A PM based on time series and deep learning framework 2.5 Estimation Method

Abstract

The present invention discloses a PM based on time series and deep learning framework 2.5 The estimation method includes obtaining the PM of the area to be estimated. 2.5 Ground observation data, AOD product data, meteorological data, land use / cover and economic and social data are used together as multi-source data; the multi-source data are encoded into a time series format to obtain time series data; the spatial module, temporal module and enhancement module of the spatiotemporal enhanced neural network model are sequentially constructed to treat the PM 2.5 Perform inversion estimation; input the data of the prediction set into the spatiotemporal enhanced neural network model that has been tested for stability and accuracy, and calculate the PM 2.5 Perform spatial continuous estimation and obtain PM with a spatial resolution of 1 km 2.5 The present invention makes full use of PM 2.5 The temporal and spatial changes of remote sensing data can generate long-term, more detailed, more reliable and higher-resolution PM 2.5 The product has important theoretical and practical significance for the prevention and control of air pollution in my country.

Description

Technical Field

[0001] The present invention belongs to the field of atmospheric pollutant inversion prediction, and specifically relates to a PM2.5 prediction method based on time series and deep learning framework. 2.5 Estimation method. Background Art

[0002] PM 2.5 Refers to particles with an aerodynamic diameter of less than or equal to 2.5 μg. According to an article in The Lancet, PM 2.5 In 2015, it became the fifth leading risk factor for death in the world (Cohen et al., 2017), causing approximately 4.2 million deaths worldwide. 2.5 The number of monitoring stations increases year by year. However, the geographical coverage of these monitoring networks remains limited and uneven, meaning that the concentrations collected are not necessarily representative of PM2.5 levels in the region. 2.5 PM concentrations are difficult to assess from ground-level measurements alone. 2.5 regional variability.

[0003] Compared with traditional ground monitoring, satellite remote sensing technology has the characteristics of macroscopic and rapid information acquisition, and has gradually become an important monitoring method. PM based on satellite remote sensing 2.5 The optical aerosol thickness (AOD) is often used in forecasts. The optical aerosol thickness describes the attenuation of light by aerosols and has been shown to be closely related to near-ground particulate matter. 2.5 The inverted remote sensing products include MODIS, MISR, etc. 2.5 and AOD remote sensing products, early studies have developed various methods for estimating PM 2.5 Statistical regression methods include simple linear regression, multivariate linear regression, and geostatistical regression considering spatiotemporal heterogeneity, such as the geographically weighted regression (GWR) model and the geographically and temporally weighted regression (GTWR) model. At the same time, machine learning models are also widely used to estimate PM. 2.5 However, PM 2.5 It is determined by a variety of related natural conditions and human factors, with strong spatial variation, correlation and time dependence. The statistical regression method has weak feature learning ability and cannot be used in PM. 2.5 and AOD; machine learning models can fit nonlinear relationships well, but they are only based on big data feature modeling, which is hidden in PM 2.5 Physical features in time series and geospatial distribution are not fully exploited.

[0004] In recent years, deep learning has achieved remarkable success in discovering complex nonlinear relationships of target objects from time series. For example, two prominent deep learning models, recurrent neural network (RNN) and its variant LSTM, are specifically designed to identify time-varying patterns of long time series and capture complex nonlinear relationships. 2.5 Therefore, some researchers have applied deep learning methods to estimate and predict PM using long time series data. 2.5 Qi et al. used the LSTM model to estimate the hourly PM 2.5 In order to further improve the spatiotemporal learning ability of neural networks, Pak et al. combined a variety of neural networks: convolutional neural network (CNN), artificial neural network (ANN) and LSTM to predict daily PM 2.5 However, the weak spatial generalization ability of deep learning models limits their application in PM 2.5 In addition, for PM 2.5 Most deep learning models for forecasting are conducted on short time scales (i.e., daily and hourly); the potential of deep learning models for long-term trend analysis has not been explored. Moreover, most RNN models based on time series ignore spatial heterogeneity and spatial correlation. Therefore, it is necessary to establish a robust remote sensing product-based model that can better handle PM 2.5 Spatial correlation and temporal dependence of PM 2.5 Estimation models are critical. Summary of the invention

[0005] The purpose of the present invention is to solve the problems existing in the prior art and provide a PM based on time series and deep learning framework. 2.5 Estimation method.

