CN114861882B - A CO2 spatio-temporal distribution reconstruction method and system - Google Patents

Abstract

The invention discloses a method and system for reconstructing _CO2 spatiotemporal distribution, comprising the following steps: S1: establishing an environmental database, the environmental database and a multi-output deep neural network model; S2: using _NO2 satellite remote sensing data and environmental data Perform initial training on the multi-output deep neural network model; S3: Use CO ₂ satellite remote sensing data and environmental data to conduct secondary training on the multi-output deep neural network model after initial training; S4: Use the environmental data and the trained The multi-output deep neural network model predicts the spatio-temporal distribution of CO ₂ and obtains the reconstruction results of the spatio-temporal distribution of CO ₂ . The present invention reconstructs the NO ₂ satellite data with higher temporal and spatial resolution data set with higher accuracy, and at the same time, assigns the fossil fuel combustion information represented by the NO ₂ satellite data to the model, and realizes the high temporal and spatial resolution of CO ₂ Reconstruction of spatiotemporal distribution.

Description

A CO2 spatio-temporal distribution reconstruction method and system

technical field

The present invention relates to the technical field of environmental monitoring, in particular to a method and system for reconfiguring the temporal and spatial distribution of CO ₂ .

Background technique

Currently, machine learning methods for reconstructing the _{spatiotemporal} distribution of CO mainly include neural network (ANN), extreme gradient booster (XGBOOST), and light gradient booster (Light-GBM). In the technical modeling of using machine learning to reconstruct the temporal and spatial distribution of CO ₂ , the dependent variable is CO ₂ satellite remote sensing data, and the independent variables mainly include: land use type, normalized difference vegetation index, meteorological conditions, population, altitude, road information, and CO ₂ Emission inventory, etc. The existing technology lacks the support of relevant data information when reconstructing the temporal and spatial distribution of CO ₂ , resulting in large deviations in the process of CO ₂ reconstruction; at the same time, because CO ₂ satellites are easily disturbed by external conditions when acquiring data, the acquisition The satellite remote sensing data sample size is small and there is sampling bias, which leads to a certain underestimation and overestimation of CO ₂ in areas where satellite data is scarce, and a certain high value of CO ₂ concentration in industrial areas with high vegetation coverage and high altitude. Underestimate the problem.

In view of the above problems, some existing technologies have sampled NO ₂ satellite remote sensing data (TROPOMI-NO ₂ ) and interpolated time and space, and added it as an independent variable into the model to reconstruct the temporal and spatial distribution of CO ₂ . However, the _coverage and spatio-temporal resolution of the original TROPOMI-NO ₂ data cannot meet the needs of high-resolution spatial-temporal distribution reconstruction of CO ₂ ; Will reduce the accuracy of _CO2 spatiotemporal reconstruction results.

In view of this, this application is proposed.

Contents of the invention

The technical problem to be solved by the present invention is: the accuracy of the results obtained by using the existing technology to reconstruct the _CO2 time-space distribution is low, and the purpose is to provide a _CO2 time-space distribution reconstruction method and system, using the satellite data of TROPOMI-NO2 As the relevant information of fossil fuel combustion, a transfer learning method based on shared parameters is used to realize the reconstruction of the global spatial and temporal distribution of _CO2 concentration.

The present invention realizes through following technical scheme:

On the one hand, the present invention provides a kind of CO _Spatio -temporal distribution reconstruction method, comprising the following steps:

S1: Establish an environmental database, which includes: TROPOMI-NO ₂ satellite remote sensing data, CO ₂ satellite remote sensing data and basic data related to the environment;

S2: Establish a multi-output deep neural network model;

S3: Using the TROPOMI-NO ₂ satellite remote sensing data and the environment-related basic data to perform initial training on the multi-output deep neural network model;

S4: Using the CO ₂ satellite remote sensing data and the environment-related basic data to perform secondary training on the multi-output deep neural network model after the initial training;

S5: Using the basic data related to the environment and the multi-output deep neural network model after the secondary training to predict the temporal and spatial distribution of CO ₂ , and obtain a reconstruction result of the temporal and spatial distribution of CO ₂ .

