An hourly near-surface ozone retrieval method based on multi-source data fusion

Abstract

A method for hourly near-surface ozone retrieval based on multi-source data fusion, relating to the field of data monitoring, includes: time-aligning ozone total column concentration characteristic data, meteorological characteristic data, spatiotemporal geographic characteristic data, and socioeconomic characteristic data based on timestamps; uniformly processing the time-aligned multi-source data into hourly data at a preset grid resolution; spatially matching the hourly data at the preset grid resolution based on monitoring station identifiers and latitude and longitude, and constructing a sample set for model training by combining it with station measured data; obtaining a near-surface ozone concentration retrieval model based on the training sample set; and retrieving ozone concentration in the target study area using the ozone concentration retrieval model to obtain target hourly ozone concentration data at a preset grid resolution within the target study area. Implementing this method can achieve high-precision fusion of multi-source data and hourly ozone concentration retrieval, effectively improving the accuracy and effectiveness of near-surface ozone monitoring.

Description

Technical Field

[0001] This application relates to the field of data monitoring, and in particular to an hourly near-surface ozone retrieval method based on multi-source data fusion. Background Technology

[0002] Ground-level ozone (O3), a typical photochemical pollutant, is formed under the complex influence of multiple factors. According to the Global Burden of Disease study, ozone-related diseases caused 489,500 premature deaths globally in 2021. Accurate assessment and prediction of ozone pollution require comprehensive consideration of multi-dimensional information, including satellite observations, meteorological conditions, and geographical environment.

[0003] Currently, ozone monitoring data comes from diverse sources, including total ozone column concentration data acquired through satellite remote sensing, meteorological reanalysis data, geographic information data, and socioeconomic data. These data record information on different factors influencing ozone formation, each with its own unique spatiotemporal resolution and data format.

[0004] However, due to significant differences in spatiotemporal scales and data structures, these heterogeneous data are difficult to integrate directly and effectively. This results in existing ozone concentration retrieval methods often only utilizing partial data sources, failing to fully leverage the synergistic effect of multi-source data. Summary of the Invention

[0005] This application provides an hourly near-surface ozone retrieval method based on multi-source data fusion, which is used to improve the accuracy of hourly near-surface ozone concentration retrieval.

[0006] Firstly, this application provides an hourly near-surface ozone retrieval method based on multi-source data fusion, applied to an hourly near-surface ozone concentration retrieval system. The method includes: collecting ozone total column concentration characteristic data, meteorological characteristic data, spatiotemporal geographic characteristic data, and socioeconomic characteristic data; performing time alignment on the ozone total column concentration characteristic data, meteorological characteristic data, spatiotemporal geographic characteristic data, and socioeconomic characteristic data based on timestamps; uniformly processing the time-aligned data into hourly data at a preset grid resolution; spatially matching the hourly data at the preset grid resolution based on monitoring station identifiers and latitude and longitude to obtain first matching data; constructing a training sample set based on the first matching data and measured data from monitoring stations; training an ozone concentration retrieval model based on the training sample set; and performing ozone concentration retrieval on the target study area using the ozone concentration retrieval model to obtain target hourly ozone concentration data at a preset grid resolution within the target study area.

[0007] In the above embodiments, the system collects four types of feature data and aligns them based on timestamps, eliminating inconsistencies in the time dimension. The aligned data is then uniformly processed into hourly data at a preset grid resolution, resolving differences in spatial scale and temporal granularity. Spatial matching is used to associate the grid data with measured data from stations to construct a training sample set, establishing a mapping relationship between multi-source features and ozone concentration. The synergy of multi-source data compensates for the limitations of a single data source, gridded processing achieves continuous spatial coverage, and the hourly scale captures diurnal variation characteristics, solving the problems of insufficient inversion accuracy and incomplete spatiotemporal coverage caused by the inability of existing methods to fully integrate heterogeneous data.

[0008] In conjunction with some embodiments of the first aspect, in some embodiments, the step of uniformly processing ozone total column concentration characteristic data, meteorological characteristic data, spatiotemporal geographic characteristic data, and socioeconomic characteristic data into hourly data at a preset grid resolution specifically includes: determining the preset grid resolution based on the spatial range of the target study area within the ozone total column concentration characteristic data; upsampling the ozone total column concentration characteristic data to the preset grid resolution using bilinear interpolation; upsampling the meteorological characteristic data and socioeconomic characteristic data to the preset grid resolution using bilinear interpolation; performing scale matching processing on the spatiotemporal geographic characteristic data based on the area-weighted average method; and converting non-hourly scale data into hourly data.

[0009] In the above embodiments, the system uses bilinear interpolation to upsample total ozone concentration, meteorological characteristics, and socioeconomic characteristics to a preset grid, achieving a smooth numerical transition through distance-weighted calculation. For spatiotemporal geographic features, an area-weighted average method is used for downsampling, calculating weighted values ​​based on the overlapping area ratio to preserve spatial statistical characteristics. Bilinear interpolation ensures the spatial continuity of the upsampled data, while the area-weighted method minimizes information loss during downsampling. A unified grid resolution and hourly time scale ensure complete spatiotemporal alignment of all data sources, eliminating scale mismatch issues in multi-source data fusion.

[0010] In conjunction with some embodiments of the first aspect, in some embodiments, the step of spatially matching hourly data at a preset grid resolution based on monitoring station identifiers and latitude and longitude to obtain first matching data, and constructing a training sample set based on the first matching data and the measured ozone concentration data of the stations, specifically includes: based on the preset grid resolution, extracting raster feature data corresponding to each monitoring station from the time-aligned multi-source hourly grid data according to the latitude and longitude location of each monitoring station; the raster feature data includes at least ozone total column features, meteorological features, spatiotemporal geographic features, and socioeconomic features; associating the raster feature data with the corresponding monitoring station identifiers and latitude and longitude to establish a spatial mapping relationship between the station location and the raster data to obtain the first matching data; and pairing the first matching data with the corresponding measured ozone concentration data of the monitoring stations according to the monitoring station identifiers to construct a training sample set.

[0011] In the above embodiments, corresponding raster feature data is extracted according to the latitude and longitude of the monitoring station, ensuring accurate spatial matching. A spatial mapping relationship is established by associating the raster features with the station identifier and latitude and longitude to obtain the first matching data. This achieves precise binding between the station location and the raster data, establishing a unique index. Then, based on the monitoring station identifier and latitude and longitude, the first matching data is paired with the measured concentration to construct a training sample set, achieving precise mapping from features to labels. Spatial matching ensures positioning accuracy, the dual association of identifier and latitude and longitude ensures data traceability and location accuracy, and feature pairing provides high-quality supervisory signals, solving the problem of difficulty in effectively associating multi-source data with ground observations.

[0012] In conjunction with some embodiments of the first aspect, in some embodiments, the step of training an ozone concentration inversion model based on a training sample set specifically includes: constructing a spatiotemporal dual-stream model architecture comprising a temporal sub-model and a spatial sub-model, wherein the temporal sub-model is used to capture the diurnal variation characteristics of ozone concentration, and the spatial sub-model is used to characterize the spatial distribution characteristics of ozone concentration; extracting total ozone column features and meteorological features from the training sample set, constructing time series features in chronological order and inputting them into the temporal sub-model; extracting spatiotemporal geographical features and socioeconomic features from the training sample set, constructing spatial features in conjunction with station spatial location information and inputting them into the spatial sub-model; training an ensemble model based on the output features of the temporal sub-model and the spatial sub-model; evaluating the performance of the ensemble model using cross-validation, and obtaining the ozone concentration inversion model through model parameter optimization.

[0013] In the above embodiments, the temporal sub-model captures the diurnal variation characteristics of ozone, the spatial sub-model represents the spatial distribution characteristics, and the dual-stream architecture achieves spatiotemporal decoupling. The total ozone column and meteorological features are extracted and input into the temporal sub-model in a temporal sequence, fully utilizing temporal evolution information. Spatiotemporal geographical features and socioeconomic features are extracted and combined with station spatial location information, inputting them into the spatial sub-model to capture spatial heterogeneity and anthropogenic influences. The ensemble model is trained based on the dual-stream output, fusing spatiotemporal multidimensional representations, cross-validating to evaluate performance, and optimizing parameters to ensure generalization ability. The synergistic effect of the dual-stream architecture, temporal modeling, spatial modeling, and feature fusion solves the problem that a single model cannot simultaneously characterize the complex spatiotemporal variations of ozone.

[0014] In conjunction with some embodiments of the first aspect, in some embodiments, after the step of performing ozone concentration inversion on the target study area according to the ozone concentration inversion model to obtain the target hourly ozone concentration data with a preset grid resolution within the target study area, the method further includes: spatiotemporally aligning the target hourly ozone concentration data with the measured data of the monitoring stations based on the monitoring station identifier, latitude and longitude, and timestamp to obtain an aligned dataset; calculating the concentration deviation between the target hourly ozone concentration data and the measured ozone concentration data of the monitoring stations in the aligned dataset to obtain station deviation data; generating a grid correction factor through spatial interpolation based on the station deviation data and incorporating factors such as topography and meteorology as covariates; correcting the target hourly ozone concentration data according to the grid correction factor to obtain corrected ozone concentration data; and outputting the corrected ozone concentration data.

