A method and system for filling atmospheric CO2 data using a feedforward neural network combined with time series changes

Abstract

The present invention proposes a method and system for filling atmospheric CO2 data using a feedforward neural network combined with time series changes. The method includes: learning the temporal and spatial variation law of CO2 in the atmosphere through a deep learning model, fitting the nonlinear relationship between auxiliary variables and CO2, and constructing a global CO2 distribution with high temporal and spatial resolution and high precision, that is, integrating satellite, ground-based and model data, which greatly improves the accuracy and coverage; adding the time variation of CO2 to the training set, the model can learn the time variation law of CO2 to a certain extent, and improve the accuracy. Use multiple simple models to construct a long-term series CO2 distribution to reduce the amount of calculation.

Description

Technical Field

[0001] The present invention belongs to the field of atmospheric CO2 data filling, and in particular relates to an atmospheric CO2 data filling method and system using a feedforward neural network combined with time series changes. Background Art

[0002] As one of the most important greenhouse gases, changes in the concentration of atmospheric CO2 have an important impact on global climate change, human living environment and even the structure of the entire social-economic system. Therefore, the monitoring of atmospheric CO2 concentration has always been a hot topic in current scientific research. Traditional atmospheric CO2 concentration monitoring is mainly based on ground-based observations, which have high observation accuracy, but there are few observation sites around the world and they are unevenly distributed, making it difficult to accurately describe the distribution of CO2 concentration. Satellite monitoring can provide global CO2 concentration observation data over a large range.

[0003] Satellite observations of CO2 are limited by factors such as clouds and aerosols. The global coverage of CO2 observations is low under the requirement of high temporal and spatial resolution. Currently, global CO2 with high temporal and spatial resolution is mainly obtained through geographic spatiotemporal interpolation methods. However, this method has low accuracy and ignores the relationship between CO2 as an atmospheric gas and other factors. Therefore, the use of deep learning models to construct nonlinear relationships between other factors (auxiliary variables such as NDVI, SIF, etc.) and CO2 has become a research hotspot. However, current research ignores the temporal changes of CO2, which greatly affects the accuracy of CO2 prediction. Summary of the invention

[0004] In order to solve the above technical problems, the present invention proposes a technical solution of using a feedforward neural network combined with a method for filling atmospheric CO2 data with time series changes to solve the above technical problems.

[0005] The first aspect of the present invention discloses a method for filling atmospheric CO2 data using a feedforward neural network combined with time series changes, the method comprising:

[0006] Step S1, obtaining satellite CO2 data, ground-based CO2 data and existing global CO2 model CO2 data, and resampling the satellite CO2 data, ground-based CO2 data and existing global CO2 model CO2 data to the same predefined spatial resolution to obtain first CO2 data; obtaining auxiliary variables of global vegetation characteristics and auxiliary variables of global vegetation characteristics, and resampling the auxiliary variables of global vegetation characteristics and auxiliary variables of global meteorological characteristics to the same predefined spatial resolution as the first CO2 data to obtain auxiliary variable data;

[0007] Step S2, according to the observation records of the satellite CO2 data, extract the corresponding auxiliary variable data to obtain a satellite auxiliary variable data set; according to the observation records of the satellite CO2 data, extract the corresponding existing global CO2 model CO2 data to obtain a satellite time series data set; the satellite auxiliary variable data set and the satellite time series data set constitute a satellite data set; according to the observation records of the ground-based CO2 data, extract the corresponding auxiliary variable data to obtain a ground-based auxiliary variable data set; according to the observation records of the ground-based CO2 data, extract the corresponding existing global CO2 model CO2 data to obtain a ground-based time series data set; the ground-based auxiliary variable data set and the ground-based time series data set constitute a ground-based data set;

[0008] Step S3, construct multiple feedforward neural network models; each feedforward neural network model is responsible for filling the atmospheric CO2 data within a period of time, and multiple models are combined to complete the atmospheric CO2 data filling in the case of a long time series;

[0009] Step S4, training the multiple feedforward neural network models using the satellite data set to obtain an initial training model;

[0010] Step S5: continue to train the initial training model using the ground-based data set to obtain a relay training model;

[0011] Step S6, according to the longitude and latitude coordinates and time of the center of the global predefined spatial resolution grid, extract the auxiliary variables of the global vegetation characteristics and the auxiliary variables of the global meteorological characteristics to obtain a grid auxiliary variable data set; according to the longitude and latitude coordinates and time of the center of the global predefined spatial resolution grid, extract the corresponding existing global CO2 model CO2 data to obtain a grid time series data set; the grid auxiliary variable data set and the grid time series data set constitute a grid data set;

[0012] Step S7, applying the relay training model and the grid dataset to generate a global CO2 dataset with a time resolution of one day and a spatial resolution of a predefined spatial resolution;

[0013] Step S8: weighted fusion is performed on the XCO2 global data of the repeated time periods, and the fusion result is calculated.

[0014] According to the method of the first aspect of the present invention, in step S1, the auxiliary variables of the global vegetation characteristics include: sunlight-induced chlorophyll fluorescence data and normalized vegetation index.

[0015] According to the method of the first aspect of the present invention, in step S1, the auxiliary variables of global meteorological characteristics include: U-shaped wind, V-shaped wind and sea level pressure.

[0016] According to the method of the first aspect of the present invention, in step S2, extracting corresponding auxiliary variable data according to the observation record of the satellite CO2 data to obtain the satellite auxiliary variable data set includes:

[0017] According to the longitude and latitude coordinates and time of the observation record of the satellite CO2 data, the auxiliary variables of global vegetation characteristics and the auxiliary variables of global meteorological characteristics corresponding to the longitude and latitude coordinates and time are extracted, and the missing values ​​of the auxiliary variables of global vegetation characteristics and the auxiliary variables of global meteorological characteristics should be set to 0 or Null to obtain a satellite auxiliary variable data set;

[0018] The extracting corresponding existing global CO2 model CO2 data according to the observation record of the satellite CO2 data to obtain the satellite time series data set includes:

[0019] According to the latitude and longitude coordinates and time of the observation record of the satellite CO2 data, the CO2 data of the day existing in the global CO2 model, as well as the CO2 data of the previous day and the CO2 data of the next day are extracted to obtain a satellite time series data set.