As a further description of the present invention,

The basic data related to the environment include: _CO2 emission data, population density data, altitude data, land use data, normalized difference vegetation index and meteorological data; the meteorological data include: surface temperature, surface air pressure, wind speed, Wind direction, relative humidity, and planetary boundary layer altitude.

As a further description of the present invention, before said S2 includes:

S11: Perform 1km grid processing and standardization processing on all the data in the environmental database;

S12: dividing the grid containing the TROPOMI-NO ₂ satellite remote sensing data processed by the S11 into an initial training set and an initial test set, and dividing the grid containing the CO ₂ satellite remote sensing data processed by the S11 into Secondary training set and secondary test set.

As a further description of the present invention, said S2 includes the following steps:

Establish the first deep neural network model and the second deep neural network model, the first deep neural network model and the second deep neural model have the same model structure, including: Dense layer, Batch Normalization layer, activation function and Dropout layer;

Establishing a corresponding data transmission chain between the layers of the first deep neural network model and the second deep neural network model, the direction of the data transmission chain is from the first deep neural network model to the second deep neural network model Two deep neural network models;

Connecting the input end of the first deep neural network model and the input end of the second deep neural network model to the output end of the autoencoder.

As a further description of the present invention, said S3 includes the following steps:

S31: Using the basic data related to the environment as an independent variable, and inputting the initial training set as a dependent variable into the multi-output deep neural network model;

S32: Use the first deep neural network model to perform deep learning on the environment-related basic data and the initial training set, and fit the environment-related basic data and the TROPOMI-NO ₂ satellite remote sensing an initial non-linear relationship between the data, and sending the fitting result of each layer of the first deep neural network to the corresponding layer of the second deep neural network;

S33: Use the second deep neural network model to perform deep learning on the environment-related basic data, the initial training set, and data from the first deep neural network model, and fit the environment-related The final nonlinear relationship between the basic data and the TROPOMI-NO ₂ satellite remote sensing data.

As a further description of the present invention,

Said S32 comprises the following steps before:

performing dimensionality reduction processing on the basic data related to the environment and the initial training set by using the autoencoder;

Carry out cross-validation on the initial training set;

After said S33, the following steps are included:

Using the initial test set to test the multi-output deep neural network model after the initial training.

As a further description of the present invention, said S4 includes:

S41: Using the basic data related to the environment as an independent variable, and inputting the secondary training set as a dependent variable into the multi-output deep neural network model;

S42: Only use the second deep neural network model trained in S33 to perform deep learning on the basic data related to the environment and the secondary training set, and fit the basic data related to the environment and the Describe the nonlinear relationship between _CO2 satellite remote sensing data.

As a further description of the present invention,

Before the S42, the following steps are included: using the autoencoder to perform dimensionality reduction processing on the basic data related to the environment and the secondary training set;

After the S42, the following steps are included: using the secondary test set to test the multi-output deep neural network model after the secondary training.

On the other hand, the present invention provides a system for reconfiguring the spatial _- temporal distribution of CO, comprising:

Database creation module for creating an environmental database containing TROPOMI-NO ₂ satellite remote sensing data, CO ₂ satellite remote sensing data and basic data related to the environment;

A model creation module for setting up a multi-output deep neural network model;

Initial training module, for utilizing said TROPOMI-NO ₂ satellite remote sensing data and said environment-related basic data to carry out initial training to said multi-output deep neural network model;

The secondary training module is used to utilize the CO ₂ satellite remote sensing data and the basic data related to the environment to carry out secondary training to the multi-output deep neural network model after the initial training;

The model prediction module is used to predict the temporal and spatial distribution of CO 2 by using the basic data related to the environment and the multi-output deep neural network model after the secondary training, and obtain the reconstruction result of the temporal and spatial distribution _{of CO 2 .}

As a further description of the present invention, the system also includes:

The data processing module is used to perform 1km grid processing and standardization processing on all data in the environmental database, and divide the processed grid containing _CO2 satellite remote sensing data into a secondary training set and a secondary testing set .

