CN-116704362-B - GRACE satellite sky window period data reconstruction method, system and equipment combining space-time characteristics

Abstract

The invention discloses a GRACE satellite sky window period data reconstruction method, system and equipment combining space-time characteristics, which are characterized in that TWSA data sets of three data processing centers are firstly collected and preprocessed, TWSA reconstruction initial values are obtained by LSTM model processing TWSA long time sequence data, TWSA initial values are subjected to trend removal processing, TWSA initial values are decomposed into TWSA random items and trend items, then BCNN models are utilized for correcting TWSA random items to obtain TWSA correction items, and TWSA trend items and correction items are combined to obtain final values of TWSA reconstruction. Aiming at the problem of 11 months of empty window period between GRACE two-generation satellites, the invention constructs an LSTM-BCNN reconstruction model with space-time characteristics, and when reconstructing the land water reserves abnormal (TWSA) data in the empty window period, the invention can extract the historical data time sequence characteristics of the land water reserves and the space characteristics of environmental factors, complete the high-precision reconstruction of TWSA data in the empty window period and provide data support for the time sequence analysis and drought monitoring of the water reserves.

Inventors

• MENG LINGKUI
• LUO JIE
• WANG XI
• YANG BEIBEI
• ZHANG WEN
• LI LINYI
• TAO CHONGXIN
• LI JUNJIE
• ZHANG ZHEN
• WANG ZHE
• LIU YATING

Assignees

• 武汉大学

Dates

Publication Date

20260508

Application Date

20230512

Claims (6)