[0020] According to the method of the first aspect of the present invention, in the step S2, extracting corresponding auxiliary variable data according to the observation record of the ground-based CO2 data to obtain a ground-based auxiliary variable data set includes: extracting auxiliary variables of global vegetation characteristics and auxiliary variables of global meteorological characteristics at corresponding longitude and latitude coordinates and time according to the observation record of the ground-based CO2 data, and the missing values ​​of the auxiliary variables of global vegetation characteristics and the auxiliary variables of global meteorological characteristics should be set to 0 or Null to obtain a ground-based auxiliary variable data set;

[0021] The extracting corresponding existing global CO2 model CO2 data according to the observation records of the ground-based CO2 data to obtain the ground-based time series data set includes:

[0022] According to the latitude and longitude coordinates and time of the observation record of the ground-based CO2 data, the CO2 data of the day existing in the global CO2 model, as well as the CO2 data of the previous day and the CO2 data of the next day, are extracted, and the ground-based time series data set.

[0023] According to the method of the first aspect of the present invention, in the step S5, the parameters of the initial training model need to be frozen during the process of continuing to train the initial training model using the ground-based data set; the number of times the initial training model is trained using the ground-based data set is less than the number of times the multiple feedforward neural network models are trained using the satellite data set.

[0024] According to the method of the first aspect of the present invention, in step S6, extracting the corresponding existing global CO2 model CO2 data according to the longitude and latitude coordinates and time of the center of the global predefined spatial resolution grid to obtain the grid time series data set includes:

[0025] According to the longitude and latitude coordinates and time of the center of the global predefined spatial resolution grid, the CO2 data of the day, the CO2 data of the previous day and the CO2 data of the next day that already exist in the global CO2 model are extracted to obtain the grid time series dataset.

[0026] According to the method of the first aspect of the present invention, in step S8, the specific method of weighted fusion of XCO2 global data of repeated time periods and calculating the fusion result includes:

[0027] In step S3, there is an overlapping time period between each two adjacent FNN models. Similarly, when using the model generation, there will be global XCO2 data with repeated time periods. For the data in the repeated time period, the specific operations are as follows:

[0028] The two adjacent FNN models are denoted as A and B, and the RMSE of the test set in step S5 is denoted as RMSE A , RMSE B The generated global XCO2 data for the repeated time period is recorded as XCO 2A 、XCO 2B , the final fusion result is recorded as XCO 2M .

[0029] (1) Calculate the corresponding weights of A and B when they are fused. Since the larger the RMSE, the worse the accuracy of the model, the weight calculation method is as follows:

[0030]

[0031] (2) Calculate the fusion results:

[0032]

[0033] The second aspect of the present invention discloses an atmospheric CO2 data filling system using a feedforward neural network combined with time series changes, the system comprising:

[0034] The first processing module is configured to obtain satellite CO2 data, ground-based CO2 data and existing global CO2 model CO2 data, and resample the satellite CO2 data, ground-based CO2 data and existing global CO2 model CO2 data to the same predefined spatial resolution to obtain first CO2 data; obtain auxiliary variables of global vegetation characteristics and auxiliary variables of global vegetation characteristics, and resample the auxiliary variables of global vegetation characteristics and auxiliary variables of global meteorological characteristics to the same predefined spatial resolution as the first CO2 data to obtain auxiliary variable data;

[0035] The second processing module is configured to extract corresponding auxiliary variable data according to the observation record of the satellite CO2 data to obtain a satellite auxiliary variable data set; extract corresponding existing global CO2 model CO2 data according to the observation record of the satellite CO2 data to obtain a satellite time series data set; the satellite auxiliary variable data set and the satellite time series data set constitute a satellite data set; extract corresponding auxiliary variable data according to the observation record of the ground-based CO2 data to obtain a ground-based auxiliary variable data set; extract corresponding existing global CO2 model CO2 data according to the observation record of the ground-based CO2 data to obtain a ground-based time series data set; the ground-based auxiliary variable data set and the ground-based time series data set constitute a ground-based data set;

[0036] The third processing module is configured to construct multiple feedforward neural network models; each feedforward neural network model is responsible for filling the atmospheric CO2 data within a period of time, and multiple models are combined to complete the atmospheric CO2 data filling in the case of a long time series;

[0037] A fourth processing module is configured to train the multiple feedforward neural network models using the satellite data set to obtain an initial training model;

[0038] A fifth processing module is configured to continue training the initial training model using the ground-based data set to obtain a relay training model;

[0039] The sixth processing module is configured to extract auxiliary variables of global vegetation characteristics and auxiliary variables of global meteorological characteristics according to the longitude and latitude coordinates and time of the center of the global predefined spatial resolution grid to obtain a grid auxiliary variable data set; extract corresponding existing global CO2 model CO2 data according to the longitude and latitude coordinates and time of the center of the global predefined spatial resolution grid to obtain a grid time series data set; the grid auxiliary variable data set and the grid time series data set constitute a grid data set;

[0040] a seventh processing module, configured to apply the relay training model and the grid dataset to generate a global CO2 dataset with a time resolution of one day and a spatial resolution of a predefined spatial resolution;

[0041] The eighth processing module is configured to perform weighted fusion on the XCO2 global data of repeated time periods and calculate the fusion result.

[0042] The third aspect of the present invention discloses an electronic device. The electronic device includes a memory and a processor, the memory stores a computer program, and when the processor executes the computer program, the steps in any one of the atmospheric CO2 data filling methods using a feedforward neural network combined with time series changes in the first aspect of the present disclosure are implemented.

