CN117216480B - Near-surface ozone remote sensing estimation method for deep coupling geographic space-time information - Google Patents
Near-surface ozone remote sensing estimation method for deep coupling geographic space-time information Download PDFInfo
- Publication number
- CN117216480B CN117216480B CN202311196916.4A CN202311196916A CN117216480B CN 117216480 B CN117216480 B CN 117216480B CN 202311196916 A CN202311196916 A CN 202311196916A CN 117216480 B CN117216480 B CN 117216480B
- Authority
- CN
- China
- Prior art keywords
- time
- space
- data
- geographic
- network
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
- CBENFWSGALASAD-UHFFFAOYSA-N Ozone Chemical compound [O-][O+]=O CBENFWSGALASAD-UHFFFAOYSA-N 0.000 title claims abstract description 68
- 238000000034 method Methods 0.000 title claims abstract description 32
- 230000008878 coupling Effects 0.000 title claims abstract description 21
- 238000010168 coupling process Methods 0.000 title claims abstract description 21
- 238000005859 coupling reaction Methods 0.000 title claims abstract description 21
- 238000013136 deep learning model Methods 0.000 claims abstract description 16
- 238000012549 training Methods 0.000 claims abstract description 12
- 238000011160 research Methods 0.000 claims abstract description 8
- 238000007781 pre-processing Methods 0.000 claims abstract description 5
- 238000004364 calculation method Methods 0.000 claims description 21
- 230000006870 function Effects 0.000 claims description 12
- 238000012795 verification Methods 0.000 claims description 11
- 238000012544 monitoring process Methods 0.000 claims description 9
- 210000002569 neuron Anatomy 0.000 claims description 9
- 238000010606 normalization Methods 0.000 claims description 6
- 238000013528 artificial neural network Methods 0.000 claims description 4
- 238000006243 chemical reaction Methods 0.000 claims description 4
- 238000002790 cross-validation Methods 0.000 claims description 4
- 208000037170 Delayed Emergence from Anesthesia Diseases 0.000 claims description 3
- 238000012952 Resampling Methods 0.000 claims description 3
- 238000010276 construction Methods 0.000 claims description 3
- 238000000605 extraction Methods 0.000 claims description 3
- 238000013507 mapping Methods 0.000 claims description 3
- 238000007670 refining Methods 0.000 claims description 3
- 230000009286 beneficial effect Effects 0.000 abstract description 2
- 238000012417 linear regression Methods 0.000 description 4
- 238000010586 diagram Methods 0.000 description 3
- 230000005855 radiation Effects 0.000 description 3
- 238000007637 random forest analysis Methods 0.000 description 3
- 238000013473 artificial intelligence Methods 0.000 description 2
- 230000000694 effects Effects 0.000 description 2
- 238000012986 modification Methods 0.000 description 2
- 230000004048 modification Effects 0.000 description 2
- 230000006978 adaptation Effects 0.000 description 1
- 239000000809 air pollutant Substances 0.000 description 1
- 231100001243 air pollutant Toxicity 0.000 description 1
- 238000007418 data mining Methods 0.000 description 1
- 238000013135 deep learning Methods 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 230000007613 environmental effect Effects 0.000 description 1
- 239000003344 environmental pollutant Substances 0.000 description 1
- 230000036541 health Effects 0.000 description 1
- 231100000719 pollutant Toxicity 0.000 description 1
- 230000000630 rising effect Effects 0.000 description 1
- 230000002123 temporal effect Effects 0.000 description 1
Classifications
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02A—TECHNOLOGIES FOR ADAPTATION TO CLIMATE CHANGE
- Y02A90/00—Technologies having an indirect contribution to adaptation to climate change
- Y02A90/10—Information and communication technologies [ICT] supporting adaptation to climate change, e.g. for weather forecasting or climate simulation
Landscapes
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
The invention relates to a near-surface ozone remote sensing estimation method of depth coupling geographic space-time information, which comprises the following steps: acquiring multi-source data of a research area and preprocessing the multi-source data; performing space-time matching on the preprocessed multi-source data, and constructing a space-time sample data set; constructing a deep learning model of coupling geographic space-time information; training the deep learning model by using a space-time sample data set, optimizing and determining model super-parameters, and verifying the overall accuracy and space-time robustness of the model; and carrying out large-scale remote sensing estimation of the near-surface ozone concentration by using the trained model. The beneficial effects of the invention are as follows: the method is used for deeply coupling geographic space-time information when estimating the near-surface ozone, improves the accuracy and space-time robustness of the remote sensing estimation model, and has important practical application value.
Description
Technical Field
The invention relates to the technical field of environmental remote sensing monitoring methods, in particular to a near-surface ozone remote sensing estimation method of depth coupling geographic space-time information.