1. 1. The GRACE satellite sky window period data reconstruction method combining the space-time characteristics is characterized by comprising the following steps of: step 1, collecting a plurality of TWSA data, and preprocessing a TWSA data set to obtain a TWSA long-time-sequence interpolation data set ; Step 2, processing TWSA the long time sequence interpolation dataset by using the LSTM model Obtaining TWSA initial values; the LSTM model is a trained LSTM model, and the training process comprises the following substeps: (1) Mascon products of a plurality of data processing centers are collected, and a TWSA dataset is preprocessed to obtain TWSA long-time sequence interpolation data ; (2) Long time sequence interpolation dataset TWSA Splitting the training set data into a training set and a testing set, cutting the training set data at different initial positions to increase model training sequences, and ensuring the length of each training sequence to be uniform through filling data; (3) When the LSTM model is trained, the input data comprises Historical data, year and month, 3 dimensions in total, next month The data is a predictive label of an LSTM model, and the number of the predictive labels is 1; (4) Ending training after the training times reach a threshold value; Step 3, removing trend from TWSA initial values, decomposing to obtain TWSA trend items And TWSA random terms ; Step 4, utilizing BCNN model to make environment factor and TWSA random item Processing to obtain TWSA correction term ; The BCNN model is a trained BCNN model, and the training process comprises the following substeps: (1) TWSA random terms obtained by combining original environmental factors and detrending Copying slices of the three-dimensional data of the original environmental factors of approximately 3 months t-1, t-2 and t-3 to proper tensor index positions so as to be used as four-dimensional tensors input by the model; (2) t month TWSA random term The model label is a three-dimensional tensor; (3) Ending training after the training times reach a threshold value; step 5, merging TWSA trend items in step 3 And TWSA correction term in step 4 The final value of TWSA reconstructions is obtained.
2. 2. The method for reconstructing data of a GRACE satellite sky window incorporating spatio-temporal features according to claim 1, wherein step 1 specifically comprises the following sub-steps: step 1.1, resampling is carried out on Mascon products of a plurality of collected data processing centers, so that the spatial resolutions of the data are consistent; Step 1.2, cutting a data set according to a research area, selecting data in a research period, and screening the data in space dimension and time dimension; Step 1.3, carrying out mean processing on the data set to obtain a mean value ; Step 1.4, adopting interpolation method to implement Filling the product missing and empty window period data to obtain an uninterrupted TWSA long time sequence interpolation data set 。
3. 3. The method for reconstructing GRACE satellite sky window data by combining space-time features according to claim 1, wherein in step 2, the LSTM model firstly provides an input tensor to an LSTM layer for encoding, obtains an output tensor containing an encoding result through indexing, then changes the output tensor from an original three-dimensional tensor shape into a two-dimensional tensor, then carries out regression operation on the output tensor through two linear transformation and a tanh activation function, and uses the output for predicting a target value.
4. 4. The method for reconstructing GRACE satellite sky window data by combining space-time features according to claim 1, wherein in step 4, the BCNN model comprises an input convolution layer, 3 downsampling modules, 7 residual blocks, 4 upsampling modules and an output convolution layer; The input convolution layer, the first residual block, the first downsampling module, the second residual block, the second downsampling module, the third residual block, the third downsampling module, the fourth residual block, the first upsampling module, the fifth residual block, the second upsampling module, the sixth residual block, the third upsampling module, the seventh residual block, the fourth upsampling module and the output convolution layer are sequentially connected; the input convolution layer and the output convolution layer are respectively arranged at the input part and the output part of the BCNN model, the input convolution layer uses a1 multiplied by 1 convolution kernel to carry out convolution operation on input characteristics, carries out preliminary processing on input image information and converts the original information into a corresponding number of characteristic diagrams, and the output convolution layer uses the 1 multiplied by 1 convolution kernel to extract image key information and restore the image key information to a final image result; The residual blocks are composed of two RCAB, the number of input and output channels is the same, each RCAB module is composed of a double 3X 3 convolution layer, a residual connection and a channel attention mechanism, the convolution kernel is 3X 3, the step length is 1, the filling is 1, and PReLu is used for activating the function; the up-sampling module comprises a deconvolution layer, a normalization layer and an activation function, wherein the deconvolution layer up-samples the width and the height of tensors to 2 times, the number of channels is changed to 1/2, the convolution kernel size is 3 multiplied by 3, the step length is 2, no filling is carried out, and PReLu activation functions are used; the downsampling module comprises a convolution layer, a batch normalization and PReLU activation function, wherein the convolution kernel size is 3 multiplied by 3, the step length is 2, and the height and the width of an output tensor are half of those of an input tensor; The method comprises the steps of splicing channels among modules of BCNN models, fusing characteristics of a bottom layer and a high layer, wherein an input tensor x is subjected to input convolution layer to obtain x1, x1 is subjected to first residual block to obtain x2, x2 is subjected to first upsampling module to obtain x3, x3 is subjected to second residual block to obtain x4, x4 is subjected to second downsampling module to obtain x5, x5 is subjected to third residual block to obtain x6, x6 is subjected to third downsampling module to obtain x7, x7 is subjected to fourth residual block to obtain x8, serial skip connection of x7 and x8 is performed, then x9 is subjected to first upsampling module to obtain serial skip connection of x9, x9 and x6 is subjected to fifth residual block to serial skip connection of x10, then x10 and x5 is subjected to second upsampling module to obtain x11, x11 and x4 is subjected to addition skip connection, then sixth residual block to obtain x12, x12 and x3 is subjected to serial skip connection, then to third upsampling module to obtain x13, x13 and x2 is subjected to addition skip connection, then seventh upsampling module is subjected to x14 and x14 is subjected to serial skip connection of x15, and output residual layer is subjected to serial skip connection to fourth skip connection to obtain x 15.
5. 5. A GRACE Wei Xingkong window period data reconstruction system incorporating spatio-temporal features for implementing the method of any of claims 1-4, characterized by comprising the following modules: a first module for collecting a plurality of TWSA data and preprocessing TWSA data set to obtain TWSA long time sequence interpolation data set ; A second module for processing TWSA the long time sequence interpolation dataset by using LSTM model Obtaining TWSA initial values; a third module for removing trend from TWSA initial values and decomposing to obtain TWSA trend terms And TWSA random terms ; A fourth module 4 for using BCNN model to make the environmental factor and TWSA random terms Processing to obtain TWSA correction term ; A fifth module 5 for merging TWSA trend items in the third module And TWSA correction terms in the fourth module The final value of TWSA reconstructions is obtained.
6. 6. A GRACE satellite sky window data reconstruction device incorporating spatio-temporal features, comprising: One or more processors; Storage means for storing one or more programs which when executed by the one or more processors cause the one or more processors to implement the method of GRACE satellite sky window data reconstruction incorporating spatio-temporal features of any of claims 1 to 4.