[0043] The fourth aspect of the present invention discloses a computer-readable storage medium. The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps in any one of the atmospheric CO2 data filling methods using a feedforward neural network combined with time series changes in the first aspect of the present disclosure are implemented.

[0044] In summary, the solution proposed in the present invention can learn the temporal and spatial variation laws of CO2 in the atmosphere through a deep learning model, fit the nonlinear relationship between auxiliary variables and CO2, and construct a global CO2 distribution with high temporal and spatial resolution and high precision; it can reduce the impact of factors such as clouds and aerosols on satellite observations, improve the accuracy of satellite observations of atmospheric CO2, and improve the spatial coverage of data. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] In order to more clearly illustrate the specific implementation methods of the present invention or the technical solutions in the prior art, the drawings required for use in the specific implementation methods or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are some implementation methods of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0046] Figure 1 A flowchart of a method for filling atmospheric CO2 data using a feedforward neural network combined with time series changes according to an embodiment of the present invention;

[0047] Figure 2 A structural diagram of an atmospheric CO2 data filling system using a feedforward neural network combined with time series changes according to an embodiment of the present invention;

[0048] Figure 3 The figure is a structural diagram of an electronic device according to an embodiment of the present invention. DETAILED DESCRIPTION

[0049] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0050] The first aspect of the present invention discloses a method for filling atmospheric CO2 data using a feedforward neural network combined with time series changes. Figure 1 FIG. 4 is a flow chart of a method for filling atmospheric CO2 data using a feedforward neural network combined with time series changes according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0051] Step S1, obtaining satellite CO2 data, ground-based CO2 data and existing global CO2 model CO2 data, and resampling the satellite CO2 data, ground-based CO2 data and existing global CO2 model CO2 data to the same predefined spatial resolution to obtain first CO2 data; obtaining auxiliary variables of global vegetation characteristics and auxiliary variables of global vegetation characteristics, and resampling the auxiliary variables of global vegetation characteristics and auxiliary variables of global meteorological characteristics to the same predefined spatial resolution as the first CO2 data to obtain auxiliary variable data;

[0052] Step S2, according to the observation records of the satellite CO2 data, extract the corresponding auxiliary variable data to obtain a satellite auxiliary variable data set; according to the observation records of the satellite CO2 data, extract the corresponding existing global CO2 model CO2 data to obtain a satellite time series data set; the satellite auxiliary variable data set and the satellite time series data set constitute a satellite data set; according to the observation records of the ground-based CO2 data, extract the corresponding auxiliary variable data to obtain a ground-based auxiliary variable data set; according to the observation records of the ground-based CO2 data, extract the corresponding existing global CO2 model CO2 data to obtain a ground-based time series data set; the ground-based auxiliary variable data set and the ground-based time series data set constitute a ground-based data set;

[0053] Step S3, construct multiple feedforward neural network models; each feedforward neural network model is responsible for filling the atmospheric CO2 data within a period of time, and multiple models are combined to complete the atmospheric CO2 data filling in the case of a long time series;

[0054] Step S4, training the multiple feedforward neural network models using the satellite data set to obtain an initial training model;

[0055] Step S5: continue to train the initial training model using the ground-based data set to obtain a relay training model;

[0056] Step S6, according to the longitude and latitude coordinates and time of the center of the global predefined spatial resolution grid, extract the auxiliary variables of the global vegetation characteristics and the auxiliary variables of the global meteorological characteristics to obtain a grid auxiliary variable data set; according to the longitude and latitude coordinates and time of the center of the global predefined spatial resolution grid, extract the corresponding existing global CO2 model CO2 data to obtain a grid time series data set; the grid auxiliary variable data set and the grid time series data set constitute a grid data set;

[0057] Step S7, applying the relay training model and the grid dataset to generate a global CO2 dataset with a time resolution of one day and a spatial resolution of a predefined spatial resolution;

[0058] Step S8: weighted fusion is performed on the XCO2 global data of the repeated time periods, and the fusion result is calculated.

[0059] In step S1, satellite CO2 data, ground-based CO2 data and existing global CO2 model CO2 data are obtained, and the satellite CO2 data, ground-based CO2 data and existing global CO2 model CO2 data are resampled to the same predefined spatial resolution to obtain first CO2 data; auxiliary variables of global vegetation characteristics and auxiliary variables of global vegetation characteristics are obtained, and the auxiliary variables of global vegetation characteristics and auxiliary variables of global meteorological characteristics are resampled to the same predefined spatial resolution as the first CO2 data to obtain auxiliary variable data.

[0060] In some embodiments, in step S1, the auxiliary variables of the global vegetation characteristics include: sunlight-induced chlorophyll fluorescence data and normalized vegetation index.

[0061] The auxiliary variables of the global meteorological characteristics include: U-shaped wind, V-shaped wind and sea level pressure.

[0062] The predefined spatial resolution is 0.1°×0.1°.

[0063] In step S2, according to the observation records of the satellite CO2 data, the corresponding auxiliary variable data are extracted to obtain a satellite auxiliary variable data set; according to the observation records of the satellite CO2 data, the corresponding existing global CO2 model CO2 data are extracted to obtain a satellite time series data set; the satellite auxiliary variable data set and the satellite time series data set constitute a satellite data set; according to the observation records of the ground-based CO2 data, the corresponding auxiliary variable data are extracted to obtain a ground-based auxiliary variable data set; according to the observation records of the ground-based CO2 data, the corresponding existing global CO2 model CO2 data are extracted to obtain a ground-based time series data set; the ground-based auxiliary variable data set and the ground-based time series data set constitute a ground-based data set.

[0064] In some embodiments, in step S2, extracting corresponding auxiliary variable data according to the observation record of the satellite CO2 data to obtain a satellite auxiliary variable data set includes:

[0065] According to the longitude and latitude coordinates and time of the observation record of the satellite CO2 data, the auxiliary variables of global vegetation characteristics and the auxiliary variables of global meteorological characteristics at the corresponding longitude and latitude coordinates and time are extracted, and the missing values ​​of the auxiliary variables of global vegetation characteristics and the auxiliary variables of global meteorological characteristics should be set to 0 or Null and should not be discarded, so as to obtain a satellite auxiliary variable data set;

[0066] The extracting corresponding existing global CO2 model CO2 data according to the observation record of the satellite CO2 data to obtain the satellite time series data set includes:

[0067] According to the latitude and longitude coordinates and time of the observation record of the satellite CO2 data, the CO2 data of the day existing in the global CO2 model, as well as the CO2 data of the previous day and the CO2 data of the next day are extracted to obtain a satellite time series data set.

[0068] The extracting corresponding auxiliary variable data according to the observation record of the ground-based CO2 data to obtain the ground-based auxiliary variable data set includes: extracting auxiliary variables of global vegetation characteristics and auxiliary variables of global meteorological characteristics at corresponding longitude and latitude coordinates and time according to the observation record of the ground-based CO2 data, and the missing values ​​of the auxiliary variables of global vegetation characteristics and the auxiliary variables of global meteorological characteristics should be set to 0 or Null and should not be discarded, so as to obtain the ground-based auxiliary variable data set;

[0069] The extracting corresponding existing global CO2 model CO2 data according to the observation records of the ground-based CO2 data to obtain the ground-based time series data set includes:

[0070] According to the latitude and longitude coordinates and time of the observation record of the ground-based CO2 data, the CO2 data of the day existing in the global CO2 model, as well as the CO2 data of the previous day and the CO2 data of the next day, are extracted, and the ground-based time series data set.

[0071] In step S3, multiple feedforward neural network models are constructed; each feedforward neural network model is responsible for filling in atmospheric CO2 data within a period of time, and multiple models are combined to complete the atmospheric CO2 data filling in the case of long time series; in particular, the two models responsible for adjacent time periods should overlap for a certain period of time, such as model A is responsible for global CO2 filling from January 1 to May 31, 2019, and model B is responsible for global CO2 filling from May 1 to October 31, 2019.

[0072] Specifically, the structure of the feedforward neural network model is an input layer, three hidden layers, and one output layer. The dropout method is used for random inactivation to reduce the training time.

[0073] In step S4, the plurality of feedforward neural network models are trained using the satellite data set to obtain an initial training model.

[0074] Specifically, within a certain period of time, each record in the satellite data set is used to train the FNN corresponding to the time, as follows:

[0075]

[0076] Among them, XCO2_sa represents the initial training model; Lat, Lon, Time represent longitude, latitude and time; Sif, Ndvi represent sunlight-induced chlorophyll fluorescence data and normalized vegetation index; U10, V10, Msl represent U-type wind, V-type wind and sea level pressure; XCO2_mo1, XCO2_mo2, XCO2_mo3 represent the CO2 data of the existing global CO2 model on the day, as well as the CO2 data of the previous day and the CO2 data of the next day.

[0077] During training, the satellite data set is randomly shuffled, 80% is used for training, and 20% is used for testing. Cross-validation is used during training, the optimizer is adam, and the error is MSE. Since the difference in CO2 concentration values ​​is mainly concentrated in 1-10 -2 , the learning rate should be as small as possible to ensure the accuracy of prediction, and early stopping should be used to avoid overfitting.

[0078] The above training is performed for all feedforward neural network models constructed.

[0079] In step S5, the initial training model is further trained using the ground-based data set to obtain a relay training model.

[0080] In some embodiments, in step S5, the parameters of the initial training model need to be frozen during the process of continuing to train the initial training model using the ground-based dataset; the number of times the initial training model is trained using the ground-based dataset is less than the number of times the multiple feedforward neural network models are trained using the satellite dataset.

[0081] Specifically,

[0082]

[0083] Among them, XCO2_gr represents the relay training model;

[0084] During the test, the accuracy performance of each model in the ground-based test set needs to be recorded, that is, the RMSE of the model in the test set needs to be recorded.

[0085]

[0086] In order to prevent the model from overfitting, the parameters of the initial training model need to be frozen during the process of continuing to train the initial training model using the ground-based data set; by adjusting the parameters of the early stopping method, the number of times the initial training model is trained using the ground-based data set is less than the number of times the multiple feedforward neural network models are trained using the satellite data set. If the initial training model stipulates that the training is stopped if the MSE of the prediction performance in the training set does not decrease significantly within 20 times, 20 should be reduced during the training process of the relay training model.

[0087] In step S6, auxiliary variables of global vegetation characteristics and auxiliary variables of global meteorological characteristics are extracted according to the longitude and latitude coordinates and time of the center of the global predefined spatial resolution grid to obtain a grid auxiliary variable data set; according to the longitude and latitude coordinates and time of the center of the global predefined spatial resolution grid, the corresponding existing global CO2 model CO2 data is extracted to obtain a grid time series data set; the grid auxiliary variable data set and the grid time series data set constitute a grid data set.

[0088] In some embodiments, in step S6, extracting the corresponding existing global CO2 model CO2 data according to the longitude and latitude coordinates and time of the center of the global predefined spatial resolution grid to obtain the grid time series data set includes:

[0089] According to the longitude and latitude coordinates and time of the center of the global predefined spatial resolution grid, the CO2 data of the day, the CO2 data of the previous day and the CO2 data of the next day that already exist in the global CO2 model are extracted to obtain the grid time series dataset.

[0090] Specifically, according to the longitude and latitude coordinates and time of the center of the global predefined spatial resolution grid, the auxiliary variables of global vegetation characteristics and the auxiliary variables of global meteorological characteristics at the corresponding longitude and latitude coordinates and time are extracted. The missing values ​​of the auxiliary variables of global vegetation characteristics and the auxiliary variables of global meteorological characteristics should be set to 0 or Null and should not be discarded, so as to obtain the grid auxiliary variable data set.

[0091] In step S7, the relay training model and the grid dataset are applied to generate a global CO2 dataset with a time resolution of one day and a spatial resolution of a predefined spatial resolution.

[0092] Specifically,

[0093]

[0094] Among them, XCO2 represents the global CO2 dataset;

[0095] In some specific embodiments, the global dataset in step S6 and the trained single FNN are used to generate a global XCO2 dataset with a spatial resolution of (0.1°x0.1° and a temporal resolution of 1 day).

[0096] In step S8, weighted fusion is performed on the XCO2 global data of the repeated time period, and the fusion result is calculated. The specific method is as follows:

[0097] In step S3, there is an overlapping time period between each two adjacent FNN models. Similarly, when using the model generation, there will be global XCO2 data with repeated time periods. For the data in the repeated time period, the specific operations are as follows:

[0098] The two adjacent FNN models are denoted as A and B, and the RMSE of the test set in step S5 is denoted as RMSE A , RMSE B The generated global XCO2 data for the repeated time period is recorded as XCO 2A 、XCO 2B , the final fusion result is recorded as XCO 2M ;

[0099] (1) Calculate the corresponding weights of A and B when they are fused. Since the larger the RMSE, the worse the accuracy of the model, the weight calculation method is as follows:

[0100]

[0101] (2) Calculate the fusion results:

[0102]

[0103] In summary, this application ensures that the predicted CO2 will not be discontinuous in spatial scale by adding longitude and latitude and the existing CO2 model; previous studies have shown that CO2 in a certain place will not change significantly in a short period of time, and adding the previous day's pattern data and the next day's pattern data to the data set can effectively improve the prediction accuracy; through relay training, the ground-based high-precision data set is integrated to further improve the prediction accuracy; by weighted fusion of the data set within the repetition time, the time continuity of the final data set on a long time scale is guaranteed. This application uses multiple simple models to construct a long-term series CO2 distribution to reduce the amount of calculation. By adding the time variation of CO2 to the training set, the model can learn the time variation law of CO2 to a certain extent and improve the accuracy. The fusion of satellite, ground-based and model data greatly improves the accuracy and coverage.

[0104] The second aspect of the present invention discloses an atmospheric CO2 data filling system using a feedforward neural network combined with time series changes. Figure 2 FIG. 1 is a structural diagram of an atmospheric CO2 data filling system using a feedforward neural network combined with time series changes according to an embodiment of the present invention; Figure 2 As shown, the system 100 includes:

[0105] The first processing module 101 is configured to obtain satellite CO2 data, ground-based CO2 data and existing global CO2 model CO2 data, and resample the satellite CO2 data, ground-based CO2 data and existing global CO2 model CO2 data to the same predefined spatial resolution to obtain first CO2 data; obtain auxiliary variables of global vegetation characteristics and auxiliary variables of global vegetation characteristics, and resample the auxiliary variables of global vegetation characteristics and auxiliary variables of global meteorological characteristics to the same predefined spatial resolution as the first CO2 data to obtain auxiliary variable data;

[0106] The second processing module 102 is configured to extract corresponding auxiliary variable data according to the observation record of the satellite CO2 data to obtain a satellite auxiliary variable data set; extract corresponding existing global CO2 model CO2 data according to the observation record of the satellite CO2 data to obtain a satellite time series data set; the satellite auxiliary variable data set and the satellite time series data set constitute a satellite data set; extract corresponding auxiliary variable data according to the observation record of the ground-based CO2 data to obtain a ground-based auxiliary variable data set; extract corresponding existing global CO2 model CO2 data according to the observation record of the ground-based CO2 data to obtain a ground-based time series data set; the ground-based auxiliary variable data set and the ground-based time series data set constitute a ground-based data set;

[0107] The third processing module 103 is configured to construct multiple feedforward neural network models; each feedforward neural network model is responsible for filling the atmospheric CO2 data within a period of time, and multiple models are combined to complete the atmospheric CO2 data filling in the case of a long time series;

[0108] The fourth processing module 104 is configured to train the multiple feedforward neural network models using the satellite data set to obtain an initial training model;

[0109] The fifth processing module 105 is configured to continue training the initial training model using the ground-based data set to obtain a relay training model;

[0110] The sixth processing module 106 is configured to extract auxiliary variables of global vegetation characteristics and auxiliary variables of global meteorological characteristics according to the longitude and latitude coordinates and time of the center of the global predefined spatial resolution grid to obtain a grid auxiliary variable data set; extract corresponding existing global CO2 model CO2 data according to the longitude and latitude coordinates and time of the center of the global predefined spatial resolution grid to obtain a grid time series data set; the grid auxiliary variable data set and the grid time series data set constitute a grid data set;

[0111] The seventh processing module 107 is configured to apply the relay training model and the grid dataset to generate a global CO2 dataset with a time resolution of one day and a spatial resolution of a predefined spatial resolution.

[0112] The eighth processing module 108 is configured to perform weighted fusion on the XCO2 global data of the repeated time periods and calculate the fusion result.

[0113] According to the system of the second aspect of the present invention, the first processing module 101 is specifically configured such that the auxiliary variables of the global vegetation characteristics include: sunlight-induced chlorophyll fluorescence data and normalized vegetation index.

[0114] The auxiliary variables of the global meteorological characteristics include: U-shaped wind, V-shaped wind and sea level pressure.

[0115] The predefined spatial resolution is 0.1°×0.1°.

[0116] According to the system of the second aspect of the present invention, the second processing module 102 is specifically configured to extract corresponding auxiliary variable data according to the observation record of the satellite CO2 data to obtain the satellite auxiliary variable data set including:

[0117] According to the longitude and latitude coordinates and time of the observation record of the satellite CO2 data, the auxiliary variables of global vegetation characteristics and the auxiliary variables of global meteorological characteristics at the corresponding longitude and latitude coordinates and time are extracted, and the missing values ​​of the auxiliary variables of global vegetation characteristics and the auxiliary variables of global meteorological characteristics should be set to 0 or Null and should not be discarded, so as to obtain a satellite auxiliary variable data set;

[0118] The extracting corresponding existing global CO2 model CO2 data according to the observation record of the satellite CO2 data to obtain the satellite time series data set includes:

[0119] According to the latitude and longitude coordinates and time of the observation record of the satellite CO2 data, the CO2 data of the day existing in the global CO2 model, as well as the CO2 data of the previous day and the CO2 data of the next day are extracted to obtain a satellite time series data set.

[0120] The extracting corresponding auxiliary variable data according to the observation record of the ground-based CO2 data to obtain the ground-based auxiliary variable data set includes: extracting auxiliary variables of global vegetation characteristics and auxiliary variables of global meteorological characteristics at corresponding longitude and latitude coordinates and time according to the observation record of the ground-based CO2 data, and the missing values ​​of the auxiliary variables of global vegetation characteristics and the auxiliary variables of global meteorological characteristics should be set to 0 or Null and should not be discarded, so as to obtain the ground-based auxiliary variable data set;

[0121] The extracting corresponding existing global CO2 model CO2 data according to the observation records of the ground-based CO2 data to obtain the ground-based time series data set includes:

[0122] According to the latitude and longitude coordinates and time of the observation record of the ground-based CO2 data, the CO2 data of the day existing in the global CO2 model, as well as the CO2 data of the previous day and the CO2 data of the next day, are extracted, and the ground-based time series data set.

[0123] According to the system of the second aspect of the present invention, the third processing module 103 is specifically configured as follows: the structure of the feedforward neural network model is an input layer, three hidden layers, and one output layer. The dropout method is used for random inactivation to reduce training time.

[0124] According to the system of the second aspect of the present invention, the fourth processing module 104 is specifically configured to train the FNN corresponding to the time using each record in the satellite data set within a certain period of time, as follows:

[0125]

[0126] Among them, XCO2_sa represents the initial training model; Lat, Lon, Time represent longitude, latitude and time; Sif, Ndvi represent sunlight-induced chlorophyll fluorescence data and normalized vegetation index; U10, V10, Msl represent U-type wind, V-type wind and sea level pressure; XCO2_mo1, XCO2_mo2, XCO2_mo3 represent the CO2 data of the existing global CO2 model on the day, as well as the CO2 data of the previous day and the CO2 data of the next day.

[0127] During training, the satellite data set is randomly shuffled, 80% is used for training, and 20% is used for testing. Cross-validation is used during training, the optimizer is adam, and the error is MSE. Since the difference in CO2 concentration values ​​is mainly concentrated in 1-10 -2 , the learning rate should be as small as possible to ensure the accuracy of prediction, and early stopping should be used to avoid overfitting.

[0128] The above training is performed for all feedforward neural network models constructed.

[0129] According to the system of the second aspect of the present invention, the fifth processing module 105 is specifically configured to freeze the parameters of the initial training model while continuing to train the initial training model using the ground-based data set; the number of times the initial training model is trained using the ground-based data set is less than the number of times the multiple feedforward neural network models are trained using the satellite data set.

[0130] Specifically,

[0131]

[0132] Among them, XCO2_gr represents the relay training model.

[0133] In order to prevent the model from overfitting, the parameters of the initial training model need to be frozen during the process of continuing to train the initial training model using the ground-based data set; by adjusting the parameters of the early stopping method, the number of times the initial training model is trained using the ground-based data set is less than the number of times the multiple feedforward neural network models are trained using the satellite data set. If the initial training model stipulates that the training is stopped if the MSE of the prediction performance in the training set does not decrease significantly within 20 times, 20 should be reduced during the training process of the relay training model.

[0134] According to the system of the second aspect of the present invention, the sixth processing module 106 is specifically configured to extract the corresponding existing global CO2 model CO2 data according to the longitude and latitude coordinates and time of the center of the global predefined spatial resolution grid, and obtain the grid time series data set including:

[0135] According to the longitude and latitude coordinates and time of the center of the global predefined spatial resolution grid, the CO2 data of the day, the CO2 data of the previous day and the CO2 data of the next day that already exist in the global CO2 model are extracted to obtain the grid time series dataset.

[0136] Specifically, according to the longitude and latitude coordinates and time of the center of the global predefined spatial resolution grid, the auxiliary variables of global vegetation characteristics and the auxiliary variables of global meteorological characteristics at the corresponding longitude and latitude coordinates and time are extracted. The missing values ​​of the auxiliary variables of global vegetation characteristics and the auxiliary variables of global meteorological characteristics should be set to 0 or Null and should not be discarded, so as to obtain the grid auxiliary variable data set.

[0137] According to the system of the second aspect of the present invention, the seventh processing module 107 is specifically configured as follows:

[0138]

[0139] Among them, XCO2 represents the global CO2 dataset;

[0140] According to the system of the second aspect of the present invention, the eighth processing module 108 is specifically configured as follows:

[0141] Perform weighted fusion of XCO2 global data for repeated time periods and calculate the fusion results;

[0142] There is an overlapped time period between each two adjacent FNN models. Similarly, when using the model generation, there will be global XCO2 data with repeated time periods. For the data in the repeated time period, the specific operations are as follows:

[0143] Let the two adjacent FNN models be A and B, and their RMSE on the test set be denoted as RMSE A , RMSE B The generated global XCO2 data for the repeated time period is recorded as XCO 2A 、XCO 2B , the final fusion result is recorded as XCO 2M .

[0144] (1) Calculate the corresponding weights of A and B when they are fused. Since the larger the RMSE, the worse the accuracy of the model, the weight calculation method is as follows:

[0145]

[0146] (2) Calculate the fusion results:

[0147]

[0148] The third aspect of the present invention discloses an electronic device. The electronic device includes a memory and a processor, the memory stores a computer program, and when the processor executes the computer program, the steps in the atmospheric CO2 data filling method using a feedforward neural network combined with time series changes in any one of the first aspects of the present invention are implemented.

[0149] Figure 3 is a structural diagram of an electronic device according to an embodiment of the present invention, such as Figure 3 As shown, the electronic device includes a processor, a memory, a communication interface, a display screen and an input device connected via a system bus. Among them, the processor of the electronic device is used to provide computing and control capabilities. The memory of the electronic device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The communication interface of the electronic device is used to communicate with an external terminal in a wired or wireless manner, and the wireless manner can be achieved through WIFI, an operator network, near field communication (NFC) or other technologies. The display screen of the electronic device can be a liquid crystal display screen or an electronic ink display screen, and the input device of the electronic device can be a touch layer covered on the display screen, or a button, a trackball or a touch pad set on the housing of the electronic device, or an external keyboard, touch pad or mouse, etc.

[0150] Those skilled in the art will understand that Figure 3 The structure shown in the figure is only a structural diagram of the part related to the technical solution of the present disclosure, and does not constitute a limitation on the electronic device to which the technical solution of the present application is applied. The specific electronic device may include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components.

[0151] The fourth aspect of the present invention discloses a computer-readable storage medium. The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps in the method for filling atmospheric CO2 data using a feedforward neural network combined with time series changes in any one of the first aspects of the present invention are implemented.

[0152] Please note that the technical features of the above embodiments can be combined arbitrarily. In order to make the description concise, all possible combinations of the technical features in the above embodiments are not described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification. The above embodiments only express several implementation methods of the present application, and their descriptions are relatively specific and detailed, but they cannot be understood as limiting the scope of the invention patent. It should be pointed out that for ordinary technicians in this field, without departing from the concept of the present application, several variations and improvements can be made, which all belong to the scope of protection of the present application. Therefore, the scope of protection of the patent in this application shall be based on the attached claims.

Claims

1. A method for filling atmospheric CO2 data using a feedforward neural network combined with time series changes, characterized in that: The method comprises: Step S1, obtaining satellite CO2 data, ground-based CO2 data and existing global CO2 model CO2 data, and resampling the satellite CO2 data, ground-based CO2 data and existing global CO2 model CO2 data to the same predefined spatial resolution to obtain first CO2 data; obtaining auxiliary variables of global vegetation characteristics and auxiliary variables of global vegetation characteristics, and resampling the auxiliary variables of global vegetation characteristics and auxiliary variables of global meteorological characteristics to the same predefined spatial resolution as the first CO2 data to obtain auxiliary variable data; Step S2, according to the observation records of the satellite CO2 data, extract the corresponding auxiliary variable data to obtain a satellite auxiliary variable data set; according to the observation records of the satellite CO2 data, extract the corresponding existing global CO2 model CO2 data to obtain a satellite time series data set; the satellite auxiliary variable data set and the satellite time series data set constitute a satellite data set; according to the observation records of the ground-based CO2 data, extract the corresponding auxiliary variable data to obtain a ground-based auxiliary variable data set; according to the observation records of the ground-based CO2 data, extract the corresponding existing global CO2 model CO2 data to obtain a ground-based time series data set; the ground-based auxiliary variable data set and the ground-based time series data set constitute a ground-based data set; Step S3, construct multiple feedforward neural network models; each feedforward neural network model is responsible for filling the atmospheric CO2 data within a period of time, and multiple models are combined to complete the atmospheric CO2 data filling in the case of a long time series; Step S4, training the multiple feedforward neural network models using the satellite data set to obtain an initial training model; Step S5: continue to train the initial training model using the ground-based data set to obtain a relay training model; Step S6, according to the longitude and latitude coordinates and time of the center of the global predefined spatial resolution grid, extract the auxiliary variables of the global vegetation characteristics and the auxiliary variables of the global meteorological characteristics to obtain a grid auxiliary variable data set; according to the longitude and latitude coordinates and time of the center of the global predefined spatial resolution grid, extract the corresponding existing global CO2 model CO2 data to obtain a grid time series data set; the grid auxiliary variable data set and the grid time series data set constitute a grid data set; Step S7, applying the relay training model and the grid dataset to generate a global CO2 dataset with a time resolution of one day and a spatial resolution of a predefined spatial resolution; Step S8: weighted fusion is performed on the XCO2 global data of the repeated time periods, and the fusion result is calculated.

2. The atmospheric CO2 data filling method using a feedforward neural network combined with time series changes according to claim 1 is characterized in that: In the step S1, the auxiliary variables of the global vegetation characteristics include: sunlight-induced chlorophyll fluorescence data and normalized vegetation index; In step S1, the auxiliary variables of global meteorological characteristics also include: U-shaped wind, V-shaped wind and sea level pressure.

3. The atmospheric CO2 data filling method using a feedforward neural network combined with time series changes according to claim 1 is characterized in that: In step S2, extracting corresponding auxiliary variable data according to the observation record of the satellite CO2 data to obtain a satellite auxiliary variable data set includes: According to the longitude and latitude coordinates and time of the observation record of the satellite CO2 data, the auxiliary variables of global vegetation characteristics and the auxiliary variables of global meteorological characteristics corresponding to the longitude and latitude coordinates and time are extracted, and the missing values ​​of the auxiliary variables of global vegetation characteristics and the auxiliary variables of global meteorological characteristics should be set to 0 or Null to obtain a satellite auxiliary variable data set; The extracting corresponding existing global CO2 model CO2 data according to the observation record of the satellite CO2 data to obtain the satellite time series data set includes: According to the latitude and longitude coordinates and time of the observation record of the satellite CO2 data, the CO2 data of the day existing in the global CO2 model, as well as the CO2 data of the previous day and the CO2 data of the next day are extracted to obtain a satellite time series data set.

4. The atmospheric CO2 data filling method using a feedforward neural network combined with time series changes according to claim 3 is characterized in that: In the step S2, the extracting corresponding auxiliary variable data according to the observation record of the ground-based CO2 data to obtain the ground-based auxiliary variable data set includes: extracting auxiliary variables of global vegetation characteristics and auxiliary variables of global meteorological characteristics at corresponding longitude and latitude coordinates and time according to the observation record of the ground-based CO2 data, and the missing values ​​of the auxiliary variables of global vegetation characteristics and the auxiliary variables of global meteorological characteristics should be set to 0 or Null to obtain the ground-based auxiliary variable data set; The extracting corresponding existing global CO2 model CO2 data according to the observation records of the ground-based CO2 data to obtain the ground-based time series data set includes: According to the latitude and longitude coordinates and time of the observation record of the ground-based CO2 data, the CO2 data of the day existing in the global CO2 model, as well as the CO2 data of the previous day and the CO2 data of the next day, are extracted, and the ground-based time series data set.

5. The atmospheric CO2 data filling method using a feedforward neural network combined with time series changes according to claim 4 is characterized in that: In the step S5, during the process of continuing to train the initial training model using the ground-based data set, the parameters of the initial training model need to be frozen; The number of times the initial training model is trained using the ground-based data set is less than the number of times the multiple feedforward neural network models are trained using the satellite data set.

6. The atmospheric CO2 data filling method using a feedforward neural network combined with time series changes according to claim 5 is characterized in that: In step S6, extracting the corresponding existing global CO2 model CO2 data according to the longitude and latitude coordinates and time of the center of the global predefined spatial resolution grid to obtain the grid time series data set includes: According to the longitude and latitude coordinates and time of the center of the global predefined spatial resolution grid, the CO2 data of the day, the CO2 data of the previous day and the CO2 data of the next day that already exist in the global CO2 model are extracted to obtain the grid time series dataset.

7. The atmospheric CO2 data filling method using a feedforward neural network combined with time series changes according to claim 6 is characterized in that: In step S8, the specific method of weighted fusion of XCO2 global data of repeated time periods and calculating the fusion result includes: There is an overlapped time period between each two adjacent FNN models. Similarly, there will be repeated time periods of global XCO2 data when using the model generation; For data in repeated time periods, the specific methods are as follows: Let the two adjacent FNN models be A and B, and their RMSE on the test set be denoted as RMSE A , RMSE B The global XCO2 data of the generated repeated time period is recorded as XCO 2A 、XCO 2B , the final fusion result is recorded as XCO 2M ; Calculate the corresponding weights of A and B when they are fused. Since the larger the RMSE, the worse the accuracy of the model, the weight calculation method is as follows: Calculate the fusion result:

8. A system for filling atmospheric CO2 data using a feedforward neural network combined with temporal changes, characterized in that: The system comprises: The first processing module is configured to obtain satellite CO2 data, ground-based CO2 data and existing global CO2 model CO2 data, and resample the satellite CO2 data, ground-based CO2 data and existing global CO2 model CO2 data to the same predefined spatial resolution to obtain first CO2 data; obtain auxiliary variables of global vegetation characteristics and auxiliary variables of global vegetation characteristics, and resample the auxiliary variables of global vegetation characteristics and auxiliary variables of global meteorological characteristics to the same predefined spatial resolution as the first CO2 data to obtain auxiliary variable data; The second processing module is configured to extract corresponding auxiliary variable data according to the observation record of the satellite CO2 data to obtain a satellite auxiliary variable data set; extract corresponding existing global CO2 model CO2 data according to the observation record of the satellite CO2 data to obtain a satellite time series data set; the satellite auxiliary variable data set and the satellite time series data set constitute a satellite data set; extract corresponding auxiliary variable data according to the observation record of the ground-based CO2 data to obtain a ground-based auxiliary variable data set; extract corresponding existing global CO2 model CO2 data according to the observation record of the ground-based CO2 data to obtain a ground-based time series data set; the ground-based auxiliary variable data set and the ground-based time series data set constitute a ground-based data set; The third processing module is configured to construct multiple feedforward neural network models; each feedforward neural network model is responsible for filling the atmospheric CO2 data within a period of time, and multiple models are combined to complete the atmospheric CO2 data filling in the case of a long time series; A fourth processing module is configured to train the multiple feedforward neural network models using the satellite data set to obtain an initial training model; A fifth processing module is configured to continue training the initial training model using the ground-based data set to obtain a relay training model; The sixth processing module is configured to extract auxiliary variables of global vegetation characteristics and auxiliary variables of global meteorological characteristics according to the longitude and latitude coordinates and time of the center of the global predefined spatial resolution grid to obtain a grid auxiliary variable data set; extract corresponding existing global CO2 model CO2 data according to the longitude and latitude coordinates and time of the center of the global predefined spatial resolution grid to obtain a grid time series data set; the grid auxiliary variable data set and the grid time series data set constitute a grid data set; A seventh processing module is configured to apply the relay training model and the grid dataset to generate a global CO2 dataset with a time resolution of one day and a spatial resolution of a predefined spatial resolution; The eighth processing module is configured to perform weighted fusion on the XCO2 global data of repeated time periods and calculate a fusion result.

9. An electronic device, characterized in that: The electronic device includes a memory and a processor, the memory stores a computer program, and when the processor executes the computer program, the steps of the atmospheric CO2 data filling method using a feedforward neural network combined with time series changes described in any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, and when the computer program is executed by the processor, the steps of the atmospheric CO2 data filling method using a feedforward neural network combined with time series changes according to any one of claims 1 to 7 are implemented.

Images