[0006] The specific technical solutions adopted by the present invention are as follows:

[0007] The present invention provides a PM based on time series and deep learning framework 2.5 The estimation method is as follows:

[0008] S1: Get the PM of the area to be estimated 2.5 Ground observation data, AOD product data, meteorological data, land use / cover and economic and social data are taken together as multi-source data; the multi-source data are encoded into a time series format to obtain time series data;

[0009] S2: The spatial module, temporal module and enhancement module of the spatiotemporal enhanced neural network model are constructed sequentially to determine the PM 2.5 Perform inverse estimation;

[0010] The construction method of the space module is as follows: 2.5 The spatial distribution of ground observation sites, and the selection of several PM near the target site 2.5 The ground observation station calculates the spatial distance between the two, and the selected PM 2.5 PM from ground-based observation stations 2.5 The concentration is spatially combined and weighted to obtain the PM concentration of several adjacent stations on the target station. 2.5 The concentration effect is recorded as a spatial signal; the spatial signal is then taken together with the time series data as space-time series data;

[0011] The construction method of the time module is as follows: build a bidirectional long short-term memory model, and assign time characteristics to the spatiotemporal sequence data to obtain the output value h t ;

[0012] The construction method of the enhancement module is as follows: construct an attention mechanism structure to output the value h t Different variables are given different weights;

[0013] S3: The data of the prediction set is input into the spatiotemporal enhanced neural network model that has been tested for stability and accuracy, and the PM in the target time of the estimated area is 2.5 Perform spatial continuous estimation and obtain PM with a spatial resolution of 1 km 2.5 Spatiotemporal distribution products.

[0014] Preferably, the region to be estimated is China.

[0015] Preferably, in step S1, PM 2.5 The ground observation data are daily ground station data; the AOD product data adopts the MODIS 1km AOD product generated by the atmospheric multi-angle correction algorithm; the meteorological data include wind speed, humidity, temperature, air pressure and planetary boundary layer height; the land use / cover and economic and social data include digital elevation model, normalized difference vegetation index, land use data and population density.

[0016] Preferably, in step S1, ArcGIS 10.3 and R language are used to pre-process AOD product data, meteorological data, land use / cover and economic and social data into data consistent with PM2.5. 2.5 The ground observation data have the same spatial and temporal scale; then the bilinear interpolation method is used to ensure that all data are at the same spatial and temporal resolution; then according to the time sequence of the multi-source data, Python is used to encode them into a time series format to obtain time series data.

[0017] Preferably, in the process of constructing the space module in step S2, the space signal SS of each target site i is i The calculation method is as follows:

[0018]

[0019]

[0020] Among them, X j refers to the PM of the neighboring target site i 2.5 PM measured at ground observation station j 2.5 Concentration; d ji is the spatial distance between site i and site j; n is the PM of the selected neighboring target site i 2.5 The number of ground observation sites is preferably 3; (x i ,y i ) and (x j ,y j ) are the coordinates of site i and site j respectively.

[0021] Furthermore, in the process of constructing the spatial module in step S2, the spatial signal at each time node t is encoded together with the time series data into a vector x t , vector x t The form is [v1, v2, ..., v m ], m is the total number of all data in the spatial signal and time series data; all vectors x within a time step T t spliced ​​into a spatiotemporal sequence data in the form of [x1, x2, ...x t , ..., x T ], where T is the time step of input to the next module.

[0022] Furthermore, during the construction of the time module in step S2, the output value h t The calculation method is as follows:

[0023] f t =σ(W f ·[h t-1 , x t ]+b f ) (3)

[0024] i t =σ(W i ·[h t-1 , x t ]+b i ) (4)

[0025] C t =f t*C t-1 +i t *tanh(W C ·[h t-1 , x t ]+b C ) (5)

[0026] o t =σ(W o ·[h t-1 , x t ]+b o ) (6)

[0027] h t =o t *tanh(C t ) (7)

[0028] Among them, f t 、i t , o t , C t Represent the forget gate, input gate, output gate and unit state vector, W f , W i , W C , W o are weights, b f 、b i 、b C 、b o Indicates the corresponding gate and unit states respectively; h t-1 and h t is the output variable at different time steps t-1 and t, x t is the current input, [h t-1 , x t ] represents merging two vectors into one vector; σ and tanh are activation functions.