Description

GRACE satellite sky window period data reconstruction method, system and equipment combining space-time characteristics Technical Field The invention belongs to the technical field of hydrological remote sensing data reconstruction, relates to a GRACE satellite sky window period data reconstruction method, system and equipment, and particularly relates to a method, system and equipment for filling up land water reserve abnormality (TERRESTRIAL WATER Storage Anomaly, TWSA) data in a GRACE satellite sky window period for 11 months between GRACE two-generation satellites (GRACE and GRACE-FO). Background The GRACE satellite includes two phases, a GRACE satellite launched in month 4 of 2002 and a GRACE-FO satellite launched in month 5 of 2018. There is an empty window of 11 months of data between the two generations of GRACE and GRACE-FO satellites, resulting in an inability to provide continuous land water reserves data. The time sequence change research of land water reserves depends on the support of a long-time sequence and high-reliability data set, and no alternative earth gravity observation data exists at present, so that various reconstruction schemes are developed to fill in blank data between two GRACE satellites. According to the different reconstruction principles, the data reconstruction scheme can be divided into two types of physical model reconstruction and mathematical model reconstruction. The physical model reconstruction method is to construct a model by using the physical information about the land water reserves provided by the swart task (Forootan et al., 2020) or satellite laser ranging (Meyer et al., 2019) and the like. The mathematical reconstruction model evaluates the change in monthly land water reserves by analyzing the timing law or correlation using meteorological data or hydrologic data. The land water reserves are sensitive to environmental changes, the data model can sensitively capture the land water reserves change detail information of the el Nino or the climate change through the characteristic constraint of environmental factors such as precipitation, temperature and the like, and therefore the mathematical model reconstruction becomes a mainstream reconstruction scheme. Various mathematical methods have been proposed by scholars to reconstruct land water reserves data. First, when the accuracy requirements for water reserves data are not high, a simple linear regression can predict the missing land water reserves value. Forootan et al (2020) proposes an Auto Regressive (AR) model for predicting total water reserves in an area to compensate for blank data between GRACE satellites for the current status of western non-hydrographic data starvation. Regression models are susceptible to outlier interference in the time series, resulting in limited prediction accuracy. The reconstruction of the data in the empty window period essentially belongs to data prediction, and the change trend of the empty window period is predicted according to the data before and after the empty window period. The time series model can be used to reconstruct the null window period data. Time series analysis is widely used in many hydrologic and climatological studies, but these applications rely largely on continuous, uninterrupted long time series data accumulation. The GRACE satellite records land water reserve data for about 20 years, the monitoring period is relatively short, and the data is interrupted due to satellite power and other self reasons, so that the development of a time sequence method in GRACE reconstruction data is limited. Furthermore, while these models can well model periodic variations in data, land water reserves vary randomly due to natural factors and do not fit exactly (Ahmed et al, 2019). To address this difficulty, many scholars have recently attempted to reconstruct the empty window data using machine learning algorithms. Zhao Dandan (2021) is used for comparing BP (Back propagation) neural network with a multiple linear regression model to reconstruct the water reserves of the yellow river basin, and the result shows that the fitting accuracy of the BP neural network is superior to that of the multiple linear regression model. Mo et al (2022) constructed bayesian convolutional neural networks (Bayesian Convolutional Neural Network, BCNN) by combining bayesian and convolutional neural network methods, and reconstructed empty window period data by training trended land water reserve data and climate data. However, the change of GRACE/GRACE-FO is affected by multiple factors of human activities and climate change, not only has long-term periodic trend signals, but also is constrained by extreme weather and human factors such as precipitation, temperature and the like, and complex change characteristics cannot be captured by simple linear trend. At present, the data reconstruction method based on the environmental factors can sensitively identify the detailed information o