[0015] In the above embodiments, the inverted data and measured data are spatiotemporally aligned based on station identifiers, latitude and longitude, and timestamps to ensure accurate matching. Concentration deviation is calculated to quantify systematic errors. Correction factors are generated through spatial interpolation using topographic and meteorological factors, extending station deviations to the entire domain. The inverted data is then corrected based on these correction factors to eliminate prediction biases. The synergistic effect of spatiotemporal alignment, deviation calculation, covariate interpolation, and correction resolves the problem of systematic biases in the model inversion results, improving data reliability.

[0016] In conjunction with some embodiments of the first aspect, in some embodiments, after the step of performing ozone concentration inversion on the target study area according to the ozone concentration inversion model to obtain the target hourly ozone concentration data with a preset grid resolution within the target study area, the method further includes: extracting high-value area data from the target hourly ozone concentration data, matching the high-value area data with wind direction and wind speed data in meteorological feature data to obtain matching results; determining the ozone pollution transmission path based on the matching results and spatiotemporal geographic feature data, and dividing regional transmission units based on the ozone pollution transmission path.

[0017] In the above embodiments, high-value area data is first extracted from the target data to accurately pinpoint areas of severe ozone pollution. Then, the high-value data is matched with wind direction and speed data from meteorological features. Wind direction and speed reflect the dynamic direction and speed of pollutant diffusion, and combined with spatiotemporal geographical features, the ozone pollution transmission path is identified. Finally, regional transmission units are divided based on the path. The synergistic effect of these features—high-value data focusing on key pollution areas, meteorological data providing the dynamic basis for transmission, and geographical features assisting in path determination—ultimately clearly identifies the pollution transmission patterns, providing a structured spatial division for accurate analysis of pollution diffusion mechanisms and enhancing the ability to control the ozone pollution transmission process.

[0018] In conjunction with some embodiments of the first aspect, in some embodiments, after the step of performing ozone concentration inversion on the target study area according to the ozone concentration inversion model to obtain target hourly ozone concentration data with a preset grid resolution in the target study area, the method further includes: collecting pollution source inventory data of the target study area, spatially allocating the pollution source inventory data according to the preset grid resolution; establishing a grid emission inventory based on the target hourly ozone concentration data and the pollution source inventory data, and determining the regional ozone pollution control target value according to the grid emission inventory.

[0019] In the above embodiments, pollution source inventory data of the target study area is first collected. This data can clearly identify the emission intensity and type of various pollution sources in the area, providing a source basis for subsequent correlation of ozone concentration. Then, the pollution source inventory data is spatially allocated according to a preset grid resolution. This preset grid is consistent with the grid scale of the previous unified processing of multi-source data, which can eliminate the difference in spatial scale of the data and ensure that the pollution source data and the target hourly ozone concentration data obtained by inversion correspond accurately in the spatial dimension. Subsequently, a grid emission inventory is established based on the two, so that the "emission-concentration" correlation of each grid is clearly presented. Finally, the regional ozone pollution control target value is determined according to the grid emission inventory. Because the inventory accurately reflects the emission and concentration correlation at the grid level, a general regional control target can be avoided.

[0020] Secondly, embodiments of this application provide a near-surface ozone concentration hourly inversion system, the near-surface ozone concentration hourly inversion system comprising: one or more processors and a memory; the memory is coupled to the one or more processors, the memory is used to store computer program code, the computer program code including computer instructions, the one or more processors calling the computer instructions to cause the near-surface ozone concentration hourly inversion system to perform the method as described in the first aspect and any possible implementation thereof.

[0021] Thirdly, embodiments of this application provide a computer program product containing instructions that, when the computer program product is run on a near-ground ozone concentration hourly inversion system, cause the near-ground ozone concentration hourly inversion system to perform the method described in the first aspect and any possible implementation thereof.

[0022] Fourthly, embodiments of this application provide a computer-readable storage medium including instructions that, when executed on a near-ground ozone concentration hourly inversion system, cause the near-ground ozone concentration hourly inversion system to perform the method described in the first aspect and any possible implementation thereof.

[0023] Understandably, the near-surface ozone concentration hourly inversion system provided in the second aspect, the computer program product provided in the third aspect, and the computer storage medium provided in the fourth aspect are all used to execute the method provided in the embodiments of this application. Therefore, the beneficial effects they can achieve can be referred to the beneficial effects in the corresponding methods, and will not be repeated here.

[0024] One or more technical solutions provided in the embodiments of this application have at least the following technical effects or advantages:

[0025] 1. This application eliminates inconsistencies in the temporal dimension by collecting four types of feature data and aligning them based on timestamps. The aligned data is then uniformly processed into hourly data at a preset grid resolution, resolving differences in spatial scale and temporal granularity. Spatial matching is used to associate the grid data with measured data from stations to construct a training sample set, establishing a mapping relationship between multi-source features and ozone concentration. The synergy of multi-source data compensates for the limitations of a single data source, gridding achieves continuous spatial coverage, and the hourly scale captures diurnal variation characteristics, solving the problems of insufficient inversion accuracy and incomplete spatiotemporal coverage caused by the inability of existing methods to fully integrate heterogeneous data.

[0026] 2. This application employs bilinear interpolation to upsample total ozone concentration, meteorological characteristics, and socioeconomic characteristics to a pre-defined grid, achieving a smooth numerical transition through distance-weighted calculation. For spatiotemporal geographic features, an area-weighted average method is used for downsampling, calculating weighted values ​​based on overlapping area ratios to preserve spatial statistical characteristics. Bilinear interpolation ensures the spatial continuity of the upsampled data, while the area-weighted method minimizes information loss during downsampling. A uniform grid resolution and hourly time scale ensure complete spatiotemporal alignment of all data sources, eliminating scale mismatch issues in multi-source data fusion.

[0027] 3. This application ensures precise spatial matching by extracting corresponding raster feature data based on the latitude and longitude of monitoring stations. A spatial mapping relationship is established by associating raster features with station identifiers and latitude / longitude to obtain the first matching data, achieving precise binding between station locations and raster data and establishing a unique index. The first matching data is then paired with measured concentrations to construct a training sample set, achieving precise mapping from features to labels. Spatial matching guarantees positioning accuracy, the dual association of identifiers and latitude / longitude ensures data traceability and location accuracy, and feature pairing provides high-quality supervisory signals, solving the problem of effectively associating multi-source data with ground observations. Attached Figure Description

[0028] Figure 1 This is a flowchart illustrating an hourly near-surface ozone retrieval method based on multi-source data fusion in this application embodiment;

[0029] Figure 2 This is another flowchart illustrating the hourly near-surface ozone retrieval method based on multi-source data fusion in this application embodiment;

[0030] Figure 3 This is a schematic diagram of the physical device structure of a near-ground ozone concentration hourly inversion system in this application embodiment. Detailed Implementation

[0031] The terminology used in the following embodiments of this application is for the purpose of describing particular embodiments only and is not intended to be limiting of this application. As used in the specification of this application, the singular expressions “a,” “an,” “the,” “the,” and “this” are intended to include the plural expressions as well, unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used in this application refers to any or all possible combinations including one or more of the listed items.

[0032] Hereinafter, the terms "first" and "second" are used for descriptive purposes only and should not be construed as implying or suggesting relative importance or implicitly indicating the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature, and in the description of the embodiments of this application, unless otherwise stated, "multiple" means two or more.

[0033] To facilitate understanding, the application scenarios of the embodiments of this application are described below.

[0034] A regional air quality management department needs to conduct real-time monitoring and health risk assessment of ozone pollution within a 300km x 300km area under its jurisdiction. This area has 23 ground-based air quality monitoring stations, but their distribution is extremely uneven – the distance between stations in densely populated urban areas is approximately 5-10km, while the distance between stations in rural areas reaches 50-100km. During a severe ozone pollution event in the summer of 2023, the management department found that relying solely on these discrete monitoring points was insufficient to accurately grasp the spatial distribution and transport patterns of pollution. Especially during the afternoon hours when sunlight is strong and temperatures are high, ozone concentrations often rise rapidly, but due to the lack of continuous observation data with high temporal resolution, it is impossible to promptly detect and issue warnings of high-value areas. This highlights the shortcomings of the existing ground-based monitoring network in terms of spatial coverage and temporal continuity, necessitating the development of new monitoring methods capable of providing high spatiotemporal resolution ozone concentration distribution.

