CN-122025172-A - Ground ozone inversion method based on machine learning and hyperspectral satellite observation

CN122025172ACN 122025172 ACN122025172 ACN 122025172ACN-122025172-A

Abstract

The invention discloses a ground ozone inversion method based on machine learning and hyperspectral satellite observation, which belongs to the technical field of ground ozone inversion and is used for ground ozone inversion and comprises the steps of obtaining data, selecting ozone sensitive wave bands, inputting a model, training the model, verifying the model and carrying out ground ozone inversion composition, wherein the ozone sensitive wave band selection comprises the steps of constructing a feature matrix and a target variable, carrying out feature importance calculation in ERT regression, obtaining important features and sequencing according to the importance to obtain 22 optimal frequency bands, directly inverting ground ozone information from hyperspectral observation of OMI satellites, avoiding errors possibly caused by intermediate products, automatically determining the ozone sensitive wave bands through importance contribution analysis, and fusing the ozone sensitive wave bands with a spectrum physical principle of ozone absorption.

Inventors

SUN LIN
FAN YULONG
WANG ZAIFEI

Assignees

山东科技大学

Dates

Publication Date: 20260512
Application Date: 20260128

Claims (6)

1. The ground ozone inversion method based on machine learning and hyperspectral satellite observation is characterized by comprising the steps of acquiring data, selecting ozone sensitive wave bands, inputting a model, training the model, verifying the model and carrying out ground ozone inversion composition; the ozone sensitive wave band selection comprises the steps of constructing a feature matrix and a target variable, calculating feature importance in ERT regression, obtaining important features, and sequencing according to the importance to obtain 22 optimal frequency bands; The feature importance calculation in ERT regression comprises the steps of carrying out random selection and a random threshold value in the node 1 to obtain the node 2 and the node 3, wherein the random threshold value comprises the steps of calculating MSE, carrying out normalization and executing node splitting; the model input includes the surface of 22 optimal frequency bands Measuring and geometrically observing to obtain surface characteristics and weather results; the model verification comprises 10 times of cross verification, overall accuracy is obtained according to sample dimensions, and prediction accuracy is obtained according to spatial dimensions and time dimensions.
2. The machine learning and hyperspectral satellite observation based ground ozone inversion method of claim 1 wherein acquiring data includes acquiring ground ozone observation data, on-board OMI hyperspectral instrument ozone observation data, surface and atmospheric assistance data.
3. The machine learning and hyperspectral satellite observation based ground ozone inversion method of claim 1 wherein ERT regression comprises: ; In the formula, Represent the first Importance score of each wave band, measure the first The degree of contribution of the individual bands to the prediction results of the extremely random tree model, Is the total number of decision trees contained in the extreme stochastic tree model, Is the first The number of decision trees is chosen such that, Is the first The number of nodes in the network is, To instruct the function, judge the node Whether or not the splitting characteristic of (2) is the first The characteristics of the device are that, Is the total number of nodes to be programmed, Is a reduced amount of non-purity.
4. The machine learning and hyperspectral satellite observation based ground ozone inversion method of claim 1 wherein model training comprises constructing each decision tree based on a complete dataset, randomly generating a plurality of splitting thresholds in a random subset of candidate features at node splitting, selecting a threshold that minimizes mean square error MSE to complete node splitting.
5. The method for ground ozone inversion based on machine learning and hyperspectral satellite observation according to claim 4 wherein the training set is set as the model training , In order to input the feature vector(s), Is a corresponding target variable, is equivalent to the ground ozone concentration true value, and in the node splitting process of each decision tree, the extreme random tree randomly selects a feature subset For each feature in the subset of features, randomly generating a set of fragmentation thresholds And calculates the mean square error of the left and right subsets obtained after splitting by this threshold.
6. The ground ozone inversion method based on machine learning and hyperspectral satellite observation according to claim 5 wherein the mean square error is: ; ; ; In the formula, Representing the sample size of the left subset, Representing the sample size of the subset on the right, Is the first The characteristics of the device are that, 、 Is a set of two threshold values for the filter, Represents the average of ground concentration truth values for the left subset, Mean value of ground concentration truth for right subset is presented.