Background
Near-surface ozone is one of six air pollutants, and seriously threatens the safety of the global ecological system and the health of people. In recent years, with the implementation of a series of policies for preventing and controlling atmospheric pollution, the concentration of PM 2.5 in China has been remarkably reduced, but at the same time, the concentration of ozone on the near-surface has a fluctuation rising trend, and the pollution season gradually extends from summer to winter and spring, so that the PM 2.5 becomes a new primary pollutant. Therefore, the satellite remote sensing is utilized to carry out large-scale dynamic monitoring on the near-surface ozone so as to clarify the time-space distribution characteristics of the near-surface ozone, and the method has important practical significance for preventing and treating ozone pollution.
The total amount of the ozone column inverted by satellite remote sensing represents the ozone concentration condition of the whole atmosphere column, and is influenced by meteorological conditions, artificial activities and the like, the ozone is not uniformly distributed in the vertical direction and continuously changes at any time, the near-surface ozone concentration and the total amount of the ozone column are in a complex nonlinear relation, and a high-performance model needs to be established to accurately estimate the near-surface ozone concentration. The current correlation models can be divided into two main types, namely a traditional multiple linear regression model and an emerging artificial intelligence model. The traditional multiple linear regression model is difficult to describe a complex nonlinear relation, so that the estimation accuracy of the near-surface ozone concentration is relatively low. Emerging artificial intelligent models such as random forests, deep neural networks and the like rely on data mining and strong nonlinear fitting capability, relatively high precision is obtained on near-surface ozone concentration estimation, but the geographic space-time heterogeneity of the near-surface ozone is not considered, so that the space-time robustness of the model is generally poor, and the precision is required to be further improved. Therefore, there is an urgent need for an artificial intelligence model that can fully incorporate geographical spatiotemporal context information to enable accurate remote sensing estimation of near-surface ozone.
Disclosure of Invention
The invention aims at overcoming the defects of the prior art, and provides a near-surface ozone remote sensing estimation method for deep coupling geographic space-time information.
In a first aspect, a near-surface ozone remote sensing estimation method for depth-coupled geographic spatiotemporal information is provided, including:
step 1, multi-source data of a research area are obtained, and the multi-source data are preprocessed;
step 2, performing space-time matching on the preprocessed multi-source data, and constructing a space-time sample data set;
Step 3, constructing a depth learning model coupling geographic space-time information, wherein the depth learning model comprises a geographic space information coding network, a geographic time sequence information coding network and a geographic space-time feature decoding network;
Step 4, training the deep learning model by utilizing the space-time sample data set, optimizing and determining model super-parameters, and verifying the overall accuracy and space-time robustness of the model;
And 5, carrying out large-scale remote sensing estimation of the near-surface ozone concentration by using the trained model.
Preferably, in step 1, the multi-source data includes: the method comprises the steps of satellite remote sensing inversion of ozone column total amount data TO3, meteorological data, normalized vegetation index NDVI data, digital elevation model DEM data and ozone monitoring data of a ground air quality station;
the pretreatment comprises the following steps: uniformly resampling meteorological data, NDVI data and DEM data TO the same spatial resolution as TO3 data inverted by satellite remote sensing; and eliminating the missing value in the ozone monitoring data of the ground air quality station and calculating the average ozone concentration of 8 hours per day.
Preferably, in step 2, the space-time matching includes:
2.1, acquiring longitude and latitude LAT and LON of a ground air quality site on a spatial scale, taking a corresponding grid as a center, extracting grid pixel values in a 7X 7 range around TO3 data and meteorological data, and extracting only center grid pixel values for NDVI and DEM data;
2.2, calculating annual product day DOY of the target day on a time scale, and extracting data of 3 days before and after the target day from the total amount data of the ozone column and the meteorological data by taking the target day as a reference, wherein NDVI data are used for extracting current month data of the target day;
And 2.3, constructing a space-time sample data set, wherein the space-time sample data set consists of a three-dimensional space-time variable and a one-dimensional dummy variable.
Preferably, in step 3, the geospatial information coding network includes: a 1 x 1 convolutional layer for mapping input data to a high-dimensional feature space; a residual module of a 1×1 convolution kernel, configured to extract a coupling relationship between features and increase nonlinear expression; a residual module of a 3 x 3 convolution kernel for depth extraction of local geospatial dependencies of features; a 7 x 7 global convolution layer for identifying global geospatial dependencies of features; the structural calculation formula of the geospatial information coding network is as follows:
FS＝Conv_7(ResBlock_3(ResBlock_1(Conv_1(x))))
In the above formula, FS is an extracted geospatial feature, x is an input three-dimensional space-time variable, conv_1 and conv_7 represent convolution layers with convolution kernel sizes of 1×1 and 7×7, and ResBlock _1 and ResBlock _3 represent residual modules with convolution kernel sizes of 1×1 and 3×3, respectively;
the residual block calculation formula is expressed as follows:
x′＝ReLU(x+BN(Conv_i(ReLU(BN(Conv_i(x))))))
in the above formula, x' is the output result of the residual module, x is the input data, conv_i represents the convolution layer with the convolution kernel size of i×i, and ReLU and BN represent the linear rectification function and the batch normalization layer respectively.