[0035] To address the aforementioned issues, the region attempted to acquire ozone distribution data using traditional satellite remote sensing methods. They employed Tropomi satellite data for ozone concentration retrieval; this satellite passes overhead once daily at 13:30 with a spatial resolution of 7km × 7km. However, in practical application, they found that: 1) due to the single daily observation, the diurnal variation characteristics of ozone concentration could not be captured, especially missing the peak ozone period from 2-4 pm; 2) satellite data quality significantly decreased under cloudy weather conditions; 3) even under clear weather conditions, the correlation coefficient (R²) between the retrieval results from satellite data and ground-based measurements was only around 0.65. Furthermore, they also attempted a statistical model method using only meteorological data, but due to the lack of crucial information such as emission source intensity, the model was unstable and exhibited significant errors during pollution transmission. These attempts demonstrate that methods relying on a single data source are insufficient to meet the demands of high-precision ozone monitoring.

[0036] By employing the multi-source data fusion method of this invention, a complete ozone concentration retrieval system was established in this region. The system first integrates hourly observation data (3.5km × 7.7km) from the GK2B-GEMS satellite, hourly meteorological reanalysis data (0.25°) from ERA5, DEM elevation data (90m), and emission inventory data. All data are then unified into a 3km grid through spatial resampling. Each grid cell contains 18 characteristic parameters, covering multiple dimensions including satellite observations, meteorological conditions, geographical features, and anthropogenic emissions. By introducing a dual-stream architecture of temporal and spatial sub-models, the system can simultaneously characterize the spatiotemporal variation of ozone. Validation results show that the correlation coefficient (R²) between the hourly retrieval results and ground observations is improved to over 0.85, with spatiotemporal resolutions reaching 1 hour and 3km, respectively. Particularly during pollution transport, due to the integration of meteorological field and topographic factors, the system can accurately track the transport path of ozone plumes, providing reliable data support for cross-regional joint prevention and control.

[0037] To facilitate understanding, the method provided in this implementation will be described in detail below, using the above scenario as an example. Please refer to [link / reference]. Figure 1 This is a flowchart illustrating an hourly near-surface ozone retrieval method based on multi-source data fusion in this application embodiment.

[0038] S101. Collect ozone total column concentration characteristic data, meteorological characteristic data, spatiotemporal geographic characteristic data, and socioeconomic characteristic data.

[0039] Among them, the total ozone column concentration characteristic data represents the vertical integral value of the number of ozone molecules in the atmospheric column, used to characterize the overall ozone content in the atmosphere. Meteorological characteristic data refers to key meteorological elements affecting ozone formation and transport, including parameters such as temperature, humidity, wind speed, and air pressure. Spatiotemporal geographic characteristic data represents the geographic environmental information of the study area, including static features such as altitude, vegetation cover, and topographic relief, as well as dynamic features reflecting temporal evolution. Socioeconomic characteristic data is used to represent the intensity of human activities and the distribution of emission sources, including indicators such as population density and pollution emissions.

[0040] Specifically, this step is performed first before ozone concentration inversion to establish a multi-source data foundation. The system acquires data by calling the API interfaces or databases of various data sources. For GK2B-GEMS satellite data, the latest observation data needs to be retrieved from the satellite data center according to a preset time window (e.g., hourly); for ERA5 meteorological data, an automatic download script needs to be set up to retrieve updates periodically; for spatiotemporal geographic feature data, it is imported all at once during system initialization; for socioeconomic feature data, it is updated periodically according to its update frequency (e.g., annually or quarterly). All data undergoes basic quality checks during the acquisition process, including data integrity verification and outlier identification.

[0041] In some embodiments, data acquisition and preprocessing can be implemented in multiple ways: Optionally, a distributed crawler system can be used to crawl the latest data from various data sources in real time, store the raw data in a distributed file system, perform preliminary cleaning through a data quality control module, and finally save the cleaned data to a time-series database in a unified format; Optionally, a data integration platform can be built to connect with various data sources through standardized data interfaces to achieve automatic data acquisition and updates, while establishing a data quality assessment system to evaluate and filter the collected data in real time to ensure data reliability. It is understood that other methods can also be used to implement the data acquisition and preprocessing process, which are not limited here.

[0042] S102. Time-align the ozone total column concentration characteristic data, meteorological characteristic data, spatiotemporal geographic characteristic data, and socioeconomic characteristic data based on timestamps.

[0043] The timestamp represents the time identifier of the recorded data acquisition moment, in the format "year-month-day hour:minute:second". Time alignment refers to the process of matching different data sources according to a unified time benchmark. The total ozone column concentration characteristic data represents the atmospheric ozone content obtained from GK2B-GEMS satellite observations, with an hourly time resolution. Meteorological characteristic data refers to hourly records of meteorological elements such as temperature, humidity, wind speed, air pressure, and precipitation provided by ERA5 reanalysis data. Spatiotemporal geographic characteristic data represents geographic environmental information such as altitude and NDVI vegetation index, as well as time markers such as the number of days in a year and the hours in a day. Socioeconomic characteristic data is used to characterize indicators related to human activities, such as population density and pollutant emissions.

[0044] Specifically, this step is performed after data acquisition is complete, aiming to eliminate inconsistencies in the time dimension across different data sources. The system assigns a unified timestamp identifier to each sample, in the format "2023-12-01 09:00:00". For GK2B-GEMS satellite data, its hourly observation times are directly extracted as timestamps. For ERA5 meteorological data, its raw time field is read and converted to a standard timestamp format. For NDVI data, since it is a 16-day composite product, the values ​​for that period need to be allocated to each hour within that period. For annual population data, the annual values ​​are allocated to all hours of that year with the same value. The system establishes a time index in hourly units, mapping records from all data sources to a unified timeline, ensuring that satellite features, meteorological features, and spatiotemporal features correspond one-to-one under the same timestamp.

[0045] In some embodiments, time alignment can be achieved in several ways: Optionally, a time standardization processing module can be established. First, the original time fields of each data source are extracted. Hourly timestamps from satellite and ERA5 data are directly converted to a unified format. For 16-day synthesized NDVI data, a numerical replication method is used to extend the values ​​of a single time period to all hours within that period. For annual population and emissions data, a static assignment method is used to assign annual values ​​to all times within that year. Finally, a multi-source data record table containing unified timestamps is generated. Optionally, a time alignment verification process can be constructed. This involves performing correlation queries on each data source according to timestamps, checking whether each hour has complete multi-source characteristic data, filling in missing timestamps with the average of preceding and following times, marking and removing abnormal timestamps, and establishing a time integrity verification mechanism to ensure that the data processed subsequently is fully aligned in the time dimension. It is understood that other methods can also be used to achieve the time alignment process, which are not limited here.

[0046] S103. Process the time-aligned data into hourly data with a preset grid resolution.

[0047] The preset grid resolution represents the spatial precision of dividing the study area into regular grids; in this invention, it is set to 3 km × 3 km. Hourly data refers to a continuous data sequence with an hourly time interval. Spatial resampling refers to the process of converting data with different spatial resolutions into a uniform grid scale. Bilinear interpolation is an interpolation method that calculates the value of the target grid point by weighting the values ​​of four adjacent grid points. Area-weighted averaging is a downsampling method that calculates a weighted average based on the proportion of the overlapping area between the original data grid and the target grid.

[0048] Specifically, this step is performed after time alignment is complete, aiming to unify all data sources to a 3-kilometer grid and an hourly time scale. The system first determines the spatial extent of the study area and the 3-kilometer grid division scheme. For GK2B-GEMS satellite data, with an original spatial resolution of 3.5 km × 7.7 km, it is upsampled to a 3-kilometer grid using bilinear interpolation. The distance weights of the four nearest neighbor satellite grid points surrounding the target grid point are calculated, and the total ozone column concentration values ​​are weighted and averaged accordingly. For ERA5 meteorological data, with an original resolution of 0.25 degrees (approximately 27.75 km), it is also upsampled to a 3-kilometer grid using bilinear interpolation. For DEM elevation data, with an original resolution of 90 meters, it is downsampled using an area-weighted average method. The number of 90-meter grids within each 3-kilometer grid and their area proportion are calculated, and the elevation values ​​are weighted and averaged according to the area proportion. For 1-kilometer resolution population data and NDVI data, the area-weighted average method is also used for downsampling to 3 kilometers. For 27-kilometer resolution emission inventory data, it is upsampled to a 3-kilometer grid using bilinear interpolation.

[0049] In some embodiments, unified data processing can be achieved in several ways: Optionally, a spatial resampling process can be established. First, a 3-kilometer regular grid matrix is ​​generated based on the boundary of the study area. For low-resolution data, a bilinear interpolation algorithm is used to calculate the target point value grid by grid. Specifically, the coordinates and values ​​of the four nearest neighbor original grid points around the target point are extracted. The interpolation weights are calculated based on the distance from the target point to the four neighboring points. The target grid value is obtained by summing the values ​​according to the weights. For high-resolution data, an area-weighted method is used to calculate the overlap area between the target grid and all overlapping original grids. The original grid values ​​are weighted and summed according to the area ratio. Finally, all data layers are aligned to a unified 3-kilometer grid framework. Optionally, a classification processing strategy can be adopted to divide the data into high-resolution and low-resolution groups according to the original resolution. For the high-resolution group, a spatial aggregation method is used to calculate the average or sum of all original grids within the target grid. For the low-resolution group, a distance-inverse weighted interpolation method is used to calculate the target grid value. The value rationality of the resampling results is checked by a quality control module, and outliers that exceed the normal range are removed to ensure that all data are unified to a 3-kilometer hourly scale in both spatial and temporal dimensions. It is understandable that other methods can be used to achieve the unified data processing process, and no specific method is specified here.