The model creation module includes:

The first model creation unit is used to create a first deep neural network model comprising a Dense layer, a Batch Normalization layer, an activation function and a Dropout layer, and connect the input of the first deep neural network model to the output of an autoencoder connect;

The second model creation unit is used to create a second deep neural network model having the same model structure as the first deep neural network model, and connect the input terminal of the second deep neural network model to the autoencoder The output terminal connection;

A data transmission chain creation unit, configured to establish a corresponding data transmission chain between the layers of the first deep neural network model and the second deep neural network model, the direction of the data transmission chain is from the first A deep neural network model to the second deep neural network model.

Compared with the prior art, the present invention has the following advantages and beneficial effects:

1. On the one hand, the present invention takes advantage of the large amount of TROPOMI-NO ₂ satellite remote sensing data and can also be used as relevant information on fossil fuel combustion, and introduces it into the reconstruction of CO ₂ temporal and spatial distribution, reflecting human activities from the side, which can solve the problem of existing technologies When reconstructing the temporal and spatial distribution of CO ₂ , there is a lack of support for relevant data information and the sample size of satellite remote sensing data acquired by CO ₂ satellites is small, resulting in large deviations in the reconstruction results; on the other hand, by establishing a multi-output deep neural network Network model, and remember the initial training and secondary training for the model in sequence, using the TROPOMI-NO ₂ satellite data to reflect the information of fossil fuel combustion and the characteristics of the same source of NO ₂ and CO ₂ emissions, the fossil fuel represented by TROPOMI-NO ₂ Combustion information is given to the model to realize the reconstruction of the spatial and temporal distribution of _CO2 with high temporal and spatial resolution.

2. The present invention performs dimensionality reduction processing on the initial independent variable dimension through the autoencoder part, which improves the learnability of the data;

3. The present invention utilizes the transfer learning method to effectively combine the relevant information of TROPOMI-NO2 and CO2 on the temporal and spatial distribution, and uses the TROPOMI- _NO2 data as auxiliary data, which solves the problem of _CO2 data sparseness and improves the _CO2 concentration prediction results accuracy.

Description of drawings

In order to more clearly illustrate the technical solutions of the exemplary embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention. Therefore, it should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can also be obtained according to these drawings without creative work.

Fig. 1 is a kind of _CO2 spatio-temporal distribution reconfiguration method schematic flowchart provided by the embodiment of the present invention;

Fig. 2 is a schematic structural diagram of a multi-output deep neural network model provided by an embodiment of the present invention;

Fig. 3 is a schematic structural diagram of a CO ₂ spatiotemporal distribution reconstruction system provided by an embodiment of the present invention.

Detailed ways

In order to make the purpose, technical solutions and advantages of the present invention clearer, the present invention will be further described in detail below in conjunction with the examples and accompanying drawings. As a limitation of the present invention.

In the following description, numerous specific details are set forth in order to provide a thorough understanding of the present invention. It will be apparent, however, to one of ordinary skill in the art that these specific details need not be employed to practice the present invention. In other instances, well-known structures, circuits, materials, or methods have not been described in detail in order to avoid obscuring the present invention.

Throughout this specification, reference to "one embodiment," "an embodiment," "an example," or "example" means that a particular feature, structure, or characteristic described in connection with the embodiment or example is included in the present invention. In at least one embodiment. Thus, appearances of the phrases "one embodiment," "an embodiment," "an example," or "example" in various places throughout this specification are not necessarily all referring to the same embodiment or example. Furthermore, particular features, structures or characteristics may be combined in any suitable combination and/or subcombination in one or more embodiments or examples. Furthermore, those of ordinary skill in the art will appreciate that the drawings provided herein are for illustrative purposes and are not necessarily drawn to scale. As used herein, the term "and/or" includes any and all combinations of one or more of the associated listed items.

