Enhanced spatiotemporal convolution gracedata time series interpolation method

Abstract

The application discloses a kind of enhanced spatiotemporal convolution GRACE data timing interpolation methods, wherein method includes: obtaining GRACE time series data and environmental time series data;It is decomposed, respectively to obtain the trend component and the detrended component corresponding to GRACE time series data and environmental time series data;The detrended component corresponding to GRACE time series data and environmental time series data is input to the ESTConvS2S model constructed in advance, and the detrended component prediction value corresponding to the missing value of the GRACE time series data output by the ESTConvS2S model is obtained;Based on the trend component and the detrended component prediction value of GRACE time series data, the reconstructed GRACE time series data is obtained, and the TWSA time series prediction value of land water storage anomaly is determined.The application improves the accuracy of TWSA prediction under complex scene.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of satellite data processing, and particularly relates to an enhanced space-time convolution GRACE data time series interpolation method. BACKGROUND

[0002] As a key component of global water resources, groundwater plays a vital role in maintaining the ecological system and sustainable development of human society. Accurate monitoring of groundwater is crucial to ensure water resource security. The Gravity Recovery and Climate Experiment (GRACE) satellite and its successor, the Gravity Recovery and Climate Experiment Follow-On (GRACE-FO) satellite, can estimate the Terrestrial Water Storage Anomaly (TWSA) by accurately measuring the changes in the Earth's gravity field, which is of great significance for groundwater monitoring. However, there are a large number of data gaps in the transition from GRACE to GRACE-FO, resulting in deviations and uncertainties in water storage prediction. Therefore, the prior art uses component replacement method, data-driven method or learning-based method to predict the data gap. However, the accuracy of these methods is poor in complex application scenarios. SUMMARY

[0003] The application provides an enhanced space-time convolution GRACE data time series interpolation method to solve the problem that the prior art uses component replacement method, data-driven method or learning-based method to predict the Terrestrial Water Storage Anomaly, and the accuracy of the prediction is poor in complex application scenarios.

[0004] In a first aspect, the application provides an enhanced space-time convolution GRACE data time series interpolation method, comprising: obtaining GRACE time series data and environmental time series data of the Gravity Recovery and Climate Experiment satellite;

[0005] Decomposing the GRACE time series data and the environmental time series data to obtain a trend component and a detrended component of the GRACE time series data, and a trend component and a detrended component of the environmental time series data;

[0006] Inputting the detrended component of the GRACE time series data and the detrended component of the environmental time series data into a pre-constructed enhanced space-time convolution sequence-to-sequence network ESTConvS2S model to obtain a predicted value of the detrended component corresponding to the missing value of the GRACE time series data output by the ESTConvS2S model;

[0007] reconstruct the GRACE time series data based on the trend component of the GRACE time series data and the detrended component prediction value, to determine a land water storage anomaly TWSA time series prediction value;

[0008] The ESTConvS2S model is trained based on a sample detrended component of sample GRACE time series data and a sample detrended component of sample environmental time series data, and a detrended component label corresponding to the sample GRACE time series data.

[0009] In some embodiments, the ESTConvS2S model includes a time module, a space module, a feature fusion module, and a prediction module.

[0010] Correspondingly, the inputting of the detrended component of the GRACE time series data and the detrended component of the environmental time series data into the pre-constructed enhanced space-time convolution sequence-to-sequence network ESTConvS2S model to obtain the detrended component prediction value corresponding to the missing value of the GRACE time series data output by the ESTConvS2S model includes:

[0011] The inputting of the detrended component of the GRACE time series data and the detrended component of the environmental time series data into the time module to obtain a time feature map output by the time module;

[0012] The inputting of the detrended component of the GRACE time series data and the detrended component of the environmental time series data into the space module to obtain a space feature map output by the space module;

[0013] The inputting of the time feature map and the space feature map into the feature fusion module to obtain a spatio-temporal feature map output by the feature fusion module;

[0014] The inputting of the spatio-temporal feature map into the prediction module to obtain the detrended component prediction value corresponding to the missing value of the GRACE time series data output by the prediction module.

[0015] In some embodiments, the time module includes a factorization convolution layer, an attention module, a fully connected layer, and a pooling layer; the space module includes a dilated convolution layer and a convolution attention module CBAM, the CBAM including a channel attention module and a spatial attention module; and the feature fusion module includes a concatenation layer, a merging convolution layer, and a regularization layer.

[0016] In some embodiments, the expression of the time module is:

[0017]

[0018] wherein, represents the time feature map after the input data of the time module is processed by the lth layer of time convolution, K t is the time convolution kernel, FactorizedConv3D represents the factorized convolution layer, represents the attention feature map, SE represents the attention module, the Sigmoid function represents the weight, FC is the full connection layer, and GlobalAveragePooling is the pooling layer. represents the final time feature map, BatchNorm represents the standardization, and ReLU and LeakyReLU are activation functions.

[0019] In some embodiments, the expression of the space module is:

[0020]

[0021] wherein, represents the space feature map after the input data of the space module is processed by the lth layer of space dilated convolution, represents the space feature map after the input data of the space module is processed by the (l-1)th layer of space dilated convolution, K s is the dilated convolution kernel, dilation_rate is the dilated rate, and DilatedConv3D represents the dilated convolution layer. represents the channel attention feature map, CA represents the channel attention module, represents the space attention feature map, SA represents the space attention module, represents the final space feature map, BatchNorm represents the standardization, and LeakyReLU is the activation function.

