CN116189796A - Machine learning-based satellite-borne short wave infrared CO 2 Column concentration estimation method - Google Patents

Abstract

The invention discloses a machine learning-based satellite-borne short wave infrared CO ₂ A method of estimating column concentration comprising: s1, extracting OCO-2 satellite wave band data to obtain 9 weakCO ₂ Band and 6O ₂ Band data; s2, 9 weakCO ₂ Band and 6O ₂ Performing feature screening on the wave band data, the NDVI normalized vegetation index, the SR surface reflectivity data, the DEM elevation topography data, the ERA5 atmosphere data, the AOD aerosol data and the TCCON station observation data, and reserving the first 31 features of screening according to importance; s3, performing correlation analysis on the first 31 screened features through a heat map to find out the correlation with CO ₂ Features of stronger correlation and weaker features of column concentration; s4, will be combined with CO ₂ Combining the characteristic with strong correlation with weak characteristic, inputting to five regression models with integrated learning, and predicting accuracy by using the different modelsComparative analysis, determining coefficient R of extreme random forest regression model ² Highest, least error, best prediction effect.

Description

Spaceborne shortwave infrared CO2 column concentration estimation method based on machine learning

Technical Field

The present invention relates to the technical field of atmospheric satellite remote sensing prediction, and in particular to a satellite-borne shortwave infrared _CO2 column concentration estimation method based on machine learning.

Background Art

_CO2 is the main greenhouse gas in the atmosphere and has a very important impact on global climate change. Since the industrial age, the concentration of _CO2 has increased to about 30% and has maintained a trend of continuous growth. Therefore, monitoring _CO2 is of great significance for the study of global warming; therefore, accurately grasping the _CO2 content and its changes in the atmosphere can provide support for climate prediction and environmental decision-making.

Although traditional ground-based atmospheric CO ₂ detection methods have the advantages of high accuracy and reliability, they are all single-point measurements and lack the ability to detect large-scale regional and global real-time detection. Therefore, it is imperative to develop satellite observation methods and technologies for CO _2. The short-wave infrared band is more sensitive to near-ground CO ₂ , so it is more suitable for monitoring the dynamic changes of ground carbon sources and sinks.

At present, most short-wave infrared CO ₂ observation data in the world use full physical inversion algorithms, which require simulation of the entire optical path, and the calculation of the radiation transfer equation is complex and time-consuming. Due to the complex influence of aerosols, water vapor and surface reflectivity on the short-wave infrared radiation process, the existing physical inversion model requires many input parameters and has uncertainty.

Summary of the invention

The purpose of the present invention is to estimate the _CO2 column concentration by using decision tree, XGBoost, ordinary random forest, extreme random forest and gradient boosting regression model respectively, and then compare and analyze to find the model with the highest estimation accuracy, and estimate the atmospheric _CO2 column concentration from satellite remote sensing. This method has the advantages of high prediction accuracy and strong interpretability, and greatly improves the prediction efficiency.

To achieve the above purpose, the technical solution of the present application is: a spaceborne shortwave infrared _CO2 column concentration estimation method based on machine learning, comprising:

S1. Obtain OCO-2 satellite band data, extract the OCO-2 satellite band data through sensitivity analysis of atmospheric carbon dioxide inversion parameters, and obtain 9 weak_CO ₂ bands and 6 O ₂ band data;

S2. The 9 weak_CO ₂ bands and 6 O ₂ bands data, NDVI normalized vegetation index, SR surface reflectance data, DEM elevation terrain data, ERA5 atmospheric data, AOD aerosol data, and TCCON station observation data were used for feature screening, and the top 31 features were retained according to importance;

S3. Perform correlation analysis on the first 31 features screened through heat maps to find out the features with strong and weak correlation with _CO2 column concentration;

S4. The features with strong and weak correlation with _CO2 column concentration are merged as the input feature data set, and then the decision tree, XGBoost, ordinary random forest, extreme random forest and gradient boosting regression models are used to estimate the average column concentration of _CO2 . By comparing and analyzing the determination coefficient ^R2 , root mean square error RMSE, mean absolute error MAE, mean relative error MRE estimated by different regression models and the prediction accuracy within the allowable error range, the model with the highest prediction accuracy is found to be the extreme random forest regression model, and the extreme random forest regression model is used to predict the average _CO2 column concentration.

Furthermore, the OCO-2 satellite band data includes longitude lon, latitude lat, zenith angle and azimuth of the sun, zenith angle and azimuth of the satellite; the ERA5 atmospheric data includes temperature, humidity, pressure, U/V component of wind, rainfall, boundary layer height (blh), cloud base height (cbh), cloud cover (tcc), total rainfall (tp), and vertical speed of wind.

