components, namely a stable term and a random term. The stable term refers to the part whose changes at the same time in different years can be ignored, for example, the monthly GNO2 concentration of a specified pixel is similar to the value of other years. The random term refers to the part that changes unpredictably at the same time in different years, such as the rapid change in the daily concentration of GNO2 in a specified pixel. According to this mechanism, the real GNO2 can be decomposed into:

[0085] GNO2(i,j)=stableGNO2(i,j)+randomGNO2(i,j)

[0086] Among them, GNO2(i,j) refers to the true GNO2 level of the specified pixel (i,j); stableGNO2(i,j) and randomGNO2(i,j) are the stable term and random term of the specified pixel (i,j).

[0087] (1) Calculate stable GNO2. The stable term represents the long-term pattern of GNO2. Choosing a model that can represent the overall change trend of GNO2 is crucial to expressing GNO2. The LGBM model is an iterative boosting tree system based on the gradient boosting decision tree (GBDT) framework launched by Microsoft in 2017. LGBM uses an enhanced histogram decision tree algorithm to divide continuous feature values ​​into K intervals before training, and traverses the entire sample from the histogram containing K for statistical accumulation to find the best split point, which improves the computational execution efficiency by half compared with GBDT. Compared with the classic GBDT, which generally only uses the loss function to fit the model, LGBM can find the optimization function through training, thereby integrating K regression trees to fit the final model, with more efficient training capabilities and more stable training results. The LGBM model is regarded as the first layer of inversion, and its formula is as follows:

[0088] stableGNO2(i,j)=LGBM[CNO2(i,j),NDVI(i,j),DEM(i,j),POP(i,j)...]

[0089] In the calculation formula of the first layer of the above inversion, is the stability term for the specified pixel (i, j); LGBM[CNO2(i, j), NDVI(i, j), DEM(i, j), POP(i, j)...] is a series of variables input into the reference LGBM model including CNO2, NDVI, DEM, POP and other meteorological factors.

[0090] (2) Calculate random GNO2. It is difficult to use only the traditional LGBM model to describe the random changes in GNO2 between different days. Therefore, the embodiment introduces the ERF model as the second layer to further optimize the inversion results of the first layer to overcome the negative impact of random effects on a single traditional machine learning model. ERF is an ensemble learning algorithm that uses a set of decision trees to improve the accuracy and stability of predictions. It is a variant of the random forest algorithm, but has stronger randomness during the training process. ERF randomly selects features for splitting at each node instead of using the best performing features based on criteria such as information gain. This random selection of the ERF model helps to maintain the random variation characteristics of the daily GNO2 level, thereby preventing overfitting and improving the generalization ability of the ensemble. Therefore, this embodiment uses the ERF model to describe the random terms in the equation, and the specific formula is as follows:

[0091]

[0092] in, is the random term of GNO2 for the specified pixel (i, j); CNO2, NDVI, DEM, POP and other meteorological factors are a series of variables input into the reference ERF model.

[0093] In addition to the direct fitting results after the statistical model training is completed, the model is also verified by a ten-fold cross-validation method, which can avoid potential overfitting problems in the model. This embodiment divides the data set into 10 equal parts, one of which is used for testing and nine are used for model fitting. By repeating this process 10 times, the optimized model parameters are selected to obtain the best model. In addition, this embodiment takes into account the influence of adjacent location information and selects sample-based cross-validation and location-based cross-validation to evaluate the generalization and spatial extrapolation of the model.

[0094] At the same time, this embodiment also selects the cross-validation determination coefficient (R 2 ), mean absolute error (MAE) and root mean square error (RMSE) are used as statistical indicators for accuracy evaluation. The calculation formula is as follows:

[0095]

[0096] Among them, y i is the actual value, is the predicted value, is the mean of the actual values, and n is the number of observations. 2 The closer the absolute value of is to 1, the better the model fit is. The closer the RMSE and MAE values ​​are to 0, the closer the predicted value is to the actual value.

[0097] The accuracy of the model is verified using the above accuracy indicators, and the results are as follows Figure 3 As shown in Figure 2, it can be seen that the overall accuracy of the DLML model is higher than that of the traditional LGBM model. 2 It can be as high as 0.97, which is 0.02 higher than the traditional LGBM model. Its MAE and RMSE are about 0.87μɡ / m lower than the LGBM model. 3 and 1.11μɡ / m 3 It can be seen that the DLML model improves the accuracy and reliability of model prediction. The R 2 All of them are above 0.86, and the R 2 It can reach 0.87, indicating that the DLML model has strong stability on different data sets. At the same time, compared with the traditional LGBM model, the R 2 The MAE and RMSE are reduced to about 4.5 μɡ / m 3 and 6.1 μɡ / m 3 The universality and effectiveness of the proposed DLML model in unknown data estimation are analyzed. The model can accurately retrieve GNO2 at non-participating sites and has good retrieval ability and spatial generalization ability.