[0022] In some embodiments, the expression of the feature fusion module is as follows:

[0023]

[0024] M final = Conv3D(Dropout(M), K merge )

[0025] wherein, M represents the time feature map and the space feature map after splicing, Concat represents the splicing layer, and M finalrepresents the final spatio-temporal feature map, Conv3D represents the merging convolutional layer, Dropout(M) represents the initial spatio-temporal feature map after regularization, Dropout represents the regularization layer, K merge is a merging convolution kernel.

[0026] In some embodiments, based on the trend component of the GRACE time series data and the detrended component prediction value, the reconstructed GRACE time series data is obtained, and the TWSA time series prediction value is determined.

[0027] Based on the detrended component prediction value, the GRACE time series data is interpolated to obtain the detrended component of the interpolated GRACE time series data.

[0028] Based on the detrended component of the interpolated GRACE time series data and the trend component of the GRACE time series data, the GRACE time series data is reconstructed to obtain the reconstructed GRACE time series data; and based on the reconstructed GRACE time series data, the TWSA time series prediction value is determined.

[0029] In some embodiments, the determination process of the ESTConvS2S model comprises:

[0030] Decomposing the sample GRACE time series data and the sample environmental time series data to obtain the sample detrended component of the sample GRACE time series data and the sample detrended component of the sample environmental time series data;

[0031] Determining the detrended component label corresponding to the sample GRACE time series data;

[0032] Based on the sample detrended component of the sample GRACE time series data and the sample detrended component of the sample environmental time series data, and the detrended component label corresponding to the sample GRACE time series data, the initial ESTConvS2S model is trained;

[0033] Iteratively optimizing the parameters of the initial ESTConvS2S model to obtain the ESTConvS2S model.

[0034] In a second aspect, the present application also provides an enhanced space-time convolution GRACE data time series interpolation device, comprising: an acquisition unit configured to acquire GRACE time series data and environmental time series data; a decomposition unit configured to decompose the GRACE time series data and the environmental time series data to obtain a trend component and a detrended component of the GRACE time series data, and a trend component and a detrended component of the environmental time series data;

[0035] a first prediction unit configured to input the detrended component of the GRACE time series data and the detrended component of the environmental time series data into a pre-constructed enhanced space-time convolution sequence-to-sequence network ESTConvS2S model to obtain a predicted value of the detrended component corresponding to a missing value of the GRACE time series data output by the ESTConvS2S model; and a second prediction unit configured to obtain reconstructed GRACE time series data based on the trend component of the GRACE time series data and the predicted value of the detrended component, and determine a TWSA time series prediction value.

[0036] The ESTConvS2S model is trained based on sample detrended components of sample GRACE time series data and sample detrended components of sample environmental time series data, and detrended component labels corresponding to the sample GRACE time series data.

[0037] In a third aspect, the present application also provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the enhanced space-time convolution GRACE data time series interpolation method.

[0038] The enhanced space-time convolution GRACE data time series interpolation method can effectively interpolate missing data, improve the accuracy of TWSA prediction, and especially improve the accuracy of TWSA prediction in complex application scenarios. BRIEF DESCRIPTION OF DRAWINGS

[0039] In order to more clearly illustrate the technical solutions in the present application or the prior art, the accompanying drawings needed to be used in the embodiments or prior art description will be briefly introduced as follows. Obviously, the accompanying drawings in the following description are some embodiments of the present application, and all other embodiments obtained by those of ordinary skill in the art based on these accompanying drawings without creative effort should be within the protection scope of the present application.

[0040] Figure 1 A flowchart of the enhanced spatiotemporal convolution GRACE data time series interpolation method provided by the embodiment of the present application is shown in the figure.

[0041] Figure 2 A structure diagram of the ESTConvS2S model provided by the embodiment of the present application is shown in the figure.

[0042] Figure 3 A flowchart of the determination process of the ESTConvS2S model provided by the embodiment of the present application is shown in the figure.

[0043] Figure 4 An enhanced spatiotemporal convolution GRACE data time series interpolation device provided by the embodiment of the present application is shown in the figure.

[0044] Figure 5 A structure diagram of an electronic device provided by the embodiment of the present application is shown in the figure. DETAILED DESCRIPTION

[0045] In order to make the objects, technical solutions and advantages of the present application clearer, the technical solutions in the present application will be described clearly and completely in combination with the accompanying drawings in the present application. Obviously, the described embodiments are some embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort should be within the protection scope of the present application.

[0046] Recently, various methods have been developed to bridge the data gap between the GRACE and GRACE-FO satellite missions to enable continuous monitoring of global or regional TWSA. These methods are divided into three categories: component replacement methods, data-driven methods, and learning-based methods. However, these methods still have the following shortcomings in capturing complex hydrological interactions and maintaining accuracy in diverse geographical environments: 1) Existing methods have difficulty accurately simulating the complex dynamic interactions between groundwater, surface water, and the atmosphere, leading to deviations in predicted hydrological changes from actual conditions; 2) Existing methods may not adequately consider unique hydrological and geographical characteristics of specific regions, such as complex topography of groundwater systems or water resource allocation in arid regions, leading to poor prediction accuracy in special environmental conditions; 3) When using data-driven or learning-based methods, insufficient training data and inadequate model generalization capabilities can result in poor performance of the model in new environments or extreme hydrological events, leading to decreased accuracy.