Furthermore, before extracting the OCO-2 satellite band data, the resampling method is used to determine the extraction range, that is, the grid is drawn according to the latitude and longitude range of the target area, and the resolution after sampling is set to 0.5°×0.5°. Through the longitude and latitude of each grid, the Euclidean distance between the center point of each grid and the center point of each pixel corresponding to the original image is obtained:

Where lon _k is the longitude of the fixed site, lat _k is the latitude of the fixed site, lon _i and lat _i are the longitude and latitude of the grid respectively.

Furthermore, the outliers in the 9 weak_CO ₂ bands and 6 O ₂ bands are processed as follows:

Where σ is the standard deviation of the data for the day, that is, all outliers outside ±3σ are eliminated, and the average of the data of each band measured multiple times at each station every day is taken.

Furthermore, the decision tree uses the Gini index to divide attributes. Assuming that the proportion of the k-th class of samples in the current sample set X is p _k (k = 1, 2, 3, ..., y), the Gini value is:

Gini(X) indicates the possibility of random sampling of inconsistency between two different types of labels; Gini impurity refers to the probability of the sample being selected multiplied by the probability of error. The smaller Gini(X), the higher the purity of the sample set X; when all samples in a node are of the same class, Gini impurity is 0.

Assuming that the discrete attribute a has v possible values, if a is used to classify the sample set X, v branch nodes will be generated. Let _Xv be the vth branch node containing all samples in the sample set X that have values on attribute a. Then the Gini index of attribute a is defined as:

The Gini index Gini(X, A) represents the uncertainty of the sample set X after the partition by A=a; the larger the Gini index, the greater the uncertainty of the sample.

表示为：Furthermore, XGBoost assumes that there are K trees in total, and F represents the tree model, then the predicted value

It is expressed as:

Where _xi is the input instance, representing the feature vector of the i-th data point; K is the number of CART trees; _fk represents the k-th CART tree;

The corresponding objective function L is:

Where l is the loss function, which represents the error between the predicted value and the true value; _yi is the true value; Ω is the regularization function to prevent the model from overfitting.

Furthermore, in ordinary random forests, for the feature parameter set X of the data set, a model h(X,θ _i ), i=1,2,…,k is established, and m features are randomly selected so that each leaf node selects the feature with the maximum information gain for splitting; the information gain is expressed as:

为第m个划分叶节点的权重值。Where i is the regression value, _pi represents the probability of the corresponding value, and w is the number of partition nodes.

is the weight value of the mth partition leaf node.

Furthermore, in extreme random forests, assuming that the generalization error of an individual learner is E _i , the weighted value of the generalization error of the learner is:

Assuming that the divergence value of the individual learner is A _i , the weighted divergence value of the learner is:

The generalization error after integration is expressed as:

Where _wi is the weight and T is the total number of decision trees with different structures.

Furthermore, the new learner obtained in each iteration of gradient boosting is fitted to the residual of the previous learner, and finally the predictions of all trees are added together to complete the prediction task; the residual is obtained as follows:

r _ni = _yi -f _n-1 ( _xi )

In the formula, _yi is the measured value of the ith sample, and fn _-1 ( _xi ) is the predicted value of the previous round of learner. The residual memory is fitted to obtain a fitted residual model _hn (x), and the regression tree is updated:

f _n (x) = f _n-1 (x) + _hn (x)

Furthermore, the determination coefficient R ² , root mean square error RMSE, mean absolute error MAE, and mean relative error MRE are obtained as follows:

为平均值。Where: N is the number of samples; _fi is the predicted value; _yi is the true value;

is the average value.

Due to the adoption of the above technical scheme, the present invention can achieve the following technical effects: This method uses different integrated learning methods to predict the average concentration of atmospheric _CO2 column through satellite, vegetation, surface topography, atmosphere, aerosol data, etc., which has practical significance. Compared with the traditional physical inversion method, this method considers sufficient features, is easy to modify, easy to explain, simple to operate, and greatly improves the prediction efficiency. The _CO2 column concentration can be better predicted, so that the integrated learning model can predict the _CO2 column concentration of different satellites more accurately, providing data support for the decision-making of the environmental protection department.

BRIEF DESCRIPTION OF THE DRAWINGS

Figure 1 is the spectrum observed by the OCO-2 satellite. The observation sample is from the Tsukuba station (36.05N°, 140.12E°) on January 1, 2019. (a) is the weak absorption band of CO ₂ , and (b) is the absorption band of O2-A.

Figure 2 is a bar chart of the importance of the first 31 features;

Figure 3 is a correlation diagram between various influencing factors in the satellite inversion of the average CO ₂ column concentration;

Figure 4 is a diagram showing the training results of five prediction models;

Figure 5 is a graph showing the difference between the predicted CO ₂ column concentration and the true value of the test set of five prediction models;

Figure 6 is a graph showing the influence of the prediction performance of the extreme random forest regression model on its own parameters;

Figure 7 is a flow chart of the spaceborne shortwave infrared CO ₂ column concentration estimation method based on machine learning.

