Near-surface nitrogen dioxide concentration inversion method based on two-stage ensemble learning

Abstract

The invention discloses a near-surface nitrogen dioxide concentration inversion method based on two-stage ensemble learning. The method comprises the following steps: acquiring ground station nitrogen dioxide, satellite nitrogen dioxide troposphere vertical column concentration and other auxiliary variable data; performing Kriging interpolation processing on the satellite data and other geographical auxiliary variable data; in the first stage, a base model is constructed by adopting ET, RF, GBRT and AdaBoost algorithms, and split node feature weights are dynamically adjusted through a feature-level attention mechanism; in the second stage, prediction results of all the models in the first stage serve as secondary features to be input into the GWR model, a spatial heterogeneity weighting matrix is constructed in combination with attention weight, and dynamic optimization of local regression parameters is achieved; and inverting the near-surface nitrogen dioxide concentration by using a two-stage integrated learning method. According to the method, the accuracy and reliability of near-surface nitrogen dioxide concentration inversion are improved through the two-stage ensemble learning model fusing the attention mechanism.

Description

Technical Field

[0001] This invention belongs to the field of satellite remote sensing technology application technology, specifically relating to a method for inverting near-ground nitrogen dioxide concentration based on two-stage integrated learning. Background Technology

[0002] With the accelerated pace of industrialization and urbanization, air pollution has become an increasingly severe problem, drawing global attention. Nitrogen dioxide (NO2), as a significant air pollutant, has had a substantial negative impact on air quality, human health, and the ecological environment. Regarding air quality, NO2 is a major precursor to secondary pollutants such as photochemical smog and acid rain, reducing visibility, affecting the air quality index, and consequently harming respiratory health, increasing the risk of respiratory diseases. In the ecological environment, NO2 damages plant photosynthesis and physiological functions, affecting the balance and stability of ecosystems. Furthermore, research indicates that NO2 also contributes to pollution formation in power grid transmission lines, potentially affecting the safe and stable operation of the power system.

[0003] Traditional ground-based monitoring stations are an important means of obtaining NO2 concentration data, but they have significant limitations. These stations are scattered and limited in number, mostly concentrated in urban areas, making it difficult to comprehensively and continuously reflect the spatial characteristics of near-surface NO2 concentrations. In contrast, satellite remote sensing technology has significant advantages such as wide coverage and strong spatiotemporal continuity, making it possible to obtain large-scale NO2 concentration information. Satellites can obtain NO2 tropospheric vertical column concentration data; however, there is a complex vertical and horizontal relationship between the satellite-retrieved NO2 tropospheric vertical column concentration and the near-surface NO2 concentration. Various factors, such as atmospheric vertical structure and surface features, can influence this relationship. Therefore, directly using satellite data for near-surface NO2 concentration retrieval has certain limitations and cannot accurately reflect the actual near-surface situation.

[0004] Currently, research on retrieving large-scale near-surface NO2 concentrations using statistical models based on satellite NO2 column concentrations and ground-based NO2 observations has made some progress. Many scholars have used different statistical models to retrieve near-surface NO2 concentrations and achieved some results. However, most of these studies use a single model or select one model after comparing multiple models, failing to fully integrate and leverage the advantages of multi-model inversion. Different models have different characteristics and applicable ranges; a single model may not be able to fully capture the complex relationship between NO2 concentration and various influencing factors, leading to certain errors in the inversion results. Therefore, developing a method that can fully integrate the advantages of multiple models and improve the accuracy and reliability of near-surface NO2 concentration retrieval is of significant practical importance.