[0047] Deep learning provides an effective method to address these challenges, introducing new methodologies to bridge data gaps and improve prediction accuracy. For example, convolutional neural networks (CNNs) have outperformed traditional hydrological models in TWSA data prediction. However, reconstructing TWSA data during the gap remains challenging, and current methods face challenges in fully capturing complex storage patterns, especially in regions with limited data and significant climate change. Additionally, many models focus on the relationship between single-variable temporal features or multivariate spatial relationships and TWSA, ignoring their comprehensive correlation in time and space.

[0048] To this end, the embodiment of the present application provides an enhanced spatio-temporal convolutional GRACE data time series interpolation method, which obtains GRACE time series data and environmental time series data, decomposes them to obtain corresponding trend components and detrended components, inputs the detrended components of the GRACE time series data and the detrended components of the environmental time series data into a pre-constructed ESTConvS2S model, combines the model output with the trend components of the GRACE time series data to obtain the predicted value of the land water storage anomaly TWSA time series. The present application can effectively interpolate missing data, improve the accuracy and accuracy of TWSA prediction, especially in complex application scenarios.

[0049] Figure 1 The flowchart of the enhanced spatio-temporal convolutional GRACE data time series interpolation method provided by the embodiment of the present application is shown in FIG. 1. Figure 1 As shown in FIG. 1, an enhanced spatio-temporal convolutional GRACE data time series interpolation method is provided, which includes the following steps: step 110, step 120, step 130, and step 140. The method flow steps are only one possible implementation of the present application.

[0050] Step 110, obtaining gravity recovery and climate experiment satellite (GRACE) time series data and environmental time series data.

[0051] Optionally, the GRACE time series data and the environmental time series data are collected from multiple data sources.

[0052] Table 1 provides an example of the GRACE time series data and the environmental time series data provided by the embodiment of the present application. As shown in Table 1, an example of the GRACE time series data and the environmental time series data is provided, wherein the environmental time series data includes relevant data such as soil moisture, evapotranspiration, temperature, and precipitation, N represents the total length of the time series data, N is a natural number greater than or equal to 1, Step is the length before and after the predicted time series data (default is 3), H represents the length of the input variable, W represents the width of the input variable, and C represents the number of channels of different types of data.

[0053] Table 1 provides an example of the time series data provided by the embodiment of the present application.

[0054] Input variable name Variable size GRACE Mascon data (N-Step+1, Step, H, W_1) Soil moisture (N-Step+1, Step, H, W C) Evapotranspiration (N-Step+1, Step, H, W C) Temperature (N-Step+1, Step, H, W, C) Precipitation (N-Step+1, Step, H, WC)

[0055] Step 120, decomposing the GRACE time series data and the environmental time series data to obtain trend components and detrended components of the GRACE time series data, and trend components and detrended components of the environmental time series data.

[0056] Optionally, singular spectrum analysis (SSA) is applied to decompose the GRACE time series data and the environmental time series data.

[0057] The trend component represents the long-term trend of the time series data over time, and the detrended component is the remaining part after removing the trend component from the time series data, mainly reflecting the seasonality, periodicity, and random fluctuations in the data.

[0058] Taking the GRACE Mascon data as an example, it is assumed that the TWSA presents a clear upward trend in 10 years. Through SSA analysis, the trend component extracted may be a smooth upward curve, indicating that the TWSA has an overall increasing trend in the past 10 years. At the same time, after removing the above trend component, the remaining part is the detrended component. This part of data may contain seasonal fluctuations (such as annual periodic changes) and random short-term fluctuations (such as some years of abnormal drought or flood events). The specific formula is:

[0059] TWSA(t) = TWSA detrend (t) + TWSA t rend (t)

[0060] Wherein, TWSA(t) is the original time series data, that is, the data form before the trend decomposition processing, TWSA detrend (t) is the corresponding detrended component, TWSA trend (t) is the corresponding trend component.

[0061] Step 130, input the detrended component of the GRACE time series data and the detrended component of the environmental time series data into the pre-constructed enhanced space-time convolution sequence to sequence network ESTConvS2S model, to obtain the detrended component prediction value corresponding to the missing value of the GRACE time series data output by the ESTConvS2S model.

[0062] Wherein, the ESTConvS2S model is trained based on the sample detrended component of the sample GRACE time series data and the sample detrended component of the sample environmental time series data, and the corresponding detrended component label of the sample GRACE time series data.

[0063] Wherein, the ESTConvS2S model combines high-dimensional space-time convolution layers, which can effectively capture and learn the interaction between various climate and geographical variables and their influence on TWSA.

[0064] Optionally, the main structure of the ESTConvS2S model is represented as:

[0065] ESTConvS2S(X)=f ESTConvS2S (X;0)#

[0066] Wherein, ESTConvS2S(X represents processing the detrended component of the GRACE time series data and the detrended component of the environmental time series data by using the ESTConvS2S model, f ESTConvS2S represents the space-time convolution operation within the model, X represents the detrended component of the GRACE time series data and the detrended component of the environmental time series data, and theta is the parameter of the ESTConvS2S model.

[0067] Figure 2 The structure diagram of the ESTConvS2S model provided by the embodiment of the present application is as follows: Figure 2As shown, in some embodiments, the ESTConvS2S model includes two key modules, a temporal module and a spatial module; specifically, the ESTConvS2S model includes: a factorized convolution layer, a dilated convolution layer, a batch normalization layer, a squeeze-and-excitation layer (SE), a convolutional block attention module (CBAM), a dropout layer, and a merging convolution layer.