Preferably, in step 3, the timing information encoding network is a modified transducer network, including: the position coding module is used for endowing the extracted geospatial information features with time sequence information which can be identified by the network; the multi-head self-attention module is used for extracting time sequence characteristics affecting near-surface ozone concentration estimation; a feed-forward network for further refining the spatiotemporal features; the structural calculation formula of the time sequence information coding network is as follows:
FST＝FFN(Multihead(Pos(FS)))
In the above formula, FST is the extracted geospatial feature, FS is the geospatial feature extracted by the geospatial information encoding network, pos represents the position encoding module, multihead represents the multi-head self-attention module, and FFN represents the feed-forward network.
Preferably, in step 3, the calculation formula of the position coding module is as follows:
In the above formula, p represents the p-th time point, i represents the i-th feature, and c is the total feature number;
The calculation formula of the multi-head self-attention module is expressed as follows:
MA(Q,K,V)＝Concat(head1,…,headh)WO
Where FS' is a position-coded geospatial feature, AndThe weights of Query (Q), key (K) and Value (V) of the ith self-attention head are represented respectively, softmax is a normalized exponential function, h is the number of multi-head self-attention, W O is the weight of an output layer, concat is a connection function, and MA is the multi-head self-attention result;
the FFN calculation formula is expressed as follows:
FST＝LN(LN(x)+FC(ReLU(FC(LN(MA)))
in the above formula, LN represents layer normalization operation, and FC represents fully connected layers.
Preferably, in step 3, the geographic spatiotemporal feature decoding network includes: the connection module is used for connecting the geographic space-time characteristics extracted by the geographic space information coding network and the time sequence information coding network together with the one-dimensional Dummy variable so as to acquire more geographic space-time characteristics; the deep neural network DNN comprises a plurality of full-connection layers and a ReLU function, and is used for performing characteristic nonlinear conversion decoding and finally realizing near-surface ozone concentration estimation; the whole structural calculation formula is expressed as follows:
O3＝DNN(Concat(FST,Dummy)
wherein DNN is represented as follows:
In the above-mentioned method, the step of, The ith neuron representing the first layer, k l represents the total number of neurons of the first layer, and W ji and b ji represent neurons, respectivelyAndThe weights and residuals between, L is the total number of layers.
Preferably, in step 4, the model super-parameters include the size of the geospatial neighborhood, the length of the neighborhood time window, the number of internal feature channels of the geospatial information coding network, the number of multi-head self-attentions of the time sequence information coding network, the number of hidden layers of the geospatial feature decoding network, and the optimal result is selected through 5-fold cross validation to determine the final parameter combination.
Preferably, in step 4, the overall accuracy of the model is verified by 5-fold cross verification based on samples, the time-based and site-based 5-fold cross verification are used for verifying the space-time robustness of the model, and the determination coefficient R 2, the root mean square error RMSE and the absolute average error MAE are used for evaluating the verification accuracy.
In a second aspect, a near-surface ozone remote sensing estimation system for deep coupling geographic space-time information is provided, and the near-surface ozone remote sensing estimation method for deep coupling geographic space-time information in the first aspect is performed, and includes:
The acquisition module is used for acquiring multi-source data of the research area and preprocessing the multi-source data;
The matching module is used for carrying out space-time matching on the preprocessed multi-source data and constructing a space-time sample data set;
The construction module is used for constructing a deep learning model coupled with the geographic space-time information, and the deep learning model comprises a geographic space information coding network, a geographic time sequence information coding network and a geographic space-time feature decoding network;
The training module is used for training the deep learning model by utilizing the space-time sample data set, optimizing and determining model super-parameters, and verifying the overall accuracy and space-time robustness of the model;
And the estimation module is used for carrying out large-scale remote sensing estimation of the near-surface ozone concentration by using the trained model.
The beneficial effects of the invention are as follows: according to the method, firstly, the multi-source data are processed into a space-time sample data set composed of three-dimensional space-time variables and one-dimensional dummy variables according to the characteristics of the multi-source data, then a targeted geographic space information coding network and a time sequence information coding network are constructed to deeply extract geographic space-time context characteristics, and then a geographic space-time characteristic decoding network is utilized to carry out nonlinear conversion decoding of the geographic space-time characteristics, so that near-surface ozone high-precision estimation is finally realized. Compared with the existing mainstream model, the method has the advantages that geographic space-time information is deeply coupled when near-surface ozone is estimated, the accuracy and space-time robustness of the remote sensing estimation model are improved, and the method has important practical application value.