[0005] For example, Chinese patent CN112884079A, published on June 1, 2021, discloses a method for estimating near-surface nitrogen dioxide concentration based on a Stacking ensemble model. This method includes: acquiring actual NO2 concentration data from monitoring stations, meteorological monitoring data, geographical data, and population data; acquiring and validating satellite remote sensing data on NO2 tropospheric column concentration; data cleaning, employing a gridded approach to average the actual NO2 concentration data from monitoring stations and the satellite remote sensing data on NO2 tropospheric column concentration, and matching it with the meteorological monitoring data to construct an NO2 dataset; modeling based on the NO2 dataset, using GA-RF, GA-GBRT, and GA-XGBoost models as base learners and ridge regression as a meta-learner to fuse them into a Stacking ensemble model; and using the Stacking ensemble model to simulate and estimate the near-surface NO2 concentration to obtain the near-surface NO2 concentration. However, since NO2 concentration has a strong geospatial effect, and the ridge regression used in this technical solution as the meta-learner cannot capture the nonlinear characteristics of NO2 concentration changing with geospatial variations; moreover, the feature selection process of the base model in this technical solution lacks a dynamic optimization mechanism, resulting in insufficient accuracy and reliability of the estimation results. Summary of the Invention

[0006] The technical problem to be solved by the present invention is to provide a method for inverting near-surface nitrogen dioxide concentration based on two-stage ensemble learning, which addresses the shortcomings of the prior art. By constructing a two-stage ensemble learning model that incorporates an attention mechanism, the geospatial effect of NO2 is fully considered and the features are effectively processed, thereby improving the accuracy and reliability of near-surface NO2 concentration inversion.

[0007] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is: a method for inverting near-ground nitrogen dioxide concentration based on two-stage ensemble learning, the method comprising the following steps: Step S1, Data Acquisition: Acquire NO2 concentration data from ground monitoring stations and NO2 vertical column concentration from the satellite troposphere, as well as other auxiliary variable data; Step S2, Data Preprocessing: Perform Kriging interpolation on satellite data and other geographic auxiliary variable data, and uniformly resample to 500m spatial resolution; Step S3, First stage modeling: The base model is constructed using Extremely Random Tree (ET), Random Forest (RF), Gradient Boosting Decision Tree (GBDT) and Adaptive Boosting (AdaBoost) algorithms, and the feature weights of each model splitting node are dynamically adjusted through a feature-level attention mechanism; Step S4, Second Stage Modeling: The prediction results of each base model in the first stage are used as secondary features to input the Geographically Weighted Regression (GWR) model. The spatial heterogeneity weighting matrix is ​​constructed by combining attention weights to achieve dynamic optimization of local regression parameters. Step S5, Concentration Inversion: Output the near-ground NO2 concentration distribution with a spatial resolution of 500m based on the optimized two-stage ensemble learning model.

[0008] The above-mentioned near-surface nitrogen dioxide concentration inversion method based on two-stage ensemble learning includes other auxiliary variable data such as meteorological elements, normalized difference vegetation index (NDVI), population grid density, land use type, and temporal and spatial elements.

[0009] The above-described near-surface nitrogen dioxide concentration inversion method based on two-stage ensemble learning uses the POMINO-TROPOMI v2.1 satellite to acquire the tropospheric NO2 vertical column concentration. The POMINO-TROPOMI v2.1 satellite data covers an area of ​​70-140°E, 15-55°N, and has a spatial resolution of 0.05°. The inversion calculation is performed using the AMF v6 software package and the LIDORT v3.6 radiative transfer model.

[0010] The above-mentioned near-surface nitrogen dioxide concentration inversion method based on two-stage ensemble learning uses meteorological data from the ERA5-Land reanalysis dataset, including 2m temperature, relative humidity, wind speed, pressure, rainfall, and surface downflow solar radiation, with a spatial resolution of 0.1° and a temporal resolution of 1 hour.

[0011] The above-mentioned near-surface nitrogen dioxide concentration inversion method based on two-stage ensemble learning uses the Normalized Difference Vegetation Index (NDVI) from the MODIS 13A1 dataset with a spatial resolution of 500m and a temporal resolution of 16 days; the population grid density data has a spatial resolution of 1km and a temporal resolution of 1 year; and the land use type data comes from the MODIS Q13A dataset with a spatial resolution of 500m and a temporal resolution of 1 year.