[0029] Furthermore, in the process of constructing the enhancement module in step S2, the time feature h t As input vector, ground PM 2.5 Ground observation data as target vector y t , by inputting the vector h t and the target vector y t The information and mutual relationship of the attention vector The Softmax function is then used to calculate the weight value α t ; Finally, based on the weight value α t , the adjustment vector is obtained by weighted summation The specific formula is as follows:

[0030]

[0031]

[0032]

[0033] Among them, W C and W s They are all model parameters, tanh is the activation function, and T is the time step of inputting the next module.

[0034] Preferably, in step S3, the stability test adopts sample-based cross-validation, spatial cross-validation and temporal cross-validation methods.

[0035] Preferably, in step S3, the accuracy evaluation uses the determination coefficient, root mean square error, mean absolute error, relative prediction error and prediction deviation to test the estimation results of the spatiotemporal enhanced neural network model.

[0036] Compared with the prior art, the present invention has the following beneficial effects:

[0037] The present invention develops a spatially generalizable deep learning model, the spatiotemporal enhanced neural network (STENN), which combines multi-temporal remote sensing products, meteorological indicators, and human activity indicators to more accurately predict PM2.5 in the study area. 2.5 By inputting spatial signals into a recurrent neural network capable of processing time series, this model provides a geographic data-driven approach to better process PM. 2.5 It is subject to spatial heterogeneity and temporal dependence, and thus performs significantly better than most previous PM 2.5 In addition, the model has strong temporal and spatial transferability and can be extrapolated to the absence of PM 2.5 Year and region of ground monitoring.

[0038] The present invention makes full use of PM 2.5 The temporal and spatial variation characteristics of remote sensing data can generate long-term, more detailed, and more reliable high-resolution PM 2.5 Products, PM 2.5 The research on pollution and its harm to human body is of great help, providing important reference for the formulation of policies and guidelines to improve China's air quality, and has the value of promotion and application. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 The comparison between the STENN model and the currently popular deep learning and machine learning models in the embodiment, where (a) to (d) are the PM predicted by the STENN model, LSTM model, RF model, and Cubist model, respectively. 2.5 PM relative to ground-based observations2.5 The verification results;

[0040] Figure 2 are validation results obtained using different cross-validation methods in the embodiments, wherein (a) is the time cross-validation result from 2015 to 2020, and (b) is the spatial cross-validation result of seven major geographical regions in China;

[0041] Figure 3 The annual average PM from 2014 to 2020 and the average PM for 7 years (2014-2020) were obtained in the embodiment. 2.5 Concentration spatiotemporal distribution product. DETAILED DESCRIPTION

[0042] The present invention is further described and illustrated below in conjunction with the accompanying drawings and specific embodiments. The technical features of each embodiment of the present invention can be combined accordingly without conflicting with each other.

[0043] PM 2.5 Affected by multiple factors such as natural and human factors, the current common PM 2.5 The estimation methods all ignore its complex time dependence and spatial correlation. Therefore, this paper builds a deep learning model with spatial generalization ability to estimate China's PM over the years. 2.5 concentration and spatial distribution, providing a PM2.5-based 2.5 The estimation method is as follows:

[0044] S1: Get the PM of the region to be estimated (such as China) 2.5 Ground observation data, AOD product data, meteorological data, land use / cover and economic and social data are used together as multi-source data. 2.5The ground observation data are daily ground station data; the AOD product data are MODIS 1km AOD products generated by the atmospheric multi-angle correction (MAIAC) algorithm; the meteorological data include wind speed (WS), relative humidity (RHU), temperature (TEMP), air pressure (P), and planetary boundary layer height (BLH); the land use / cover and economic and social data include digital elevation model (DEM), normalized difference vegetation index (NDVI), land use data (LUCC), and population density (POP).

[0045] The above multi-source data are then encoded into a time series format to obtain time series data, as follows:

[0046] Using ArcGIS 10.3 and R language, AOD product data, meteorological data, land use / cover and economic and social data were preprocessed into PM2.5-related 2.5 The ground observation data have the same spatiotemporal scale. The meteorological data, land use / cover and economic and social data of the 1km grid unit were then resampled using the bilinear interpolation method to calculate the average values ​​of AOD and BLH data to ensure that all data are at the same spatiotemporal resolution. Then, according to the temporal order of the multi-source data obtained, Python was used to encode them into a time series format: that is, a code was written in Python to encode the individual annual products together into a file every few years, and finally the time series data was obtained.