Drawings
FIG. 1 is a flow chart of a near-surface ozone remote sensing estimation method of depth coupling geographic space-time information provided by the invention;
FIG. 2 is a schematic diagram of a geospatial information encoding network architecture provided by the present invention;
FIG. 3 is a schematic diagram of a timing information encoding network architecture according to the present invention;
FIG. 4 is a schematic diagram of a geographic spatiotemporal feature decoding network architecture provided by the present invention;
FIG. 5 is a graph comparing the cross-validation results of the method provided by the present invention with the mainstream method.
Detailed Description
The invention is further described below with reference to examples. The following examples are presented only to aid in the understanding of the invention. It should be noted that it will be apparent to those skilled in the art that modifications can be made to the present invention without departing from the principles of the invention, and such modifications and adaptations are intended to be within the scope of the invention as defined in the following claims.
Example 1:
As shown in fig. 1, embodiment 1 of the present application provides a near-surface ozone remote sensing estimation method for deep coupling geographic space-time information, which includes:
Step 1, multi-source data of a research area are obtained, and the multi-source data are preprocessed.
Specifically, the multi-source data includes: the method comprises the steps of satellite remote sensing inversion of ozone column total amount data TO3, meteorological data, normalized vegetation index NDVI data, digital elevation model DEM data and ozone monitoring data of a ground air quality station; the meteorological data comprise air temperature T2M at the position of 2 meters on the earth surface, dew point temperature D2M at the position of 2 meters on the earth surface, earth surface pressure SP, weft wind speed U10 at the position of 10 meters on the earth surface, warp wind speed V10 at the position of 10 meters on the earth surface, atmospheric layer top height BLH, relative humidity RH, O3 mass mixing ratio OMR, earth surface net solar radiation SSR, earth surface net heat radiation STR and earth surface downlink ultraviolet radiation UVB.
The pretreatment comprises the following steps: uniformly resampling meteorological data, NDVI data and DEM data TO the same spatial resolution as TO3 data inverted by satellite remote sensing; and eliminating the missing value in the ozone monitoring data of the ground air quality station and calculating the average ozone concentration of 8 hours per day.
And step 2, performing space-time matching on the preprocessed multi-source data, and constructing a space-time sample data set.
In step 2, the space-time matching includes:
2.1, acquiring longitude and latitude LAT and LON of a ground air quality site on a spatial scale, taking a corresponding grid as a center, extracting grid pixel values in a 7X 7 range around TO3 data and meteorological data, and extracting only center grid pixel values for NDVI and DEM data;
2.2, calculating annual product day DOY of the target day on a time scale, and extracting data of 3 days before and after the target day from the total amount data of the ozone column and the meteorological data by taking the target day as a reference, wherein NDVI data are used for extracting current month data of the target day;
And 2.3, constructing a space-time sample data set, wherein the space-time sample data set consists of a three-dimensional space-time variable and a one-dimensional dummy variable. In particular, three-dimensional space-time variables include TO3, T2M, D2M, SP, U10, V10, BLH, RH, OMR, SSR, STR and UVB, each variable having dimensions of 7 x 7, 7 x 7 neighborhood spatiotemporal information representing adjacent 7 days; one-dimensional dummy variables include NDVI, DEM, LAT, LON and DOY.
And 3, constructing a depth learning model coupling the geographic space-time information, wherein the depth learning model comprises a geographic space information coding network, a geographic time sequence information coding network and a geographic space-time feature decoding network.
And 4, training the deep learning model by using the space-time sample data set, optimizing and determining model super-parameters, and verifying the overall accuracy and space-time robustness of the model.
And 5, carrying out large-scale remote sensing estimation of the near-surface ozone concentration by using the trained model.
Example 2:
Based on embodiment 1, embodiment 2 of the present application provides a more specific near-surface ozone remote sensing estimation method for depth-coupled geographic spatiotemporal information, which comprises:
Step 1, multi-source data of a research area are obtained, and the multi-source data are preprocessed.
And step 2, performing space-time matching on the preprocessed multi-source data, and constructing a space-time sample data set.
And 3, constructing a depth learning model coupling the geographic space-time information, wherein the depth learning model comprises a geographic space information coding network, a geographic time sequence information coding network and a geographic space-time feature decoding network.
In step 3, as shown in fig. 2, the geospatial information encoding network includes: a1 x 1 convolutional layer for mapping input data to a high-dimensional feature space; a residual module of a1×1 convolution kernel, configured to extract a coupling relationship between features and increase nonlinear expression; a residual module of a3 x3 convolution kernel for depth extraction of local geospatial dependencies of features; a 7 x 7 global convolution layer for identifying global geospatial dependencies of features; the structural calculation formula of the geospatial information coding network is as follows:
FS＝Conv_7(ResBlock_3(ResBlock_1(Conv_1(x))))
In the above formula, FS is an extracted geospatial feature, x is an input three-dimensional space-time variable, conv_1 and conv_7 represent convolution layers with convolution kernel sizes of 1×1 and 7×7, and ResBlock _1 and ResBlock _3 represent residual modules with convolution kernel sizes of 1×1 and 3×3, respectively;
the residual block calculation formula is expressed as follows:
x′＝ReLU(x+BN(Conv_i(ReLU(BN(Conv_i(x))))))
in the above formula, x' is the output result of the residual module, x is the input data, conv_i represents the convolution layer with the convolution kernel size of i×i, and ReLU and BN represent the linear rectification function and the batch normalization layer respectively.