[0068] It can be understood that, by combining the dilated convolution layer and the factorized convolution layer and the attention module (such as CBAM and SE layer), the ESTConvS2S model can more effectively capture and learn the interaction between various climate and geographical variables and their impact on TWSA, especially in areas with complex terrain and severe hydrological dynamics.

[0069] Step 140, based on the trend component and the detrended component prediction value of the GRACE time series data, the reconstructed GRACE time series data is obtained, and the TWSA time series prediction value of the land water storage anomaly is determined.

[0070] In the embodiments of the present application, by acquiring the time series data to be interpolated, the time series data includes GRACE satellite remote sensing data and environmental data, the time series data is decomposed to obtain the trend component and the detrended component of the time series data, the detrended component of the time series data is input into the ESTConvS2S model constructed in advance, and the output result of the model is combined with the trend component of the GRACE time series data to obtain the TWSA time series prediction value of the land water storage anomaly, which can effectively interpolate the missing data and improve the accuracy and accuracy of TWSA prediction, especially the accuracy of TWSA prediction in complex application scenarios.

[0071] In some embodiments, the ESTConvS2S model includes a temporal module, a spatial module, a feature fusion module, and a prediction module.

[0072] Correspondingly, step 130 inputs the detrended component of the GRACE time series data and the detrended component of the environmental time series data into the ESTConvS2S model constructed in advance, to obtain the detrended component prediction value corresponding to the missing value of the GRACE time series data output by the ESTConvS2S model, including steps 131, 132, 133 and 134.

[0073] Step 131, input the detrended component of the GRACE time series data and the detrended component of the environmental time series data into a time module to obtain a time feature map output by the time module.

[0074] The time module is configured to extract time features of the detrended component of the time series data. The time feature map contains key information of the detrended component of the time series data in the time dimension, which is crucial for understanding the time dependence of the data.

[0075] Step 132, input the detrended component of the GRACE time series data and the detrended component of the environmental time series data into a space module to obtain a space feature map output by the space module.

[0076] The space module is configured to extract spatial features of the detrended component of the time series data, such as the performance of the data in different geographical locations and their mutual relationship. The space feature map contains key information of the detrended component of the time series data in the space dimension, which can show the distribution and variation of the data in space.

[0077] Step 133, input the time feature map and the space feature map into a feature fusion module to obtain a spatio-temporal feature map output by the feature fusion module.

[0078] Step 134, input the spatio-temporal feature map into a prediction module to obtain a predicted value of the detrended component corresponding to the missing value of the GRACE time series data output by the prediction module.

[0079] It can be understood that by respectively adopting the time module and the space module to extract features of the detrended component of the GRACE time series data and the detrended component of the environmental time series data, the time feature map and the space feature map are obtained, the feature fusion module is adopted to fuse the time feature map and the space feature map to obtain the spatio-temporal feature map, which can more comprehensively capture the complex factors affecting the TWSA change, and the prediction module is adopted to predict the detrended component of the missing value of the GRACE time series data according to the spatio-temporal feature map, which can effectively interpolate the missing value, so as to combine with the trend component of the GRACE time series data, and on this basis, obtain the TWSA time series prediction value, thereby improving the prediction accuracy.

[0080] In some embodiments, the time module includes a factorization convolution layer, an attention module, a fully connected layer, and a pooling layer; the space module includes a dilated convolution layer and a convolution attention module CBAM, the CBAM includes a channel attention module and a spatial attention module; the feature fusion module includes a concatenation layer, a merging convolution layer, and a regularization layer.

[0081] In the embodiments of the present application, the design of the time module and the space module enables the ESTConvS2S model to more comprehensively extract key information in the input data by integrating different types of convolutional layers and attention mechanisms, and the design of the feature fusion module enables the ESTConvS2S model to effectively integrate feature information from different modules, and the combination of these modules not only optimizes the structure of the ESTConvS2S model, but also improves the ability of the ESTConvS2S model to process complex data.

[0082] In some embodiments, the expression of the time module is:

[0083]

[0084] wherein, represents the time feature map after the input data of the time module is processed by the lth layer of time convolution, K t is the time convolution kernel, FactorizedConv3D represents a factorized convolutional layer, represents the attention feature map, SE represents the attention module, the Sigmoid function represents the weight, FC is a fully connected layer, and GlobalAveragePooling is a pooling layer, represents the final time feature map, BatchNorm represents standardization, and ReLU and LeakyReLU are activation functions.

[0085] Optionally, the time module includes a series of 3D factorized convolutional layers with a kernel configuration of t x 1 x 1.

[0086] It should be noted that in the time module, the convolution operation focuses on the time dimension and ignores the change in the spatial dimension, each layer in the module uses Batch Normalization and LeakyReLU as a nonlinear activation function, the module uses causal convolution to maintain the causal relationship of the time series, which means that when calculating the output of the current time step, future information will not be used, and at the same time, the time module uses the SE module for feature recalibration.

[0087] It can be understood that the factorized convolutional layer in the time module can effectively reduce the number of parameters of the ESTConvS2S model, thereby reducing the computational complexity, and the attention module enables the ESTConvS2S model to pay more attention to valuable parts of the input data and improve the feature processing efficiency.