DETAILED DESCRIPTION

The embodiments of the present invention are implemented on the premise of the technical solution of the present invention, and detailed implementation methods and specific operation processes are given, but the protection scope of the present invention is not limited to the following embodiments.

Example 1

This embodiment provides a spaceborne shortwave infrared CO ₂ column concentration estimation method based on machine learning, including:

S1. Obtain OCO-2 satellite band data, and extract the OCO-2 satellite band data through sensitivity analysis of atmospheric carbon dioxide inversion parameters; since the CO ₂ weak absorption band is greatly affected by water vapor, the corresponding absorption channel is selected at the strong absorption band (1.61 μm) to obtain 9 weak_CO ₂ bands and 6 O ₂ bands data;

The OCO-2 satellite band data includes longitude lon, latitude lat, the zenith angle and azimuth of the sun, and the zenith angle and azimuth of the satellite.

It should be noted that before extracting the OCO-2 satellite band data, the extraction range is determined by resampling, that is, a grid is drawn according to the latitude and longitude range of the target area, and the resolution after sampling is set to 0.5°×0.5°. Through the longitude and latitude of each grid, the Euclidean distance between the center point of each grid and the center point of each pixel corresponding to the original image is obtained:

Where lon _k is the longitude of the fixed site, lat _k is the latitude of the fixed site, lon _i and lat _i are the longitude and latitude of the grid respectively.

Preferably, the outliers in the 9 weak_CO ₂ bands and 6 O ₂ bands are processed as follows:

Where σ is the standard deviation of the data for the day, that is, all outliers outside ±3σ are eliminated, and the average of the data of each band measured multiple times at each station every day is taken.

S2. The 9 weak_CO ₂ bands and 6 O ₂ bands data, NDVI normalized vegetation index, SR surface reflectance data, DEM elevation terrain data, ERA5 atmospheric data, AOD aerosol data, and TCCON station observation data were used for feature screening, and the top 31 features were retained according to importance;

Specifically, feature screening is performed through the above-mentioned resampling method.

S3. Perform correlation analysis on the first 31 features screened through heat maps to find out the features with strong and weak correlation with _CO2 column concentration;

S4. Merge the features with strong correlation with _CO2 column concentration and the features with weak correlation with CO2 column concentration, input them into five ensemble learning regression models, and output the predicted _CO2 column concentration respectively;

Specifically, in extreme random forests, assuming that the generalization error of an individual learner is E _i , the weighted value of the generalization error of the learner is:

Assuming that the divergence value of the individual learner is A _i , the weighted divergence value of the learner is:

The generalization error after integration can be expressed as:

Where _wi is the weight and T is the total number of decision trees with different structures.

Using these different models to predict the accuracy, through comparative analysis, the extreme random forest regression model has the highest coefficient of determination ^R2 , the smallest error, and the best prediction effect, which is significantly better than the prediction results of other models; the relevant data of the four evaluation indicators are shown in the following table:

Table 1 Related data of four evaluation indicators

The embodiments of the present invention have better practicability and are not intended to limit the present invention in any form. The technical features or combinations of technical features described in the embodiments of the present invention should not be considered isolated, and they can be combined with each other to achieve better technical effects. The scope of the preferred embodiments of the present invention may also include other implementations, and should be understood by those skilled in the art of the invention embodiments.

Claims (10)

1. A spaceborne shortwave infrared CO ₂ column concentration estimation method based on machine learning, characterized by comprising: S1. Obtain OCO-2 satellite band data, extract the OCO-2 satellite band data through sensitivity analysis of atmospheric carbon dioxide inversion parameters, and obtain 9 weak_CO ₂ bands and 6 O ₂ band data; S2. The 9 weak_CO ₂ bands and 6 O ₂ bands data, NDVI normalized vegetation index, SR surface reflectance data, DEM elevation terrain data, ERA5 atmospheric data, AOD aerosol data, and TCCON station observation data were used for feature screening, and the top 31 features were retained according to importance; S3. Perform correlation analysis on the first 31 features screened through heat maps to find out the features with strong and weak correlation with _CO2 column concentration; S4. The features with strong and weak correlation with _CO2 column concentration are merged as the input feature data set, and then the decision tree, XGBoost, ordinary random forest, extreme random forest and gradient boosting regression models are used to estimate the average column concentration of _CO2 . By comparing and analyzing the determination coefficient ^R2 , root mean square error RMSE, mean absolute error MAE, mean relative error MRE estimated by different regression models and the prediction accuracy within the allowable error range, the model with the highest prediction accuracy is found to be the extreme random forest regression model, and the extreme random forest regression model is used to predict the average _CO2 column concentration. 2. According to the method for estimating the concentration of a satellite-borne shortwave infrared _CO2 column based on machine learning in claim 1, it is characterized in that the OCO-2 satellite band data includes longitude lon, latitude lat, zenith angle and azimuth of the sun, zenith angle and azimuth of the satellite; the ERA5 atmospheric data includes temperature, humidity, pressure, U/V component of wind, rainfall, boundary layer height, cloud base height, cloud cover, total rainfall, and vertical speed of wind. 3. The satellite-borne shortwave infrared _CO2 column concentration estimation method based on machine learning according to claim 1 is characterized in that the extraction range is determined by resampling before extracting the OCO-2 satellite band data, that is, a grid is drawn according to the longitude and latitude range of the target area, and the resolution after sampling is set to 0.5°×0.5°. The Euclidean distance between the center point of each grid and the center point of each pixel corresponding to the original image is obtained by the longitude and latitude of each grid:

Where lon _k is the longitude of the fixed site, lat _k is the latitude of the fixed site, lon _i and lat _i are the longitude and latitude of the grid respectively. 4. According to the method for estimating the spaceborne shortwave infrared CO ₂ column concentration based on machine learning in claim 1, it is characterized in that the abnormal values in the 9 weak_CO ₂ bands and 6 O ₂ bands are processed as follows:

In the formula, σ is the standard deviation of the data for that day, that is, all outliers outside ±3σ are eliminated. 5. According to the method for estimating the concentration of shortwave infrared CO ₂ column based on spaceborne machine learning in claim 1, it is characterized in that the decision tree uses the Gini index to divide the attributes. Assuming that the proportion of the k-th class of samples in the current sample set X is p _k (k=1,2,3,…,y), the Gini value is:

Gini(X) indicates the probability of random sampling of inconsistencies between two different types of labels; Assuming that the discrete attribute a has v possible values, if a is used to classify the sample set X, v branch nodes will be generated. Let _Xv be the vth branch node containing all samples in the sample set X that have values on attribute a. Then the Gini index of attribute a is defined as:

The Gini index Gini(X, A) represents the uncertainty of the sample set X after the partition by A=a; the larger the Gini index, the greater the uncertainty of the sample. 6. According to the method for estimating the spaceborne shortwave infrared CO ₂ column concentration based on machine learning in claim 1, it is characterized in that, in XGBoost, it is assumed that there are a total of K trees, F represents the tree model, and the predicted value

It is expressed as:

Where _xi is the input instance, representing the feature vector of the i-th data point; K is the number of CART trees; _fk represents the k-th CART tree; The corresponding objective function L is:

Where l is the loss function, which represents the error between the predicted value and the true value; _yi is the true value; Ω is the regularization function to prevent the model from overfitting. 7. The method for estimating the spaceborne shortwave infrared CO ₂ column concentration based on machine learning according to claim 1 is characterized in that, in a common random forest, for the feature parameter set X of the data set, a model h(X,θ _i ), i=1,2,…,k is established, and m features are randomly selected so that each leaf node selects the feature with the maximum information gain for splitting; wherein the information gain is expressed as:

Where i is the regression value, _pi represents the probability of the corresponding value, and w is the number of partition nodes.

is the weight value of the mth partition leaf node. 8. The method for estimating the spaceborne shortwave infrared CO ₂ column concentration based on machine learning according to claim 1, characterized in that, in the extreme random forest, assuming that the generalization error of the individual learner is E _i , the weighted value of the generalization error of the learner is:

Assuming that the divergence value of the individual learner is A _i , the weighted divergence value of the learner is:

The generalization error after integration is expressed as:

Where _wi is the weight and T is the total number of decision trees with different structures. 9. The method for estimating the concentration of _CO2 column from satellite-borne shortwave infrared based on machine learning according to claim 1 is characterized in that the new learner obtained in each iteration of gradient boosting is fitted with the residual of the previous learner, and finally the predictions of all trees are added together to complete the prediction task; the residual is obtained in the following manner: r _ni = _yi -f _n-1 ( _xi ) In the formula, _yi is the measured value of the ith sample, and fn _-1 ( _xi ) is the predicted value of the previous round of learner. The residual memory is fitted to obtain a fitted residual model _hn (x), and the regression tree is updated: f _n (x)=f _n-1 (x)+h _n (x). 10. The method for estimating the spaceborne shortwave infrared CO ₂ column concentration based on machine learning according to claim 1, characterized in that the determination coefficient R ² , root mean square error RMSE, mean absolute error MAE, and mean relative error MRE are obtained in the following manner:

Where: N is the number of samples; _fi is the predicted value; _yi is the true value;

is the average value.

Images