[0012] The aforementioned near-ground nitrogen dioxide concentration inversion method based on two-stage ensemble learning is implemented using the feature-level attention mechanism in step S3 as follows: For each split node of the model, calculate the features. attention weights :

[0013] in, Features With target variable Mutual information entropy, The total number of features, weights Information gain calculation used to adjust node splitting:

[0014] in, The entropy of the target variable, Let be the conditional entropy.

[0015] The above-mentioned method for inverting near-ground nitrogen dioxide concentration based on two-stage ensemble learning, the method for constructing the spatial heterogeneity weighting matrix in step S4 is as follows: For spatial location The GWR model, its weighting matrix Defined as:

[0016] in, It is × A diagonal matrix that reflects the different features in position Spatial weights at each location; This represents the number of secondary features; Indicates the first Secondary features in The normalized weights at each location are calculated using the following formula:

[0017] in, For the first Secondary features at position The average attention weights at each location are generated by the first-stage model through a feature-level attention mechanism. To sum the attention weights for all secondary features, weight normalization is achieved.

[0018] The above-mentioned method for retrieving near-surface nitrogen dioxide concentration based on two-stage ensemble learning is characterized in that the geographically weighted regression (GWR) model in step S4 uses a Gaussian kernel function to calculate spatial weights.

[0019] in For position The regression coefficients, For position and location The spatial weight matrix elements between them Indicates position The dependent variable observation at the location, Indicates position The coefficient of the independent variable at the location, This represents the number of samples.

[0020] Compared with existing technologies, this invention has the following advantages: By constructing a two-stage ensemble learning model that incorporates an attention mechanism, this invention effectively overcomes the limitations of single-model or simple multi-model ensemble methods in existing technologies; compared with traditional ensemble methods that use ridge regression as a meta-learner, the second stage of this invention uses a geographically weighted regression model, which can fully consider the nonlinear characteristics of NO2 concentration variation with geographic space, and achieves dynamic optimization of local regression parameters by constructing a spatially heterogeneous weighted matrix, significantly improving the spatial adaptability of the inversion results; simultaneously, the first stage integrates extreme random trees, random forests, gradient boosting decision trees, and adaptive boosting four... This invention employs a heterogeneous base model and dynamically adjusts the feature weights of each model's split nodes through a feature-level attention mechanism. This allows the model to automatically focus on features more strongly correlated with NO2 concentration during training, solving the problem of lacking dynamic optimization in feature selection in existing technologies. Furthermore, this invention uses the prediction results of each base model in the first stage as secondary features input into the GWR model. This not only preserves the global information capture capability of multiple models but also strengthens the correlation of local features through a spatial weighting mechanism, achieving an organic fusion of global and local information. This effectively improves the accuracy and reliability of near-surface NO2 concentration inversion, and is particularly suitable for monitoring areas with significant spatial heterogeneity.

[0021] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0022] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0023] like Figure 1 As shown, the near-surface nitrogen dioxide concentration inversion method based on two-stage ensemble learning of the present invention includes the following steps: Step S1, Data Acquisition: Acquire NO2 concentration data from ground monitoring stations and NO2 vertical column concentration from the satellite troposphere, as well as other auxiliary variable data; Step S2, Data Preprocessing: Perform Kriging interpolation on satellite data and other geographic auxiliary variable data, and uniformly resample to 500m spatial resolution; Step S3, First stage modeling: The base model is constructed using Extremely Random Tree (ET), Random Forest (RF), Gradient Boosting Decision Tree (GBDT) and Adaptive Boosting (AdaBoost) algorithms, and the feature weights of each model splitting node are dynamically adjusted through a feature-level attention mechanism; Step S4, Second Stage Modeling: The prediction results of each base model in the first stage are used as secondary features to input the Geographically Weighted Regression (GWR) model. The spatial heterogeneity weighting matrix is ​​constructed by combining attention weights to achieve dynamic optimization of local regression parameters. Step S5, Concentration Inversion: Output the near-ground NO2 concentration distribution with a spatial resolution of 500m based on the optimized two-stage ensemble learning model.