[0088] In some embodiments, the expression of the space module is:

[0089]

[0090] wherein, is the spatial feature map after the i-th layer of spatial dilated convolution processing of the input data of the spatial module, is the spatial feature map after the i-1-th layer of spatial dilated convolution processing of the input data of the spatial module, K s is the dilated convolution kernel, dilation rate is the dilated rate, and DilatedConv3D represents the dilated convolution layer, is the channel attention feature map, and CA represents the channel attention module, is the spatial attention feature map, and SA represents the spatial attention module, is the final spatial feature map, BatchNorm represents normalization, and LeakyReLU is the activation function.

[0091] Optionally, the spatial module is composed of a 3D dilated convolution layer, and the kernel size used is 1xdxd.

[0092] Optionally, each layer of the spatial module uses the channel attention and spatial attention mechanism of CBAM, as well as batch normalization and LeakyReLU activation function.

[0093] It should be noted that the design of the spatial module enables it to learn the features and dependencies in the space while keeping the time dimension unchanged, and at the same time, it performs spatial dilated convolution processing on the input data to expand the receptive field of the model, thereby improving the efficiency of processing high-dimensional data. The application of CBAM further enhances the ability of the ESTConvS2S model to identify key features.

[0094] It can be understood that by introducing the dilated convolution layer in the spatial module and combining it with the CBAM attention mechanism, the ESTConvS2S model can more effectively capture long-distance dependencies in spatial features and highlight important features through the attention mechanism, thereby improving the prediction performance of the ESTConvS2S model.

[0095] In some embodiments, the expression of the feature fusion module is as follows:

[0096]

[0097] M final = Conv3D(Dropout(M), K merge )

[0098] where M represents the initial spatiotemporal feature map obtained by splicing the temporal feature map and the spatial feature map , Concat represents the splicing layer, and M finaldenotes the final spatio-temporal feature map, Conv3D denotes a merge convolution layer, Dropout(M) denotes the initial spatio-temporal feature map after regularization, Dropout denotes a regularization layer, K merge is a merge convolution kernel.

[0099] It should be noted that, in the feature fusion module, the use of the splicing layer provides a flexible way to combine multiple feature maps, which is highly adaptive, and the merge convolution layer is responsible for effectively integrating the features, and the regularization layer helps to prevent overfitting, thereby enhancing the generalization ability of the ESTConvS2S model.

[0100] In some embodiments, based on the trend component and the detrended component prediction value of the GRACE time series data, the reconstructed GRACE time series data is obtained, and the TWSA time series prediction value of the land water storage anomaly is determined, including:

[0101] Based on the detrended component prediction value, the GRACE time series data is interpolated to obtain the interpolated detrended component of the GRACE time series data;

[0102] Based on the interpolated detrended component of the GRACE time series data and the trend component of the GRACE time series data, the GRACE time series data is reconstructed to obtain the reconstructed GRACE time series data;

[0103] Based on the reconstructed GRACE time series data, the TWSA time series prediction value of the land water storage anomaly is determined.

[0104] Optionally, the GWSA value is calculated based on the TWSA time series prediction value.

[0105] In the embodiments of the present application, by interpolating the detrended component of the GRACE time series data based on the detrended component prediction value of the missing value of the GRACE time series data, combining the trend component of the GRACE time series data, and reconstructing the GRACE time series data, the data gap between GRACE and GRACE-FO can be filled, continuous data support is provided for hydrological research, and finally the TWSA time series is accurately predicted, which provides support for effective monitoring and management of regional water resources.

[0106] Figure 3 The flowchart of the determination process of the ESTConvS2S model provided in the embodiments of the present application is shown in Figure 3 As shown in some embodiments, the determination process of the ESTConvS2S model includes:

[0107] Step 310, decomposing the sample GRACE time series data and the sample environment time series data to obtain a sample detrended component of the sample GRACE time series data and a sample detrended component of the sample environment time series data; Step 320, determining a detrended component label corresponding to the sample GRACE time series data;

[0108] Step 330, training an initial ESTConvS2S model based on the sample detrended component of the sample GRACE time series data and the sample detrended component of the sample environment time series data, and the detrended component label corresponding to the sample GRACE time series data;

[0109] Step 340, iteratively optimizing parameters of the initial ESTConvS2S model to obtain an ESTConvS2S model.

[0110] Optionally, the sample detrended component of the sample GRACE time series data and the sample detrended component of the sample environment time series data are input into a pre-constructed initial ESTConvS2S model to obtain a detrended component prediction value corresponding to a missing value of the GRACE time series data output by the initial ESTConvS2S model.

[0111] Optionally, the ESTConvS2S model comprises an initial time module, an initial space module, an initial feature fusion module, and an initial prediction module.

[0112] Correspondingly, the sample detrended component of the sample GRACE time series data and the sample detrended component of the sample environment time series data are input into a pre-constructed initial ESTConvS2S model to obtain a detrended component prediction value corresponding to a missing value of the GRACE time series data output by the initial ESTConvS2S model, comprising:

[0113] The sample detrended component of the sample GRACE time series data and the sample detrended component of the sample environment time series data are input into the initial time module to obtain a sample time feature map output by the initial time module;

[0114] The sample detrended component of the sample GRACE time series data and the sample detrended component of the sample environment time series data are input into the initial space module to obtain a sample space feature map output by the initial space module;

[0115] The sample time feature map and the sample space feature map are input into the initial feature fusion module to obtain a sample spatiotemporal feature map output by the initial feature fusion module;

[0116] The sample spatiotemporal feature map is input into the initial prediction module to obtain a predicted value of a detrended component corresponding to a missing value of the GRACE time series data output by the initial prediction module.