[0050] In some embodiments, this step specifically includes: determining a preset grid resolution based on the spatial range of the ozone total column concentration characteristic data for the target study area; upsampling the ozone total column concentration characteristic data to the preset grid resolution using bilinear interpolation; upsampling meteorological characteristic data and socioeconomic characteristic data to the preset grid resolution using bilinear interpolation; performing scale matching processing on the spatiotemporal geographic characteristic data based on the area-weighted average method; and converting the time-aligned data into hourly data.

[0051] The target study area refers to the geographical range within which near-surface ozone concentration inversion needs to be performed; in this invention, the entire territory of China is used as the target study area. The preset grid resolution refers to the spatial precision of dividing the study area into regular grids; in this invention, it is set to 3 km × 3 km. Bilinear interpolation is a spatial interpolation method that calculates the target point's value by weighting the values ​​of the four nearest neighbor grid points around the point to be interpolated. Upsampling refers to the process of converting lower spatial resolution data into higher resolution data. The area-weighted average method is a downsampling method that calculates a weighted average based on the overlap area ratio between the original data grid and the target grid. Hourly data represents a continuous data sequence with an hourly time interval.

[0052] This step is performed after the time alignment of multi-source data is completed. Its purpose is to unify all data sources to a 3-kilometer grid and hourly scale in both spatial and temporal dimensions. First, a 3-kilometer grid resolution is determined as a unified standard within the coverage area of ​​the GK2B-GEMS ozone total column concentration data, based on the entire border of China. For GK2B-GEMS satellite data, the original resolution is 3.5 km × 7.7 km. It is upsampled to a 3-kilometer grid using bilinear interpolation. Specifically, for any target's 3-kilometer grid center point, the ozone total column concentration values ​​of its four nearest-neighbor original satellite grid points are extracted and denoted as V1, V2, V3, and V4, respectively. The Euclidean distances from the target point to the four neighbors are calculated and denoted as d1, d2, d3, and d4, respectively. The interpolation weights are calculated inversely proportional to these distances. The parameters w1=1 / d1, w2=1 / d2, w3=1 / d3, and w4=1 / d4 are used, with normalized weights W1=w1 / (w1+w2+w3+w4), W2=w2 / (w1+w2+w3+w4), W3=w3 / (w1+w2+w3+w4), and W4=w4 / (w1+w2+w3+w4). The total ozone column concentration value of the target grid is calculated as V=W1×V1+W2×V2+W3×V3+W4×V4. For ERA5 meteorological data, the original resolution is 0.25 degrees, approximately 27.75 kilometers. Bilinear interpolation is also used to upsample meteorological elements such as temperature, humidity, wind speed, air pressure, and precipitation to a 3-kilometer grid, with the interpolation calculation method being the same as for satellite data. For ABaCAS-EI emission inventory data, the original resolution is 27 km. Bilinear interpolation is used to upsample NOx, VOC, and SO2 emissions to a 3 km grid. For DEM elevation data, the original resolution is 90 m. An area-weighted average method is used to downsample to 3 km. Specifically, the boundary of each 3 km target grid is determined, and all 90 m original grid cells within that grid are counted. The overlap area between each 90 m grid and the 3 km grid is calculated. Let Hi be the elevation of the i-th 90 m grid, and Si be its overlap area with the target grid. The elevation value of the target grid is calculated as H = (ΣHi × Si) / ΣSi. For 1 km resolution population density and NDVI data, the area-weighted average method is also used for downsampling. The area ratio of all 1 km grid cells within each 3 km grid is calculated, and the population density and NDVI values ​​are weighted and averaged according to the area ratio. For the time dimension, the time resolution of all data sources is checked. GK2B-GEMS and ERA5 data are already hourly. For 16-day synthesized NDVI data, the values ​​of that period are copied and distributed to each hour within the period. For annual population and emissions data, the annual values ​​are distributed to all hourly moments of that year with the same values. Finally, all data sources are converted to hourly series in the time dimension.

[0053] S104. Based on the monitoring station identifier and latitude and longitude, perform spatial matching on the hourly data of the preset grid resolution to obtain the first matching data, and construct a training sample set based on the first matching data and the measured data of the monitoring station.

[0054] Here, the monitoring station identifier represents the unique number of each ground monitoring station in the China National Environmental Monitoring Centre platform. Latitude and longitude refer to the geographical coordinates of the monitoring station. Spatial matching refers to the process of determining the location of a station within a 3-kilometer grid based on its coordinates. The first matching data represents the association results between the monitoring station and the multi-source feature data of its grid. The measured data from the monitoring station refers to the actual hourly ozone concentration observations recorded by ground monitoring equipment. The training sample set represents the set of paired data between input features and output labels used for training the machine learning model.

[0055] Specifically, this step is performed after data gridding is completed, with the aim of establishing a training mapping relationship between grid features and ground observations. The system first reads the latitude and longitude coordinates and unique identifiers of the monitoring stations, and uses a coordinate transformation algorithm to determine the location of each station within a 3-kilometer grid cell. For each station, all feature data of its grid at a specific time are extracted, including the GK2B-GEMS total ozone column concentration, ERA5 meteorological parameters, DEM altitude, NDVI vegetation index, population density, emissions, and temporal characteristics. This feature data is associated with the monitoring station identifier to form the first matching data. Then, based on the timestamp, the feature vectors in the first matching data are paired with the measured ozone concentration value of the station at the same time. For example, if the measured concentration at a station at 14:00 on August 1, 2023, is 120 micrograms per cubic meter, then all feature values ​​of the grid where that station is located at that time are extracted as input, and 120 micrograms per cubic meter is used as the output label, forming a training sample. Samples from all stations at all times are organized in this way to form a complete training sample set.

[0056] In some embodiments, spatial matching and sample construction can be implemented in several ways: Optionally, a spatial indexing module is established. First, the coordinates of the monitoring stations are transformed to the same projected coordinate system as the grid. The row and column indices of the grid corresponding to the station coordinates are calculated by integer division. All feature data of the grid are directly extracted based on the grid index. The feature data is bound with the station identifier and timestamp to form the first matching data. Then, the ozone concentration observation values ​​of the corresponding station and timestamp are queried from the station's measured database. The feature vectors and observation values ​​are paired to form a single sample record. The sample records of all stations are merged in chronological order to generate a training dataset. Optionally, a batch processing strategy is adopted. The coordinate information of all stations is input into the spatial matching algorithm in batches. The grid positions of all stations are calculated at once. The feature data matrix of all grids in the whole time period is extracted in batches. An index table is established through the station identifier and timestamp. The feature data and measured concentration values ​​are quickly associated. The times with missing measured data are marked and removed. Finally, a complete training sample set without missing values ​​is generated. It is divided into a training set and a validation set in chronological order or randomly. It is understood that other methods can also be used to implement the spatial matching and sample construction process, which are not limited here.

[0057] In some embodiments, this step specifically includes: based on a preset grid resolution, extracting raster feature data corresponding to each monitoring station from time-aligned multi-source hourly grid data according to the latitude and longitude of each monitoring station; the raster feature data includes at least ozone total column features, meteorological features, spatiotemporal geographic features, and socioeconomic features; associating the raster feature data with the corresponding monitoring station identifier and latitude and longitude to establish a spatial mapping relationship between the station location and the raster data, obtaining first matching data; and pairing the first matching data with the corresponding measured ozone concentration data of the monitoring station according to the monitoring station identifier to construct a training sample set.

[0058] The preset grid resolution represents the spatial division accuracy of the study area, set to 3 km × 3 km. Multi-source hourly grid data refers to the set of satellite, meteorological, geographic, and socioeconomic data after time alignment and spatial unification processing. The latitude and longitude location of the monitoring station is its geographic coordinate information. Raster feature data represents the multi-dimensional feature vector corresponding to the grid where the station is located, including ozone total column features, meteorological features, spatiotemporal geographic features, and socioeconomic features. The monitoring station identifier is a unique code for each monitoring station. The spatial mapping relationship represents the precise correspondence between station locations (including identifiers and latitude and longitude) and raster data. The first matching data represents the association result between raster features and station identifiers and latitude and longitude. The training sample set represents a set of paired data containing input features and output labels.

