A method for mapping atmospheric pollutant concentrations that takes into account multi-dimensional geographical similarity

Abstract

This application discloses an atmospheric pollutant concentration mapping method that takes into account multi-dimensional geographical similarity, addressing the problems of insufficient interpretability and difficulty in measuring nonlinear correlations in existing atmospheric pollutant concentration mapping techniques. The method includes: acquiring long-term geographical environmental statistics and using similarity calculation methods to measure the geographical scene similarity between the monitored pixels and ground-based observation stations within the monitored area, obtaining a similarity matrix; dividing the area into a non-Euclidean neighborhood monitoring station set at the pixel scale based on the similarity matrix; acquiring historical observation data of the monitored area, constructing a graph topology network structure and training dataset for the monitored area based on the non-Euclidean neighborhood monitoring station set and historical observation data; and training a target mapping model using the graph topology network structure and training dataset. This target mapping model is used to generate a seamless spatial atmospheric pollutant concentration map with hourly resolution across the entire day based on real-time observation data of the monitored area.

Description

Technical Field

[0001] This application relates to the field of meteorological data processing, and in particular to a method for mapping atmospheric pollutant concentrations that takes into account multi-dimensional geographical similarity. Background Technology

[0002] Air pollutants refer to harmful substances produced by natural processes and human activities that have negative impacts on the environment or humans, including ozone and PM2.5. 10 PM 2.5 Air pollution has significant negative impacts on various aspects of the natural environment, socio-economic development, and people's lives and livelihoods. While air quality monitoring stations can provide real-time, high-quality data on air pollutant concentrations, these stations are mainly located in urban areas and are sparsely distributed, resulting in insufficient overall spatial coverage and failing to meet the requirements for real-time assessment of area air quality. Satellite optical remote sensing is easily affected by weather factors such as clouds and fog, leading to spatial gaps in estimations. Furthermore, the lack of satellite optical remote sensing data at night makes it impossible to estimate nighttime air pollutant concentrations. The spatial resolution and data accuracy of numerical model data are also insufficient to meet the requirements for near-real-time area air pollutant concentration monitoring.

[0003] In existing technologies, the mainstream method for producing seamless, gridded atmospheric pollutant concentration data across all times is spatial interpolation, which infers unknown points from observation data from known sparse stations. However, for regions with sparsely distributed observation stations, such as western China, traditional spatial interpolation methods based on spatial proximity exhibit low interpolation accuracy. Furthermore, while many improved algorithms for traditional spatial interpolation measure the similarity of attributes between known and unknown points using variables with high correlation and apply this similarity as interpolation weights to some extent, they cannot simulate the complex nonlinear relationships between the interpolation point and neighboring monitoring stations due to the complexity of geographical scene similarities. This prevents further improvement in the data accuracy of the interpolation results and limits their interpretability, ultimately leading to low mapping accuracy and poor interpretability of atmospheric pollutant concentrations. Summary of the Invention

[0004] This application provides an atmospheric pollutant concentration mapping method that takes into account multi-dimensional geographical similarity, in order to solve the problems of insufficient interpretability and difficulty in measuring nonlinear correlation in existing atmospheric pollutant concentration mapping techniques.

[0005] The first aspect of this application provides a method for mapping atmospheric pollutant concentrations that takes into account multi-dimensional geographical similarity, including:

[0006] Long-term geographic environment statistics are obtained, and similarity calculation methods are used to measure the geographic scene similarity between the monitored pixels in the monitored area and the ground-based observation station to obtain a similarity matrix;

[0007] The non-Euclidean neighborhood monitoring station set at the pixel scale is obtained based on the similarity matrix.

[0008] Historical observation data of the area to be monitored are obtained, including short-term meteorological reanalysis data and time series of atmospheric pollutant concentration observation data;

[0009] Based on the non-Euclidean neighborhood monitoring station set and the historical observation data, a graph topology network structure and a training dataset for the area to be monitored are constructed. The training dataset includes a pixel-scale neighborhood monitoring station feature dataset and a station-scale atmospheric pollutant concentration observation target dataset.

[0010] A target mapping model is trained using the graph topology network structure and the training dataset. The target mapping model is used to generate a seamless spatial atmospheric pollutant concentration map with hourly resolution throughout the day based on real-time observation data of the area to be monitored.

[0011] Optionally, the step of using a similarity calculation method to measure the geographical scene similarity between the pixels to be monitored within the monitoring area and the ground-based observation station to obtain a similarity matrix includes:

[0012] Determine the geographic environmental feature vectors of each pixel to be monitored and all ground-based observation stations within the monitoring area, and calculate statistical indicators to evaluate the differences in features based on the geographic environmental feature vectors;

[0013] The sequence of statistical indicators is input into the similarity calculation method to obtain similarity index values ​​that measure the similarity of geographical environment scenes;

[0014] A similarity matrix is ​​generated for each pixel to be monitored within the monitored area based on the similarity index value.

[0015] Optionally, the step of calculating statistical indicators for evaluating feature differences based on the geographic environment feature vector includes:

[0016] For each pixel to be monitored, calculate the normalized statistical index of the geographic environment feature vector and the geographic environment feature matrix of all ground-based observation stations;

[0017] The step of inputting the sequence of statistical indicators into the similarity calculation method to obtain similarity index values ​​that measure the similarity of geographical environment scenes includes:

[0018] The sequence of the normalized statistical indicators is input into the mathematical statistical model for evaluating geographical similarity;

[0019] The output value of the mathematical statistics model is determined as the similarity index value for measuring the similarity of geographical environment scenes.

[0020] Optionally, when the atmospheric pollutant concentration map is a PM2.5 concentration map, the geographic environmental feature vector includes at least one of the following: historical meteorological variable Met, historical PM2.5 concentration, historical AOD data, surface elevation data DEM, land use type data LULC, nighttime light data NLT, and normalized difference vegetation index data NDVI.