In the description of the present invention, the terms "front", "rear", "left", "right", "upper", "lower", "vertical", "horizontal", "higher", "lower", "inner ", "outside" and other indicated orientations or positional relationships are based on the orientations or positional relationships shown in the drawings, which are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the referred device or element must have a specific Orientation, construction and operation in a particular orientation, therefore, should not be construed as limiting the scope of the invention.

Example 1

Because the existing technology lacks the support of relevant data information when reconstructing the temporal and spatial distribution of CO ₂ , and CO ₂ satellites are easily disturbed by external conditions when acquiring data, resulting in large deviations in the process of CO ₂ reconstruction, this paper The embodiment provides a method for reconstructing the temporal and spatial distribution of CO ₂ , the process flow of which is shown in Figure 1, using the satellite data of TROPOMI-NO ₂ as the relevant information of fossil fuel combustion, and based on the transfer learning method of shared parameters, the CO ₂ Reconstruction of spatial and temporal distribution of concentration in the global domain. The main modeling process is to pre-train the deep learning model with TROPOMI-NO ₂ as the dependent variable, so that the model can fit the nonlinear relationship between the independent variable and TROPOMI-NO ₂ , and then use CO ₂ as the dependent variable to pre-train The final model is then trained. Through the above method, while reconstructing the TROPOMI-NO ₂ high-temporal-spatial resolution data set with higher accuracy, the model also endows the information of fossil fuel combustion represented by TROPOMI-NO ₂ into the model, realizing the high-level analysis of CO ₂ Reconstruction of spatiotemporal distributions at spatiotemporal resolution. The implementation steps are as follows:

S1: Establish an environmental database, which includes: TROPOMI-NO ₂ satellite remote sensing data, CO ₂ satellite remote sensing data and basic data related to the environment. Among them, the environment-related basic data include: _CO2 emission data, population density data, altitude data, land use data, normalized difference vegetation index and meteorological data; the meteorological data include: surface temperature, surface pressure, wind speed, wind direction , relative humidity and altitude of the planetary boundary layer.

Taking advantage of the large amount of TROPOMI-NO ₂ satellite remote sensing data which can also be used as relevant information on fossil fuel combustion, it is introduced into the reconstruction of the temporal and spatial distribution of CO ₂ to reflect human activities from the side, which can solve the problems of existing technologies in the analysis of the temporal and spatial distribution of CO ₂ The lack of support for relevant data information and the small sample size of satellite remote sensing data acquired by CO ₂ satellites during reconstruction lead to large deviations in reconstruction results. Accurate full coverage of CO ₂ concentration prediction results can provide a basis for carbon emission statistics and accounting. and provide data support for the formulation of carbon reduction policies.

S2: Perform 1 km grid processing and standardization processing on all the data in the environmental database. Specifically include the following steps:

S21: Take the area weighted average method for the population density data, altitude data, land use data, and normalized difference vegetation index variables to convert the spatial resolution into a 1km grid;

S22: CO2 emission inventory and meteorological data are resampled to a 1km grid using elevation cokriging interpolation.

S23: After processing the above data to the same spatial scale, standardize them respectively, so that each independent variable is at the same data scale, which is beneficial to the training of the deep learning model.

S24: Carry out 1 km grid processing on the TROPOMI-NO ₂ satellite remote sensing data and CO ₂ satellite remote sensing data, and divide the grid containing the TROPOMI-NO ₂ satellite remote sensing data after the processing from step (1) to step (3) It is an initial training set and an initial test set; the grid containing CO ₂ satellite remote sensing data processed from step (1) to step (3) is divided into a secondary training set and a secondary test set.

S3: Establish a multi-output deep neural network model. Include the following steps:

S31: Establish a first deep neural network model and a second deep neural network model, the first deep neural network model and the second deep neural model have the same model structure, including: Dense layer, BatchNormalization layer, activation function and Dropout layer;

S32: Establish a corresponding data transmission chain between the layers of the first deep neural network model and the second deep neural network model, the direction of the data transmission chain is from the first deep neural network model to the Describe the second deep neural network model;

S33: Connect the input end of the first deep neural network model and the input end of the second deep neural network model to the output end of the autoencoder.