[0117] Optionally, the initial temporal module comprises an initial factor decomposition convolutional layer, an initial attention module, an initial fully connected layer and an initial pooling layer; the initial spatial module comprises an initial dilated convolutional layer and an initial convolutional attention module CBAM, the initial CBAM comprising an initial channel attention module and an initial spatial attention module; and the initial feature fusion module comprises an initial concatenation layer, an initial merging convolutional layer and an initial regularization layer.

[0118] In the embodiment of the present application, the sample GRACE time series data and the sample environmental time series data are decomposed to obtain corresponding sample detrended components, the detrended component labels corresponding to the sample GRACE time series data are determined, the sample detrended components of the sample GRACE time series data and the sample detrended components of the sample environmental time series data are used as training data, the detrended component labels corresponding to the sample GRACE time series data are used as training labels, the initial ESTConvS2S model is trained, the parameters of the initial ESTConvS2S model are iteratively optimized, the prediction accuracy, the accuracy and the generalization of the ESTConvS2S model are improved, and the ESTConvS2S model can be applied to complex application scenarios.

[0119] The enhanced spatiotemporal convolutional GRACE data time series interpolation device provided in the embodiment of the present application is described below, and the enhanced spatiotemporal convolutional GRACE data time series interpolation device described below can be correspondingly referred to the enhanced spatiotemporal convolutional GRACE data time series interpolation method described above.

[0120] Figure 4 The structural diagram of the enhanced spatiotemporal convolutional GRACE data time series interpolation device provided in the embodiment of the present application is shown in Figure 4 As shown in the figure, the enhanced spatiotemporal convolutional GRACE data time series interpolation device 400 comprises:

[0121] The acquisition unit 410 is configured to acquire the gravity recovery and climate experiment satellite GRACE time series data and environmental time series data.

[0122] The decomposition unit 420 is configured to decompose the GRACE time series data and the environmental time series data to obtain a trend component and a detrended component of the GRACE time series data and a trend component and a detrended component of the environmental time series data; the first prediction unit 430 is configured to input the detrended component of the GRACE time series data and the detrended component of the environmental time series data into a pre-constructed enhanced space-time convolution sequence-to-sequence network ESTConvS2S model to obtain a detrended component prediction value corresponding to a missing value of the GRACE time series data output by the ESTConvS2S model;

[0123] The second prediction unit 440 is configured to obtain reconstructed GRACE time series data based on the trend component of the GRACE time series data and the detrended component prediction value, and determine a TWSA time series prediction value.

[0124] The ESTConvS2S model is trained based on sample detrended components of sample GRACE time series data and sample detrended components of sample environmental time series data, and detrended component labels corresponding to the sample GRACE time series data.

[0125] Optionally, the ESTConvS2S model includes a time module, a space module, a feature fusion module, and a prediction module; correspondingly, the input of the detrended component of the GRACE time series data and the detrended component of the environmental time series data into the pre-constructed enhanced space-time convolution sequence-to-sequence network ESTConvS2S model to obtain a detrended component prediction value corresponding to a missing value of the GRACE time series data output by the ESTConvS2S model includes:

[0126] The input of the detrended component of the GRACE time series data and the detrended component of the environmental time series data into the time module to obtain a time feature map output by the time module;

[0127] The input of the detrended component of the GRACE time series data and the detrended component of the environmental time series data into the space module to obtain a space feature map output by the space module;

[0128] The input of the time feature map and the space feature map into the feature fusion module to obtain a spatiotemporal feature map output by the feature fusion module;

[0129] The input of the spatiotemporal feature map into the prediction module to obtain a detrended component prediction value corresponding to a missing value of the GRACE time series data output by the prediction module.

[0130] Optionally, the time module comprises a factorized convolutional layer, an attention module, a fully connected layer and a pooling layer; the space module comprises a dilated convolutional layer and a convolutional attention module CBAM, the CBAM comprises a channel attention module and a spatial attention module; and the feature fusion module comprises a concatenation layer, a merging convolutional layer and a regularization layer.

[0131] Optionally, the expression of the time module is as follows:

[0132]

[0133] wherein, represents a time feature map after the input data of the time module is processed by the lth layer of time convolution, K t is a time convolution kernel, FactorizedConv3D represents a factorized convolutional layer, represents an attention feature map, SE represents an attention module, Sigmoid function represents a weight, FC is a fully connected layer, and GlobanAveragePooling is a pooling layer, represents a final time feature map, BatchNorm represents standardization, and ReLU and LeakyReLU are activation functions.

[0134] Optionally, the expression of the space module is as follows:

[0135]

[0136] wherein, represents a spatial feature map after the input data of the space module is processed by the lth layer of spatial dilated convolution, represents a spatial feature map after the input data of the space module is processed by the (l-1)th layer of spatial dilated convolution, K s is a dilated convolution kernel, dilationrate is a dilated rate, and DilatedConv3D represents a dilated convolutional layer, represents a channel attention feature map, CA represents a channel attention module, represents a spatial attention feature map, SA represents a spatial attention module, represents a final spatial feature map, BatchNorm represents standardization, and LeakyReLU is an activation function.