[0047] S2: Build a spatiotemporal enhanced neural network (STENN) model to treat PM in the estimated area 2.5 The model is composed of three parts: space module, time module and enhancement module. The spatiotemporal enhanced neural network (STENN) model is based on multi-phase multi-source remote sensing products and PM 2.5 The ground observation data is used as input, the spatial signal is obtained according to the spatial correlation, the spatiotemporal features are obtained through the LSTM model, and the spatiotemporal features are enhanced through the attention mechanism, and the next time step is used as the PM 2.5 The STENN model has strong spatial generalization ability and can generalize the PM at the site level.2.5 The forecasts are extended to 1km resolution at national scale.

[0048] The specific operation steps for each module are as follows:

[0049] S21 (Building Space Module): Based on PM 2.5 The spatial distribution of ground observation sites, and the selection of several PM near the target site 2.5 The ground observation station calculates the spatial distance between the two, and the selected PM 2.5 PM from ground-based observation stations 2.5 The concentration is spatially combined and weighted to obtain the PM concentration of several adjacent stations on the target station. 2.5 The influence of concentration is recorded as spatial signal (SS).

[0050] The spatial signal considers the influence of neighboring observation points j (j=1, 2..., n) around the target point i, and uses distance weighting to enhance or weaken the influence of neighboring observation points.

[0051] In this operation, the spatial signal SS of each target site i i The calculation method is as follows:

[0052]

[0053]

[0054] In the formula, X j refers to the PM of the neighboring target site i 2.5 PM measured at ground observation station j 2.5 Concentration; d ji is the spatial distance between site i and site j (i.e., Euclidean distance); n is the PM of the selected neighboring target site i 2.5 The number of ground observation sites, the performance of the model has no significant relationship with the size of n, and the model performance is best when n is 3; (x i ,y i ) and (x j ,y j ) are the coordinates of site i and site j respectively.

[0055] The spatial signal is then taken together with other variables (i.e., time series data) as spatiotemporal series data. , as follows:

[0056] The spatial signal at each time node t is encoded into a vector x together with other preprocessed environmental variables (i.e., time series data) t , vector x t The form is [v1, v2, ..., v m], m is the total number of all data in the spatial signal and time series data. Then all vectors x within a time step T are t spliced ​​into a spatiotemporal sequence data in the form of [x1, x2, ...x t , ..., x T ], where T is the time step of input to the next module.

[0057] S22 (Building time module): Build a Bidirectional long short-term memory (BiLSTM) model and assign time features to the spatiotemporal sequence data obtained through S21 to obtain the output value h t .

[0058] The Bidirectional Long Short-Term Memory (BiLSTM) model is an extended version of the Long Short-Term Memory (LSTM) model. LSTM makes up for the problem of gradient vanishing or gradient exploding in the RNN model when processing long time series data. Its definition is as follows:

[0059] The LSTM model uses a "gate" structure, which determines whether the state of the LSTM module is updated and to what extent. Each LSTM module contains a forget gate f t , which determines how much information needs to be retained from time step t-1; an input gate i t , which determines how much information is stored starting from the current time step t; a unit state C t , which is used to update the state of the current unit; an output gate o t , which determines the current cell state C t How much information to output.

[0060] The equation of the LSTM model at time step t is as follows:

[0061] f t =σ(W f ·[h t-1 , x t ]+b f ) (3)

[0062] i t =σ(W i ·[h t-1 , x t ]+b i ) (4)

[0063] C t =f t *C t-1 +i t *tanh(W C ·[ht-1 , x t ]+b C ) (5)

[0064] o t =σ(W o ·[h t-1 , x t ]+b o ) (6)

[0065] h t =o t *tanh(C t ) (7)

[0066] Among them, f t 、i t , o t , C t Represent the forget gate, input gate, output gate and unit state vector, W f , W i , W C , W o are weights, b f 、b i 、b C 、b o Indicates the corresponding gate and unit states respectively; h t-1 and h t is the output variable at different time steps t-1 and t, x t is the current input, [h t-1 , x t ] represents merging two vectors into one vector; σ and tanh are activation functions.

[0067] The Bidirectional LSTM (Bi-LSTM) is an extension of LSTM. It changes the defect of LSTM that the time series only flows forward and allows it to flow forward and backward. In the Bidirectional LSTM, the output vector at time t is expanded into two vectors, namely and The two vectors flowing in opposite time directions are combined to form the final output vector That is h t .