[0021] The mathematical and statistical model used to assess geographical similarity is:

[0022] P ij =||R(F ij ),(1-NRMSE(F ij ))||2

[0023] F ij ={AOD,PM 2.5 ,Met,DEM,LULC,NLT,NDVI}

[0024] Where i and j represent the pixel to be monitored and the number of a specific ground-based observation station, respectively, P ij The similarity index value is represented by ||·||², which represents the Euclidean norm, R is the pre-defined correlation coefficient, NRMSE is the normalized root mean square error, and F... ij This represents the geographic environment feature vector.

[0025] Optionally, the step of dividing the non-Euclidean neighborhood monitoring station set at the pixel scale according to the similarity matrix includes:

[0026] Based on the similarity matrix of each pixel to be monitored, a neighborhood monitoring station is defined for each pixel to be monitored. The neighborhood monitoring station is a number of stations arranged in descending order of the similarity index value.

[0027] The neighborhood monitoring stations are integrated to obtain a non-Euclidean neighborhood monitoring station set for each pixel to be monitored.

[0028] Optionally, the non-Euclidean neighborhood monitoring station set is represented as:

[0029]

[0030] Among them, S i This represents the set of non-Euclidean neighborhood monitoring stations corresponding to the pixel i to be monitored; topK(·) represents the set of independent variable points with the largest K values, i.e., a function that selects the K values ​​with the greatest geographical similarity; K is the number of neighborhood monitoring stations corresponding to each pixel i to be monitored. represents the similarity matrix corresponding to the pixel i to be monitored, and N is the total number of ground-based observation stations.

[0031] Optionally, the step of constructing the graph topology network structure and training dataset of the area to be monitored based on the non-Euclidean neighborhood monitoring station set and the historical observation data includes:

[0032] The geographical similarity between the pixel to be monitored and the neighboring monitoring stations is dynamically calculated based on the non-Euclidean neighborhood monitoring station set and the historical observation data.

[0033] Using the geographical similarity as edge weights and the geographical features of the pixel to be monitored and the neighboring monitoring stations as nodes, a graph topology network structure for the area to be monitored is constructed.

[0034] Based on the graph network topology and the observed atmospheric pollutant concentration data of the neighboring monitoring stations, a pixel-scale neighboring monitoring station feature dataset is formed, and based on the atmospheric pollutant concentration observation data of the monitoring stations, a station-scale atmospheric pollutant concentration observation target dataset is formed.

[0035] A second aspect of this application provides a system for mapping atmospheric pollutant concentrations that takes into account multi-dimensional geographical similarity, comprising:

[0036] The computing unit is used to acquire long-term geographic environment statistics and use similarity calculation methods to measure the geographic scene similarity between the monitored pixels in the monitored area and the ground-based observation station to obtain a similarity matrix.

[0037] A partitioning unit is used to partition a non-Euclidean neighborhood monitoring station set at the pixel scale based on the similarity matrix;

[0038] The acquisition unit is used to acquire historical observation data of the area to be monitored, including short-term meteorological reanalysis data and time series of atmospheric pollutant concentration observation data.

[0039] The construction unit is used to construct the graph topology network structure and training dataset of the area to be monitored based on the non-Euclidean neighborhood monitoring station set and the historical observation data. The training dataset includes a pixel-scale neighborhood monitoring station feature dataset and a station-scale atmospheric pollutant concentration observation target dataset.

[0040] The training unit is used to train a target mapping model using the graph topology network structure and the training dataset. The target mapping model is used to generate a seamless spatial atmospheric pollutant concentration map with hourly resolution for all day based on real-time observation data of the area to be monitored.

[0041] A third aspect of this application provides an apparatus for mapping atmospheric pollutant concentrations that takes into account multi-dimensional geographical similarity, the apparatus comprising:

[0042] Processor, memory, input / output units, and bus;

[0043] The processor is connected to the memory, the input / output unit, and the bus;

[0044] The memory stores a program that the processor invokes to execute the first aspect and any optional method of the first aspect for mapping atmospheric pollutant concentrations that takes into account multi-dimensional geographical similarity.

[0045] The fourth aspect of this application provides a computer-readable storage medium storing a program that, when executed on a computer, performs the first aspect and any optional method of the first aspect for mapping atmospheric pollutant concentrations taking into account multi-dimensional geographical similarity.

[0046] As can be seen from the above technical solutions, this application has the following advantages:

[0047] First, long-term geographic environmental statistics are acquired, and similarity calculation methods are used to measure the geographic scene similarity between the monitored pixels and ground-based observation stations within the monitored area. This constructs a similarity matrix and divides the area into a non-Euclidean neighborhood monitoring station set at the pixel scale. This process effectively captures the similar characteristics between different regions, ensuring that the selection of neighborhood monitoring stations considers more layers of geographic and meteorological environmental influences, avoiding the singular impact of simple spatial proximity relationships on the results. Next, historical short-term meteorological reanalysis data and atmospheric pollutant concentration observation data are combined to construct a graph topology network structure and training dataset. Based on this, a target mapping model is trained. This target mapping model employs graph attention network technology, which can consider the potential nonlinear relationships between geographic objects, enhancing its ability to handle complex relationships between geographic objects compared to existing spatial interpolation techniques.