[0024] In this embodiment, the other auxiliary variable data include meteorological elements, normalized difference vegetation index (NDVI), population grid density, land use type, and temporal and spatial elements.

[0025] In this embodiment, the NO2 concentration data of the ground monitoring station mentioned in step S1 comes from the China National Environmental Monitoring Centre. The near-ground NO2 concentration is determined by chemiluminescence method. In order to match the passing time of the Sentinel-5P satellite TROPOMI sensor at about 13:30 local time, the observation values ​​at 13:00 and 14:00 every day are selected, and the arithmetic mean is calculated as the model label value. In this embodiment, the tropospheric NO2 vertical column concentration is acquired using the POMINO-TROPOMI v2.1 satellite. The POMINO-TROPOMI v2.1 satellite data covers an area of ​​70-140°E and 15-55°N, with a spatial resolution of 0.05°. The inversion calculation is performed using the AMF v6 software package and the LIDORT v3.6 radiative transfer model. The POMINO-TROPOMI algorithm is an optimization and improvement based on the POMINO algorithm previously developed for ozone monitoring instruments (OMI). The main adjustments are made to instrument-related issues, such as the available spectral coverage, spatial resolution, and details of cloud detection. In this embodiment, the meteorological data comes from the ERA5-Land reanalysis dataset, which is a subset of the ERA5 reanalysis dataset, but with higher resolution output for land surface variables. The selected parameters include 2m temperature, relative humidity, wind speed, pressure, rainfall, and downward solar radiation at the surface, with a spatial resolution of 0.1° and a temporal resolution of 1 hour. Similar to the NO2 concentration data from ground monitoring stations, the arithmetic mean of the data at 13:00 and 14:00 Beijing time is selected as the model input feature. In this embodiment, the Normalized Difference Vegetation Index (NDVI) is from the MODIS 13A1 dataset, with a spatial resolution of 500m and a temporal resolution of 16 days; the population grid density data has a spatial resolution of 1km and a temporal resolution of 1 year; and the land use type data is from the MODIS Q13A dataset, with a spatial resolution of 500m and a temporal resolution of 1 year. In this embodiment, the time and space elements include the number of working days, the number of days in a year, longitude, and latitude.

[0026] In this embodiment, step S2 involves performing Kriging interpolation on satellite data and other geographic auxiliary variable data, uniformly resampling to a spatial resolution of 500m, and calculating the unknown location using the Kriging interpolation formula based on the known spatial location and values ​​of the data points. The estimated value The Kriging interpolation formula is:

[0027] in, The weighting coefficients are determined through semi-variogram optimization. Representative position Observations at that location This represents the number of neighboring observation points involved in the interpolation.

[0028] In this embodiment, the first stage of training each base model in step S3 is as follows: Step S301, Base Model Selection and Construction: Construct four base models respectively: Extreme Random Tree (ET), Random Forest (RF), Gradient Boosting Decision Tree (GBDT) and Adaptive Boosting (AdaBoost), and initialize the hyperparameters of each model, including the number of trees, maximum depth, learning rate and minimum number of samples for split nodes; Step S302, Feature-level attention mechanism implementation: For each split node of the model, calculate the features. attention weights :

[0029] in, Features With target variable Mutual information entropy, The total number of features, weights Information gain calculation used to adjust node splitting:

[0030] in, The entropy of the target variable, Let be the conditional entropy.