In step 3, the timing information encoding network is an improved transform network, which eliminates the decoding module of the conventional transform network, simplifies the encoding module, and can effectively improve the estimation efficiency of the model, as shown in fig. 3, and specifically includes: the position coding module is used for endowing the extracted geospatial information features with time sequence information which can be identified by the network; the multi-head self-attention module is used for extracting time sequence characteristics affecting near-surface ozone concentration estimation; a feed-forward network for further refining the spatiotemporal features; the structural calculation formula of the time sequence information coding network is as follows:
FST＝FFN(Multihead(Pos(FS)))
In the above formula, FST is the extracted geospatial feature, FS is the geospatial feature extracted by the geospatial information encoding network, pos represents the position encoding module, multihead represents the multi-head self-attention module, and FFN represents the feed-forward network.
In step 3, the calculation formula of the position coding module is expressed as follows:
In the above formula, p represents the p-th time point, i represents the i-th feature, and c is the total feature number;
The calculation formula of the multi-head self-attention module is expressed as follows:
MA(Q,K,V)＝Concat(head1,…,headh)WO
Where FS' is a position-coded geospatial feature, AndThe weights of Query (Q), key (K) and Value (V) of the ith self-attention head are represented respectively, softmax is a normalized exponential function, h is the number of multi-head self-attention, W O is the weight of an output layer, concat is a connection function, and MA is the multi-head self-attention result;
the FFN calculation formula is expressed as follows:
FST＝LN(LN(x)+FC(ReLU(FC(LN(MA)))
in the above formula, LN represents layer normalization operation, and FC represents fully connected layers.
In step 3, the geographic space-time feature decoding network is as shown in fig. 4, and includes: the connection module is used for connecting the geographic space-time characteristics extracted by the geographic space information coding network and the time sequence information coding network together with the one-dimensional Dummy variable so as to acquire more geographic space-time characteristics; the deep neural network DNN comprises a plurality of full-connection layers and a ReLU function, and is used for performing characteristic nonlinear conversion decoding and finally realizing near-surface ozone concentration estimation; the whole structural calculation formula is expressed as follows:
O3＝DNN(Concat(FST,Dummy)
wherein DNN is represented as follows:
In the above-mentioned method, the step of, The ith neuron representing the first layer, k l represents the total number of neurons of the first layer, and W ji and b ji represent neurons, respectivelyAndThe weights and residuals between, L is the total number of layers.
And 4, training the deep learning model by using the space-time sample data set, optimizing and determining model super-parameters, and verifying the overall accuracy and space-time robustness of the model.
In step 4, the model super parameters comprise the size of the geographic space neighborhood, the length of the neighborhood time window, the number of characteristic channels in the geographic space information coding network, the number of multi-head self-attentions of the time sequence information coding network and the number of hidden layers of the geographic space-time characteristic decoding network, and the optimal result is selected through 5-fold cross verification to determine the final parameter combination.
In step 4, the overall accuracy of the model is verified by 5-fold cross verification based on samples, the time-based and site-based 5-fold cross verification are respectively adopted for verifying the space-time robustness of the model, and the determination coefficient R 2, the root mean square error RMSE and the absolute average error MAE are adopted for evaluating the verification accuracy.
And 5, carrying out large-scale remote sensing estimation of the near-surface ozone concentration by using the trained model.
The effect of the present invention is further analyzed by specific experimental results as follows:
Firstly, collecting multisource data of 1 month 1 day to 31 months 12 years in a long triangular region 2022, wherein the multisource data comprises ozone column total amount data of sentinel No. 5 TROPOMI, analysis data meteorological data of European medium weather forecast center ERA5, NDVI data of MODIS of the American national aviation and aerospace agency, DEM data of USGS of the American geological survey agency and ground site monitoring data of national environment monitoring center, and preprocessing and constructing a required space-time sample data set according to the method.
Then according to the method, a pytorch deep learning tool is used for constructing a deep learning model for coupling geographic space-time information, training of the model is carried out by utilizing a space-time sample data set, super-parameters are determined, and overall accuracy and space-time robustness are verified. As shown in FIG. 5, the sample-based five-fold cross validation shows that the model decision coefficient R 2, the root mean square error RMSE and the absolute average error MAE are respectively 0.94, 10.25 mug/m 3 and 7.52 mug/m 3, which are higher than the mainstream random forest and the multiple linear regression model, and the model has excellent near-surface ozone estimation accuracy. R 2 of the cross verification result based on the site and time of the model reaches 0.94 and 0.76 respectively, and is also obviously higher than the mainstream random forest and multiple linear regression model, so that the space-time robustness of the model to near-surface ozone estimation is proved.