[0068] S23 (Building Enhancement Module): Build the attention mechanism structure, and use the output of the S22 time module as the input of this module, which is the output value h obtained by S22 tDifferent variables in the dataset are assigned different weights. That is, the attention mechanism assigns different weights to different environmental variables by learning the relationship between data, so as to focus on more important environmental variables.

[0069] In the attention mechanism, we first pass the input vector h t and the target vector y t The information and mutual relationship of the attention vector Then use the Softmax function to calculate the weight value α t Finally, based on the weight value α t , the adjustment vector is obtained by weighted summation The specific formula is as follows:

[0070]

[0071]

[0072]

[0073] Among them, W C and W s are the model parameters required for calculation, and tanh is the activation function.

[0074] S3: Perform stability test and accuracy evaluation on the model in S2 to ensure PM 2.5 The prediction data is then input into the spatiotemporal enhanced neural network model that has undergone stability testing and accuracy evaluation to estimate the PM in the target time in the estimated area. 2.5 Perform spatial continuous estimation and obtain PM with a spatial resolution of 1 km 2.5 Spatiotemporal distribution products.

[0075] The stability test uses sample-based cross-validation, spatial cross-validation and temporal cross-validation methods. The accuracy evaluation uses the determination coefficient, root mean square error, mean absolute error, relative prediction error and prediction bias to test the estimation results of the spatiotemporal enhanced neural network model.

[0076] The specific methods of the three cross-validation are as follows:

[0077] (1) Sample-based cross validation: This is the more common 10-fold cross validation. The data set is randomly divided into 10 parts, 9 of which are used to build and train the model, and the remaining 1 is used for model validation. The two parts are then rotated alternately until all 10 parts are used for model validation.

[0078] (2) Time cross-validation: One year’s data is used for model validation, while data from other years are used for model fitting and rotated until all years are used for validation.

[0079] (3) Spatial cross-validation: Based on the geographical location of the data, the dataset is divided into seven parts according to China’s seven major geographical divisions. One part is used for model validation, and the remaining six parts are used for model training, which are then rotated for validation in turn.

[0080] Five evaluation indicators: determination coefficient (R 2 ), root mean square error (RMSE), mean absolute error (MAE), relative prediction error (PRE,%), and predicted bias (Predicted bias) are calculated as follows:

[0081]

[0082]

[0083]

[0084]

[0085] Predicted Bias = f i -y i (15)

[0086] Where n is the total number of data samples, f i ,y i The PM observed on the ground is 2.5 and the PM estimated by the model 2.5 .

[0087] Example

[0088] China was selected as the research area and the above method of the present invention was used for analysis. The specific steps are as follows:

[0089] Step 1) Data acquisition: daily PM from 2013 to 2020 2.5The ground observation data comes from the China National Environmental Monitoring Center (CNEMC, http: / / www.cnemc.cn / ). Between 2013 and 2020, the number of monitoring stations increased from 835 to 1700; multi-temporal (2011-2019) MODIS 1-km AOD data is a data product generated by the MODIS satellite through the multi-angle atmospheric correction (MAIAC) algorithm, with a spatial resolution of 1km. After the data is geometrically corrected, the AOD data with a wavelength of 550nm is selected, and data covering the whole of China is obtained through reprojection and splicing; meteorological data includes five variables such as wind speed (WS), humidity (RHU), temperature (TEMP), air pressure (P), and planetary boundary layer height (BLH). The meteorological data is generated by the European Center for Medium-Term Weather Forecasts (ECMWF) Fifth Generation Climate and Weather Products (ERA5) Global Atmospheric Reanalysis Tool; land use / cover and economic and social data include digital elevation model (DEM), normalized difference vegetation index (NDVI), land use data (LUCC), and population density (POP). The data source is the Resource and Environmental Science and Data Center of the Chinese Academy of Sciences. The time coverage of meteorological, land use / cover and economic and social data is from 2011 to 2019.

[0090] Step 2) Data preprocessing: Data processing was performed using ArcGIS 10.3 and R language to preprocess multi-source remote sensing products such as AOD product data, meteorological data, land use / cover, and economic and social data to achieve the same temporal and spatial scale as PM2.5 ground observation data. The meteorological data, land use / cover, and economic and social data of 1 km grid cells were then resampled using the bilinear interpolation method, and the average values ​​of AOD and BLH data were calculated to ensure that all data were at the same temporal resolution (year by year) and spatial resolution (1 km).

