Grace satellite air window period data reconstruction method, system and equipment combined with space-time characteristics

Abstract

The application discloses a kind of spatiotemporal feature combined GRACE satellite air window period data reconstruction method, system and equipment, first collect three data processing center TWSA dataset and carry out pre-processing;LSTM model processes TWSA long time series data and obtains TWSA reconstruction initial value;TWSA initial value is treated to remove tendency, and TWSA initial value is decomposed into TWSA random term and trend term;Then, TWSA random term is corrected using BCNN model to obtain TWSA correction term;TWSA trend term and correction term are combined to obtain the final value of TWSA reconstruction.The application aims at the 11-month air window period problem between GRACE two generations of satellites, constructs a spatiotemporal feature LSTM-BCNN reconstruction model, when reconstructing air window period land water storage anomaly (TWSA) data, it can extract land water storage historical data time sequence characteristics, and can also extract environmental factor spatial characteristics, complete the high-precision reconstruction of air window period TWSA data, provide data support for water storage time series analysis and drought monitoring.

Description

Technical Field

[0001] This invention belongs to the field of hydrological remote sensing data reconstruction technology, and relates to a method, system and equipment for reconstructing data during the GRACE satellite gap period. Specifically, it relates to a method, system and equipment for filling in the 11-month monitoring gap period of Terrestrial Water Storage Anomaly (TWSA) data between the two generations of GRACE satellites (GRACE and GRACE-FO). Background Technology

[0002] The GRACE satellite system comprises two phases: the GRACE satellite launched in April 2002 and the GRACE-FO satellite launched in May 2018. An 11-month data gap exists between the two generations of satellites, preventing them from providing continuous data on terrestrial water storage. Studies of temporal variations in terrestrial water storage rely on long-term, highly reliable datasets. Currently, there are no other alternative Earth gravity observation data available, so various reconstruction schemes have been developed to fill the data gap between the two GRACE satellites. Based on different reconstruction principles, data reconstruction schemes can be divided into two types: physical model reconstruction and mathematical model reconstruction. Physical model reconstruction methods utilize physical information about terrestrial water storage changes provided by the Swarm mission (Forootan et al., 2020) or satellite laser ranging (Meyer et al., 2019) to construct a model. Mathematical reconstruction models assess monthly changes in terrestrial water storage by analyzing temporal patterns or correlations using meteorological or hydrological data. Terrestrial water storage is sensitive to environmental changes. Data models can sensitively capture detailed information about changes in terrestrial water storage caused by El Niño or climate change by using characteristic constraints of environmental factors such as precipitation and temperature. Therefore, mathematical model reconstruction has become the mainstream reconstruction method. Currently, scholars have proposed a variety of mathematical methods to reconstruct terrestrial water storage data.

[0003] First, when the accuracy requirements for water storage data are not high, simple linear regression can predict missing terrestrial water storage values. Forootan et al. (2020) proposed an autoregressive (AR) model to predict total regional water storage in West Africa, addressing the scarcity of hydrological data, to fill gaps in GRACE satellite data. However, regression models are susceptible to outliers in time series data, leading to limited prediction accuracy. Data reconstruction during gap periods is essentially a form of data prediction, forecasting the changing trends of the gap period based on data before and after it. Therefore, time series models can be used to reconstruct gap period data. Time series analysis is widely used in many hydrological and climatological studies, but these applications largely rely on continuous, uninterrupted long-term data accumulation. GRACE satellites record approximately 20 years of terrestrial water storage data, a relatively short monitoring period, and data interruptions occur due to satellite power issues, limiting the development of time series methods for GRACE data reconstruction. Furthermore, while these models can effectively simulate the periodic changes in data, they cannot accurately fit the random variations in terrestrial water storage caused by natural factors (Ahmed et al., 2019). To address this challenge, many scholars have recently attempted to reconstruct window period data using machine learning algorithms. Zhao Dandan (2021) compared the reconstruction of water storage in the Yellow River Basin using a back propagation (BP) neural network and a multiple linear regression model, showing that the BP neural network had better fitting accuracy than the multiple linear regression model. Mo et al. (2022) constructed a Bayesian Convolutional Neural Network (BCNN) by combining Bayesian and convolutional neural network methods, and reconstructed window period data by training detrended terrestrial water storage and climate data. However, the changes in GRACE / GRACE-FO are influenced by multiple factors, including human activities and climate change. They not only exhibit long-term periodic trend signals but are also constrained by extreme weather events such as precipitation and temperature, as well as anthropogenic factors. Simple linear trends cannot capture these complex variation characteristics.

[0004] Currently, data reconstruction methods based on environmental factors can sensitively identify detailed information about the impact of climate change on terrestrial water storage, but they underestimate the temporal trend characteristics of terrestrial water storage. In contrast, data reconstruction methods based on time-series analysis can remember long-term annual and seasonal trends, but they ignore the stochastic influence of environmental factors. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention provides a window period reconstruction scheme that takes into account both long-term temporal trend characteristics and short-term spatial characteristics of environmental factors, thereby providing long-term, continuous water storage monitoring data for more accurate analysis of interannual variations and long-term trend characteristics of terrestrial water storage.

[0006] The technical solution adopted by the method of the present invention is: a method for reconstructing GRACE satellite gap data by combining spatiotemporal characteristics, comprising the following steps:

[0007] Step 1: Collect a number of TWSA datasets and preprocess the TWSA datasets to obtain a TWSA long-time-series interpolation dataset. ;

[0008] Step 2: Process the TWSA long-time-series interpolation dataset using an LSTM model The initial value of TWSA is obtained;

[0009] Step 3: After detrending the initial TWSA values, decompose them to obtain the TWSA trend term. and TWSA random items ;

[0010] Step 4: Use the BCNN model to analyze environmental factors and TWSA random terms. After processing, the TWSA correction term is obtained. ;

[0011] Step 5: Merge the TWSA trend items from Step 3 and the TWSA correction item in step 4 This yields the final value of the TWSA reconstruction.

[0012] Preferably, step 1 specifically includes the following sub-steps:

[0013] Step 1.1: Resample the Mascon products collected from several data processing centers to ensure consistent spatial resolution.

[0014] Step 1.2: Clip the dataset based on the study area and select data within the study period, then screen the data for spatial and temporal dimensions;

[0015] Step 1.3: Perform mean processing on the dataset to obtain the mean. This reduces uncertainty between data.

[0016] Step 1.4: Use interpolation methods to... By filling in product gaps and idle period data, a continuous TWSA long-time-series interpolated dataset is obtained. .

[0017] Preferably, in step 2, the LSTM model first provides the input tensor to the LSTM layer for encoding, and obtains the output tensor containing the encoding result through indexing; then, the output tensor is transformed from the original three-dimensional tensor shape into a two-dimensional tensor, and then a regression operation is performed on it through two linear transformations and a tanh activation function, and the output is used to predict the target value.

[0018] Preferably, the LSTM model is a pre-trained LSTM model; the training process includes the following sub-steps:

[0019] (1) Collect Mascon products from several data processing centers, and perform preprocessing such as cropping, averaging, and interpolation on the TWSA dataset to obtain TWSA long time series interpolated data. ;

[0020] (2) Transform the TWSA long-time-series interpolation dataset The dataset is split into training and testing sets. For the training set data, it is pruned at different starting positions to increase the number of training sequences for the model. Data padding is used to ensure that the length of each training sequence is uniform.

[0021] (3) When training the LSTM model, the input data includes Historical data, including year and month, totaling three dimensions; next month The data consists of predicted labels from an LSTM model, with one dimension.

[0022] (4) Training is stopped after the number of training iterations reaches the preset value, and the MSE loss function is used.

[0023] Preferably, in step 3, the BCNN model includes an input convolutional layer, 3 downsampling modules, 7 residual blocks, 4 upsampling modules, and an output convolutional layer.

[0024] The input convolutional 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 third upsampling module, the seventh residual block, the fourth upsampling module, and the output convolutional layer are connected sequentially.

[0025] The input convolutional layer and the output convolutional layer are respectively set in the input and output parts of the BCNN model; the input convolutional layer uses a 1×1 convolutional kernel to perform convolution operations on the input features, performs preliminary processing on the input image information, and converts the original information into a corresponding number of feature maps; the output convolutional layer uses a 1×1 convolutional kernel to extract key information of the image and restore the final image result.

[0026] Each residual block consists of two RCABs with the same number of input and output channels. Each RCAB module consists of two 3×3 convolutional layers, a residual connection, and a channel attention mechanism. The convolutional kernel size is 3×3, the stride is 1, the padding is 1, and the PReLu activation function is used.

[0027] The upsampling module includes a deconvolution layer, a normalization layer, and an activation function; the deconvolution layer upsamples the width and height of the tensor to twice the size, and the number of channels is halved; the kernel size is 3×3, the stride is 2, there is no padding, and the PReLu activation function is used.

[0028] The downsampling module includes a convolutional layer, a batch normalization, and a PReLU activation function; the convolutional kernel size is 3×3, the stride is 2, and the height and width of the output tensor are half of the input tensor; the paddings=1 parameter is used to ensure that the output tensor has the same size as the input tensor in the height and width directions.

[0029] The BCNN model performs channel concatenation between modules to fuse features from lower and higher levels. The input tensor x passes through an input convolutional layer to obtain x1; x1 passes through a first residual block to obtain x2; x2 passes through a first downsampling module to obtain x3; x3 passes through a second residual block to obtain x4; x4 passes through a second downsampling module to obtain x5; x5 passes through a third residual block to obtain x6; x6 passes through a third downsampling module to obtain x7; x7 passes through a fourth residual block to obtain x8; x7 and x8 are concatenated and then pass through a first upsampling module to obtain x9; x9 and... x6 is added and connected, then passes through the fifth residual block to obtain x10; x10 and x5 are connected in series and connected, then pass through the second upsampling module to obtain x11; x11 and x4 are added and connected, then pass through the sixth residual block to obtain x12; x12 and x3 are connected in series and connected, then pass through the third upsampling module to obtain x13; x13 and x2 are added and connected, then pass through the seventh residual block to obtain x14; x14 and x1 are connected in series and connected, then pass through the fourth upsampling module to obtain x15; x15 passes through the output convolutional layer to obtain the output tensor y.

[0030] Preferably, the BCNN model is a pre-trained BCNN model; the training process includes the following sub-steps:

[0031] (1) Combine the original environmental factors (precipitation, temperature, evaporation, runoff and soil moisture) with the detrended TWSA random terms. The time-series data is reorganized into the input format required by the BCNN model. The original three-dimensional environmental factor data for the past three months (t-1, t-2, and t-3) is sliced ​​and copied to the appropriate tensor index positions to serve as the four-dimensional tensor input to the model.

[0032] (2) t month TWSA random item The label is for the model, and the tensor is for the three-dimensional model.

[0033] (3) Training is stopped after the number of training iterations reaches the preset value, and the MSE loss function is used.

[0034] The technical solution adopted by the system of this invention is: a GRACE satellite gap period data reconstruction system combining spatiotemporal characteristics, comprising the following modules:

[0035] The first module is used to collect a number of TWSA data sets and preprocess the TWSA datasets to obtain a TWSA long-time-series interpolated dataset. ;

[0036] The second module is used to process the TWSA long-time-series interpolation dataset using an LSTM model. The initial value of TWSA is obtained;

[0037] The third module is used to decompose the initial TWSA values ​​after detrending to obtain the TWSA trend term. and TWSA random items ;

[0038] Module 4 is used to apply the BCNN model to environmental factors and TWSA random terms. After processing, the TWSA correction term is obtained. ;

[0039] Module 5 is used to merge the TWSA trend items in Module 3. And the TWSA correction item in Module 4 This yields the final value of the TWSA reconstruction.

[0040] The technical solution adopted by the device of this invention is: a GRACE satellite gap period data reconstruction device combining spatiotemporal characteristics, comprising:

[0041] One or more processors;

[0042] A storage device for storing one or more programs, which, when executed by one or more processors, enable the one or more processors to implement the GRACE satellite window data reconstruction method incorporating spatiotemporal features.

[0043] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0044] 1. To address the TWSA reconstruction problem during the GRACE satellite gap period, two issues need to be addressed. First, it's necessary to memorize, store, and transmit periodic information from the TWSA time-series data monitored by GRACE / GRACE-FO. Second, since TWSA is influenced by driving factors such as precipitation, temperature, and evaporation, the TWSA during the gap period is highly correlated with environmental factors. Therefore, it's also necessary to extract environmental features from these driving factors for auxiliary prediction. For the former, LSTM can be used to learn long-term dependent characteristics to extract the periodic information of TWSA. For the latter, a BCNN model, sensitive to spatial features, is selected to extract spatial features from the driving factors. Finally, the two methods are combined to output a more stable and accurate TWSA reconstruction during the GRACE satellite gap period.

[0045] 2. Terrestrial water storage plays a crucial role in regulating the global hydrological cycle and is a primary indicator of water resource availability, as well as an important indicator for drought disaster prediction. The GRACE satellite provides reliable terrestrial water storage data, improving the reliability and availability of TWSA data during off-peak periods. Attached Figure Description

[0046] Figure 1 Flowchart for reconstructing TWSA data during the GRACE / GRACE-FO window using the LSTM-BCNN model in this embodiment of the invention;

[0047] Figure 2 Flowcharts of embodiments of the present invention;

[0048] Figure 3 This is a diagram of the LSTM network structure of the LSTM-BCNN model used in an embodiment of the present invention.

[0049] Figure 4 This is a diagram of the BCNN network structure using the LSTM-BCNN model in an embodiment of the present invention;

[0050] Figure 5 The timing variations are for embodiments of the present invention;

[0051] Figure 6 This is a magnified time-series comparison diagram of an embodiment of the present invention. Detailed Implementation

[0052] To facilitate understanding and implementation of the present invention by those skilled in the art, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0053] This invention combines a Long Short-Term Memory (LSTM) neural network and a Bayesian Convolutional Neural Network (BCNN) to construct an LSTM-BCNN model to improve the accuracy of TWSA (Time-Wide Short-Term Memory) gap period data reconstruction. LSTM can remember the long-term dynamic temporal characteristics of TWSA and predict gap period data based on historical data trends; BCNN is good at extracting spatial features from the data and considers the short-term impact of driving factors on TWSA. In this experiment, initial interpolation processing was performed on Mascon data from three institutions to obtain rough estimates. Then, LSTM was used to obtain initial TWSA values ​​through LSTM-LSM training and detrended to obtain TWSA random and trend terms. The TWSA random term, along with environmental factors from three months ago, was used as data input to BCNN to complete TWSA correction; finally, the TWSA trend term and the corrected TWSA values ​​were merged to obtain the final TWSA reconstruction value.

[0054] Please see Figure 1 and Figure 2 The present invention provides a method for reconstructing GRACE satellite window data by combining spatiotemporal characteristics, comprising the following steps:

[0055] Step 1: Collect Mascon products from several data processing centers, and perform preprocessing on the TWSA dataset, including cropping, averaging, and interpolation, to obtain the TWSA long-time-series interpolated dataset. ;

[0056] In this embodiment, step 1 specifically includes the following sub-steps:

[0057] Step 1.1: Resample the Mascon products collected from several data processing centers to ensure consistent spatial resolution.

[0058] Step 1.2: Clip the dataset based on the study area and select data within the study period, then screen the data for spatial and temporal dimensions;

[0059] Step 1.3: Perform mean processing on the dataset to obtain the mean. This reduces uncertainty between data;

[0060] Step 1.4: Use interpolation methods to... By filling in product gaps and idle period data, a continuous TWSA long-time-series interpolated dataset is obtained. .

[0061] Step 2: Process the TWSA long-time-series interpolation dataset using an LSTM model The initial value of TWSA is obtained;

[0062] Please see Figure 3The LSTM model used in this embodiment is modeled using torch.nn.LSTM(), with an intermediate layer dimension of 8 and a layer count of 1. The model first provides the input tensor to the LSTM layer for encoding, and retrieves the output tensor containing the encoding results through indexing. Next, the output tensor is transformed from its original three-dimensional tensor shape into a two-dimensional tensor, and then it is subjected to regression operation through two linear transformations and a tanh activation function. The output is then used to predict the target value.

[0063] The LSTM model performs gradient descent on all network parameters except for the initial learning rate of 1e-2, and uses the MSE loss function during training.

[0064] The LSTM model used in this embodiment is a pre-trained LSTM model. This embodiment collects monthly TWSA data for GRACE / GRACE-FO from April 2002 to March 2022. Within the study area of ​​60°S-84°N and 180°W-180°E, each input data is resampled to a spatial resolution of 0.5°×0.5°, obtaining image data with height H=288 and width W=280. Monthly TWSA data from April 2002 to March 2016 is selected as the training set data. The training set data comprises 168 months, approximately 70% of the total dataset. The remaining 61 months of dataset after excluding the window period are used as the test set, including two phases: April 2016 to June 2017 and June 2018 to March 2022. The TWSA window period reconstruction process can be divided into two main phases: initial LSTM reconstruction and BCNN correction reconstruction.

[0065] The training process includes the following sub-steps:

[0066] (1) Collect Mascon products from three data processing centers and perform preprocessing such as cropping, averaging and interpolation on the three sets of TWSA datasets to obtain TWSA long time series interpolated data;

[0067] This embodiment selects GRACE / GRACE-FO Mascons data products from three institutions: JPL, CSR, and GSFC, and performs preprocessing. First, the CSR Mascons product, with its higher grid resolution, is upscaled and resampled to 0.5° × 0.5° to maintain the same spatial resolution as JPL and GSFC. Then, the three products are cropped to obtain study area data for the 60°S–84°N and 180°W–180°E regions. Third, based on the experimental requirements of this paper, data from the study period from April 2002 to March 2022 are selected, and other data are deleted. Then, interpolation methods are used to supplement the 22-month gap period data. This process yields interpolated data for an 11-month window period, but the accuracy is poor. Finally, the mean of the three data products is calculated to obtain the TWSA data.

[0068] (2) Transform the TWSA long-time-series interpolation dataset The dataset is split into training and testing sets. For the training set data, it is pruned at different starting positions to increase the number of training sequences for the model. Data padding is used to ensure that the length of each training sequence is uniform.

[0069] (3) When training the LSTM model, the input data includes Historical data, including year and month, totaling three dimensions; next month The data consists of predicted labels from an LSTM model, with one dimension.

[0070] (4) Training is terminated after 200 iterations, and the MSE loss function is used during training.

[0071] In this embodiment, the initial reconstruction process of TWSA by LSTM is as follows: Figure 3 As shown. Since the annual variation in terrestrial water storage is periodic, the year and month are also used as two dimensions for training labels, resulting in an input dimension of 3, consisting of the TWSA values ​​for the year, month, and current month. The TWSA values ​​for the following month are used as the predicted label data; therefore, the output data has a dimension of 1, denoted as the initial reconstruction value. The hidden layer is set to 1 layer with 8 hidden units, and the epoch is 200. After preprocessing, the TWSA data is fed into the LSTM network. The LSTM needs to maintain a consistent time step, which is set to 167 based on the training set length. This setting allows more historical data to participate in training. The batch processing parameter is set to batchsize = 48. During batch processing, the training data is pruned at different starting positions to obtain data at different starting pruning points. Data padding ensures a uniform training sequence length.

[0072] Step 3: Detrend the initial TWSA values ​​by decomposing the initial TWSA values ​​to obtain the TWSA trend term. and TWSA random items ;

[0073] This embodiment uses a linear detrending method to obtain the trend term. and random items . It has long-term cyclical trend signals. It is a complex signal affected by external environmental interference and needs to be corrected by combining environmental factors.

[0074] Step 4: Correct the TWSA random terms using the BCNN model TWSA correction term obtained ;

[0075] Please see the BCNN model in this embodiment. Figure 4 It includes an input convolutional layer, three downsampling modules, seven residual blocks, four upsampling modules, and an output convolutional layer;

[0076] In this embodiment, the input convolutional 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 third upsampling module, the seventh residual block, the fourth upsampling module, and the output convolutional layer are connected sequentially.

[0077] In this embodiment, the input convolutional layer and the output convolutional layer are respectively set in the input and output parts of the BCNN model; the input convolutional layer uses a 1×1 convolutional kernel to perform convolution operations on the input features, performs preliminary processing on the input image information, and converts the original information into a corresponding number of feature maps; the output convolutional layer uses a 1×1 convolutional kernel to extract key information of the image and restore the final image result.

[0078] In this embodiment, each residual block consists of two RCABs with the same number of input and output channels. Each RCAB module consists of two 3×3 convolutional layers, a residual connection, and a channel attention mechanism. The convolutional kernel size is 3×3, the stride is 1, the padding is 1, and the PReLu activation function is used.

[0079] The upsampling module in this embodiment includes a deconvolution layer, a normalization layer, and an activation function; the deconvolution layer upsamples the width and height of the tensor to twice the size, and the number of channels is halved; the kernel size is 3×3, the stride is 2, there is no padding, and the PReLu activation function is used;

[0080] The downsampling module in this embodiment includes a convolutional layer, a batch normalization, and a PReLU activation function; the convolutional kernel size is 3×3, the stride is 2, and the height and width of the output tensor are half of the input tensor; the paddings=1 parameter is used to ensure that the output tensor has the same size as the input tensor in the height and width directions;

[0081] In this embodiment of the BCNN model, channels are concatenated between modules to fuse features from lower and higher layers. The input tensor x passes through an input convolutional layer to obtain x1; x1 passes through a first residual block to obtain x2; x2 passes through a first downsampling module to obtain x3; x3 passes through a second residual block to obtain x4; x4 passes through a second downsampling module to obtain x5; x5 passes through a third residual block to obtain x6; x6 passes through a third downsampling module to obtain x7; x7 passes through a fourth residual block to obtain x8; x7 and x8 are cascaded and then passed through a first upsampling module to obtain x9; x9 and x6 are added and connected, then pass through the fifth residual block to obtain x10; x10 and x5 are connected in series and connected, then pass through the second upsampling module to obtain x11; x11 and x4 are added and connected, then pass through the sixth residual block to obtain x12; x12 and x3 are connected in series and connected, then pass through the third upsampling module to obtain x13; x13 and x2 are added and connected, then pass through the seventh residual block to obtain x14; x14 and x1 are connected in series and connected, then pass through the fourth upsampling module to obtain x15; x15 passes through the output convolutional layer to obtain the output tensor y.

[0082] The BCNN model in this embodiment defines a multi-layer convolutional network, including five basic modules, each outputting 48 feature layers (C=48). The feature map sizes H and W are halved sequentially in the first three convolutional modules to reduce the number of model parameters. The first four residual modules are downsampled using pooling, with sizes of 1 / 2, 1 / 4, 1 / 8, and 1 / 16 of H×W, respectively. Then, in the last three residual modules, they are upsampled sequentially to restore the feature maps to an H×W layer. Skip connections are added between the convolutional layers. These skip connections address the vanishing gradient problem and significantly improve network performance by interacting with feature information between different layers. Additive skip connections are used to... This means that feature maps of the same shape but different layers can be added together while keeping the number of channels unchanged. Serial jump connections are used. This means connecting features of the same shape but different layers together to increase the number of channels.

[0083] This embodiment corrects the TWSA random terms using a BCNN model. For each calendar month, select the first 3 months. The BCNN reconstruction model is constructed using driving factors as inputs. The BCNN model employs a Bayesian variational inference algorithm, using probability distributions in the weights of convolutional and fully connected layers. The input driving factors include precipitation, temperature, evaporation, runoff, and soil moisture data. The correction term is obtained through the BCNN model and is denoted as... .

[0084] The BCNN model in this embodiment is a pre-trained BCNN model; the training process includes the following sub-steps:

[0085] The TWSA random term correction stage in this embodiment is completed based on the BCNN model constructed by Mo et al. (2022). During the training of the BCNN network, the input data consists of TWSA and historical data from the past three months for five driving data points, i.e., detrended TWSA. The data includes five driving factors (precipitation, temperature, runoff, evaporation, and soil moisture) back two months, one month, and the current month. Therefore, the BCNN model input data has 18 feature layers. Thus, the input dimension of BCNN is 18×288×720. The training set batch size is 12. The initial learning rate is 0.0025, and the prediction accuracy is quantified using the mean squared error loss function. The example training is performed 200 times, using the MSE loss function during training.

[0086] Step 5: Merge the TWSA trend items from Step 3 and the TWSA correction item in step 4 This yields the final value of the TWSA reconstruction.

[0087] In this embodiment, three evaluation metrics—correlation coefficient, Nash efficiency coefficient, and standard root mean square error (NRMSE)—are used to assess the accuracy of reconstructed data at both the grid and watershed scales. The correlation coefficient (CC) measures the correlation between predicted data and actual observations. The Nash-Sutcliffe efficiency coefficient (NSE) is commonly used to verify the accuracy of hydrological model simulation results. The standard root mean square error (NRMSE) eliminates the influence of dataset and model scales by normalizing the root mean square error.

[0088]

[0089]

[0090]

[0091]

[0092] in, These are the i-th observation and the fitted value, respectively. and These are the observed average and the fitted average, respectively, and n is the total number of test datasets.

[0093] To minimize the influence of randomness in single results, the trained model was run 20 times for the reconstruction test period using TWSA, and the arithmetic mean of the 20 simulations was taken as the final model fit value. First, to more intuitively compare the model reconstruction accuracy, the difference between the fitted value and the observed value in the test period was compared using a visual interpretation method. Then, the difference between the fitted value and the observed value was calculated to quantify the model bias. The larger the absolute value of the error, the lower the model accuracy; the smaller the absolute value, the higher the model accuracy. Due to space limitations, TWSA data from four months out of the 61 test periods were selected for visualization and comparison: February 2017, May 2019, August 2020, and November 2021. Of the four sets of data selected, one set was before the window period, and three sets were after the window period. These four periods are evenly distributed throughout the test period and belong to different quarters, allowing for a comprehensive and objective evaluation of model accuracy.

[0094] Finally, a long-term variation analysis of GRACE / GRACE-FO was performed in this experiment.

[0095] The example plots the temporal variations of the fitted values ​​of three reconstruction models—LSTM, BCNN, and LSTM=BCNN—across 40 watersheds with observations from the GRACE task. Figure 5 As shown in the figure, the solid black line represents the TWSA time series data monitored by GRACE, and the gray shaded area represents the monitoring gaps of GRACE and GRACE-FO. The solid gray line within the gap is a rough estimate obtained by interpolation from GRACE. The TWSA data reconstructed by LSTM, BCNN, and LSTM-BCNN are plotted using gray dotted lines (.), gray dashed lines (-.), and black dashed lines (-), respectively. Differences in amplitude and magnitude between different time series may lead to poor consistency; therefore, these watershed average time series have been standardized to make them comparable.

[0096] Figure 6 This is a magnified time-series comparison of the Lake Chad and Amazon River basins. The graph shows that the LSTM model's fitted values ​​exhibit some overestimation or underestimation, with peaks and troughs. The BCNN model shows a high degree of fit during the testing phase, but fails to capture periodic features during the idle period. The LSTM-BCNN model effectively combines the characteristics of two individual models, improving accuracy during reconstruction.

[0097] It should be understood that the above description of the preferred embodiments is quite detailed, but it should not be considered as a limitation on the scope of protection of this invention. Those skilled in the art, under the guidance of this invention, can make substitutions or modifications without departing from the scope of protection of the claims of this invention, and all such substitutions or modifications fall within the scope of protection of this invention. The scope of protection of this invention should be determined by the appended claims.

Claims

1. A method for reconstructing GRACE satellite gap data by combining spatiotemporal characteristics, characterized in that, Includes the following steps: Step 1: Collect a number of TWSA datasets and preprocess the TWSA datasets to obtain a TWSA long-time-series interpolation dataset. ; Step 2: Process the TWSA long-time-series interpolation dataset using an LSTM model The initial value of TWSA is obtained; The LSTM model mentioned is a pre-trained LSTM model; the training process includes the following sub-steps: (1) Collect Mascon products from several data processing centers and preprocess the TWSA dataset to obtain TWSA long time series interpolated data. ; (2) Transform the TWSA long-time-series interpolation dataset The dataset is split into training and testing sets. For the training set data, it is pruned at different starting positions to increase the number of training sequences for the model. Data padding is used to ensure that the length of each training sequence is uniform. (3) When training the LSTM model, the input data includes Historical data, including year and month, totaling three dimensions; next month The data consists of predicted labels from an LSTM model, with one dimension. (4) Training ends after the number of training sessions reaches the threshold; Step 3: After detrending the initial TWSA values, decompose them to obtain the TWSA trend term. and TWSA random items ; Step 4: Use the BCNN model to analyze environmental factors and TWSA random terms. After processing, the TWSA correction term is obtained. ; The BCNN model mentioned is a pre-trained BCNN model; the training process includes the following sub-steps: (1) Combine the original environmental factors and the TWSA random terms obtained by detrending. The time series data is reorganized into the input format required by the BCNN model; slices of the original three-dimensional environmental factor data for the past three months (t-1, t-2, and t-3) are copied to the appropriate tensor index positions to serve as four-dimensional tensors for model input. (2) t month TWSA random item The model is labeled; the 3D tensor is used. (3) Training ends after the number of training sessions reaches the threshold; Step 5: Merge the TWSA trend items from Step 3 and the TWSA correction item in step 4 This yields the final value of the TWSA reconstruction.

2. The GRACE satellite gap period data reconstruction method combining spatiotemporal characteristics according to claim 1, characterized in that, Step 1 specifically includes the following sub-steps: Step 1.1: Resample the Mascon products collected from several data processing centers to ensure consistent spatial resolution. Step 1.2: Clip the dataset based on the study area and select data within the study period, then screen the data for spatial and temporal dimensions; Step 1.3: Perform mean processing on the dataset to obtain the mean. ; Step 1.4: Use interpolation methods to... By filling in product gaps and idle period data, a continuous TWSA long-time-series interpolated dataset is obtained. .

3. The GRACE satellite gap period data reconstruction method combining spatiotemporal characteristics according to claim 1, characterized in that: In step 2, the LSTM model first provides the input tensor to the LSTM layer for encoding, and obtains the output tensor containing the encoding result through indexing; then, the output tensor is transformed from the original three-dimensional tensor shape into a two-dimensional tensor, and then a regression operation is performed on it through two linear transformations and a tanh activation function, and the output is used to predict the target value.

4. The GRACE satellite gap period data reconstruction method combining spatiotemporal characteristics according to claim 1, characterized in that: In step 4, the BCNN model includes an input convolutional layer, 3 downsampling modules, 7 residual blocks, 4 upsampling modules, and an output convolutional layer. The input convolutional 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 third upsampling module, the seventh residual block, the fourth upsampling module, and the output convolutional layer are connected sequentially. The input convolutional layer and the output convolutional layer are respectively set in the input and output parts of the BCNN model; the input convolutional layer uses a 1×1 convolutional kernel to perform convolution operations on the input features, performs preliminary processing on the input image information, and converts the original information into a corresponding number of feature maps; the output convolutional layer uses a 1×1 convolutional kernel to extract key information of the image and restore the final image result. Each residual block consists of two RCABs with the same number of input and output channels. Each RCAB module consists of two 3×3 convolutional layers, a residual connection, and a channel attention mechanism. The convolutional kernel size is 3×3, the stride is 1, the padding is 1, and the PReLu activation function is used. The upsampling module includes a deconvolution layer, a normalization layer, and an activation function; the deconvolution layer upsamples the width and height of the tensor to twice the size, and the number of channels is halved; the kernel size is 3×3, the stride is 2, there is no padding, and the PReLu activation function is used. The downsampling module includes a convolutional layer, a batch normalization, and a PReLU activation function; the convolutional kernel size is 3×3, the stride is 2, and the height and width of the output tensor are half of the input tensor; the paddings=1 parameter is used to ensure that the output tensor has the same size as the input tensor in the height and width directions. The BCNN model performs channel concatenation between modules to fuse features from lower and higher levels. The input tensor x passes through an input convolutional layer to obtain x1; x1 passes through a first residual block to obtain x2; x2 passes through a first downsampling module to obtain x3; x3 passes through a second residual block to obtain x4; x4 passes through a second downsampling module to obtain x5; x5 passes through a third residual block to obtain x6; x6 passes through a third downsampling module to obtain x7; x7 passes through a fourth residual block to obtain x8; x7 and x8 are concatenated and then pass through a first upsampling module to obtain x9; x9 and... x6 is added and connected, then passes through the fifth residual block to obtain x10; x10 and x5 are connected in series and connected, then pass through the second upsampling module to obtain x11; x11 and x4 are added and connected, then pass through the sixth residual block to obtain x12; x12 and x3 are connected in series and connected, then pass through the third upsampling module to obtain x13; x13 and x2 are added and connected, then pass through the seventh residual block to obtain x14; x14 and x1 are connected in series and connected, then pass through the fourth upsampling module to obtain x15; x15 passes through the output convolutional layer to obtain the output tensor y.

5. A GRACE satellite gap period data reconstruction system incorporating spatiotemporal characteristics, used to implement the method described in any one of claims 1-4; characterized in that, Includes the following modules: The first module is used to collect a number of TWSA data sets and preprocess the TWSA datasets to obtain a TWSA long-time-series interpolated dataset. ; The second module is used to process the TWSA long-time-series interpolation dataset using an LSTM model. The initial value of TWSA is obtained; The third module is used to decompose the initial TWSA values ​​after detrending to obtain the TWSA trend term. and TWSA random items ; Module 4 is used to apply the BCNN model to environmental factors and TWSA random terms. After processing, the TWSA correction term is obtained. ; Module 5 is used to merge the TWSA trend items in Module 3. And the TWSA correction item in Module 4 This yields the final value of the TWSA reconstruction.

6. A data reconstruction device for the GRACE satellite gap period that combines spatiotemporal characteristics, characterized in that, include: One or more processors; A storage device for storing one or more programs, which, when executed by one or more processors, cause the one or more processors to implement the GRACE satellite window data reconstruction method incorporating spatiotemporal features as described in any one of claims 1 to 4.

Images