In this embodiment, the same or similar parts as those in embodiment 1 may be referred to each other, and will not be described in detail in the present disclosure.
Example 3:
Based on embodiments 1 and 2, embodiment 3 of the present application provides a near-surface ozone remote sensing estimation system for depth-coupled geographic spatial and temporal information, which comprises:
The acquisition module is used for acquiring multi-source data of the research area and preprocessing the multi-source data;
The matching module is used for carrying out space-time matching on the preprocessed multi-source data and constructing a space-time sample data set;
The construction module is used for constructing a deep learning model coupled with the geographic space-time information, and the deep learning model comprises a geographic space information coding network, a geographic time sequence information coding network and a geographic space-time feature decoding network;
The training module is used for training the deep learning model by utilizing the space-time sample data set, optimizing and determining model super-parameters, and verifying the overall accuracy and space-time robustness of the model;
And the estimation module is used for carrying out large-scale remote sensing estimation of the near-surface ozone concentration by using the trained model.
Specifically, the system provided in this embodiment is a system corresponding to the method provided in embodiments 1 and 2, so that the portions in this embodiment that are the same as or similar to those in embodiments 1 and 2 may be referred to each other, and will not be described in detail in this disclosure.
Claims (6)
1. A near-surface ozone remote sensing estimation method of depth coupling geographic space-time information is characterized by comprising the following steps:
step 1, multi-source data of a research area are obtained, and the multi-source data are preprocessed;
step 2, performing space-time matching on the preprocessed multi-source data, and constructing a space-time sample data set;
Step 3, constructing a depth learning model coupling geographic space-time information, wherein the depth learning model comprises a geographic space information coding network, a geographic time sequence information coding network and a geographic space-time feature decoding network;
In step 3, the geospatial information encoding network includes: a1 x1 convolutional layer for mapping input data to a high-dimensional feature space; a residual module of a1×1 convolution kernel, configured to extract a coupling relationship between features and increase nonlinear expression; a residual module of a 3 x 3 convolution kernel for depth extraction of local geospatial dependencies of features; a 7 x 7 global convolution layer for identifying global geospatial dependencies of features; the structural calculation formula of the geospatial information coding network is as follows:
FS＝Conv_7(ResBlock_3(ResBlock_1(Conv_1(x))))
In the above formula, FS is an extracted geospatial feature, x is an input three-dimensional space-time variable, conv_1 and conv_7 represent convolution layers with convolution kernel sizes of 1×1 and 7×7, and ResBlock _1 and ResBlock _3 represent residual modules with convolution kernel sizes of 1×1 and 3×3, respectively;
the residual block calculation formula is expressed as follows:
x′＝ReLU(x+BN(Conv_i(ReLU(BN(Conv_i(x))))))
In the above formula, x' is the output result of the residual error module, x is input data, conv_i represents a convolution layer with a convolution kernel size of i×i, and ReLU and BN represent a linear rectification function and a batch normalization layer respectively;
In step 3, the timing information encoding network is a modified transducer network, including: the position coding module is used for endowing the extracted geospatial information features with time sequence information which can be identified by the network; the multi-head self-attention module is used for extracting time sequence characteristics affecting near-surface ozone concentration estimation; a feed-forward network for further refining the spatiotemporal features; the structural calculation formula of the time sequence information coding network is as follows:
FST＝FFN(Multihead(Pos(FS)))
In the above formula, FST is the extracted geospatial features, FS is the geospatial features extracted by the geospatial information coding network, pos represents the position coding module, multihead represents the multi-head self-attention module, and FFN represents the feedforward network;
In step 3, the calculation formula of the position coding module is expressed as follows:
In the above formula, p represents the p-th time point, i represents the i-th feature, and c is the total feature number;
The calculation formula of the multi-head self-attention module is expressed as follows:
MA(Q,K,V)＝Concat(head1,…,headh)WO
Where FS' is a position-coded geospatial feature, AndThe weights of Query (Q), key (K) and Value (V) of the ith self-attention head are represented respectively, softmax is a normalized exponential function, h is the number of multi-head self-attention, W O is the weight of an output layer, concat is a connection function, and MA is the multi-head self-attention result;
the FFN calculation formula is expressed as follows:
FST＝LN(LN(x)+FC(ReLU(FC(LN(MA)))
In the above formula, LN represents layer normalization operation, and FC represents a full connection layer;
In step 3, the geographic spatiotemporal feature decoding network includes: the connection module is used for connecting the geographic space-time characteristics extracted by the geographic space information coding network and the time sequence information coding network together with the one-dimensional Dummy variable so as to acquire more geographic space-time characteristics; the deep neural network DNN comprises a plurality of full-connection layers and a ReLU function, and is used for performing characteristic nonlinear conversion decoding and finally realizing near-surface ozone concentration estimation; the whole structural calculation formula is expressed as follows:
O3＝DNN(Concat(FST,Dummy)
wherein DNN is represented as follows:
In the above formula, Y i l represents the ith neuron of the first layer, k l represents the total number of neurons of the first layer, and W ji and b ji represent neurons Y i l and b ji, respectively The weight and residual error between the two layers, L is the total layer number;
Step 4, training the deep learning model by utilizing the space-time sample data set, optimizing and determining model super-parameters, and verifying the overall accuracy and space-time robustness of the model;
And 5, carrying out large-scale remote sensing estimation of the near-surface ozone concentration by using the trained model.