[0048] By employing multi-dimensional geographical similarity analysis, the division of non-Euclidean neighborhood monitoring station sets, and the application of graph attention network technology, the problems of insufficient interpretability and difficulty in nonlinear correlation measurement in existing atmospheric pollutant concentration mapping technologies have been solved, improving the accuracy and flexibility of mapping results and successfully achieving all-day, hourly resolution, and spatially seamless atmospheric pollutant concentration monitoring. Attached Figure Description

[0049] To more clearly illustrate the technical solutions in this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0050] Figure 1 A schematic flowchart of an embodiment of the atmospheric pollutant concentration mapping method that takes into account multi-dimensional geographical similarity provided in this application;

[0051] Figure 2 A schematic diagram of the method architecture for mapping atmospheric pollutant concentrations that takes into account multi-dimensional geographical similarity provided in this application;

[0052] Figure 3 The method for mapping atmospheric pollutant concentrations that takes into account multi-dimensional geographical similarity provided in this application targets PM2.5. 2.5 A schematic diagram of an embodiment of a concentration graph;

[0053] Figure 4 A scatter plot of atmospheric pollutant concentration mapping accuracy in the method for mapping atmospheric pollutant concentrations that takes into account multi-dimensional geographical similarity provided in this application;

[0054] Figure 5 A schematic diagram of an embodiment of the atmospheric pollutant concentration mapping system that takes into account multi-dimensional geographical similarity provided in this application;

[0055] Figure 6 A schematic diagram of an embodiment of the apparatus for mapping atmospheric pollutant concentrations that takes into account multi-dimensional geographical similarity provided in this application. Detailed Implementation

[0056] This application provides an atmospheric pollutant concentration mapping method that takes into account multi-dimensional geographical similarity, in order to solve the problems of insufficient interpretability and difficulty in measuring nonlinear correlation in existing atmospheric pollutant concentration mapping techniques.

[0057] It should be noted that the atmospheric pollutant concentration mapping method provided in this application, which takes into account multi-dimensional geographical similarity, can be applied to terminals as well as servers. For example, terminals can be smartphones, computers, tablets, smart TVs, smartwatches, portable computer terminals, or fixed terminals such as desktop computers. For ease of explanation, this application uses terminals as the implementing entity for illustrative purposes.

[0058] Please see Figure 1 , Figure 1 An embodiment of the method for mapping atmospheric pollutant concentrations that takes into account multi-dimensional geographical similarity provided in this application includes:

[0059] 101. Obtain long-term geographic environment statistics and use similarity calculation methods to measure the geographic scene similarity between the pixels to be monitored and the ground-based observation stations in the monitoring area, and obtain a similarity matrix;

[0060] In atmospheric pollutant concentration mapping, traditional spatial interpolation methods mainly select reference observation stations based on Euclidean distance or spatial proximity. However, this method ignores the differences in geographical, meteorological, and environmental characteristics between different regions, resulting in insufficient interpretability and accuracy of the interpolation results. Therefore, this embodiment first requires collecting long-term geographic environmental statistics for the area to be monitored, i.e., data reflecting long-term changes in the area's geography, climate, pollutant concentrations, and land use. These geographic environmental statistics include, but are not limited to, atmospheric pollutant concentrations, atmospheric parameters highly correlated with atmospheric pollutant concentrations, meteorological variables, and surface elevation data. All data undergoes preprocessing such as spatiotemporal matching. It should be noted that the long-term timescale includes, but is not limited to, monthly or annual scales. The spatiotemporal matching preprocessing includes: matching all input data according to the production time; unifying the spatial range and spatial resolution of the input data; and normalizing the numerical range of the input data to the [0,1] interval.

[0061] Subsequently, using collected geographic environment statistics and similarity calculation methods, the similarity between each pixel to be monitored in the monitored area and all ground-based observation stations is measured. The output value of the similarity calculation method is used as an indicator to measure the similarity of geographic environment scenes. For each pixel to be monitored in the monitored area, a geographic similarity matrix between the pixel and the set of ground-based observation stations needs to be calculated, ultimately obtaining a pixel-scale similarity matrix in the monitored area used to characterize the geographic environment.

[0062] 102. A set of non-Euclidean neighborhood monitoring stations at the pixel scale is obtained by partitioning based on the similarity matrix;

[0063] Each element in the similarity matrix represents the similarity value between a pixel to be monitored and a ground-based observation station. For each pixel to be monitored, based on the data in its similarity matrix, a certain number of neighboring observation stations can be selected. This ensures that the selection of neighboring monitoring stations depends not only on spatial distance but also on the similarity between the pixel and each ground-based observation station. By selecting appropriate neighboring monitoring stations for each pixel to be monitored, a non-Euclidean neighboring monitoring station set containing all pixels to be monitored and their corresponding neighboring monitoring stations can be obtained.

[0064] 103. Obtain historical observation data for the area to be monitored, including short-term meteorological reanalysis data and time series of atmospheric pollutant concentration observation data;

[0065] The distribution of air pollutant concentrations is greatly influenced by meteorological conditions, such as temperature, humidity, and wind speed, which can directly affect the diffusion and concentration changes of pollutants. To accurately map air pollutant concentrations in the monitored area, in addition to considering spatial geographical characteristics, it is also necessary to fully consider time-series data, specifically continuous historical meteorological reanalysis data and time series of air pollutant concentration observation data at short time scales. These historical observation data can reveal the patterns of pollutant concentration changes over time.

[0066] Specifically, meteorological reanalysis data is typically obtained by combining data from meteorological satellites, numerical weather prediction models, and ground-based meteorological stations, providing accurate meteorological data over a specific time frame. Meanwhile, atmospheric pollutant concentration observation data comes from ground monitoring stations or satellite remote sensing data, primarily including PM2.5 concentrations. 2.5 PM 10 Concentration values ​​of pollutants such as NO2, SO2, CO, and O3. The time scale for this historical observation data can be selected as hourly or daily, depending on the monitoring requirements.

[0067] 104. Construct a graph topology network structure and training dataset for the area to be monitored based on the non-Euclidean neighborhood monitoring station set and historical observation data. The training dataset includes a pixel-scale neighborhood monitoring station feature dataset and a station-scale atmospheric pollutant concentration observation target dataset.

[0068] Using the non-Euclidean neighborhood monitoring station set and historical observation data obtained in the previous calculation, a graph topology network structure of the area to be monitored is constructed, and feature datasets and target datasets are generated for training the all-day atmospheric pollutant concentration mapping model.