[0059] This step is performed after the time alignment and spatial unification of multi-source data are completed. Its purpose is to establish the correspondence between grid features and ground observations. First, the latitude and longitude coordinates of the monitoring stations are read, and their 3-kilometer grid locations are determined using a coordinate transformation algorithm. The coordinate transformation employs a nearest neighbor matching method, calculating the grid row and column indices corresponding to the station coordinates (lon, lat). The row index i = floor((lat - lat_min) / 0.027), and the column index j = floor((lon - lon_min) / 0.027), where 0.027 degrees is approximately equal to 3 kilometers. After determining the grid location, all feature data for that grid at a specific time are extracted from the multi-source grid data. For example, a certain station (identified as ST0123, with latitude and longitude of 120.15°E, 30.28°N) is located in the 1245th row and 876th column of the grid. Features of this grid at 14:00 on August 1, 2023 are extracted, including 18 parameters such as total ozone concentration (350 DU), temperature (305.2 K), humidity (45%), wind speed (3.2 m / s), altitude (50 m), NDVI value (0.42), population density (8500 people / km²), and NOx emissions (12 tons / year). These 18 feature values ​​are combined into a vector and bound to the station identification code and latitude and longitude information to establish a spatial mapping relationship between the station location and the raster data, forming the first matching data. Then, based on the station identification, latitude and longitude information, and timestamp, the measured ozone concentration value of this station at the same moment is queried from the station database. The feature vector is used as input X, and the measured value is used as output Y to form the training sample (X, Y). Data from all stations at all times is processed in this manner. After removing missing or abnormal records, the valid samples are divided into training, validation, and test sets in a 7:2:1 ratio to finally construct a complete training sample set.

[0060] S105. Train the ozone concentration inversion model based on the training sample set.

[0061] The training process refers to the iterative calculation process of optimizing model parameters using sample data. The ozone concentration inversion model represents a mathematical model that, after training, can convert multi-source feature data into near-surface ozone concentration. Model training refers to the process of optimizing model parameters by minimizing the error between predicted and measured values.

[0062] Specifically, this step is performed after the training sample set is constructed, and its purpose is to establish a mapping relationship between feature data and ozone concentration. The system first needs to construct a two-stream model architecture, comprising a temporal sub-model and a spatial sub-model. The temporal sub-model specifically handles temporal features, capturing the diurnal variation patterns of ozone concentration; the spatial sub-model is responsible for extracting spatial distribution features, characterizing the spatial diffusion patterns of pollutants. During training, the system evaluates model performance using cross-validation and optimizes hyperparameters using techniques such as grid search. It is particularly important to note that model training also needs to consider feature importance analysis to understand the contribution of different data sources to the prediction results.

[0063] In some embodiments, model training can be implemented in several ways: Optionally, an adaptive learning rate adjustment strategy can be adopted, firstly ranking the importance of features, constructing a two-stream model architecture based on LightGBM, preventing overfitting through an early stopping strategy, introducing a regularization term to control model complexity, and finally improving prediction stability through model ensemble; Optionally, a deep learning framework can be established, using recurrent neural networks to process temporal features, graph convolutional networks to extract spatial features, achieving feature fusion through an attention mechanism, and employing transfer learning methods to improve the model's generalization ability in different regions. It is understood that other methods can also be used to implement the model training process, which are not limited here.

[0064] In some embodiments, this step specifically includes:

[0065] A two-stream model architecture is constructed, which includes a temporal sub-model and a spatial sub-model. The temporal sub-model is used to capture the diurnal variation characteristics of ozone concentration, and the spatial sub-model is used to characterize the spatial distribution characteristics of ozone concentration.

[0066] The dual-stream model architecture refers to a model structure that includes two parallel processing branches. The temporal sub-model is the model component that processes time-series features. The spatial sub-model is the model component that processes spatial distribution features. Diurnal variation features represent the periodic variation pattern of ozone concentration over 24 hours. Spatial distribution features refer to the spatial variation pattern of ozone concentration within the study area.

[0067] The construction process of the two-stream model architecture is as follows: The temporal sub-model adopts a Long Short-Term Memory (LSTM) network structure, containing one input layer, two LSTM layers, and one fully connected layer. The input layer receives the feature sequence within a sliding time window. The first LSTM layer contains 64 neurons, and the second LSTM layer contains 32 neurons. The fully connected layer outputs an 8-dimensional vector representing the diurnal variation features. The spatial sub-model adopts a Graph Convolutional Network (GCN) structure, containing two graph convolutional layers and one pooling layer. The graph convolutional layers contain 32 and 16 convolutional kernels, respectively, used to extract spatial adjacency features. The pooling layer uses max pooling to compress the feature map into a fixed-dimensional vector. The output features of the two sub-models are connected through a feature fusion layer to form the complete two-stream architecture.

[0068] Ozone total column features and meteorological features are extracted from the training sample set, time series features are constructed in chronological order and input into the time series sub-model, spatiotemporal geographical features and socioeconomic features are extracted from the training sample set, and spatial features are constructed by combining station spatial location information and input into the spatial sub-model.

[0069] The ozone column characteristics refer to the ozone concentration values ​​in the atmospheric column observed by the GK2B-GEMS satellite. Meteorological characteristics include ERA5 reanalysis data parameters such as temperature, humidity, wind speed, and air pressure. Time series characteristics represent continuous observation data arranged in chronological order. The time series sub-model is a model component using a Long Short-Term Memory (LSTM) network structure to capture the temporal evolution of ozone concentration. Spatiotemporal geographic characteristics include altitude, vegetation cover (NDVI), topographic relief, and time markers (month, hour, etc.). Socioeconomic characteristics include population density and pollutant emissions (NOx, VOCs, etc.). Station spatial location information includes the latitude and longitude coordinates of the monitoring station. Spatial characteristics represent feature vectors containing spatially heterogeneous information such as geographical location, topography, and human activities. The spatial sub-model is a model component using a fully connected neural network structure to characterize the spatial distribution characteristics of ozone concentration.

[0070] This step is performed after the training sample set is constructed, and its purpose is to prepare input data for the dual-stream model. First, the total ozone column features and meteorological features of each sample are extracted from the training sample set. The total ozone column features include the ozone concentration value for the current hour, and the meteorological features include parameters such as temperature, humidity, wind speed, air pressure, and precipitation for the current hour. Then, these features are sorted according to timestamps to construct a sliding time window. The time window length is set to 24 hours. For any time t, the total ozone column and meteorological feature data for the 24 hours preceding that time are extracted. For example, for the sample at 14:00 on August 1, 2023, feature values ​​for 24 times from 13:00 and 12:00 on August 1 to 14:00 the previous day are extracted. Each time point contains one total ozone column value and five meteorological parameters, forming a 24×6 feature matrix. The matrix is ​​used as the input layer of the time series sub-model. The LSTM network captures the changing trends of ozone concentration and meteorological conditions over the past 24 hours through memory units. The 64 neurons in the first LSTM layer extract the short-term time dependence, and the 32 neurons in the second LSTM layer extract the long-term time dependence. Finally, an 8-dimensional feature vector is output to represent the daily variation pattern of ozone concentration.

[0071] The spatiotemporal geographic features and socioeconomic features of each sample are extracted from the training sample set. Spatiotemporal geographic features include the elevation of the grid where the station is located (e.g., 50 meters), the NDVI value of vegetation cover (e.g., 0.42), the terrain relief (e.g., plain / hill level coding), and the time identifier (e.g., month code 8 for August, hour code 14 for 2 PM). Socioeconomic features include the population density of the grid where the station is located (e.g., 8500 people / km²) and pollutant emissions (e.g., NOx emissions 12 tons / year, VOCs emissions 8 tons / year). These features are combined with the station's latitude and longitude coordinates (e.g., 120.15°E, 30.28°N) to construct a spatial feature vector. For example, for a sample from a certain site at 14:00 on August 1, 2023, the spatial feature vector contains 10 parameters: longitude 120.15, latitude 30.28, altitude 50, NDVI 0.42, topographic code 2, month 8, hour 14, population density 8500, NOx emissions 12, and VOCs emissions 8. This vector is input into the input layer of the spatial sub-model. The first layer of the fully connected neural network (128 neurons) extracts basic spatial features, the second layer (64 neurons) extracts abstract spatial patterns, and the third layer (32 neurons) performs feature compression, ultimately outputting an 8-dimensional feature vector representing the spatial distribution characteristics of the site. The 8-dimensional temporal feature vector output from the temporal sub-model is concatenated with the 8-dimensional spatial feature vector output from the spatial sub-model to form a 16-dimensional fused feature vector, which is then input into the ensemble model for final ozone concentration prediction.

[0072] The ensemble model is trained based on the output features of the temporal and spatial sub-models.

[0073] An ensemble model is an overall model that combines the outputs of multiple sub-models. Output features represent the feature vectors obtained after processing by the sub-models. Model training refers to the process of optimizing model parameters using data.

[0074] The ensemble model was trained as follows: First, the 8-dimensional diurnal variation feature vector output by the time-series sub-model and the 16-dimensional spatial feature vector output by the spatial sub-model were concatenated to obtain a 24-dimensional combined feature vector. Then, a three-layer feedforward neural network was constructed as the ensemble model. The first layer contained 128 neurons, the second layer contained 64 neurons, and the output layer corresponded to the prediction target. The Adam optimizer was used for parameter updates, with the learning rate initially set to 0.001 and dynamically adjusted using a cosine annealing strategy. The root mean square error (RMSE) was used as the loss function, and the model parameters were optimized through backpropagation.