[0091] Step 3) Model construction: The STENN model was built using the deep learning API Keras in Python. Before entering the model, the modeling set data was first standardized. The time module of the STENN model contains three BiLSTM layers with 512 hidden units, each with a batch size of 72, and an attention layer in the enhancement module. To prevent overfitting, a dropout layer was set in the enhancement module, the number of network iterations was set to 300, and the gradient descent-based optimizer Adam and the learning rate scheduler were used, where the learning rate was set to decrease by 1 / 10 every 50 iterations.

[0092] Step 4) Model validation: Three cross-validation methods were used to validate the performance of the STENN model. At the same time, the popular deep learning / machine learning models: LSTM, random forest (RF), and Cubist models were used for comparative studies. Figure 1 and 2 shown.

[0093] Figure 1 Prediction PM for STENN (a), LSTM (b), RF (c) and Cubist (d) 2.5 Relative to ground-based PM 2.5 The dotted line represents the best fit line of linear regression, and the solid line represents the 1:1 line. The units of RMSE and MAE in the figure are μg / m -3 .from Figure 1 As can be seen from the figure, the STENN model performs best in all indicators and has the highest R 2 and the lowest RMSE, MAE, and RPE, which were 0.89 and 5.84 μg m -3 , 3.97 μg m -3 , 12.74%. It can also be seen from the distribution of the scattered points that the fitted regression line of the STENN model is closest to the 1:1 line. The performance of the other three models is basically the same. Although they all have good accuracy, their ability to predict abnormally high or low values ​​is poor.

[0094] The cross-validation results of the STENN model in China from 2015 to 2020 are as follows: Figure 2 (a) The performance of the STENN model is not consistent every year. The model has the lowest uncertainty in 2020 and performs best. Overall, the predicted results are consistent with the surface observations in different years (R 2 =0.78-0.87) The spatial cross-validation results of the STENN model in seven geographical regions of China (Northwest, Northeast, North, Central, East, South and Southwest) are as follows Figure 2 (b) The model performs slightly differently in different regions. The model prediction performance is best in East China and Central China. 2 The model performance is the worst in Northwest China. There are two possible reasons for the difference in model performance. First, the density and distribution of monitoring stations determine the calculation of spatial signals and their effect on target PM. 2.5 response; secondly, different climatic conditions and land cover lead to differences in the performance of the model in different regions.

[0095] Step 5): Input the data of the prediction set into the established STENN model to obtain the 1km PM 2.5Spatially distributed products. Figure 3 The following table shows the average PM in China from 2014 to 2020. 2.5 Spatial distribution of concentration (ag), 7-year average PM 2.5 Spatial distribution of concentration (h). It can be seen that this product has a high spatial resolution and high prediction accuracy, based on which China’s PM 2.5 Analyze fine-scale spatiotemporal changes to provide a basis for formulating relevant pollution prevention and control policies.

[0096] The present invention makes full use of PM 2.5 The temporal and spatial changes of remote sensing data can generate long-term, more detailed, more reliable and higher-resolution PM 2.5 The product has important theoretical and practical significance for the prevention and control of air pollution in my country.

[0097] The above-described embodiment is only a preferred solution of the present invention, but it is not intended to limit the present invention. A person skilled in the relevant technical field may make various changes and modifications without departing from the spirit and scope of the present invention. Therefore, any technical solution obtained by equivalent replacement or equivalent transformation falls within the protection scope of the present invention.