[0137] Optionally, the expression of the feature fusion module is as follows:

[0138]

[0139] M final = Conv3D(Dropout(M), K merge )

[0140] wherein M represents a time feature map and a spatial feature map The initial spatio-temporal feature map obtained after splicing, Concat represents a splicing layer, M final represents the final spatio-temporal feature map, Conv3D represents a merging convolution layer, Dropout (M represents the initial spatio-temporal feature map after regularization, Dropout represents a regularization layer, K merge is a merging convolution kernel.

[0141] Optionally, based on the trend component and the detrended component prediction value of the GRACE time series data, the reconstructed GRACE time series data is obtained, and the TWSA time series prediction value of the land water storage anomaly is determined, comprising:

[0142] Based on the detrended component prediction value, the GRACE time series data is interpolated to obtain the detrended component of the interpolated GRACE time series data;

[0143] Based on the detrended component of the interpolated GRACE time series data and the trend component of the GRACE time series data, the GRACE time series data is reconstructed to obtain the reconstructed GRACE time series data;

[0144] Based on the reconstructed GRACE time series data, the TWSA time series prediction value of the land water storage anomaly is determined.

[0145] Optionally, the determination process of the ESTConvS2S model comprises:

[0146] The sample GRACE time series data and the sample environmental time series data are decomposed to obtain the sample detrended component of the sample GRACE time series data and the sample detrended component of the sample environmental time series data;

[0147] The detrended component label corresponding to the sample GRACE time series data is determined;

[0148] Based on the sample detrended component of the sample GRACE time series data and the sample detrended component of the sample environmental time series data, and the detrended component label corresponding to the sample GRACE time series data, the initial ESTConvS2S model is trained; the parameters of the initial ESTConvS2S model are iteratively optimized to obtain the ESTConvS2S model.

[0149] It should be noted that the enhanced space-time convolution GRACE data time series interpolation device provided by the embodiment of the present application can realize all the method steps realized by the enhanced space-time convolution GRACE data time series interpolation method embodiment, and can achieve the same technical effects. The same parts and beneficial effects in the method embodiment will not be described in detail here.

[0150] Figure 5 The structural schematic diagram of the electronic device provided by the embodiment of the present application is shown as follows, Figure 5 The electronic device can include a processor 510, a communications interface 520, a memory 530, and a communications bus 540, wherein the processor 510, the communications interface 520, and the memory 530 can communicate with each other through the communications bus 540. The processor 510 can call the logic instructions in the memory 530 to execute the enhanced space-time convolution GRACE data time series interpolation method, which includes: obtaining gravity recovery and climate experiment satellite GRACE time series data and environmental time series data; decomposing the GRACE time series data and the environmental time series data to obtain the trend component and the detrended component of the GRACE time series data, and the trend component and the detrended component of the environmental time series data; inputting the detrended component of the GRACE time series data and the detrended component of the environmental time series data into a pre-constructed enhanced space-time convolution sequence-to-sequence network ESTConvS2S model to obtain the detrended component prediction value corresponding to the missing value of the GRACE time series data output by the ESTConvS2S model; based on the trend component and the detrended component prediction value of the GRACE time series data, obtaining the reconstructed GRACE time series data, determining the TWSA time series prediction value, wherein the ESTConvS2S model is trained based on the sample detrended component of the sample GRACE time series data and the sample detrended component of the sample environmental time series data, and the detrended component label corresponding to the sample GRACE time series data.

[0151] In addition, the logic instructions in the memory 530 described above can be implemented in the form of software function units and sold or used as independent products, and can be stored in a computer readable storage medium. Based on such understanding, the technical solutions of the present application essentially or the parts that contribute to the prior art or parts of the technical solutions can be embodied in the form of a software product. The computer software product is stored in a storage medium, and includes several instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present application. The aforementioned storage medium includes various media that can store program codes, such as a U disk, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, etc.

[0152] The device embodiments described above are only schematic, wherein the units illustrated as separate components can or can not be physically separate, and the components illustrated as units can or can not be physical units, i.e., can be located in one place, or can be distributed on multiple network units. Part or all of the modules can be selected to achieve the purpose of the present embodiment according to actual needs. Those skilled in the art can understand and implement without creative labor.