The deep learning model used is a multi-output deep neural network based on an autoencoder. As shown in Figure 2, the model consists of three modules, including an autoencoder and two sub-deep learning modules.

The model uses an autoencoder to reduce the dimensionality of the number of independent variables. In this embodiment, the first deep neural network model and the second deep neural network model are respectively constructed using the reduced dimensionality variables. The fully connected layers in the first deep neural network model and the second deep neural network model include the most basic Dense layer of the neural network, and use BatchNormalization, activation function and Dropout. And there is a one-way connection between the two deep neural network models, that is, the output of each layer of the first deep neural network model will be added to the output of each layer of the second deep neural network model to serve as the output of the second deep neural network model. A layer of input data.

S34: The present invention uses a joint loss function to optimize the model. The loss function corresponding to each output is based on the root mean square error (Mean Squared Error, MSE). The specific function expression is as follows:

output one:

Loss _min = min(Loss ₂ , Loss ₃ )

Output two:

Output three:

Overall output: Loss _all = λ _min Loss ₁ + (1-λ) Loss ₂ + λLoss ₃

λ _min = min(λ, 1-λ)

分别为输出一、输出二和输出三的模型预测值；y₁、y₂、y₃分别为输出一、输出二和输出三对应的真实值；λ、(1-λ)和λ_min为施加在各损失函数上的权重，值域为(0，1)，ω为输出一最后一层神经网络节点的权重值。通过调节λ的值来设置模型训练的侧重点，当λ＜0.5时，模型中Loss₂在训练过程中占有更大的权重；当λ＞0.5时，Loss₃在训练过程中占有更大的权重。利用λ_min对Loss₁进行调节，λ_min为λ和1-λ间的较小值，并在Loss₁中结合Loss₂和Loss₃之间的较小值Loss_min。当Loss_min值较大时，Loss₁在梯度下降过程中具有较高的权重，从而使自编码器模块能够更快趋于全局最优值；当Loss_min值逐渐变小时，Loss₁权重值也相应变小，使由输入数据和输出一所构成的自编码器内部参数变动更小，为子深度学习模块①和模块②提供更为稳定的自变量降维结果，同时加上了ω的二范数值来对输出一的结果施加L2正则化，从而使模型中的自编码器结构具有更强的泛化能力。本发明利用整体的损失函数Loss_all对模型进行优化，实现模型内部参数的更新。in:

are the model prediction values of output 1, output 2 and output 3 respectively; y ₁ , y ₂ , y ₃ are the real values corresponding to output 1, output 2 and output 3 respectively; λ, (1-λ) and λ _min are the applied The weight on each loss function has a range of (0, 1), and ω is the weight value of the output-last layer neural network node. Set the focus of model training by adjusting the value of λ. When λ<0.5, Loss ₂ in the model occupies a greater weight in the training process; when λ>0.5, Loss ₃ occupies a greater weight in the training process . Use λ _min to adjust Loss ₁ , where λ _min is a smaller value between λ and 1-λ, and combine Loss ₁ with a smaller value Loss _min between Loss ₂ and Loss ₃ . When the Loss _min value is large, Loss ₁ has a higher weight in the gradient descent process, so that the autoencoder module can tend to the global optimal value faster; when the Loss _min value gradually becomes smaller, the weight value of Loss ₁ also increases. Correspondingly, the internal parameters of the autoencoder composed of input data and output 1 change less, and provide more stable independent variable dimensionality reduction results for the sub-deep learning module ① and module ②. The norm value is used to apply L2 regularization to the result of output one, so that the autoencoder structure in the model has stronger generalization ability. The present invention utilizes the overall loss function Loss _all to optimize the model and realize the update of the internal parameters of the model.