[0031] Step S303: Hyperparameter optimization of each base model: A gridded optimization strategy is adopted to set the number of trees, maximum depth, learning rate and minimum number of samples for split nodes in the search space. The performance of the parameter combination is evaluated based on ten-fold cross-validation. The root mean square error (RMSE) is used as the optimization index to select the optimal hyperparameter combination for each model. Step S304, Cross-validation Training and Prediction: The base model is trained using a ten-fold cross-validation framework. The main process includes: The training set is randomly divided into 10 mutually exclusive subsets; Nine subsets are polled to train the base model with attention mechanism, and the remaining subset is used for validation; Repeat 10 times to ensure all data are used in the validation, and record the average RMSE and R. 2 index; Retrain the model using the optimal hyperparameter combination and the full training data; Each base model outputs preliminary NO2 concentration predictions, which serve as input features for the second-stage GWR.

[0032] In this embodiment, the second-stage model training process in step S4 is as follows: Step S401, Secondary Feature Input Construction: The preliminary prediction results of NO2 concentration output by the four base models trained in the first stage are used as secondary features to form a feature vector; Step S402, Generation of spatial heterogeneity weighted matrix: For spatial location The GWR model, its weighting matrix Defined as:

[0033] in, It is × A diagonal matrix that reflects the different features in position Spatial weights at each location; This represents the number of secondary features; Indicates the first Secondary features in The normalized weights at each location are calculated using the following formula:

[0034] in, For the first Secondary features at position The average attention weights at each location are generated by the first-stage model through a feature-level attention mechanism. To sum the attention weights for all secondary features, weight normalization is achieved; Step S403, Geographically Weighted Regression (GWR) Modeling: Spatial weights are calculated using the Gaussian kernel function.

[0035] in For position The regression coefficients, For position and location The spatial weight matrix elements between them Indicates position The dependent variable observation at the location, Indicates position The coefficient of the independent variable at the location, For the sample size, For the regression coefficients, this formula estimates the regression coefficients at each location by minimizing the weighted sum of squared residuals.

[0036] In practice, , Indicates position and location The Euclidean distance between them This represents the bandwidth parameter. The formula calculates spatial weights based on the Gaussian kernel function. The closer the distance, the greater the weight, and the farther the distance, the smaller the weight.

[0037] In this embodiment, step S5, concentration inversion: outputs the near-ground NO2 concentration distribution with a spatial resolution of 500m based on the optimized two-stage ensemble learning model.

[0038] In summary, this invention overcomes the limitations of traditional single-model and simple multi-model fusion methods in near-surface NO2 concentration inversion. It effectively solves the problems of scattered coverage and limited number of traditional ground monitoring stations, which make it difficult to comprehensively reflect spatial characteristics, and the inaccuracy caused by complex vertical and horizontal relationships affecting near-surface NO2 concentration inversion using satellite data. Regarding model construction, it innovatively adopts a two-stage ensemble learning strategy. In the first stage, a feature-level attention mechanism dynamically adjusts the feature weights of the split nodes of each base model, fully leveraging the advantages of various machine learning algorithms. In the second stage, the prediction results of the base models are used as secondary features input to a geographic weighted regression (GW) system. The R model, combined with attention weights to construct a spatial heterogeneity weighted matrix, achieves dynamic optimization of local regression parameters, fully considers the geospatial effects of NO2, and can accurately capture the nonlinear characteristics of NO2 concentration variation with geospatial data. At the same time, it dynamically optimizes the feature processing process, significantly improving the accuracy and reliability of near-surface NO2 concentration inversion. While taking into account high-precision inversion and effective utilization of multi-source data features, it achieves a fine characterization of the spatial distribution of near-surface NO2 concentration, which has significant technological progress and wide application value. It can be widely used in many fields such as air quality monitoring, ecological environment assessment, and power system safety and stability operation assurance.

[0039] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any way. Any simple modifications, alterations, or equivalent structural changes made to the above embodiments based on the technical essence of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A method for inverting near-ground nitrogen dioxide concentration based on two-stage ensemble learning, characterized in that, The method includes the following steps: Step S1, Data Acquisition: Acquire NO2 concentration data from ground monitoring stations and NO2 vertical column concentration from the satellite troposphere, as well as other auxiliary variable data; Step S2, Data Preprocessing: Perform Kriging interpolation on satellite data and other geographic auxiliary variable data, and uniformly resample to 500m spatial resolution; Step S3, First stage modeling: The base model is constructed using extreme random trees, random forests, gradient boosting decision trees and adaptive boosting algorithms, and the feature weights of each model splitting node are dynamically adjusted through a feature-level attention mechanism; Step S4, Second Stage Modeling: The prediction results of each base model in the first stage are used as secondary features and input into the geographic weighted regression model. The spatial heterogeneity weighting matrix is ​​constructed by combining attention weights to achieve dynamic optimization of local regression parameters. Step S5, Concentration Inversion: Output the near-ground NO2 concentration distribution with a spatial resolution of 500m based on the optimized two-stage ensemble learning model.

2. The near-surface nitrogen dioxide concentration inversion method based on two-stage ensemble learning according to claim 1, characterized in that, Other auxiliary variable data include meteorological elements, normalized vegetation index, population grid density, land use type, and temporal and spatial elements.

3. The near-surface nitrogen dioxide concentration inversion method based on two-stage ensemble learning according to claim 1 or 2, characterized in that, The tropospheric NO2 vertical column concentration was obtained using the POMINO-TROPOMI v2.1 satellite, which covers an area of ​​70-140°E, 15-55°N, with a spatial resolution of 0.05°. The inversion calculation was performed using the AMF v6 software package and the LIDORT v3.6 radiative transfer model.

4. The near-surface nitrogen dioxide concentration inversion method based on two-stage ensemble learning according to claim 2, characterized in that, The meteorological data are derived from the ERA5-Land reanalysis dataset, including 2m temperature, relative humidity, wind speed, pressure, rainfall, and surface downdraft solar radiation, with a spatial resolution of 0.1° and a temporal resolution of 1 hour.

5. The near-surface nitrogen dioxide concentration inversion method based on two-stage ensemble learning according to claim 2, characterized in that, The normalized vegetation index is from the MODIS 13A1 dataset, with a spatial resolution of 500m and a temporal resolution of 16 days; the population grid density data has a spatial resolution of 1km and a temporal resolution of 1 year; the land use type data is from the MODIS Q13A dataset, with a spatial resolution of 500m and a temporal resolution of 1 year.

6. The near-surface nitrogen dioxide concentration inversion method based on two-stage ensemble learning according to claim 1 or 2, characterized in that, The feature-level attention mechanism in step S3 is implemented as follows: For each split node of the model, calculate the features. Attention weights : in, Features With target variable Mutual information entropy, The total number of features, weights Information gain calculation used to adjust node splitting: in, The entropy of the target variable, Let be the conditional entropy.

7. The near-surface nitrogen dioxide concentration inversion method based on two-stage ensemble learning according to claim 1, characterized in that, The method for constructing the spatial heterogeneity weighted matrix in step S4 is as follows: For spatial location The GWR model, its weighting matrix Defined as: in, It is × A diagonal matrix that reflects the different features in position Spatial weights at each location; The number of secondary features; Indicates the first Secondary features in The normalized weights at each location are calculated using the following formula: in, For the first Secondary features at position The average attention weights at each location are generated by the first-stage model through a feature-level attention mechanism. To sum the attention weights for all secondary features, weight normalization is achieved.

8. The near-surface nitrogen dioxide concentration inversion method based on two-stage ensemble learning according to claim 1, characterized in that, The geographically weighted regression model in step S4 uses a Gaussian kernel function to calculate spatial weights: in For position The regression coefficients, For position and location The spatial weight matrix elements between them Indicates position The dependent variable observation at the location, Indicates position The coefficient of the independent variable at the location, This represents the number of samples.

Images