[0069] Using pixels in the area to be monitored and neighboring monitoring stations as nodes, a graph topology network structure is constructed based on the similarity relationships between these nodes. Each node in the graph represents a pixel or an observation station, and the edges between nodes represent the spatial-temporal associations or similarities between them. This graph topology network structure can effectively describe the geographical and meteorological similarities between nodes, as well as the potential dependencies of pollutant concentrations, providing a framework for subsequent pollutant concentration prediction and interpolation.

[0070] The generated training dataset specifically includes a pixel-scale neighborhood monitoring station feature dataset and a station-scale atmospheric pollutant concentration observation target dataset. The neighborhood monitoring station feature dataset will serve as input features in the graph network, representing the geographic environment and pollutant concentration characteristics of the pixel to be monitored and its neighboring monitoring stations. The station-scale atmospheric pollutant concentration observation target dataset contains pollutant concentration data for each observation station at different time points, representing the observed pollutant concentration values ​​at the station scale. This dataset will serve as the target dataset for model training and prediction, i.e., for fitting the pollutant concentration values ​​of the stations, serving as the target variable for the mapping model. Specifically, when constructing the neighborhood monitoring station feature dataset, feature data of neighboring monitoring stations related to the pixel to be monitored can be extracted from historical observation data based on the graph topology network structure. This feature data includes information such as meteorological conditions and pollutant concentrations, and represents the observation data of each neighboring monitoring station at the current time. These feature data are then combined to form the pixel-scale neighborhood monitoring station feature dataset.

[0071] 105. A target mapping model is trained using a graph topology network structure and a training dataset. The target mapping model is used to generate a seamless spatial atmospheric pollutant concentration map with hourly resolution throughout the day based on real-time observation data of the area to be monitored.

[0072] Based on the constructed graph topology network structure and training datasets—namely, the neighborhood monitoring station feature dataset and the atmospheric pollutant concentration observation target dataset—an initial mapping model is trained. Specifically, this initial mapping model employs:

[0073] P = f r (G;s)

[0074] Where P represents the estimated concentration of air pollutants at the target time; f r (·) represents an atmospheric pollutant concentration estimation model with fitting characteristics to achieve regression, including but not limited to graph convolutional neural networks, random forest regression, support vector machines, etc.; G represents the graph topology network structure data; s represents the measured concentration of atmospheric pollutants at the current time. 10-20% of the training data is randomly selected as the validation set for model estimation cross-validation to test the model's generalization ability.

[0075] The training dataset contains neighborhood monitoring station feature datasets, providing the model with information on the geographic environment, meteorological characteristics, and pollutant concentrations of the pixels to be monitored and their neighboring monitoring stations. The target dataset provides the model with target pollutant concentration data, helping the model learn the relationship between features and pollutant concentrations. The training dataset is then input into the graph attention model for parameter training. Sensitivity analysis of the modeling variables is conducted; by calculating the relative importance of the input variables, variables that do not significantly contribute to the modeling accuracy are gradually eliminated, optimizing the model structure until the model converges, thus obtaining the target mapping model.

[0076] In actual mapping, real-time meteorological reanalysis data and ground-based atmospheric pollutant concentration data are used, combined with long-term geographic environmental statistics and similarity matrices of the area to be monitored, to construct a pixel-scale neighborhood monitoring station feature dataset of the area to be monitored. This dataset is used as input to the target mapping model, which outputs an all-day hourly resolution atmospheric pollutant concentration map of the area to be monitored. This map has high timeliness and spatial accuracy, and each area in the map can reflect the precise pollutant concentration.

[0077] In this embodiment, long-term geographic environmental statistics are first acquired, and a similarity calculation method is used to measure the geographic scene similarity between the monitored pixels and ground-based observation stations within the monitored area. This constructs a similarity matrix and divides the area into a non-Euclidean neighborhood monitoring station set at the pixel scale. This process effectively captures the similar characteristics between different regions, ensuring that the selection of neighborhood monitoring stations considers more layers of geographic and meteorological environmental influences, avoiding the singular impact of simple spatial proximity relationships on the results. Subsequently, historical short-term meteorological reanalysis data and atmospheric pollutant concentration observation data are combined to construct a graph topology network structure and training dataset. Based on this, a target mapping model is trained. This target mapping model employs graph attention network technology, which can consider potential nonlinear relationships between geographic objects, enhancing the ability to handle complex relationships between geographic objects compared to existing spatial interpolation techniques.

[0078] By employing multi-dimensional geographical similarity analysis, the division of non-Euclidean neighborhood monitoring station sets, and the application of graph attention network technology, the problems of insufficient interpretability and difficulty in nonlinear correlation measurement in existing atmospheric pollutant concentration mapping technologies have been solved, improving the accuracy and flexibility of mapping results and successfully achieving all-day, hourly resolution, and spatially seamless atmospheric pollutant concentration monitoring.

[0079] Please see Figure 2 and Figure 3 , Figure 2 A schematic diagram illustrating the framework of the atmospheric pollutant concentration mapping method that takes into account multi-dimensional geographical similarity provided in this application. Figure 3 The method for mapping atmospheric pollutant concentrations that takes into account multi-dimensional geographical similarity provided in this application targets PM2.5. 2.5 A schematic diagram of an embodiment of concentration map creation is shown below. Figure 2 and Figure 3 This application provides a detailed description of the method for mapping atmospheric pollutant concentrations that takes into account multi-dimensional geographical similarity. This method is also applicable to other atmospheric pollutant concentration mapping. The method specifically includes:

[0080] 301. Obtain long-term statistical data on the geographical environment;

[0081] In some specific implementations, it is necessary to obtain the average monthly geographic environmental statistics over the past three years. The data types specifically include: historical meteorological variables (Met), historical PM2.5, etc. 2.5 Concentration, historical AOD data, surface elevation data (DEM), land use type data (LULC), nighttime light data (NLT), and normalized difference vegetation index (NDVI) data were collected. All statistical data underwent normalization preprocessing, were resampled to the monitored area, and assigned corresponding geographic environmental statistical data based on the latitude and longitude of the ground-based observation stations. For example, historical meteorological variable Met was obtained from the European Centre for Medium-Range Weather Forecasts (ECMWF) Generation 5 Reanalysis dataset; historical PM2.5 concentrations were also collected. 2.5 Concentration data and historical AOD data were obtained from the LGHAP v2 dataset; surface elevation data (DEM) were obtained from the SRTM project; land use type data (LULC) were obtained from the CLCD v01 dataset; nighttime light data (NLT) were obtained from NPP-VIIRS satellite data; and normalized difference vegetation index (NDVI) data were obtained from the MOD13A3 product.

[0082] 302. Determine the geographic environmental feature vectors of each pixel to be monitored and all ground-based observation stations within the monitoring area, and calculate statistical indicators to evaluate the differences in features based on the geographic environmental feature vectors;

[0083] Each monitored pixel and the ground-based observation station can have a geographic environmental feature vector defined based on their geographic environmental characteristics. This geographic environmental feature vector contains various parameters reflecting differences in the geographic environment. These geographic environmental characteristics include, but are not limited to, historical meteorological variables and historical PM2.5. 2.5 Concentration, historical AOD data, surface elevation data, land use type data, nighttime light data, and normalized vegetation index data.

[0084] For each geographic environmental feature vector between the pixel to be monitored and all ground-based observation stations, the differences between them can be measured by calculating different statistical indicators. Therefore, it is necessary to calculate a variety of statistical indicators to evaluate the differences in features between the two. These statistical indicators include, but are not limited to, root mean square error, correlation coefficient, and mean deviation.

[0085] It should be noted that for each pixel to be monitored and each ground-based observation station's geographic environmental feature vector, it is recommended to calculate at least two statistical indicators. By using multiple statistical indicators to complement each other, the similarity between the two can be assessed more comprehensively and holistically.

[0086] 303. Input the sequence of statistical indicators into the similarity calculation method to obtain the similarity index value that measures the similarity of geographical environment scenes;

[0087] Each statistical indicator calculated in step 302 represents an assessment of geographical environmental differences from different perspectives. These indicators are normalized, and the normalized statistical indicators are combined into a feature sequence. This feature sequence contains multi-dimensional information to describe the differences between the monitored pixel and various ground-based observation stations. The feature sequence is input into a similarity calculation method to integrate the various statistical indicators, resulting in a single similarity index value for each monitored pixel and ground-based observation station. This index value reflects the degree of similarity between the two geographical environmental scenes and can effectively measure the geographical environmental similarity between the monitored pixel and various ground-based observation stations. Specifically, this similarity calculation method can be expressed as follows:

[0088] P ij =f p ({R1(Geo i Geo j ),R2(Geo i Geo j ),...,R n (Geo i Geo j )})

[0089] Where i and j represent the pixel to be monitored and the number of a specific ground-based observation station, respectively; P ij This represents the similarity index value of the geographical environment scene between the monitored pixel i and the ground-based observation station j; f p (·) represents a similarity algorithm that measures the similarity of vector features, capable of measuring the similarity of geographical environment scenes, including but not limited to arithmetic mean, Euclidean norm, cosine similarity, etc.; R k (·), k∈[1,n] represent various statistical indicator calculation methods; Geo i With Geo j These represent the geographic environmental characteristic variables corresponding to the pixel i to be monitored and the ground-based observation station j, respectively.

[0090] In some specific embodiments, for each pixel to be monitored, a normalized statistical index can be calculated between the geographic environment feature vector and the geographic environment feature matrix of all ground-based observation stations; the sequence of normalized statistical indices is input into a mathematical statistical model for evaluating geographic similarity; and the output value of the mathematical statistical model is determined as the similarity index value for measuring the similarity of geographic environment scenes. For example, taking the correlation coefficient and normalized root mean square error as the normalized statistical indices, and the Euclidean norm as the mathematical statistical model, the mathematical statistical model for evaluating geographic similarity can be expressed as:

[0091] P ij =||R(F ij ),(1-NRMSE(Fij ))||2

[0092] F ij ={AOD,PM 2.5 ,Met,DEM,LULC,NLT,NDVI}

[0093] Where i and j represent the pixel to be monitored and the number of a specific ground-based observation station, respectively, P ij The similarity index value is represented by ||·||², which represents the Euclidean norm, R is the pre-set correlation coefficient, RMSE is the root mean square error, and F is the value of the similarity index. ij This represents a vector of geographical environmental features.

[0094] 304. Generate a similarity matrix for each pixel to be monitored within the monitored area based on the similarity index values;

[0095] Based on the similarity index values ​​calculated above, a similarity matrix can be generated for each pixel within the monitoring area. This similarity matrix reflects the geographical environmental similarity between each pixel and all ground-based observation stations, providing support for subsequent selection of neighboring monitoring stations, pollutant concentration interpolation, and mapping models. Assuming the monitoring area contains M pixels and the total number of ground-based observation stations is N, the similarity matrix will be an M×N matrix, where each element P... ij This represents the similarity index value between the monitored pixel i and the j-th ground-based observation station.

[0096] 305. Define the neighborhood monitoring stations for each pixel to be monitored based on the similarity matrix of each pixel to be monitored. The neighborhood monitoring stations are a number of stations arranged in descending order of similarity index value.

[0097] Based on the generated similarity matrix, neighboring monitoring stations can be identified for each pixel to be monitored. By sorting the ground-based observation stations from highest to lowest similarity index value, the most similar stations can be selected as neighboring monitoring stations. This ensures that the pollutant concentration estimation for each pixel depends on the observation stations with the most representative similar geographical environments, thereby improving the accuracy and interpretability of the mapping results. The number of neighboring monitoring stations can be adjusted according to actual needs and is not limited here. For example, the 32 stations with the highest similarity index values ​​can be defined as the neighboring monitoring stations for that pixel.

[0098] 306. Integrate the neighborhood monitoring stations to obtain a non-Euclidean neighborhood monitoring station set for each pixel to be monitored;

[0099] By integrating these neighboring monitoring stations, a non-Euclidean neighboring monitoring station set is obtained for each pixel to be monitored. The relationship between these neighboring monitoring stations no longer relies solely on traditional Euclidean spatial distance. Through this integration, the similarity between the pixel to be monitored and its neighboring monitoring stations can be measured under multi-dimensional geographical environmental features, and interpolation weights can be calculated based on this similarity. This method breaks through the traditional Euclidean neighborhood method, introducing more diverse geographical scene features into the interpolation process, thereby improving the accuracy and interpretability of atmospheric pollutant concentration mapping that takes into account multi-dimensional geographical similarity.

[0100] The non-Euclidean neighborhood monitoring site set is specifically represented as follows:

[0101]

[0102] Among them, S i Let represent the set of non-Euclidean neighborhood monitoring stations corresponding to the pixel i to be monitored; topK(·) represents the set of independent variable points with the largest K values, i.e., a function that selects the K values ​​with the greatest geographical similarity; K is the number of neighborhood monitoring stations corresponding to each pixel to be monitored. represents the similarity matrix corresponding to the pixel i to be monitored, and N is the total number of ground-based observation stations.

[0103] 307. Obtain historical observation data for the area to be monitored, including short-term meteorological reanalysis data and atmospheric pollutant concentration observation data;

[0104] In this embodiment, step 307 is similar to step 103 in the previous embodiment, and will not be described again here.

[0105] 308. Dynamically calculate the geographical similarity between the pixel to be monitored and the neighboring monitoring stations based on the non-Euclidean neighborhood monitoring station set and historical observation data;

[0106] Based on a non-Euclidean neighborhood monitoring station set, the geographical environmental similarity between the monitored pixel and its neighboring monitoring stations is dynamically calculated. This is combined with short-term meteorological reanalysis data and time series of atmospheric pollutant concentration observation data to achieve real-time updates of the similarity distance. For each monitored pixel, the geographical environmental similarity between the pixel and its neighboring monitoring stations is first calculated using short-term meteorological reanalysis data and atmospheric pollutant concentration observation data. In this step, short-term meteorological reanalysis data from historical observation data is used to replace the long-term meteorological data time series used in step 303 for calculating geographical environmental similarity, enabling dynamic updates to the similarity calculation and ensuring that the calculation reflects the impact of the current moment and recent historical observation data on the monitored area.

[0107] Taking a historical 5-hour time step as an example, the geographical similarity distance between the pixel to be monitored and neighboring monitoring stations is calculated. This dynamic similarity distance calculation method adopts the following approach:

[0108] W=||R(F′ ij ),(1-NRMSE(F′ ij ))||2

[0109]

[0110] Where F′ ij A dynamic geographic feature sequence representing the pixel to be monitored and its neighboring monitoring stations; The time series of meteorological reanalysis variables represents 5 hours of historical data; the meanings of the remaining symbols are consistent with those in step 303.

[0111] 309. Using geographical similarity as edge weights and the geographical characteristics of the pixels to be monitored and neighboring monitoring stations as nodes, construct a graph topology network structure for the area to be monitored.

[0112] Using the calculated geographical similarity as edge weights in the graph, and considering the geographical features of the pixel to be monitored and its neighboring monitoring stations as nodes, a graph topology network structure for the monitored area is constructed. This graph structure can effectively represent the geographical relationship between each pixel in the monitored area and its neighboring monitoring stations, and provides a foundation for subsequent interpolation and atmospheric pollutant concentration mapping models.

[0113] The specific methods for constructing the network topology shown in this diagram can be:

[0114] G = {υ|E,W};

[0115] Wherein, G represents the network topology of the neighboring monitoring stations graph; υ represents the geographic environment feature matrix of the pixel and the neighboring monitoring stations, and the geographic environment features include short-term meteorological reanalysis data and other long-term geographic environment statistics; E represents whether there is a mutual influence relationship between the pixel and the neighboring monitoring stations. Here, it is assumed that there is mutual influence between the observed pixel and the neighboring monitoring stations, that is, a fully connected graph network topology is constructed; W represents the correlation strength between the pixel and the neighboring monitoring stations, that is, the geographic environment similarity distance.

[0116] Furthermore, the correlation strength between a pixel and neighboring monitoring stations is calculated using the following method:

[0117] W i ={f p ({R1(Geo′ i Geo′ j ),R2(Geo′ i Geo′ j ),...,Rn (Geo′ i Geo′ j )})|j

[0118] ∈[1,k]}

[0119] Where j represents the ID of a specific station in the neighborhood monitoring station set corresponding to pixel i; k represents the total number of neighborhood monitoring stations; f p (·) represents the mathematical and statistical model for measuring geographical similarity in step 303; R k (·), k∈[1,n] represents the various statistical indicator calculation methods used in steps 301-303; Geo′ represents the long-term geographical environmental characteristic data used in steps 301-303, wherein the long-term meteorological variable data used are replaced with short-term meteorological reanalysis data.

[0120] 310. Based on the graph network topology and the observed atmospheric pollutant concentration data of the neighboring monitoring stations, a pixel-scale neighboring monitoring station feature dataset is formed, and based on the atmospheric pollutant concentration observation data of the stations to be monitored, a station-scale atmospheric pollutant concentration observation target dataset is formed.

[0121] Based on the previously constructed graph network topology and the observed atmospheric pollutant concentration data from neighboring monitoring stations, two datasets are generated:

[0122] Pixel-scale neighborhood monitoring station feature dataset: This dataset will contain relevant information about the pixel to be monitored and its neighboring monitoring stations, specifically including the atmospheric pollutant concentration observation data (e.g., PM2.5) of each pixel's neighboring monitoring stations at the current time. 2.5 The data includes concentrations of nitrogen, nitrogen, nitrogen, and sulfur dioxide (NO2, SO2, etc.), as well as other meteorological characteristic data from neighboring monitoring stations (such as wind speed, temperature, humidity, etc.). The observation data from these neighboring monitoring stations are weighted according to their geographical similarity to the pixels to be monitored. The more similar the stations are, the greater the weight of the given observation data in the calculation.

[0123] Site-scale atmospheric pollutant concentration observation target dataset: This dataset contains actual observation data from each ground-based observation station, i.e., the atmospheric pollutant concentration (e.g., PM2.5) at a given time. 2.5 Concentrations such as NO2 concentration are used as target values ​​when training the model to predict the pollutant concentrations of the pixels to be monitored.

[0124] 311. A target mapping model is trained using a graph topology network structure and a training dataset. The target mapping model is used to generate a seamless spatial atmospheric pollutant concentration map with hourly resolution throughout the day based on real-time observation data of the area to be monitored.

[0125] In this embodiment, step 311 is similar to step 105 in the previous embodiment, and will not be described again here.

[0126] Please see Figure 4 , Figure 4 The scatter plot of atmospheric pollutant concentration mapping accuracy in the method for mapping atmospheric pollutant concentrations that takes into account multi-dimensional geographical similarity provided in this application, where the correlation coefficient R = 0.91 indicates that the PM2.5 concentration estimated by the mapping method is... 2.5 The concentration showed a high positive correlation with the actual observed data, indicating that it can accurately reflect the actual pollutant concentration distribution to a certain extent. The root mean square error (RMSE) was 17.75 μg / m³. 3 , indicating PM 2.5 There is a certain error between the estimated and observed values, but the overall error is relatively small, indicating that the accuracy is reliable. Overall, this shows that the estimation effect of the atmospheric pollutant concentration mapping method provided in this application is quite ideal.

[0127] Please see Figure 5 , Figure 5 An embodiment of the atmospheric pollutant concentration mapping system that takes into account multi-dimensional geographical similarity provided in this application, the system includes:

[0128] The calculation unit 501 is used to obtain long-term geographic environment statistics and use similarity calculation methods to measure the geographic scene similarity between the pixels to be monitored in the monitoring area and the ground observation station to obtain a similarity matrix.

[0129] The partitioning unit 502 is used to partition a non-Euclidean neighborhood monitoring station set at the pixel scale according to the similarity matrix;

[0130] The acquisition unit 503 is used to acquire historical observation data of the area to be monitored, including short-term meteorological reanalysis data and time series of atmospheric pollutant concentration observation data;

[0131] Construction unit 504 is used to construct the graph topology network structure and training dataset of the area to be monitored based on the non-Euclidean neighborhood monitoring station set and the historical observation data. The training dataset includes a neighborhood monitoring station feature dataset at the pixel scale and an atmospheric pollutant concentration observation target dataset at the station scale.

[0132] Training unit 505 is used to train a target mapping model using the graph topology network structure and the training dataset. The target mapping model is used to generate a seamless spatial atmospheric pollutant concentration map with hourly resolution for all day based on real-time observation data of the area to be monitored.

[0133] In this embodiment, the functions of each unit are the same as described above. Figure 1 or Figure 3 The steps in the method embodiments shown correspond to those in the examples, and will not be repeated here.

[0134] This application also provides an apparatus for mapping atmospheric pollutant concentrations that takes into account multi-dimensional geographical similarity; please refer to [link to relevant documentation]. Figure 6 , Figure 6 One embodiment of the apparatus for mapping atmospheric pollutant concentrations that takes into account multi-dimensional geographical similarity provided in this application includes:

[0135] Processor 601, memory 602, input / output unit 603, bus 604;

[0136] The processor 601 is connected to the memory 602, the input / output unit 603, and the bus 604;

[0137] The memory 602 stores a program, which the processor 601 calls to execute any of the above methods for mapping atmospheric pollutant concentrations that take into account multi-dimensional geographical similarity.

[0138] This application also relates to a computer-readable storage medium on which a program is stored, which, when run on a computer, causes the computer to perform any of the above methods for mapping atmospheric pollutant concentrations that takes into account multi-dimensional geographical similarity.

[0139] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0140] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be an indirect coupling or communication connection between apparatuses or units through some interfaces, and may be electrical, mechanical, or other forms.

[0141] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0142] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0143] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

Claims

1. A method for mapping atmospheric pollutant concentrations that takes into account multi-dimensional geographical similarity, characterized in that, The method includes: Long-term geographic environment statistics are obtained, and similarity calculation methods are used to measure the geographic scene similarity between the monitored pixels in the monitored area and the ground-based observation station to obtain a similarity matrix; The non-Euclidean neighborhood monitoring station set at the pixel scale is obtained based on the similarity matrix. Historical observation data of the area to be monitored are obtained, including short-term meteorological reanalysis data and time series of atmospheric pollutant concentration observation data; Based on the non-Euclidean neighborhood monitoring station set and the historical observation data, a graph topology network structure and a training dataset for the area to be monitored are constructed. The training dataset includes a pixel-scale neighborhood monitoring station feature dataset and a station-scale atmospheric pollutant concentration observation target dataset. A target mapping model is trained using the graph topology network structure and the training dataset. The target mapping model is a graph attention network model. The target mapping model is used to generate a seamless spatial atmospheric pollutant concentration map with all-day hourly resolution based on real-time observation data of the area to be monitored. The method of using similarity calculation to measure the geographical scene similarity between the pixels to be monitored and the ground-based observation station within the monitoring area yields a similarity matrix, including: The geographic environmental feature vectors of each pixel to be monitored and all ground-based observation stations within the monitoring area are determined, and statistical indicators for evaluating feature differences are calculated based on the geographic environmental feature vectors. When the atmospheric pollutant concentration map is a PM2.5 concentration map, the geographic environmental feature vectors include historical meteorological variable Met, historical PM2.5 concentration, historical AOD data, surface elevation data DEM, land use type data LULC, nighttime light data NLT, and normalized difference vegetation index data NDVI. The sequence of statistical indicators is input into the similarity calculation method to obtain similarity index values ​​that measure the similarity of geographical environment scenes; A similarity matrix is ​​generated for each pixel to be monitored within the monitored area based on the similarity index value; The process of constructing the graph topology network structure and training dataset for the area to be monitored based on the non-Euclidean neighborhood monitoring station set and the historical observation data includes: The geographical similarity between the pixel to be monitored and the neighboring monitoring stations is dynamically calculated based on the non-Euclidean neighborhood monitoring station set and the historical observation data. Using the geographical similarity as edge weights and the geographical features of the pixel to be monitored and the neighboring monitoring stations as nodes, a graph topology network structure for the area to be monitored is constructed. Based on the graph topology network structure and the observed atmospheric pollutant concentration data of the neighboring monitoring stations, a pixel-scale neighboring monitoring station feature dataset is formed, and based on the atmospheric pollutant concentration observation data of the monitoring stations, a station-scale atmospheric pollutant concentration observation target dataset is formed.

2. The method according to claim 1, characterized in that, The statistical indicators for evaluating feature differences based on the geographic environment feature vector include: For each pixel to be monitored, calculate the normalized statistical index of the geographic environment feature vector and the geographic environment feature matrix of all ground-based observation stations; The step of inputting the sequence of statistical indicators into the similarity calculation method to obtain similarity index values ​​that measure the similarity of geographical environment scenes includes: The sequence of the normalized statistical indicators is input into the mathematical statistical model for evaluating geographical similarity; The output value of the mathematical statistics model is determined as the similarity index value for measuring the similarity of geographical environment scenes.

3. The method according to claim 2, characterized in that, The mathematical and statistical model used to assess geographical similarity is: Where i and j represent the pixel to be monitored and the number of a specific ground-based observation station, respectively. This represents the similarity index value. R represents the Euclidean norm, R is the correlation coefficient, and NRMSE is the normalized root mean square error. This represents the geographic environment feature vector.

4. The method according to claim 1, characterized in that, The non-Euclidean neighborhood monitoring station set obtained by partitioning according to the similarity matrix includes: Based on the similarity matrix of each pixel to be monitored, a neighborhood monitoring station is defined for each pixel to be monitored. The neighborhood monitoring station is a number of stations arranged in descending order of the similarity index value. The neighborhood monitoring stations are integrated to obtain a non-Euclidean neighborhood monitoring station set for each pixel to be monitored.

5. The method according to claim 4, characterized in that, The non-Euclidean neighborhood monitoring station set is represented as follows: in, Indicates the pixel to be monitored The corresponding set of non-Euclidean neighborhood monitoring stations; Representing the largest The set of independent variable points with the highest geographical similarity is selected. A function with values; The number of neighboring monitoring stations corresponding to each of the pixels to be monitored. Represents the pixel to be monitored The corresponding similarity matrix, This represents the total number of ground-based observation stations.

6. A system for mapping atmospheric pollutant concentrations that takes into account multi-dimensional geographical similarity, characterized in that, The system includes: The computing unit is used to acquire long-term geographic environment statistics and use similarity calculation methods to measure the geographic scene similarity between the monitored pixels in the monitored area and the ground-based observation station to obtain a similarity matrix. A partitioning unit is used to partition a non-Euclidean neighborhood monitoring station set at the pixel scale based on the similarity matrix; The acquisition unit is used to acquire historical observation data of the area to be monitored, including short-term meteorological reanalysis data and time series of atmospheric pollutant concentration observation data. The construction unit is used to construct the graph topology network structure and training dataset of the area to be monitored based on the non-Euclidean neighborhood monitoring station set and the historical observation data. The training dataset includes a pixel-scale neighborhood monitoring station feature dataset and a station-scale atmospheric pollutant concentration observation target dataset. The training unit is used to train a target mapping model using the graph topology network structure and the training dataset. The target mapping model is a graph attention network model. The target mapping model is used to generate a seamless spatial atmospheric pollutant concentration map with all-day hourly resolution based on real-time observation data of the area to be monitored. The computing unit is specifically used for: The geographic environmental feature vectors of each pixel to be monitored and all ground-based observation stations within the monitoring area are determined. Statistical indicators for evaluating feature differences are calculated based on these geographic environmental feature vectors. When the atmospheric pollutant concentration map is a PM2.5 concentration map, the geographic environmental feature vectors include historical meteorological variable Met, historical PM2.5 concentration, historical AOD data, surface elevation data DEM, land use type data LULC, nighttime light data NLT, and normalized difference vegetation index data NDVI. The sequence of these statistical indicators is input into a similarity calculation method to obtain similarity index values ​​that measure the similarity of geographic environmental scenes. A similarity matrix is ​​generated for each pixel to be monitored within the monitored area based on the similarity index value; The building unit is specifically used for: The geographical similarity between the pixel to be monitored and the neighboring monitoring stations is dynamically calculated based on the non-Euclidean neighborhood monitoring station set and the historical observation data. Using the geographical similarity as edge weights and the geographical features of the pixel to be monitored and the neighboring monitoring stations as nodes, a graph topology network structure for the area to be monitored is constructed. Based on the graph topology network structure and the observed atmospheric pollutant concentration data of the neighboring monitoring stations, a pixel-scale neighboring monitoring station feature dataset is formed, and based on the atmospheric pollutant concentration observation data of the monitoring stations, a station-scale atmospheric pollutant concentration observation target dataset is formed.

7. An apparatus for mapping atmospheric pollutant concentrations that takes into account multi-dimensional geographical similarity, characterized in that, The device includes: Processor, memory, input / output units, and bus; The processor is connected to the memory, the input / output unit, and the bus; The memory stores a program, which the processor invokes to perform the method as described in any one of claims 1 to 5.

8. A computer-readable storage medium having a program stored thereon, the program performing the method as described in any one of claims 1 to 5 when executed on a computer.

Images