S4: Initially train the multi-output deep neural network model by using the TROPOMI-NO ₂ satellite remote sensing data and the environment-related basic data. Include the following steps:

S41: Using the basic data related to the environment as an independent variable, and inputting the initial training set as a dependent variable into the multi-output deep neural network model;

S42: Using the autoencoder to perform dimensionality reduction processing on the environment-related basic data and the initial training set;

S43: Use the first deep neural network model to perform deep learning on the environment-related basic data and the initial training set, and fit the environment-related basic data and the TROPOMI-NO ₂ satellite remote sensing an initial non-linear relationship between the data, and sending the fitting result of each layer of the first deep neural network to the corresponding layer of the second deep neural network;

S44: Use the second deep neural network model to perform deep learning on the environment-related basic data, the initial training set, and data from the first deep neural network model, and fit the environment-related The final non-linear relationship between the basic data and the TROPOMI-NO ₂ satellite remote sensing data;

S45: Use the initial test set to test the multi-output deep neural network model after the initial training.

For the model to have a better generalization ability, in this embodiment, the grid containing TROPOMI-NO ₂ satellite remote sensing data is divided into an initial training set and an initial test set, and cross-validation is used for the initial training set to determine the model The optimal parameters of ; the initial test set is only used to evaluate the generalization ability of the model, and does not participate in the pre-training process of the model.

During initial training, CO2 emission inventory, population density data, altitude data, land use data, normalized difference vegetation index, and meteorological data are used as input data for the model. Output 1 is the input data itself, and the output of the first deep neural network model and the second deep neural network model are both TROPOMI-NO ₂ satellite remote sensing data, thus taking advantage of the large amount of TROPOMI-NO ₂ data to complete the parameters in the model initial optimization.

Through the initial training of the model, the first deep neural network model and the second deep neural network model respectively construct the relationship between two different integrated data related to the environment (hereinafter referred to as independent variables) and TROPOMI-NO ₂ satellite remote sensing data. non-linear relationship. The decoder part and the first deep neural network model in the model reflect the nonlinear relationship between independent variables and TROPOMI-NO ₂ satellite remote sensing data, and the second deep neural network model combines the TROPOMI -NO ₂ satellite remote sensing data to construct the non-linear relationship between independent variables and TROPOMI-NO ₂ satellite remote sensing data.

In addition, the dimensionality of the initial independent variable is reduced through the autoencoder part, which improves the learnability of the data.

The relationship between independent variables and TROPOMI-NO ₂ satellite remote sensing data is: Y _TROPOMI-NO2 = f(x ₁ , x ₂ ..., x _n ).

S5: Using the CO ₂ satellite remote sensing data and the environment-related basic data to perform secondary training on the multi-output deep neural network model after the initial training. Include the following steps:

S51: Using the basic data related to the environment as an independent variable, and inputting the secondary training set as a dependent variable into the multi-output deep neural network model;

S52: Using the autoencoder to perform dimensionality reduction processing on the environment-related basic data and the secondary training set;

S53: Only use the second deep neural network model trained in S33 to perform deep learning on the basic data related to the environment and the secondary training set, and fit the basic data related to the environment and the Describe the nonlinear relationship between _CO2 satellite remote sensing data;

S54: Using the secondary test set to test the secondary trained multi-output deep neural network model.

In this embodiment, the grid containing _CO satellite remote sensing data is divided into a secondary training set and a secondary testing set. The secondary training set is used to determine the optimal parameters of the model, and the secondary testing set is only used to evaluate the generalization of the model. ability.

When the model is trained for the second time, the parameters of the autoencoder and the first deep neural network model in the model are frozen so that the parameters of the autoencoder and the first deep neural network model do not occur during the second training process Change. The variable type of the input data of the model is the same as that of the initial training, and only the second deep neural network model is involved in the training to construct the nonlinear relationship between the independent variable and CO ₂ .

Since the first deep neural network model can provide information related to TROPOMI-NO ₂ satellite remote sensing data during the initial training process, the second deep neural network model can combine TROPOMI-NO ₂ satellite remote sensing data to achieve an effective prediction of CO ₂ . The following formula is the relationship between the independent variable constructed by the second depth neural network model and the CO ₂ satellite remote sensing data: Y _CO2 =f(x ₁ , x _{2 .} . . , x _n , Y _TROPOMI-NO2 ).

S6: Using the basic data related to the environment and the multi-output deep neural network model after the second training to predict the temporal and spatial distribution of CO ₂ , and obtain a reconstruction result of the temporal and spatial distribution of CO ₂ .

After completing the initial training and secondary training of the model, the independent variables within the region are input into the trained multi-output deep neural network model, and the spatial resolution of the first deep neural network model is 1km, and the temporal resolution is The reconstruction results of the temporal and spatial distribution of TROPOMI-NO ₂ in the day, through the second deep neural network model output the reconstruction results of the temporal and spatial distribution of CO ₂ with a spatial resolution of 1km and a temporal resolution of the day, realizing the regional TROPOMI-NO ₂ and _CO2 high-resolution spatiotemporal distribution reconstruction.

To sum up, this example provides a method for reconstructing the temporal and spatial distribution of CO ₂ . In the absence of data directly related to human activities, the NO ₂ satellite remote sensing data TROPOMI-NO ₂ is used to reflect human activities and reconstruct CO ₂ High-resolution spatio-temporal distribution and accurate full-coverage CO ₂ concentration prediction results can provide a basis for statistical accounting of carbon emissions and provide data support for the formulation of carbon reduction policies. Moreover, in this method, the dimensionality of the initial independent variable is reduced through the autoencoder part, which improves the learnability of the data. The transfer learning method is used to effectively combine the relevant information of TROPOMI-NO ₂ and CO ₂ on the temporal and spatial distribution, and use the TROPOMI-NO ₂ data as auxiliary data to improve the accuracy of CO ₂ concentration prediction results.

Example 2

This embodiment provides a system corresponding to the method for reconfiguring CO _spatio- temporal distribution described in Embodiment 1, as shown in Figure 3, including:

Database creation module for creating an environmental database containing TROPOMI-NO ₂ satellite remote sensing data, CO ₂ satellite remote sensing data and basic data related to the environment;

A model creation module for setting up a multi-output deep neural network model;

Initial training module, for utilizing said TROPOMI-NO ₂ satellite remote sensing data and said environment-related basic data to carry out initial training to said multi-output deep neural network model;

The secondary training module is used to utilize the CO ₂ satellite remote sensing data and the basic data related to the environment to carry out secondary training to the multi-output deep neural network model after the initial training;

The model prediction module is used to predict the temporal and spatial distribution of CO 2 by using the basic data related to the environment and the multi-output deep neural network model after the secondary training, and obtain the reconstruction result of the temporal and spatial distribution _{of CO 2 .}

The data processing module is used to perform 1km grid processing and standardization processing on all data in the environmental database, and divide the processed grid containing _CO2 satellite remote sensing data into a secondary training set and a secondary testing set .

Among them, the model creation module includes:

The first model creation unit is used to create a first deep neural network model comprising a Dense layer, a Batch Normalization layer, an activation function and a Dropout layer, and connect the input of the first deep neural network model to the output of an autoencoder connect;

The second model creation unit is used to create a second deep neural network model having the same model structure as the first deep neural network model, and connect the input terminal of the second deep neural network model to the autoencoder The output terminal connection;

A data transmission chain creation unit, configured to establish a corresponding data transmission chain between the layers of the first deep neural network model and the second deep neural network model, the direction of the data transmission chain is from the first deep neural network model to the second deep neural network model

The specific embodiments described above have further described the purpose, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above descriptions are only specific embodiments of the present invention and are not intended to limit the scope of the present invention. Protection scope, within the spirit and principles of the present invention, any modification, equivalent replacement, improvement, etc., shall be included in the protection scope of the present invention.