[0153] From the above description of the embodiments, those skilled in the art can clearly understand that the embodiments can be implemented by means of software and the necessary general hardware platform, and of course can also be implemented by hardware. Based on such understanding, the above technical solutions essentially or the parts that contribute to the prior art can be embodied in the form of a software product. The computer software product can be stored in a computer readable storage medium, such as a ROM / RAM, a magnetic disk, an optical disk, etc., and includes several instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0154] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present application, and not to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that: it can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement to some technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A method for enhanced spatiotemporal convolutional GRACE data time series interpolation, characterized in that, The method comprises the following steps: obtaining gravity recovery and climate experiment satellite (GRACE) time series data and environmental time series data; decomposing the GRACE time series data and the environmental time series data to obtain a trend component and a detrended component of the GRACE time series data and a trend component and a detrended component of the environmental time series data; inputting the detrended component of the GRACE time series data and the detrended component of the environmental time series data into a pre-constructed enhanced space-time convolution sequence-to-sequence network (ESTConvS2S) model to obtain a detrended component prediction value corresponding to a missing value of the GRACE time series data output by the ESTConvS2S model; based on the trend component of the GRACE time series data and the detrended component prediction value, obtaining reconstructed GRACE time series data and determining a terrestrial water storage anomaly (TWSA) time series prediction value; wherein the ESTConvS2S model is trained based on sample detrended components of sample GRACE time series data and sample detrended components of sample environmental time series data and detrended component labels corresponding to the sample GRACE time series data; the ESTConvS2S model comprises a time module, a space module, a feature fusion module and a prediction module; correspondingly, the inputting the detrended component of the GRACE time series data and the detrended component of the environmental time series data into a pre-constructed enhanced space-time convolution sequence-to-sequence network (ESTConvS2S) model to obtain a detrended component prediction value corresponding to a missing value of the GRACE time series data output by the ESTConvS2S model comprises: inputting the detrended component of the GRACE time series data and the detrended component of the environmental time series data into the time module to obtain a time feature map output by the time module; inputting the detrended component of the GRACE time series data and the detrended component of the environmental time series data into the space module to obtain a space feature map output by the space module; inputting the time feature map and the space feature map into the feature fusion module to obtain a spatio-temporal feature map output by the feature fusion module; inputting the spatio-temporal feature map into the prediction module to obtain a detrended component prediction value corresponding to a missing value of the GRACE time series data output by the prediction module; the expression of the time module is: ； ； ； ； wherein, represents the time feature map after the time convolution processing of the input data of the time module, is a time convolution kernel, represents a factorization convolution layer, represents an attention feature map, represents an attention module, a function represents a weight, is a fully connected layer, is a pooling layer, represents a final time feature map, represents standardization, and is an activation function.​ the expression of the space module is: ； ； ； ； wherein, represents the spatial feature map after the input data of the spatial module is processed by the first layer spatial atrous convolution, represents the feature map after the input data of the spatial module is processed by the first layer spatial atrous convolution, is an atrous convolution kernel, is an atrous rate, represents an atrous convolution layer, represents a channel attention feature map, represents a channel attention module, represents a spatial attention feature map, represents a spatial attention module, represents a final spatial feature map, represents normalization, is an activation function.

2. The enhanced spatiotemporal convolution GRACE data temporal interpolation method according to claim 1, characterized in that, the time module comprises a factorized convolution layer, an attention module, a fully connected layer and a pooling layer; the space module comprises a dilated convolution layer and a convolution attention module (CBAM), and the CBAM comprises a channel attention module and a spatial attention module; the feature fusion module comprises a concatenation layer, a merging convolution layer and a regularization layer.

3. The enhanced spatio-temporal convolution-based time series data imputation method according to claim 2, characterized in that, the expression of the feature fusion module is as follows: ； ； wherein, denotes a temporal feature map denotes a spatial feature map an initial spatio-temporal feature map obtained after splicing, denotes the splicing layer, denotes a final spatio-temporal feature map, denotes the merging convolution layer, denotes a regularized initial spatio-temporal feature map, denotes the regularization layer, is a merging convolution kernel.

4. The enhanced spatiotemporal convolution GRACE data temporal interpolation method according to any one of claims 1-3, characterized in that, The trend component based on the GRACE time series data and the detrended component prediction value are used to obtain reconstructed GRACE time series data, and a land water storage anomaly TWSA time series prediction value is determined, including: Based on the detrended component prediction value, the GRACE time series data is interpolated to obtain the detrended component of the interpolated GRACE time series data; Based on the detrended component of the interpolated GRACE time series data and the trend component of the GRACE time series data, the GRACE time series data is reconstructed to obtain reconstructed GRACE time series data; Based on the reconstructed GRACE time series data, the land water storage anomaly TWSA time series prediction value is determined.

5. The enhanced spatiotemporal convolution GRACE data temporal interpolation method of claim 1, wherein, The determination process of the ESTConvS2S model includes: Decompose the sample GRACE time series data and sample environmental time series data to obtain sample detrended components of the sample GRACE time series data and sample detrended components of the sample environmental time series data; Determine the detrended component label corresponding to the sample GRACE time series data; Based on the sample detrended components of the GRACE time series data and the sample detrended components of the sample environmental time series data, and the detrended component label corresponding to the sample GRACE time series data, the initial ESTConvS2S model is trained; Iterative optimization is performed on the parameters of the initial ESTConvS2S model to obtain the ESTConvS2S model.

6. An apparatus for enhanced spatiotemporal convolutional GRACE data time series interpolation, comprising: The device is used to execute the enhanced space-time convolution GRACE data time series interpolation method according to any one of claims 1-5, including: An acquisition unit is configured to acquire gravity recovery and climate experiment satellite (GRACE) time series data and environmental time series data; A decomposition unit is configured to decompose the GRACE time series data and the environmental time series data to obtain trend components and detrended components of the GRACE time series data, and trend components and detrended components of the environmental time series data; A first prediction unit is configured to input the detrended components of the GRACE time series data and the detrended components of the environmental time series data into a pre-constructed enhanced space-time convolution sequence-to-sequence network (ESTConvS2S) model to obtain a detrended component prediction value corresponding to a missing value of the GRACE time series data output by the ESTConvS2S model; A second prediction unit is configured to obtain reconstructed GRACE time series data based on the trend component of the GRACE time series data and the detrended component prediction value, and determine a land water storage anomaly (TWSA) time series prediction value; The ESTConvS2S model is trained based on sample detrended components of sample GRACE time series data and sample detrended components of sample environmental time series data, and a detrended component label corresponding to the sample GRACE time series data.

7. An electronic device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, The processor implements the enhanced spatiotemporal convolution GRACE data time series interpolation method of any one of claims 1 to 5 when executing the program.

Images