2. The method for estimating the near-surface ozone remote sensing of the deep coupled geographical spatiotemporal information of claim 1, wherein in step 1, the multi-source data comprises: the method comprises the steps of satellite remote sensing inversion of ozone column total amount data TO3, meteorological data, normalized vegetation index NDVI data, digital elevation model DEM data and ozone monitoring data of a ground air quality station;
the pretreatment comprises the following steps: uniformly resampling meteorological data, NDVI data and DEM data TO the same spatial resolution as TO3 data inverted by satellite remote sensing; and eliminating the missing value in the ozone monitoring data of the ground air quality station and calculating the average ozone concentration of 8 hours per day.
3. The method for estimating near-surface ozone remote sensing of deep coupled geographic spatiotemporal information of claim 2, wherein in step 2, the spatiotemporal matching comprises:
2.1, acquiring longitude and latitude LAT and LON of a ground air quality site on a spatial scale, taking a corresponding grid as a center, extracting grid pixel values in a 7X 7 range around TO3 data and meteorological data, and extracting only center grid pixel values for NDVI and DEM data;
2.2, calculating annual product day DOY of the target day on a time scale, and extracting data of 3 days before and after the target day from the total amount data of the ozone column and the meteorological data by taking the target day as a reference, wherein NDVI data are used for extracting current month data of the target day;
And 2.3, constructing a space-time sample data set, wherein the space-time sample data set consists of a three-dimensional space-time variable and a one-dimensional dummy variable.
4. The method of claim 3, wherein in step 4, the model super-parameters include the size of the geospatial neighborhood, the length of the neighborhood time window, the number of channels of features in the geospatial information encoding network, the number of multi-headed self-attentiveness of the time sequence information encoding network, the number of hidden layers of the geospatial feature decoding network, and the optimal result is selected by 5-fold cross-validation to determine the final parameter combination.
5. The method for estimating the near-surface ozone remote sensing of the deep coupling geographic space-time information according to claim 4, wherein in the step 4, the model overall accuracy is verified by adopting sample-based 5-fold cross verification, the model space-time robustness is verified by adopting time-based and site-based 5-fold cross verification respectively, and the verification accuracy is evaluated by adopting a decision coefficient R 2, a root mean square error RMSE and an absolute average error MAE.
6. A near-surface ozone remote sensing estimation system for depth-coupled geographic spatiotemporal information, for performing the near-surface ozone remote sensing estimation method for depth-coupled geographic spatiotemporal information of any of claims 1 to 5, comprising:
The acquisition module is used for acquiring multi-source data of the research area and preprocessing the multi-source data;
The matching module is used for carrying out space-time matching on the preprocessed multi-source data and constructing a space-time sample data set;
The construction module is used for constructing a deep learning model coupled with the geographic space-time information, and the deep learning model comprises a geographic space information coding network, a geographic time sequence information coding network and a geographic space-time feature decoding network;
The training module is used for training the deep learning model by utilizing the space-time sample data set, optimizing and determining model super-parameters, and verifying the overall accuracy and space-time robustness of the model;
And the estimation module is used for carrying out large-scale remote sensing estimation of the near-surface ozone concentration by using the trained model.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202311196916.4A CN117216480B (en) | 2023-09-18 | 2023-09-18 | Near-surface ozone remote sensing estimation method for deep coupling geographic space-time information |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202311196916.4A CN117216480B (en) | 2023-09-18 | 2023-09-18 | Near-surface ozone remote sensing estimation method for deep coupling geographic space-time information |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN117216480A CN117216480A (en) | 2023-12-12 |
| CN117216480B true CN117216480B (en) | 2024-06-28 |
Family
ID=89045757
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202311196916.4A Active CN117216480B (en) | 2023-09-18 | 2023-09-18 | Near-surface ozone remote sensing estimation method for deep coupling geographic space-time information |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN117216480B (en) |
Families Citing this family (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN118212548B (en) * | 2024-03-21 | 2024-09-13 | 宁波大学 | Cascade depth network-based particulate matter and ozone remote sensing cooperative estimation method |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111310386A (en) * | 2020-02-13 | 2020-06-19 | 北京中科锐景科技有限公司 | Near-surface ozone concentration estimation method |