Claims (7)

1. CO (carbon monoxide) ₂ The space-time distribution reconstruction method is characterized by comprising the following steps of:

s1: establishing an environment database, wherein the environment database comprises: TROPOMI-NO ₂ Satellite remote sensing data, CO ₂ Satellite remote sensing data and environmental related base data;

s2: establishing a multi-output deep neural network model;

s3: using said TROPOMI-NO ₂ Satellite remote sensing data and said environment-related basis data are deep for said multiple outputsInitial training is carried out on the neural network model;

s4: by using the CO ₂ The satellite remote sensing data and the basic data related to the environment perform secondary training on the multi-output deep neural network model after the initial training;

s5: CO is performed by utilizing the basic data related to the environment and the multi-output deep neural network model after the secondary training ₂ Predicting the space-time distribution to obtain CO ₂ Reconstructing a result of space-time distribution;

the step S3 comprises the following steps:

s31: inputting the basic data related to the environment into the multi-output deep neural network model by taking the initial training set as an independent variable;

s32: deep learning the environment-related base data and the initial training set by using a first deep neural network model to fit the environment-related base data and the TROPOMI-NO ₂ The initial nonlinear relation between the satellite remote sensing data is achieved, and the fitting result of each layer of the first depth neural network is sent to the corresponding layer of the second depth neural network;

s33: deep learning the environment-related base data, the initial training set and the data from the first deep neural network model by using the second deep neural network model to fit the environment-related base data and the TROPOMI-NO ₂ A final nonlinear relationship between satellite remote sensing data.

2. A CO according to claim 1 ₂ A method for reconstructing space-time distribution is characterized in that,

the context-dependent underlying data includes: CO ₂ Emission data, population density data, altitude data, land utilization data, normalized vegetation index, and meteorological data; the meteorological data includes: surface temperature, surface air pressure, wind speed, wind direction, relative humidity, and planet boundary layer height.

3. A CO according to claim 2 ₂ The space-time distribution reconstruction method is characterized in that the method comprises the following steps:

s11: carrying out 1km gridding treatment and standardization treatment on all data in the environment database;

s12: treating the treated sample with S11 to obtain a sample containing TROPOMI-NO ₂ The grid of the satellite remote sensing data is divided into an initial training set and an initial testing set, and the grid of the satellite remote sensing data is processed by the S11 and contains CO ₂ The grid of satellite remote sensing data is divided into a secondary training set and a secondary testing set.

4. A CO according to claim 3 ₂ The space-time distribution reconstruction method is characterized in that the S2 comprises the following steps:

establishing a first deep neural network model and a second deep neural network model, wherein the first deep neural network model and the second deep neural network model have the same model structure and comprise the following steps: a Dense layer, a Batch Normalization layer, an activation function and a Dropout layer;

establishing a data transmission chain corresponding to each layer of the first deep neural network model and the second deep neural network model, wherein the direction of the data transmission chain is from the first deep neural network model to the second deep neural network model;

and connecting the input end of the first depth neural network model and the input end of the second depth neural network model with the output end of the self-encoder.

5. A CO according to claim 4 ₂ The space-time distribution reconstruction method is characterized by comprising the following steps before the step S32:

performing dimension reduction processing on the basic data related to the environment and the initial training set by using the self-encoder;

performing cross validation on the initial training set;

the step S33 includes the following steps:

and testing the multi-output deep neural network model after the initial training by using the initial test set.

6. A CO according to claim 4 or 5 ₂ The space-time distribution reconstruction method is characterized in that the S4 comprises the following steps:

s41: inputting the basic data related to the environment into the multi-output deep neural network model by taking the secondary training set as an independent variable;

s42: deep learning the environment-related basic data and the secondary training set by using only the second deep neural network model trained by the S33, and fitting the environment-related basic data and the CO ₂ Nonlinear relationships between satellite remote sensing data.

7. A CO according to claim 6 ₂ A method for reconstructing space-time distribution is characterized in that,

before S42, the method includes the following steps: performing dimension reduction processing on the basic data related to the environment and the secondary training set by using the self-encoder;

after S42, the method includes the following steps: and testing the multi-output deep neural network model after the secondary training by using the secondary test set.

Images