[0098] 3. Perform standard spatial gridding on all modeling data, and use the established two-layer machine learning model to invert the GNO2 concentration value of each pixel in the region to obtain a comprehensive set of GNO2 data products at different spatial and temporal scales in the region. Specifically, in order to better display the inversion results, the GNO2 concentration inversion model of the two-layer machine learning model built in this embodiment is used to perform a surface inversion of the GNO2 concentration in the Chinese region. Figure 4 The annual mean value map of GNO2 from 2018 to 2023 was drawn, and 5 typical regions were selected as comparison objects. By comparing with the measured data products of ground monitoring stations, the differences between the obtained data set and the measured data were analyzed. Figure 4 It can be seen that by comparing the inversion data of typical regional site monitoring, the GNO2 data inverted by this embodiment has a consistent spatial distribution with the inversion data of site monitoring, and the spatial distribution of the obtained data is complete, with good adaptability and spatial extrapolation ability, and can perform more accurate inversion of GNO2 concentration in large areas.

[0099] However, due to the short life cycle of GNO2 concentration, analyzing only the annual average total amount may obscure the details of these changes and fail to understand the overall annual changes. Therefore, a spatial distribution map of the annual average value of GNO2 from 2018 to 2023 was drawn, as shown in Figure 2. Figure 5 As shown, the GNO2 concentration change rate over the five years was calculated in order to provide a more comprehensive and scientific reference for governance policies.

[0100] The model was used to draw the quarterly mean graph of GNO2 concentration from 2018 to 2023 and the five-year time series change graph, as shown in the figure below: Figure 6 , 7 It can be seen that the GNO2 concentration distribution trend inverted in this embodiment is roughly the same as the annual average concentration distribution trend, and the good model performance in the time dimension can perform accurate inversion.

[0101] In addition, from Figure 7 Analysis shows that both Wuhan and the Yangtze River Delta experienced abnormal phenomena in December 2021, reaching the highest value in the past five years, almost exceeding 35μɡ / m 3 After that, the GNO2 level dropped sharply. In order to analyze the cause of the abnormal phenomenon, the specific time point of the abnormal phenomenon was screened out, and the concentration distribution diagram was drawn ( Figure 8 ).pass Figure 8 It can be seen that the GNO2 concentration daily data inverted by this model and the inverted data monitored at the site are consistent with each other, and their spatial variation trends are basically consistent, which can prove that this embodiment has good model performance and strong GNO2 concentration inversion capability.

[0102] It should be noted that the modules involved in the embodiments of the present invention may be implemented by software or hardware, and the modules described may also be set in a processor. The names of these modules do not, in some cases, limit the modules themselves.

[0103] This embodiment provides a near-ground nitrogen dioxide concentration inversion system, including:

[0104] A data acquisition module is used to obtain a data set, wherein the data set includes reanalysis data of the European Center for Medium-Term Weather Forecasts, tropospheric nitrogen dioxide column concentration data, MODIS satellite sensor data, global terrain elevation data and global population distribution data; based on the reanalysis data of the European Center for Medium-Term Weather Forecasts, the tropospheric nitrogen dioxide column concentration data is corrected, and the data set and the corrected tropospheric nitrogen dioxide column concentration data are spatially aligned to obtain a modeling data set;

[0105] The nitrogen dioxide concentration inversion module is used to construct a nitrogen dioxide concentration inversion model based on a lightweight gradient boosting machine model and an extreme random forest model, train the nitrogen dioxide concentration inversion model based on the modeling data set, perform nitrogen dioxide concentration inversion in the area to be inverted based on the trained nitrogen dioxide concentration inversion model, and obtain nitrogen dioxide concentration data in the area to be inverted at different time and space scales.

[0106] This embodiment provides an electronic device, including a memory and a processor, wherein the memory is used to store a computer program, and the processor runs the computer program to enable the electronic device to perform the method for inverting near-ground nitrogen dioxide concentration.

[0107] This embodiment provides a computer-readable storage medium storing a computer program. When the computer program is executed by a processor, the method for inverting the near-ground nitrogen dioxide concentration is implemented.

[0108] The above is only a preferred specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed in the present application should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

Claims

1. A method for inverting near-surface nitrogen dioxide concentration, characterized in that: include: Acquire a data set, wherein the data set includes reanalysis data of the European Centre for Medium-Range Weather Forecasts, tropospheric nitrogen dioxide column concentration data, MODIS satellite sensing data, global terrain elevation data, and global population distribution data; Based on the reanalysis data of the European Centre for Medium-Range Weather Forecasts, the tropospheric nitrogen dioxide column concentration data is corrected to obtain near-ground nitrogen dioxide concentration data; the data set and the near-ground nitrogen dioxide concentration data are spatially aligned to obtain a modeling data set; A nitrogen dioxide concentration inversion model is constructed based on a lightweight gradient boosting machine model and an extreme random forest model, the nitrogen dioxide concentration inversion model is trained based on the modeling data set, and the nitrogen dioxide concentration inversion of the area to be inverted is performed based on the trained nitrogen dioxide concentration inversion model to obtain nitrogen dioxide concentration data in the area to be inverted at different time and space scales.

2. A method for inverting near-surface nitrogen dioxide concentration according to claim 1, characterized in that: The correcting of the tropospheric nitrogen dioxide column concentration data specifically includes: Simulating vertical variation data of tropospheric nitrogen dioxide concentration based on reanalysis data of the European Centre for Medium-Range Weather Forecasts, and calculating data on the ratio of nitrogen dioxide concentration below the boundary layer to the overall tropospheric concentration based on the vertical variation data; The tropospheric nitrogen dioxide column concentration data is resampled, and the resampled tropospheric nitrogen dioxide column concentration data is height-corrected based on the ratio data to obtain near-ground nitrogen dioxide concentration data.

3. A method for inverting near-surface nitrogen dioxide concentration according to claim 2, characterized in that: The resampled tropospheric nitrogen dioxide column concentration data is highly corrected based on the ratio data, and the specific calculation formula is: In the formula, is the height correction value of the satellite-observed tropospheric NO2 column concentration (XNO2), For satellite observation of XNO2 data, is the NO2 concentration at different layers / heights (r), and BLH is the boundary layer height obtained from the ERA5 reanalysis data.

4. The method for inverting near-surface nitrogen dioxide concentration according to claim 1, characterized in that: The process of constructing the modeling dataset specifically includes: The data set and the near-ground nitrogen dioxide concentration data are spatially aligned, and based on a characteristic correlation analysis method, data having a strong correlation with the nitrogen dioxide concentration is selected, and the modeling data set is constructed based on the data selected by the characteristic correlation analysis method.

5. The method for inverting near-surface nitrogen dioxide concentration according to claim 1, characterized in that: The training of the nitrogen dioxide concentration inversion model based on the modeling data set specifically includes: The modeling data set is input into the nitrogen dioxide concentration inversion model for concentration inversion, and training is performed according to the physical change mechanism of near-ground nitrogen dioxide to obtain a trained nitrogen dioxide concentration inversion model.

6. A method for inverting near-surface nitrogen dioxide concentration according to claim 5, characterized in that: The inputting the modeling data set into the nitrogen dioxide concentration inversion model to perform concentration inversion specifically includes: The modeling data set is input into the lightweight gradient boosting machine model for concentration prediction to obtain a first concentration inversion result, the residual between the first concentration inversion result and the corresponding measured value is input into the extreme random forest model for optimization to obtain a second concentration inversion result, and the total nitrogen dioxide concentration inversion result is constructed based on the first concentration inversion result and the second concentration inversion result to complete the concentration inversion.

7. A method for inverting near-surface nitrogen dioxide concentration according to claim 6, characterized in that: The total nitrogen dioxide concentration inversion result is constructed based on the first concentration inversion result and the second concentration inversion result. The specific calculation formula is: GNO2(i,j)=stableGNO2(i,j)+randomGNO2(i,j) Where GNO2(i,j) is the near-ground nitrogen dioxide level of the specified pixel; stableGNO2(i,j) and randomGNO2(i,j) are the stable term and random term of the specified pixel.

8. A near-surface nitrogen dioxide concentration inversion system, characterized in that: include: A data acquisition module is used to obtain a data set, wherein the data set includes reanalysis data of the European Center for Medium-Term Weather Forecasts, tropospheric nitrogen dioxide column concentration data, MODIS satellite sensor data, global terrain elevation data and global population distribution data; based on the reanalysis data of the European Center for Medium-Term Weather Forecasts, the tropospheric nitrogen dioxide column concentration data is corrected, and the data set and the corrected tropospheric nitrogen dioxide column concentration data are spatially aligned to obtain a modeling data set; The nitrogen dioxide concentration inversion module is used to construct a nitrogen dioxide concentration inversion model based on a lightweight gradient boosting machine model and an extreme random forest model, train the nitrogen dioxide concentration inversion model based on the modeling data set, perform nitrogen dioxide concentration inversion in the area to be inverted based on the trained nitrogen dioxide concentration inversion model, and obtain nitrogen dioxide concentration data in the area to be inverted at different time and space scales.

9. An electronic device, characterized in that: It comprises a memory and a processor, the memory is used to store a computer program, and the processor runs the computer program to enable the electronic device to perform a near-ground nitrogen dioxide concentration inversion method according to any one of claims 1-7.

10. A computer-readable storage medium, characterized in that: It stores a computer program, which, when executed by a processor, implements a near-ground nitrogen dioxide concentration inversion method as described in any one of claims 1 to 7.

Images