Description

Ground ozone inversion method based on machine learning and hyperspectral satellite observation Technical Field The invention discloses a ground ozone inversion method based on machine learning and hyperspectral satellite observation, and belongs to the technical field of ground ozone inversion. Background Ground ozone (O 3) pollution poses a significant threat to ecosystems and public health, and while satellite remote sensing provides valuable area-scale ozone monitoring capability and effectively supplements sparse ground stations, current inversion algorithms are still limited by the significant reliance on total column ozone and ozone precursor products (e.g., nitrogen oxides and volatile organic compounds). These products inherit significant uncertainties of conventional radiation delivery methods, rely on assumptions on surface properties and atmospheric conditions, and a priori knowledge of ozone distribution. Disclosure of Invention The invention aims to provide a ground ozone inversion method based on machine learning and hyperspectral satellite observation, which aims to solve the problem of low ground ozone inversion precision in the prior art. The ground ozone inversion method based on machine learning and hyperspectral satellite observation comprises the steps of obtaining data, selecting ozone sensitive wave bands, inputting models, training the models, verifying the models and carrying out ground ozone inversion composition; the ozone sensitive wave band selection comprises the steps of constructing a feature matrix and a target variable, calculating feature importance in ERT regression, obtaining important features, and sequencing according to the importance to obtain 22 optimal frequency bands; The feature importance calculation in ERT regression comprises the steps of carrying out random selection and a random threshold value in the node 1 to obtain the node 2 and the node 3, wherein the random threshold value comprises the steps of calculating MSE, carrying out normalization and executing node splitting; the model input includes the surface of 22 optimal frequency bands Measuring and geometrically observing to obtain surface characteristics and weather results; the model verification comprises 10 times of cross verification, overall accuracy is obtained according to sample dimensions, and prediction accuracy is obtained according to spatial dimensions and time dimensions. The data acquisition comprises the steps of acquiring ground ozone observation data and satellite-borne OMI hyperspectral instrument ozone observation data and ground surface and atmosphere auxiliary data. ERT regression includes: ; In the formula, Represent the firstImportance score of each wave band, measure the firstThe degree of contribution of the individual bands to the prediction results of the extremely random tree model,Is the total number of decision trees contained in the extreme stochastic tree model,Is the firstThe number of decision trees is chosen such that,Is the firstThe number of nodes in the network is,To instruct the function, judge the nodeWhether or not the splitting characteristic of (2) is the firstThe characteristics of the device are that,Is the total number of nodes to be programmed,Is a reduced amount of non-purity. Model training includes constructing each decision tree based on a complete data set, randomly generating a plurality of splitting thresholds in a random subset of candidate features during node splitting, and selecting a threshold capable of minimizing a mean square error MSE to complete node splitting. During model training, the training set is set as,In order to input the feature vector(s),Is a corresponding target variable, is equivalent to the ground ozone concentration true value, and in the node splitting process of each decision tree, the extreme random tree randomly selects a feature subsetFor each feature in the subset of features, randomly generating a set of fragmentation thresholdsAnd calculates the mean square error of the left and right subsets obtained after splitting by this threshold. The mean square error is: ; ; ; In the formula, Representing the sample size of the left subset,Representing the sample size of the subset on the right,Is the firstThe characteristics of the device are that,、Is a set of two threshold values for the filter,Represents the average of ground concentration truth values for the left subset,Mean value of ground concentration truth for right subset is presented. Compared with the prior art, the method has the advantages that ground ozone information is inverted directly from hyperspectral observation of OMI satellites, errors possibly caused by intermediate products are avoided, an ozone sensitive wave band is automatically determined through importance contribution analysis and is fused with a spectrum physical principle of ozone absorption, and the accuracy and the interpretability of ground ozone estimation are effectively improved. Draw