[0075] The performance of the ensemble model was evaluated using cross-validation, and the ozone concentration inversion model was obtained by optimizing the model parameters.

[0076] Cross-validation refers to the method of evaluating model performance by repeatedly splitting the training and validation sets. Model parameter optimization refers to the process of adjusting model parameters to improve performance. The ozone concentration inversion model is the final model used for prediction after training.

[0077] The model evaluation and optimization process is as follows: A 5-fold cross-validation method is used, randomly dividing the dataset into 5 parts. Four parts are used alternately as the training set, and one part as the validation set. RMSE and correlation coefficient (R) are calculated for each fold of validation. Then, Bayesian optimization is used to tune the model's hyperparameters, including learning rate, batch size, and number of neurons. The optimization objective is to minimize the RMSE on the validation set. After 100 rounds of iterative optimization, the parameter configuration with the best performance is selected as the final model. The entire evaluation and optimization process ensures the model's generalization ability and prediction accuracy.

[0078] In some embodiments, the model training and inversion process is as follows:

[0079] (1) Feature integration: The merged dataset is used as the input of the model, including features from multiple data sources such as GEMS original features and current hour ERA5 features.

[0080] (2) Model selection: Select a suitable machine learning model for processing high-dimensional, nonlinear data, such as gradient booster (LightGBM or XGBoost) or neural network, and train the inversion model.

[0081] (3) Model output: The trained model can invert the near-ground ozone concentration corresponding to the station based on the input satellite and meteorological characteristics.

[0082] The model evaluation metrics, namely the coefficient of determination (R²) and the root mean square error (RMSE), are used to assess the inversion accuracy, as shown in the following formulas:

[0083]

[0084]

[0085] S106. Based on the ozone concentration inversion model, perform ozone concentration inversion on the target study area to obtain hourly ozone concentration data of the target study area with a preset grid resolution.

[0086] The target study area refers to the specific geographical region from which ozone concentration inversion needs to be performed. The target hourly ozone concentration data refers to the gridded spatiotemporal distribution data of ozone concentration predicted by the model. Ozone concentration inversion refers to the process of converting input features into ozone concentration using a trained model. The data with a preset grid resolution represents a dataset organized according to a uniform spatial scale.

[0087] Specifically, this step is performed after model training is complete, and its purpose is to obtain high-resolution ozone concentration distribution in the study area. The system first needs to prepare complete input feature data for the target area, including all necessary features such as satellite observations, meteorological conditions, and geographic information. During the inversion process, the system processes the data of each grid cell individually, while considering the spatial correlation between grids. Crucially, the system also needs to evaluate the reliability of the inversion results in real time; for areas with high uncertainty, additional correction mechanisms may be required. The inversion results also need to be compared and analyzed with historical data to ensure temporal continuity and spatial consistency.

[0088] In some embodiments, ozone concentration inversion can be achieved in several ways: Optionally, a distributed computing framework can be used to divide the study area into multiple sub-regions for parallel processing, calculate the prediction confidence interval for each grid cell, correct the initial prediction results based on the topographic-meteorological coupling factor, and finally fill any possible data gaps using spatial interpolation methods; Optionally, an integrated inversion system can be established to run multiple models with different structures simultaneously for prediction, fuse the results of each model using a Bayesian model averaging method, and correct abnormal prediction values ​​using an expert knowledge system. It is understood that other methods can also be used to achieve the ozone concentration inversion process, which are not limited here.

[0089] The following provides a more detailed description of the process of the method provided in this implementation. Please refer to [link / reference]. Figure 2 This is another flowchart illustrating the hourly near-surface ozone retrieval method based on multi-source data fusion in this application embodiment.

[0090] S201. Based on the ozone concentration inversion model, ozone concentration inversion is performed on the target study area to obtain hourly ozone concentration data of the target study area with a preset grid resolution.

[0091] The ozone concentration inversion model is a mathematical model trained on a gradient booster machine, used to map multi-source features to ozone concentration values. The target study area refers to a well-defined geographic spatial range, typically based on administrative divisions or natural geographic units. The preset grid resolution divides the study area into uniformly sized grid cells, such as 3km × 3km. The target hourly ozone concentration data refers to a complete gridded concentration field data for each hour output by the model.

[0092] In practice, the inversion process first divides the study area into regular grids according to a preset resolution. For each grid cell, corresponding feature values ​​are extracted from different data sources, including satellite observations, meteorological parameters, geographical parameters, and socioeconomic indicators for that grid. These feature data are then processed according to the format determined during training and input into the inversion model. The model calculates for each grid cell one by one, outputting the predicted ozone concentration for that grid cell at a specific time. For a 100km × 100km study area, if a 3km grid resolution is used, approximately 1100 grid cells need to be calculated separately. During the calculation process, all input features are standardized to the [-1, 1] interval. After obtaining the concentration prediction value through the model, the result is then converted back to the actual concentration unit (μg / m³).

[0093] S202. Based on the monitoring station identifier, latitude and longitude, and timestamp, the hourly ozone concentration data of the target is spatiotemporally aligned with the measured data of the monitoring station to obtain the aligned dataset.

[0094] The monitoring station identifier is a unique code for each ground-based air quality monitoring station. Latitude and longitude refer to the geographical coordinates of the monitoring station. Spatiotemporal alignment refers to the process of establishing a one-to-one correspondence between data from different sources in time and space. The aligned dataset is a standardized collection of data after spatiotemporal matching has been completed.

[0095] The spatiotemporal alignment process first pairs the station's measured data and inversion results according to a unified timestamp in the time dimension, ensuring that the time identifier of each data record is completely consistent. For example, for the measured value of a monitoring station at 14:00 on August 1, 2023, the inversion data for that moment is extracted. After time alignment, spatial matching is performed. A spatial index is established, converting the latitude and longitude coordinates of the monitoring stations into a grid index. For each station, its grid cell is determined by its latitude and longitude coordinates, and the inversion results of that grid and its eight adjacent grids are extracted to form a 3×3 spatial window. This process generates a structured dataset containing the station's measured values, the corresponding grid's inversion values, and spatiotemporal identifiers.

[0096] S203. Calculate the concentration deviation between the hourly ozone concentration data of the central target in the aligned dataset and the measured ozone concentration data of the monitoring station to obtain the station deviation data.

[0097] Concentration bias refers to the numerical difference between the retrieved value and the measured value. Site bias data represents the prediction error distribution for each grid cell. Measured ozone concentration data from monitoring sites are the actual ozone concentration observations recorded by the ground monitoring stations.

[0098] The deviation calculation process is as follows: For each monitoring station, the difference between its measured value and the corresponding grid inversion value is calculated. The specific calculation formula is: Deviation value = Inversion value - Measured value. For each time moment, the deviation values ​​of all stations are summarized to generate a complete grid deviation field. For example, if the measured value of a station at 14:00 is 120 μg / m³, and the corresponding grid inversion value is 135 μg / m³, then the deviation of that grid at this time is +15 μg / m³. These deviation values ​​will be used for subsequent model calibration. By analyzing the spatiotemporal distribution characteristics of the deviation, systematic errors in model predictions are identified. For grids without monitoring stations, their deviation values ​​are extrapolated from known deviation values ​​in the surrounding area using spatial interpolation methods.

[0099] S204. Based on the station deviation data, and combined with factors such as terrain and meteorology as covariates, a grid correction factor is generated through spatial interpolation. The hourly ozone concentration data of the target is corrected according to the grid correction factor to obtain the corrected ozone concentration data.

[0100] The grid correction factor represents the correction coefficient used to correct the original inversion results. The topographic-meteorological coupling factor is a comprehensive indicator reflecting the impact of the interaction between topographic features and meteorological conditions on pollutant dispersion. The corrected ozone concentration data refers to the more accurate gridded concentration field after correction.

[0101] The correction process first calculates the topographic-meteorological coupling factor for each grid, which is obtained by comprehensively calculating parameters such as topographic relief, wind speed, and atmospheric stability. The topographic complexity index (TCI) is calculated using a differentiated method based on the terrain type. For mountainous terrain, the TCI formula is TCI=(Hmax-Hmin) / (L×S), where Hmax and Hmin are the highest and lowest elevations within the grid, L is the grid side length, and S is the slope coefficient ranging from 0.5 to 2.0. For plains, the TCI formula is TCI=σh / H̄, where σh is the standard deviation of elevation within the grid, and H̄ is the average elevation. For basins, the TCI formula is TCI=(Hb-H0) / R, where Hb is the average elevation of the basin edge, H0 is the elevation of the basin bottom, and R is the basin radius. The meteorological condition parameter weights include a wind speed weight coefficient Wv = exp(-v / v0), where v is the actual wind speed and v0 is a reference wind speed of 3 m / s; an atmospheric stability weight Ws = (1 + Si) / 2, where Si is the normalized atmospheric stability index; and a mixing layer height weight Wh = h0 / h, where h is the actual mixing layer height and h0 is a reference height of 1000 m. The topographic-meteorological coupling factor is calculated using the formula F = TCI × (Wv × Ws × Wh), and is adjusted according to the season: 1.2 for summer, 0.8 for winter, and 1.0 for spring and autumn. Then, the grid bias values ​​are weighted and combined with the coupling factor to generate a correction factor. The calculation of the correction factor considers the magnitude, duration, and spatial correlation of the bias. For example, the correction weight is increased for areas with complex topography and atmospheric stability. Finally, the original inversion value is multiplied by the correction factor to obtain the corrected concentration value. For a region with 100×100 grids, the entire correction process requires processing data from 10,000 grid points simultaneously to ensure the spatial continuity and physical rationality of the correction results.

[0102] S205. Output the corrected ozone concentration data.

[0103] This step is performed after the ozone concentration data correction is completed. Its purpose is to standardize the output of the correction results for subsequent use. First, the corrected concentration data undergoes a quality check to verify that the numerical range is within a reasonable range. Ozone concentration values ​​should be between 0 and 500 μg / m³, and outliers outside this range are marked. Then, the output results are organized according to a standard data format. Each record includes fields such as grid row and column index, latitude and longitude coordinates, timestamp, corrected ozone concentration value, original concentration value, and correction factor value. For example, the output record for a certain grid at 14:00 on August 1, 2023 is: Row index 1245, Column index 876, Longitude 116.4, Latitude 39.9, Time 2023-08-01 14:00:00, Corrected concentration 125 μg / m³, Original concentration 135 μg / m³, Correction factor 0.926. All grid correction data is stored in NetCDF format in chronological order, with the file containing three dimensions: longitude, latitude, and time, and ozone concentration as the data variable. Simultaneously, visualization results are generated, plotting the corrected concentration data as a spatial distribution map, using color scales to represent concentration levels, and outputting as PNG or GeoTIFF image files for use by air quality management departments and the public.

[0104] S206. Extract high-value area data from the hourly ozone concentration data of the target, and match the high-value area data with the wind direction and wind speed data in the meteorological characteristic data to obtain the matching results.

[0105] High-value area data refers to a gridded dataset where ozone concentrations exceed a specific threshold. Wind direction and speed data are meteorological elements describing atmospheric motion. Matching results represent the associated datasets of high-concentration areas and their corresponding meteorological conditions.

[0106] The extraction process first establishes high-value criteria, typically selecting areas where the daily maximum 8-hour moving average exceeds the national air quality standard (160 μg / m³). For the concentration field at each time point, grid cells exceeding the threshold and their clustered areas are marked. Then, meteorological field data corresponding to these high-value grids are extracted, including 10-meter wind direction and wind speed. By calculating the divergence and curl of the wind field, pollutant transport channels are analyzed. For example, in a high-value area, if a persistent northeasterly wind with a speed of 3-5 m / s is observed, it indicates potentially significant pollutant transport. These matching results are used to identify ozone pollution transport pathways and causal analysis. For large-scale pollution events, the spatiotemporal evolution characteristics of hundreds of high-value grids need to be processed simultaneously.

[0107] S207. Determine the ozone pollution transmission path based on the matching results and spatiotemporal geographic feature data, and divide the regional transmission units based on the ozone pollution transmission path.

[0108] Ozone pollution transport pathways describe the spatial migration and diffusion trajectories of ozone and its precursors. Spatiotemporal geographic feature data includes information on factors influencing pollutant transport, such as topography and land use. Regional transport units refer to a set of regions with similar pollution transport characteristics.

[0109] The transmission path identification process is based on the Lagrange trajectory model, combined with the spatiotemporal evolution characteristics of high-concentration areas and meteorological field data. First, a three-dimensional spatial grid is established for the study area. Tracer particles are deployed in each high-concentration grid to track their trajectories under the influence of wind. The calculation process considers topographic blocking effects and turbulent diffusion, with a trajectory calculation step size of 1 hour. Based on cluster analysis of a large number of trajectory lines, the main transmission channels are identified. For example, in areas with complex terrain, pollutants are often transported along valley corridors. After determining the transmission paths, the study area is divided into several transmission units based on the similarity of transmission characteristics. The division comprehensively considers topographic units, administrative boundaries, and pollution transmission patterns to ensure that each transmission unit has relatively consistent pollution transmission characteristics.

[0110] Transport path identification employs a clustering-based approach. First, the time series of high-concentration events are standardized to extract meteorological fields at key time points. Then, the HDBSCAN algorithm is used for trajectory clustering, with the density threshold set to twice the standard deviation of the local concentration field and the minimum number of trajectory points set to 12. Seasonal analysis of transport channels includes identifying photochemical pollution transport paths and calculating diurnal variations in transport flux during summer, as well as analyzing the impact of inversion layers and determining pollutant accumulation mechanisms during winter. Regional transport unit division is based on topographic units (watershed analysis of DEM data) and administrative units (management boundaries). Then, the inter-unit transport flux matrix is ​​calculated, and spectral clustering is applied to determine the final transport units. Cross-regional transport flux calculation employs two methods: the interface flux method (setting transport cross-sections to calculate net flux) and the mass balance method (establishing regional material budget equations to solve for transport contribution rates).

[0111] S208. Collect pollution source inventory data for the target study area and spatially allocate the pollution source inventory data according to the preset grid resolution.

[0112] Pollution source inventory data is a dataset that records information on atmospheric pollutant emission sources. Spatial allocation refers to the process of distributing emission source data to a predefined grid. The predefined grid resolution represents the standardized size of the spatial computing unit.

[0113] The spatial allocation process employs a top-down approach, distributing emissions from various pollution sources to grid cells. For point sources (such as industrial enterprises), their grid assignment is directly determined by their geographic coordinates; for area sources (such as agricultural activities), allocation is based on land use type and activity level; and for mobile sources (such as motor vehicles), allocation is based on road network density and traffic flow. The allocation algorithm considers a time allocation factor, breaking down annual emissions into hourly data. For example, for a 3km × 3km grid with an annual NOx emission of 100 tons, the hourly emission contribution can be obtained by considering industrial activity patterns. This ultimately forms a gridded emissions inventory database containing various pollutants.

[0114] S209. Establish a grid emission inventory based on the target hourly ozone concentration data and pollution source inventory data, and determine the regional ozone pollution control target value based on the grid emission inventory.

[0115] A gridded emissions inventory is a spatialized dataset of pollutant emissions. Pollution control targets refer to the pollutant emission reduction targets set to improve air quality. Target hourly ozone concentration data represents a gridded ozone concentration field.

[0116] The target value determination process first establishes the response relationship between ozone concentration and precursor emissions. Ozone concentration changes under different emission reduction scenarios are simulated using a chemical transport model to establish an emission-concentration sensitivity matrix. Based on this matrix, and combined with the grid-based emission inventory, the emission reductions required to achieve the air quality improvement targets are calculated. The calculation considers the synergistic control effects of NOx and VOCs to avoid the negative effects caused by reducing emissions of a single pollutant. For example, if the ozone concentration in a certain grid exceeds the standard by 20%, sensitivity analysis determines that NOx emissions need to be reduced by 30% and VOCs emissions by 25% simultaneously. The refined control targets based on the grid scale provide a scientific basis for the formulation of differentiated emission reduction schemes. Finally, a list of emission reduction tasks by region and industry is formed.

[0117] The emission source time allocation coefficient is set with specific time patterns for different source types. The coefficient for industrial sources is 1.2 during 6:00-18:00, 0.8 during other times, 0.7 on weekends, and 1.1 on weekdays; the coefficient for motor vehicle sources is 1.5 during morning and evening peak hours (7:00-9:00 and 17:00-19:00), 1.0 during off-peak hours (9:00-17:00), and 0.5 during low-peak hours; the coefficient for residential sources is 1.8 during cooking hours (6:00-8:00, 11:00-13:00, 17:00-19:00), and 0.6 during other times. VOCs species allocation is determined based on emission source characteristics: industrial sources are allocated species based on process and raw material composition; motor vehicle emission factors are determined using the IVE model; and solvent usage is allocated species based on composition and activity. In NOx allocation, the NO2 / NOx ratio is 0.05 for industrial sources and 0.1 for motor vehicle sources, taking into account the conversion rate affected by temperature and light. The emission source classification system includes point sources (large industrial enterprises, power plants, waste incineration plants), area sources (residential, small-scale industrial and commercial, agricultural activities), and mobile sources (road vehicles, non-road mobile machinery, ships). The following describes the hourly near-surface ozone concentration inversion system in the embodiments of this invention from a hardware processing perspective; please refer to [link to relevant documentation]. Figure 3 This is a schematic diagram of the physical device structure of a near-ground ozone concentration hourly inversion system in this application embodiment.

[0118] It should be noted that, Figure 3 The structure of the near-surface ozone concentration hourly inversion system shown is merely an example and should not impose any limitations on the functionality and scope of use of the embodiments of the present invention.

[0119] like Figure 3 As shown, the near-surface ozone concentration hourly inversion system includes a central processing unit (CPU) 301, which can perform various appropriate actions and processes based on a program stored in read-only memory (ROM) 302 or a program loaded from storage section 308 into random access memory (RAM) 303, such as performing the methods described in the above embodiments. The RAM 303 also stores various programs and data required for system operation. The CPU 301, ROM 302, and RAM 303 are interconnected via a bus 304. An input / output (I / O) interface 305 is also connected to the bus 304.

[0120] The following components are connected to I / O interface 305: input section 306 including audio input devices, push-button switches, etc.; output section 307 including a liquid crystal display (LCD) and audio output devices, indicator lights, etc.; storage section 308 including a hard disk, etc.; and communication section 309 including a network interface card such as a LAN (Local Area Network) card, modem, etc. Communication section 309 performs communication processing via a network such as the Internet. Drive 310 is also connected to I / O interface 305 as needed. Removable media 311, such as a disk, optical disk, magneto-optical disk, semiconductor memory, etc., are installed on drive 310 as needed so that computer programs read from them can be installed into storage section 308 as needed.

[0121] In particular, according to embodiments of the present invention, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments of the present invention include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing computer programs for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via communication section 309, and / or installed from removable medium 311. When the computer program is executed by central processing unit (CPU) 301, it performs the various functions defined in the present invention.

[0122] It should be noted that specific examples of computer-readable storage media may include, but are not limited to: electrical connections having one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM), flash memory, optical fiber, portable compact disc read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this invention, a computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device.

[0123] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of the present invention. Each block in a flowchart or block diagram may represent a module, program segment, or portion of code, which contains one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those shown in the drawings.

[0124] Specifically, the near-surface ozone concentration hourly inversion system of this embodiment includes a processor and a memory. The memory stores a computer program. When the computer program is executed by the processor, it implements the hourly near-surface ozone inversion method based on multi-source data fusion provided in the above embodiment.

[0125] In another aspect, the present invention also provides a computer-readable storage medium, which may be included in the near-surface ozone concentration hourly inversion system described in the above embodiments; or it may exist independently and not incorporated into the near-surface ozone concentration hourly inversion system. The storage medium carries one or more computer programs, which, when executed by a processor of the near-surface ozone concentration hourly inversion system, cause the near-surface ozone concentration hourly inversion system to implement the hourly near-surface ozone inversion method based on multi-source data fusion provided in the above embodiments.

[0126] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit it. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application.

[0127] As used in the above embodiments, depending on the context, the term "when..." can be interpreted as meaning "if...", "after...", "in response to determining...", or "in response to detecting...". Similarly, depending on the context, the phrase "when determining..." or "if (the stated condition or event) is interpreted as meaning "if determining...", "in response to determining...", "when (the stated condition or event) is detected", or "in response to detecting (the stated condition or event)".

[0128] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. This program can be stored in a computer-readable storage medium, and when executed, it can include the processes described in the above method embodiments. The aforementioned storage medium includes various media capable of storing program code, such as ROM or random access memory (RAM), magnetic disks, or optical disks.

Claims

1. A method for hourly near-surface ozone retrieval based on multi-source data fusion, characterized in that, The method, applied to an hourly inversion system for near-surface ozone concentration, includes: Collect ozone total column concentration characteristic data, meteorological characteristic data, spatiotemporal geographic characteristic data, and socioeconomic characteristic data; Time-aligned ozone total column concentration characteristic data, meteorological characteristic data, spatiotemporal geographic characteristic data, and socioeconomic characteristic data based on timestamps; The time-aligned data is then uniformly processed into hourly data with a preset grid resolution; Based on the monitoring station identifier and latitude and longitude, the hourly data of the preset grid resolution are spatially matched to obtain the first matching data. A training sample set is constructed based on the first matching data and the measured data of the monitoring station. An ozone concentration inversion model was trained based on the training sample set. The ozone concentration inversion model is used to invert the ozone concentration in the target study area to obtain hourly ozone concentration data of the target study area with a preset grid resolution.

2. The method according to claim 1, characterized in that, The step of uniformly processing the time-aligned data into hourly data with a preset grid resolution specifically includes: The preset grid resolution is determined based on the spatial range of the ozone total column concentration characteristic data for the target study area. The ozone total column concentration characteristic data are upsampled to the preset grid resolution using bilinear interpolation. The meteorological and socioeconomic data are upsampled to the preset grid resolution using bilinear interpolation. The spatiotemporal geographic feature data are subjected to scale matching processing based on the area-weighted average method. The time-aligned data is converted into hourly data.

3. The method according to claim 1, characterized in that, The step of spatially matching hourly data with a preset grid resolution based on monitoring station identifiers and latitude and longitude to obtain first matching data, and constructing a training sample set based on the first matching data and the measured ozone concentration data of the stations, specifically includes: Based on the preset grid resolution, from the time-aligned multi-source hourly grid data, according to the latitude and longitude of each monitoring station, the grid feature data corresponding to each monitoring station is extracted. The grid feature data includes at least ozone total column features, meteorological features, spatiotemporal geographic features, and socioeconomic features. The raster feature data is associated with the corresponding monitoring station identifier and latitude and longitude to establish a spatial mapping relationship between the station location and the raster data, thus obtaining the first matching data; Based on the monitoring station identifier, the first matching data is paired with the corresponding measured ozone concentration data of the monitoring station to construct a training sample set.

4. The method according to claim 1, characterized in that, The step of training the ozone concentration inversion model based on the training sample set specifically includes: A spatiotemporal dual-stream model architecture is constructed, which includes a temporal sub-model and a spatial sub-model, wherein the temporal sub-model is used to capture the diurnal variation characteristics of ozone concentration, and the spatial sub-model is used to characterize the spatial distribution characteristics of ozone concentration; Ozone total column features and meteorological features are extracted from the training sample set, time series features are constructed in chronological order and input into the time series sub-model, spatiotemporal geographic features and socioeconomic features are extracted from the training sample set, and spatial features are constructed by combining station spatial location information and input into the spatial sub-model; Based on the output features of the temporal sub-model and the spatial sub-model, an ensemble model is trained; The performance of the ensemble model was evaluated using cross-validation, and the ozone concentration inversion model was obtained by optimizing the model parameters.

5. The method according to claim 4, characterized in that, After the step of performing ozone concentration inversion on the target study area according to the ozone concentration inversion model to obtain the target hourly ozone concentration data with a preset grid resolution within the target study area, the method further includes: Based on the monitoring station identifier, latitude and longitude, and timestamp, the hourly ozone concentration data of the target is spatiotemporally aligned with the measured data of the monitoring station to obtain an aligned dataset. Calculate the concentration deviation between the hourly ozone concentration data of the target in the aligned dataset and the measured ozone concentration data of the monitoring stations to obtain the station deviation data; Based on the site deviation data, and combined with factors such as terrain and meteorology as covariates, a grid correction factor is generated through spatial interpolation. The hourly ozone concentration data of the target is then corrected according to the grid correction factor to obtain the corrected ozone concentration data. Output the corrected ozone concentration data.

6. The method according to claim 1, characterized in that, After the step of performing ozone concentration inversion on the target study area according to the ozone concentration inversion model to obtain the target hourly ozone concentration data with a preset grid resolution within the target study area, the method further includes: High-value area data are extracted from the hourly ozone concentration data of the target, and the high-value area data is matched with the wind direction and wind speed data in the meteorological feature data to obtain the matching result; Based on the matching results and the spatiotemporal geographic feature data, ozone pollution transmission paths are determined, and regional transmission units are divided based on the ozone pollution transmission paths.

7. The method according to claim 1, characterized in that, After the step of performing ozone concentration inversion on the target study area according to the ozone concentration inversion model to obtain the target hourly ozone concentration data with a preset grid resolution within the target study area, the method further includes: Collect pollution source inventory data for the target study area, and spatially allocate the pollution source inventory data according to the preset grid resolution; A grid emission inventory is established based on the hourly ozone concentration data of the target and the pollution source inventory data, and the regional ozone pollution control target value is determined according to the grid emission inventory.

8. A near-ground ozone concentration hourly inversion system, characterized in that, The near-surface ozone concentration hourly inversion system includes: one or more processors and a memory; the memory is coupled to the one or more processors, the memory is used to store computer program code, the computer program code including computer instructions, and the one or more processors call the computer instructions to cause the near-surface ozone concentration hourly inversion system to perform the method as described in any one of claims 1-7.

9. A computer-readable storage medium comprising instructions, characterized in that, When the instruction is run on the near-ground ozone concentration hourly inversion system, the near-ground ozone concentration hourly inversion system performs the method as described in any one of claims 1-7.

10. A computer program product, characterized in that, When the computer program product is run on the near-ground ozone concentration hourly inversion system, the near-ground ozone concentration hourly inversion system performs the method as described in any one of claims 1-7.

Images