Claims

1. A PM based on time series and deep learning framework 2.5 The estimation method is characterized in that The details are as follows: S1: Get the PM of the area to be estimated 2.5 Ground observation data, AOD product data, meteorological data, land use / cover and economic and social data are taken together as multi-source data; the multi-source data are encoded into a time series format to obtain time series data; S2: The spatial module, temporal module and enhancement module of the spatiotemporal enhanced neural network model are constructed sequentially to determine the PM 2.5 Perform inverse estimation; The construction method of the space module is as follows: 2.5 The spatial distribution of ground observation sites, and the selection of several PM near the target site 2.5 The ground observation station calculates the spatial distance between the two, and the selected PM 2.5 PM from ground-based observation stations 2.5 The concentration is spatially combined and weighted to obtain the PM concentration of several adjacent stations on the target station. 2.5 The concentration effect is recorded as a spatial signal; the spatial signal is then taken together with the time series data as space-time series data; The construction method of the time module is as follows: build a bidirectional long short-term memory model, and assign time characteristics to the spatiotemporal sequence data to obtain the output value h t ; The construction method of the enhancement module is as follows: construct an attention mechanism structure to output the value h t Different weights are assigned to different variables; S3: The data of the prediction set is input into the spatiotemporal enhanced neural network model that has been tested for stability and accuracy, and the PM in the target time of the estimated area is 2.5 Perform spatial continuous estimation and obtain PM with a spatial resolution of 1 km 2.5 Spatiotemporal distribution products; In the process of constructing the space module in step S2, the space signal SS of each target site i i The calculation method is as follows: Among them, X j refers to the PM of the neighboring target site i 2.5 PM measured at ground observation station j 2.5 Concentration; d ji is the spatial distance between site i and site j; n is the PM of the selected neighboring target site i 2.5 The number of ground observation sites, value is 3; (x i ,y i ) and (x j ,y j ) are the coordinates of site i and site j respectively; In the process of constructing the spatial module in step S2, the spatial signal at each time node t is encoded together with the time series data into a vector x t , vector x t The form is [v1,v2,…,v m ], m is the total number of all data in the spatial signal and time series data; all vectors x within a time step T t spliced ​​into a spatiotemporal sequence data in the form of [x1,x2,…x t ,…,x T ], where T is the time step of input to the next module.

2. The PM based on time series and deep learning framework as claimed in claim 1 2.5 The estimation method is characterized in that The area to be estimated is China.

3. The PM based on time series and deep learning framework as claimed in claim 1 2.5 The estimation method is characterized in that In step S1, PM 2.5 The ground observation data are daily ground station data; the AOD product data adopts the MODIS 1km AOD product generated by the atmospheric multi-angle correction algorithm; the meteorological data include wind speed, humidity, temperature, air pressure and planetary boundary layer height; the land use / cover and economic and social data include digital elevation model, normalized difference vegetation index, land use data and population density.

4. The PM based on time series and deep learning framework as claimed in claim 1 2.5 The estimation method is characterized in that In step S1, ArcGIS 10.3 and R language are used to pre-process AOD product data, meteorological data, land use / cover and economic and social data into PM2.5 2.5 The ground-based observation data have the same spatial and temporal scales; Then, the bilinear interpolation method is used to ensure that all data are at the same spatial and temporal resolution; then, according to the temporal sequence of the multi-source data, Python is used to encode them into a time series format to obtain time series data.

5. The PM based on time series and deep learning framework as claimed in claim 1 2.5 The estimation method is characterized in that During the construction of the time module in step S2, the output value h t The calculation method is as follows: f t =σ(W f ·[h t-1 ,x t ]+b f ) (3) i t =σ(W i ·[h t-1 ,x t ]+b i ) (4) C t =f t *C t-1 +i t *tanh(W C ·[h t-1 ,x t ]+b C ) (5) the t =σ(W o ·[h t-1 ,x t ]+b o ) (6) h t =o t *tanh(C t ) (7) Among them, f t 、i t , o t , C t Represent the forget gate, input gate, output gate and unit state vector, W f , W i , W C , W o are weights, b f 、b i 、b C 、b o Indicates the corresponding gate and unit states respectively; h t-1 and h t is the output variable at different time steps t-1 and t, x t is the current input, [h t-1 ,x t ] represents merging two vectors into one vector; σ and tanh are activation functions.

6. PM based on time series and deep learning framework as claimed in claim 5 2.5 The estimation method is characterized in that In the process of building the enhancement module in step S2, the time feature h t As input vector, ground PM 2.5 Ground observation data as target vector y t , by inputting the vector h t and the target vector y t The information and mutual relationship of the attention vector The Softmax function is then used to calculate the weight value α t ; Finally, based on the weight value α t , the adjustment vector is obtained by weighted summation The specific formula is as follows: Among them, W C and W s They are all model parameters, tanh is the activation function, and T is the time step of inputting the next module.

7. The PM based on time series and deep learning framework as claimed in claim 1 2.5 The estimation method is characterized in that In step S3, the stability test adopts sample-based cross-validation, spatial cross-validation and temporal cross-validation methods.

8. The PM based on time series and deep learning framework as claimed in claim 1 2.5 The estimation method is characterized in that In step S3, the accuracy evaluation uses the determination coefficient, root mean square error, mean absolute error, relative prediction error and prediction deviation to test the estimation results of the spatiotemporal enhanced neural network model.

Images