| CN112287294A (en) * | 2020-09-10 | 2021-01-29 | 河海大学 | Time-space bidirectional soil water content interpolation method based on deep learning |
Family Cites Families (10)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112287978B (en) * | 2020-10-07 | 2022-04-15 | 武汉大学 | Hyperspectral remote sensing image classification method based on self-attention context network |
| CN112905560B (en) * | 2021-02-02 | 2022-10-11 | 中国科学院地理科学与资源研究所 | Air pollution prediction method based on multi-source time-space big data deep fusion |
| CN113189014B (en) * | 2021-04-14 | 2023-05-02 | 西安交通大学 | Ozone concentration estimation method integrating satellite remote sensing and ground monitoring data |
| CN113324923B (en) * | 2021-06-07 | 2023-07-07 | 郑州大学 | Remote sensing water quality inversion method combining space-time fusion and deep learning |
| CN113971477A (en) * | 2021-09-26 | 2022-01-25 | 浙江大学 | PM based on time series and deep learning framework2.5Estimation method |
| CN114021436B (en) * | 2021-10-26 | 2024-07-26 | 武汉大学 | Near-ground ozone inversion method based on near-ground ultraviolet radiation |
| CN114611792B (en) * | 2022-03-11 | 2023-05-02 | 南通大学 | Atmospheric ozone concentration prediction method based on mixed CNN-converter model |
| CN114926749B (en) * | 2022-07-22 | 2022-11-04 | 山东大学 | Near-surface atmospheric pollutant inversion method and system based on remote sensing image |
| CN115601281A (en) * | 2022-11-04 | 2023-01-13 | 吉林大学（Cn） | Remote sensing image space-time fusion method and system based on deep learning and electronic equipment |
| CN116705186A (en) * | 2023-06-12 | 2023-09-05 | 成都信息工程大学 | Deep learning-based wind cloud satellite near-surface air temperature inversion method |
-
2023
- 2023-09-18 CN CN202311196916.4A patent/CN117216480B/en active Active
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111310386A (en) * | 2020-02-13 | 2020-06-19 | 北京中科锐景科技有限公司 | Near-surface ozone concentration estimation method |
| CN112287294A (en) * | 2020-09-10 | 2021-01-29 | 河海大学 | Time-space bidirectional soil water content interpolation method based on deep learning |
Also Published As
| Publication number | Publication date |
|---|---|
| CN117216480A (en) | 2023-12-12 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109214592B (en) | Multi-model-fused deep learning air quality prediction method | |
| CN110346517B (en) | Smart city industrial atmosphere pollution visual early warning method and system | |
| CN112070234B (en) | Water chlorophyll and phycocyanin land-based remote sensing machine learning algorithm under complex scene | |
| CN111626518A (en) | Urban daily water demand online prediction method based on deep learning neural network | |
| CN114707688A (en) | Photovoltaic power ultra-short-term prediction method based on satellite cloud chart and space-time neural network | |
| CN114897250B (en) | Space and statistical feature fused CNN-GRU ozone concentration prediction model establishment method, prediction method and model | |
| CN117216480B (en) | Near-surface ozone remote sensing estimation method for deep coupling geographic space-time information | |
| CN112163375A (en) | Long-time sequence near-surface ozone inversion method based on neural network | |
| CN113344149B (en) | PM2.5 hourly prediction method based on neural network | |
| CN115237896B (en) | Data preprocessing method and system based on deep learning forecast air quality | |
| CN114970184B (en) | Synchronous inversion high-resolution artificial CO 2 Emission and natural CO 2 Flux assimilation method and system | |
| CN113987912A (en) | Pollutant on-line monitoring system based on geographic information | |
| CN112183625A (en) | PM based on deep learning2.5High-precision time-space prediction method | |
| CN113063737A (en) | Ocean heat content remote sensing inversion method combining remote sensing and buoy data | |
| CN113836808A (en) | PM2.5 deep learning prediction method based on heavy pollution feature constraint | |
| CN110852493A (en) | Atmospheric PM2.5 concentration prediction method based on multiple model comparisons | |
| CN118298222A (en) | Short-term precipitation prediction method based on multi-scale attention and convolution fusion | |
| CN117974401A (en) | Ecological restoration area intelligent identification method based on multi-source data and model integration | |
| CN110046756B (en) | Short-term weather forecasting method based on wavelet denoising and Catboost | |
| Li et al. | Deepphysinet: Bridging deep learning and atmospheric physics for accurate and continuous weather modeling | |
| CN115169646A (en) | Small-scale earth surface ozone concentration remote sensing method utilizing sunflower satellite data | |
| CN117589646B (en) | Method, device, equipment and medium for monitoring concentration of atmospheric fine particulate matters | |
| CN109905190B (en) | Modeling method for low-frequency ground wave propagation time delay variation characteristic | |
| CN107274024A (en) | A kind of meteorological station measures daily global radiation radiant exposure prediction optimization method | |
| CN117972625A (en) | Attention neural network data assimilation method based on four-dimensional variation constraint |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |