Methods, apparatus, equipment and media for reconstructing changes in global land water storage

Abstract

This invention provides a method, apparatus, equipment, and medium for reconstructing global land water storage changes. The method includes: acquiring first training data and second training data. The first training data represents grid data of global land water storage changes over a past time span. The second training data includes global land evapotranspiration data, global land precipitation data, global runoff data, and land surface model TWS data corresponding to multiple time data points. A target model is trained based on the first training data, the second training data, and the Unet deep learning model. Complete grid data of global land water storage changes over the time span is determined based on at least one time data point, the second training data corresponding to that time data point, and the target model. This method solves the problems of inaccurate data, low fidelity, and poor scalability in existing technologies.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a method, apparatus, equipment and medium for reconstructing changes in global land water storage. Background Technology

[0002] Terrestrial water storage (TWS) is a key component of the Earth's water cycle, and its changes are important indicators for measuring regional water resources and climate change. The Gravity Recovery and Climate Experiment (GRACE) and its follow-on mission (GRACE-FO) satellites are currently the only technology capable of directly measuring Earth's mass migration, providing a unified and systematic observational benchmark for monitoring global terrestrial water storage change (TWSC).

[0003] However, the time-series data acquired by the GRACE / GRACE-FO mission is incomplete. Due to satellite battery management issues, technical malfunctions, and an 11-month gap between the two missions, a total of 33 months of data are missing from the 2002 to 2022 data series. These data gaps severely hinder the accurate analysis of long-term trends in terrestrial water storage, seasonal cycles, and extreme hydrological events such as droughts and floods, potentially introducing biases into related studies. To address this issue, existing technologies have proposed several data reconstruction methods. One type is the autoregressive model, such as the CEEMDAN-LSTM method based on Empirical Mode Decomposition and Long Short-Term Memory (LSTM) networks. These methods primarily rely on the time-series characteristics of the GRACE / GRACE-FO data itself, predicting or filling in missing values ​​by analyzing historical data. However, these methods have significant limitations: First, autoregressive methods perform poorly in areas with complex hydrological signals or high noise levels (such as the Yangtze River basin). First, it cannot explain the complex changes driven by external physical processes such as precipitation and evapotranspiration. It treats the signal merely as a mathematical sequence and cannot accurately reconstruct the complex TWSC signal. Second, autoregressive methods are typically modeled for a single time series (such as the average value of a single watershed). To apply them to every grid point globally, signal decomposition and model training would need to be performed separately for each grid, which would consume enormous computational resources and manpower, resulting in poor scalability.

[0004] It is evident that existing technologies suffer from problems such as inaccurate data, low fidelity, and poor scalability. Summary of the Invention

[0005] This invention provides a method, apparatus, equipment, and medium for reconstructing changes in global land water storage, which solves the problems of inaccurate data, low fidelity, and poor scalability in existing technologies.

[0006] This invention provides a method for reconstructing global land water storage changes, comprising: The first training data and the second training data were acquired. The first training data represents gridded data of global land water storage changes over a past time span, including multiple temporal data, multiple spatial data, and multiple land water storage change data. At least one temporal data point has missing spatial data and land water storage change data. The second training data includes global land evapotranspiration data, global land precipitation data, global runoff data, and land surface model (TWS) data corresponding to the multiple temporal data points.

[0007] The target model is trained based on the first training data, the second training data, and the Unet deep learning model.

[0008] Based on at least one time data point, second training data corresponding to the at least one time data point, and the target model, determine the complete grid data of global land water storage changes over the time span.

[0009] Optionally, the Unet deep learning model includes a channel attention module and a spatial attention module. Based on the first training data, the second training data, and the Unet deep learning model, the target model is trained, including: The channel attention module assigns corresponding first weights to the data types of the first training data and / or the second training data.

[0010] The spatial attention module assigns corresponding second weights to the spatial locations of the first and / or second training data.

[0011] The output data is obtained by convolution based on the first training data, the first weight, the second training data, and the second weight.

[0012] The output data is processed using the loss function, optimizer, and validation strategy from the Unet deep learning model to obtain the target model.

[0013] Optionally, the Unet deep learning model also includes a selective attention mask module. Before obtaining the output data after convolution based on the first training data, the first weights, the second training data, and the second weights, it also includes: The ocean region data in the first and second training data are set to zero using a selective attention masking module.

[0014] Optionally, after obtaining the output data through convolution based on the first training data, the first weights, the second training data, and the second weights, the method further includes: The ocean region data in the output data is set to zero using the selective attention masking module.

[0015] Optionally, the loss function is L1Loss (mean absolute error). The optimizer is the AdamW optimizer. The validation strategy uses K-fold cross-validation.

[0016] Optionally, before training the target model based on the first training data, the second training data, and the Unet deep learning model, the following steps are also included: The first training data and the second training data are preprocessed to ensure that the first training data and the second training data have the same spatial resolution, the same temporal resolution, and the same equivalent water level in the same units.

[0017] Optionally, the Unet deep learning model uses the GELU Gaussian error linear unit as the activation function.

[0018] The present invention also provides a global land water storage change reconstruction device, comprising: The data acquisition module is used to acquire first training data and second training data. The first training data represents gridded data of global land water storage changes over a past time span. This gridded data includes multiple temporal data, multiple spatial data, and multiple land water storage change data. At least one temporal data point has missing spatial data and land water storage change data. The second training data includes global land evapotranspiration data, global land precipitation data, global runoff data, and land surface model (TWS) data corresponding to the multiple temporal data points.

[0019] The target model training module is used to train the target model based on the first training data, the second training data, and the Unet deep learning model.

[0020] The grid data output module is used to determine complete global land water storage change grid data over a time span based on at least one time data, second training data corresponding to the at least one time data, and the target model.

[0021] The present invention also provides an electronic device, including a processor and a memory. Memory is used to store code and related data.

[0022] A processor for executing code in memory to implement the methods involved in the above embodiments and possible implementations.

[0023] The present invention also provides a storage medium having a computer program stored thereon, which, when executed by a processor, implements the methods involved in the above embodiments and possible implementations.

[0024] The advantages of this invention compared to the prior art are: 1. The traditional U-net processing flow involves inputting a matrix containing spatial information, typically an image (e.g., a matrix with RGB channels). This is processed by the network, including feature extraction via an encoder on the left, spatial resolution restoration via a decoder on the right, and preservation of spatial details via skip connections. The output is a matrix with the same size as the input image but only one channel. This invention innovatively applies the traditional U-net model to the field of hydrology, particularly to the specific geophysical problem of Total Land Water Storage Change (TWSC). Specifically, this invention can not only accurately fill in at least one data gap corresponding to a given time period, but also utilize a second training data and the trained target model to obtain extended data for longer time series, such as extending the TWSC record forward to before the launch of the GRACE satellite, thus providing a complete dataset with a longer time span. Furthermore, this invention is not merely a data imputation tool; through deep learning, it successfully constructs a precise and robust nonlinear mapping relationship between easily accessible water cycle components (such as precipitation and evapotranspiration) and the difficult-to-measure land water storage. This is equivalent to creating a "virtual sensor" based on multi-source data, providing a new observation paradigm for future acquisition of changes in global land water storage, and may reduce reliance on continuous satellite gravity missions to some extent.

[0025] 2. In the assessment of major global watersheds, the target model of this invention has shown excellent performance in key indicators such as root mean square error (RMSE), Pearson correlation coefficient (CC), and Nash-Sutcliffe efficiency (NSE) of the reconstruction results. Its reconstruction error is much smaller than the uncertainty of the GRACE / GRACE-FO data itself, with high fidelity and data accuracy, achieving near-lossless data reconstruction.

[0026] 3. Since different data products have different spatial coverage and data gaps, such as MODIS data gaps in desert and polar regions, and GLEAM data gaps in high-latitude regions, this invention integrates multiple data sources as secondary training data, so that multiple data sources can form complementary resources, ensuring that the target model can obtain effective input information at any grid point on the global landmass, thereby enhancing the global usability and robustness of the model.

[0027] 4. The embodiments of the present invention integrate a selective attention mask module, which can eliminate noise and artifacts generated by marine area data based on the land mask generated by the land-sea boundary, and ensure the cleanliness of the target model. Attached Figure Description

[0028] Figure 1 This is a schematic diagram of the process structure of the global land water storage change reconstruction method according to an embodiment of the present invention. Figure 1 ; Figure 2 This is a schematic diagram of the process structure of the global land water storage change reconstruction method according to an embodiment of the present invention. Figure 2 ; Figure 3 This is a schematic diagram of the process structure of step S2 in the global land water storage change reconstruction method of this invention. Figure 1 ; Figure 4 This is a schematic diagram of the process structure of step S2 in the global land water storage change reconstruction method of this invention. Figure 2 ; Figure 5 This is a time series comparison chart of the reconstructed data on changes in land water storage in nine major river basins worldwide using embodiments of the present invention, and the actual data. Figure 6 This is a schematic diagram of the global land water storage change reconstruction device according to an embodiment of the present invention; Figure 7 This is a schematic diagram of the structure of an electronic device according to an embodiment of the present invention. Detailed Implementation

[0029] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0030] The terms “first,” “second,” “third,” “fourth,” etc. (if present) in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms “comprising” and “having,” and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0031] The technical solution of the present invention will be described in detail below with reference to specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0032] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments of the present invention will be described in further detail below with reference to the accompanying drawings.

[0033] To more clearly illustrate the technical effects of this solution, the technical terms that may be involved in this invention are explained here.

[0034] Land Water Storage Variation Grid Data is a spatiotemporal continuous dataset that organizes Land Water Storage Variation (TWSA) into an isotropic latitude and longitude network. Each grid point provides an equivalent water column height variation value (usually in cm or mm), which reflects the surplus or deficit of the total mass of all forms of land water (soil water, groundwater, snow, surface water, and canopy water) within that 0.25° / 0.5° grid relative to the multi-year average.

[0035] In deep learning, convolution is essentially a weighted moving average. A small matrix (the kernel) is used to translate point by point on the input feature map. Each time it stops, the kernel is multiplied element by element with the corresponding region and then summed to get a value. After scanning all of them, a new feature map is formed.

[0036] Robustness refers to the ability of a system to maintain stable operation and expected performance when faced with abnormal inputs, faults, or uncertainties.

[0037] Network depth: The number of "layers" in a neural network. Usually, only layers with learnable parameters (convolutional layers, fully linear layers, LSTM, etc.) are counted, while parameterless operations such as ReLU, BN, and Dropout are called "layers" but are not included in the depth.

[0038] U-net is a fully convolutional network (FCN) architecture designed for biomedical image segmentation. It was proposed by Ronneberger et al. in 2015. The name comes from its symmetrical "U"-shaped structure: the left side is the shrinking path (encoder), the right side is the expanding path (decoder), and the middle is connected by skip connections to concatenate feature maps of the same level, which preserves both high-level semantics and low-level spatial details.

[0039] FCN (Fully Convolutional Network) is the first deep learning architecture proposed by the Berkeley team in 2015 to perform pixel-level predictions "end-to-end". It replaces all the fully connected layers at the end of the traditional CNN with convolutional layers, allowing the network to receive inputs of arbitrary size and output a "heatmap" corresponding to the input space, thereby completing dense classification in one go (giving each pixel a label).

[0040] Selective attention mask: A type of learnable, dynamically generated or hand-designed mask matrix that masks out "irrelevant" or "unimportant" positions before / during attention calculation, allowing the model to focus only on the regions most relevant to the current task.

[0041] This invention provides a method, apparatus, equipment, and medium for reconstructing global land water storage changes, involving the reconstruction of geophysical data using deep learning and convolutional neural networks, or, more broadly, a method for filling gaps in watershed-scale satellite gravity data based on deep learning. Please refer to [link to details]. Figures 1-4 Methods for reconstructing changes in global land water storage include: Step S1: Obtain the first training data and the second training data. The first training data represents the grid data of global land water storage changes over a past time span. The global land water storage change grid data includes multiple time data, multiple spatial data, and multiple land water storage change data. Among these, the spatial data and land water storage change data corresponding to at least one of the multiple time data are missing. The second training data includes global land evapotranspiration data, global land precipitation data, global runoff data, and land surface model TWS data corresponding to the multiple time data.

[0042] Step S2: Train the target model based on the first training data, the second training data, and the Unet deep learning model.

[0043] Step S3: Based on at least one time data, the second training data corresponding to the at least one time data, and the target model, determine the complete global land water storage change grid data within the time span.

[0044] The first and second training data can be understood as driving data, such as high-quality, labeled training data, real-time collected input data, or online log data, etc. In one example implementation, the second training data includes at least one evapotranspiration data product, at least one precipitation data product, at least one runoff data product, and at least one surface water storage data product calculated based on a land surface assimilation model.

[0045] In one implementation, the first training data uses the RL06 Mascon product calculated by NASA's Goddard Space Flight Center (GSFC) as the reconstruction target. This product was chosen because, in uncertainty analyses of multiple mainstream GRACE products, the GSFC Mascon product exhibits the lowest uncertainty and the best signal-to-noise ratio in most terrestrial watersheds, providing high-quality ground truth labels for model training. The data spans from April 2002 to December 2022, with an original spatial resolution of [missing information]. The second training data set includes ten sets of driving data as input features for the model, specifically: four sets of evapotranspiration (ET) data, from the Global Land Evapotranspiration Amsterdam Model (GLEAM), the ECMWF ReanAlysis v5 (ERA5) dataset, the Global Land Data Assimilation System (GLDAS), and the Moderate Resolution Imaging Spectroradiometer (MODIS) product, respectively; two sets of precipitation (P) data, from the Integrated Multi-satellite Retrievals for GPM (IMERG) and the Global Precipitation Climatology Project (GPCP) product, respectively. The Tropical Rainfall Measuring Mission (TRMM) is the predecessor to the GPM mission. Two sets of runoff data, from ERA5 and GLDAS products, respectively. Two sets of TWS (Transient Surface Data) land surface model data, based on the GLDAS Noah model, were calculated using two different methods (the Classic method and the Water Balance Equation (WBE) method). The advantage of using multiple sources of the same type of data (e.g., four sets of evapotranspiration data) is that different data products have different spatial coverage and data gaps; for example, MODIS data has gaps in desert and polar regions, while GLEAM data has gaps in high-latitude regions. By fusing multiple data sources, complementarity can be achieved, ensuring that the model obtains effective input information at any grid point on the global landmass, thereby enhancing the model's global applicability and robustness.

[0046] For further implementation methods, please refer to Figure 2 Before step S2, step S4 is included: preprocessing the first training data and the second training data so that the first training data and the second training data have the same spatial resolution, the same temporal resolution and the same equivalent water level in the same unit.

[0047] Step S4 can be understood as follows: In order to ensure data consistency and compatibility, a strict preprocessing process is performed on all acquired data to obtain training data with uniform temporal and spatial resolution.

[0048] Specifically, this includes spatiotemporal unification, which involves processing all driving data (e.g., 10 sets) and target GSFC data into a unified form through resampling or aggregation. The data uses a global spatial grid (uniform spatial resolution) and a uniform monthly time step (uniform temporal resolution). The processed data covers the period from June 2000 to December 2022, a total of 271 months. Unit conversion is also included: all data is converted to a uniform unit, namely equivalent water height (mm). This process involves complex calculations for different products, such as splitting MODIS's 8-day synthetic product proportionally by day and accumulating it into a monthly total, or multiplying ERA5's daily average rate by the number of days in the month to convert it into a monthly cumulative amount.

[0049] Those skilled in the art will understand that steps S4 and S2 can be selectively performed for different types of data products. For example, climate factor products do not require watershed trimming, while some hydrological model products or GRACE Level-3 products have already completed anomaly calculations when they are released, and only the same time period needs to be calculated in step S4.

[0050] Those skilled in the art will understand that the traditional U-net processing flow involves inputting a matrix containing spatial information, typically an image (e.g., a matrix containing RGB channels), which is then processed by the network (including feature extraction via an "encoder" on the left, spatial resolution restoration via a "decoder" on the right, and preservation of spatial details via "skip connections"), outputting a matrix with the same size as the input image but only one channel. This invention is entirely different. Through the aforementioned spatiotemporal unification processing, the two-dimensional geospatial grid is mathematically equivalent to a "pixel matrix" similar to an image, and various hydrological variables (such as precipitation and evapotranspiration) are equivalent to multi-channel features similar to RGB. These data are then used as input data for the U-Net network.

[0051] The target model in step S2 is a deeply modified U-shaped network (Unet) structure, namely SAM-CBAM-Unet, with the main body of the model adopting the Unet architecture. Unet was chosen because its symmetrical encoder-decoder structure and skip connection design allow it to maintain good performance even when processing small sample datasets, and it can effectively fuse multi-scale features and accurately preserve spatial detail information, which is crucial for regression tasks requiring accurate grid value prediction. In one example implementation, the network depth is set to 5 layers.

[0052] In one implementation, step S2 includes: Step S21: Assign corresponding first weights to the data types of the first training data and / or the second training data through the channel attention module.

[0053] In one exemplary implementation, the module learns to assign different weights to the 10 input feature channels (i.e., 10 sets of driving data). For example, the model can learn to give higher weight to precipitation data in the Amazon basin during the wet season, and higher weight to evapotranspiration data at the edge of the Sahara Desert during the dry season.

[0054] Step S22: Assign corresponding second weights to the spatial locations of the first training data and / or the second training data through the spatial attention module.

[0055] In one exemplary implementation, the module learns to assign weights along the spatial dimensions of the feature map, enabling it to focus on more important geographical locations. For example, when processing runoff data, the model can learn to pay more attention to major river network areas.

[0056] Step S23: Based on the first training data, the first weight, the second training data, and the second weight, convolve to obtain the output data.

[0057] Step S24: Process the output data using the loss function, optimizer, and validation strategy in the Unet deep learning model to obtain the target model.

[0058] In one implementation, the loss function is L1Loss (mean absolute error); the optimizer is the AdamW optimizer; and the validation strategy uses K-fold cross-validation. L1Loss (mean absolute error) is chosen instead of the more commonly used mean squared error (MSELoss). This choice is based on a thorough analysis of the characteristics of the GRACE data. The GRACE TWSC data exhibits a "heavy-tailed" distribution, meaning it contains many extreme outliers. L1Loss is less sensitive to these outliers than MSELoss because it does not amplify large errors by squaring them, thus making the trained model more robust and less susceptible to being misled by a few extreme data points.

[0059] The AdamW (Adam with Weight Decay) optimizer was selected. Compared with the standard Adam (Adaptive Moment Estimation) optimizer, AdamW implements weight decay (L2 regularization) more effectively. It decouples the weight decay step from the gradient update step, provides better regularization, and can effectively prevent overfitting on a relatively small dataset (216 months).

[0060] K-Fold Cross-Validation is employed. The training data is divided into K subsets, and training and validation are performed K times. Each time, a different subset is used as the validation set, and the remaining K-1 subsets are used as the training set. The optimal models obtained from these K training iterations are then ensembled to form the reconstructed model. For example, 216 months of valid data are randomly divided into 6 subsets (36 months per subset). The model is trained 6 times, using 5 subsets (180 months) as the training set each time, and the remaining subset as the validation set. This approach ensures that data from each month is both used in training and evaluated as validation data. The final reconstructed model is the ensemble average of the optimal models obtained from these 6 training iterations. This strategy maximizes the use of valuable observational data and avoids the bias that may arise from simply dividing the data into training and validation sets, thus providing a more accurate and reliable assessment of the model's generalization ability.

[0061] In one implementation, the Unet deep learning model uses the GELU (Gaussian Error Linear Unit) as the activation function. This can be understood as follows: the convolutional layers in the network use the Gaussian Error Linear Unit (GELU) as the activation function, replacing the commonly used ReLU (Rectified Linear Unit) function. The smoothness of the GELU function and its non-full saturation in the negative region help alleviate the vanishing gradient problem, making the gradient descent process more stable. The last layer of the network uses a... The convolutional layers do not use any activation functions, directly mapping the multi-channel feature maps to single-channel TWSC predictions, thus ensuring that the range of output values ​​is unrestricted and can completely match the real observations.

[0062] Those skilled in the art will understand that traditional U-nets, while the present invention aims to predict continuous changes in terrestrial water storage, describe specific implementation methods for modifying the network output layer in specific embodiments to achieve this objective. For example, the last layer of the network uses a... The convolutional layers, without using any activation functions, directly map multi-channel feature maps to single-channel TWSC predictions, thus ensuring that the range of output values ​​is unrestricted and can completely match the true observations. This specific implementation transforms the output data of U-net from discrete classification to continuous regression to achieve the final data requirements of this invention, overcoming the processing limitations of U-net, which is limited to image segmentation.

[0063] Furthermore, in order to adapt to the complex water cycle process and the characteristics of GRACE data, especially considering the "heavy-tailed" distribution of GRACE data with many extreme outliers, this application abandons the traditional mean squared error (MSELoss) and instead adopts the L1Loss (mean absolute error) function, which avoids the square amplification of large errors and improves the robustness of the model to hydrological outliers, taking into account the hydrological heterogeneity at different latitudes and longitudes around the world.

[0064] For further implementation methods, please refer to Figure 4 The U-Net deep learning model also includes a selective attention masking module. Before step S23, step S25 is included: zeroing out the ocean region data in the first and second training data using the selective attention masking module. This step operates on the model input, ensuring that the model processes only land signals from the outset. After step S23, step S26 is included: zeroing out the ocean region data in the output data using the selective attention masking module. This step operates on the model output, masking the final output grid to remove noise or artifacts that may be generated in the ocean region, ensuring a clean final product. Alternatively, this invention creatively integrates a channel attention module and a spatial attention module (SAM-CBAM) into U-Net, enabling the model to dynamically assign weights to different hydrological data types (such as different driving factors in rainforest or desert environments) and spatial locations. It also introduces a selective attention masking module specifically for zeroing out ocean region data at both the input and output. This ensures that parameter updates are driven only by land region errors, eliminating ocean noise.

[0065] Furthermore, this selective attention masking module also operates during the loss function calculation process. For example, when calculating the loss between the model output and the ground truth, both are masked simultaneously. This means that the model's parameter updates are driven only by errors in the land region, and any prediction biases in the ocean region will not affect the training process.

[0066] The following is a specific example illustrating the global land water storage change reconstruction method of the present invention: The trained SAM-CBAM-Unet model was applied to a complete 271-month driving dataset (June 2000 to December 2022) to generate a spatiotemporally complete reconstruction of global land water storage changes. This product not only fills in 33 months of data gaps during the GRACE / GRACE-FO mission but also extends the time series forward by 22 months. The model performance was comprehensively evaluated by comparing the 216 months of data observed with GRACE in the reconstruction results with the ground truth values ​​from the GSFC Mascon.

[0067] Among them, the spatial distribution fidelity is relatively good. The reconstruction results after processing by the SAM module are almost completely consistent with the original observation data in terms of global spatial distribution pattern and numerical magnitude, with only minor differences in some complex land-sea boundary areas.

[0068] The time series has good consistency, such as Figure 5 As shown, in nine typical river basins worldwide, the time series reconstructed by the model perfectly match the actual observations, accurately capturing signals of seasonal amplitudes, long-term trends, and extreme events. The model's imputation results for missing data periods seamlessly connect with the preceding and following time series, demonstrating extremely high reliability.

[0069] In one implementation, the statistical indicators are as follows: Root mean square error (RMSE): In most river basins globally, the RMSE between the reconstructed results and the true values ​​is less than 20 mm, and in many large inland river basins, the RMSE is even less than 10 mm. This value is far smaller than the 20-30 mm uncertainty range of the GRACE / GRACE-FO satellite products themselves, indicating extremely high reconstruction accuracy.

[0070] Pearson correlation coefficient (CC): In most land areas of the world, a correlation coefficient higher than 0.98 indicates that the reconstructed results are almost completely consistent with the actual data in terms of the trend of change.

[0071] Nash efficiency coefficient (NSE): In most land areas globally, an NSE value higher than 0.95 indicates that the model can perfectly capture most of the variability in the observed data, demonstrating extremely high predictive power and reliability. The global average NSE is greater than 0.99, far exceeding other existing reconstruction methods.

[0072] To more specifically demonstrate the superior performance of this invention, Table 1 below lists the RMSE, CC, and NSE indices between the reconstructed results and the actual observations in 39 large watersheds worldwide.

[0073] Table 1: Performance Indicators of Reconstruction Results for 39 Large Watersheds Worldwide As shown in Table 1, the reconstruction results of the model constructed in this invention show that the CC and NSE values ​​are all above 0.98 and the RMSE is less than 9 mm for all 39 large watersheds. These data irrefutably demonstrate that the method proposed in this invention has extremely high accuracy, reliability, and universality on a global scale.

[0074] As can be seen, this invention, based on the traditional U-net algorithm, addresses the specific problem of global terrestrial water storage change in the geosciences field, particularly by spatially isomorphizing multi-source hydrological data and reconstructing the input, core network architecture, output layer, and loss function of the U-net model. This allows the invention to not only accurately fill in data gaps corresponding to at least one time period but also to obtain extended data for longer time series using second training data and the trained target model. For example, it can extend the TWSC record forward to before the launch of the GRACE satellite, thus providing a complete dataset with a longer time span. Furthermore, this invention is not merely a data imputation tool; through deep learning, it successfully constructs an accurate and robust nonlinear mapping relationship between easily obtainable water cycle components (such as precipitation and evapotranspiration) and terrestrial water storage, which is difficult to measure directly. This is equivalent to creating a "virtual sensor" based on multi-source data, providing a new observation paradigm for future acquisition of global terrestrial water storage changes and potentially reducing reliance on continuous satellite gravity missions to some extent.

[0075] Furthermore, in the assessment of major global watersheds, the target model of this invention exhibits excellent performance in key indicators such as root mean square error (RMSE), Pearson correlation coefficient (CC), and Nash-Sutcliffe efficiency (NSE) of the reconstruction results. Its reconstruction error is much smaller than the uncertainty of the GRACE / GRACE-FO data itself, demonstrating high fidelity and data accuracy, and achieving near-lossless data reconstruction.

[0076] This invention also provides a device for reconstructing global land water storage changes, such as... Figure 6 As shown, it includes: The data acquisition module 21 is used to acquire first training data and second training data. The first training data represents the grid data of global land water storage changes over a past time span. The global land water storage change grid data includes multiple time data, multiple spatial data, and multiple land water storage change data. Among them, the spatial data and land water storage change data corresponding to at least one of the multiple time data are missing. The second training data includes global land evapotranspiration data, global land precipitation data, global runoff data, and land surface model TWS data corresponding to the multiple time data.

[0077] The target model training module 22 is used to train the target model based on the first training data, the second training data, and the Unet deep learning model.

[0078] The grid data output module 23 is used to determine complete global land water storage change grid data within a time span based on at least one time data, second training data corresponding to at least one time data, and a target model.

[0079] The present invention also provides an electronic device 200, such as... Figure 7 As shown, the electronic device includes a processor 201 and a memory 203 for storing executable instructions of the processor 201.

[0080] The processor 201 is configured to execute the methods described above by executing executable instructions.

[0081] The processor 201 can communicate with the memory 203 via the bus 202.

[0082] This invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the methods described above.

[0083] Those skilled in the art will understand that all or part of the steps of the above-described method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When executed, the program performs the steps of the above-described method embodiments. The aforementioned storage medium includes various media capable of storing program code, such as ROM, RAM, magnetic disks, or optical disks.

[0084] The preferred embodiments of the present invention have been described in detail above. It should be understood that those skilled in the art can make numerous modifications and variations based on the concept of the present invention without creative effort. Therefore, all technical solutions that can be obtained by those skilled in the art based on the concept of the present invention through logical analysis, reasoning, or limited experimentation on the basis of existing technology should be within the scope of protection defined by the claims.

Claims

1. A method for reconstructing global terrestrial water storage changes, characterized by, include: Acquire first training data and second training data. The first training data represents grid data of global land water storage changes over a past time span. The global land water storage change grid data includes multiple temporal data, multiple spatial data, and multiple land water storage change data. Among the multiple temporal data, the spatial data and land water storage change data corresponding to at least one temporal data are missing. The second training data includes global land evapotranspiration data, global land precipitation data, global runoff data, and land surface model TWS data corresponding to the multiple temporal data. The target model is trained based on the first training data, the second training data, and the Unet deep learning model. In the Unet deep learning model, the last layer of the network uses a 1×1 convolutional layer without any activation function, directly mapping the multi-channel feature map to the single-channel TWSC prediction value. Based on the at least one time data, the second training data corresponding to the at least one time data, and the target model, complete global land water storage change grid data within the time span is determined.

2. The global terrestrial water storage change reconstruction method of claim 1, wherein, The Unet deep learning model includes a channel attention module and a spatial attention module; The step of training the target model based on the first training data, the second training data, and the Unet deep learning model includes: The channel attention module assigns a corresponding first weight to the data types of the first training data and / or the second training data. The spatial attention module assigns corresponding second weights to the spatial locations of the first training data and / or the second training data. The output data is obtained by convolution based on the first training data, the first weight, the second training data, and the second weight; The output data is processed using the loss function, optimizer, and validation strategy in the Unet deep learning model to obtain the target model.

3. The global terrestrial water storage change reconstruction method of claim 2, wherein, The Unet deep learning model also includes a selective attention mask module; Before obtaining the output data after convolution based on the first training data, the first weight, the second training data, and the second weight, the method further includes: The selective attention mask module sets the ocean region data in the first training data and the second training data to zero.

4. The global terrestrial water storage change reconstruction method of claim 3, wherein, After obtaining the output data through convolution based on the first training data, the first weight, the second training data, and the second weight, the method further includes: The selective attention mask module sets the ocean region data in the output data to zero.

5. The global terrestrial water storage change reconstruction method of claim 2, wherein, The loss function is L1Loss mean absolute error; the optimizer is AdamW optimizer; and the validation strategy uses K-fold cross-validation.

6. The global terrestrial water storage change reconstruction method of claim 1, wherein, Before training the target model based on the first training data, the second training data, and the Unet deep learning model, the method further includes: The first training data and the second training data are preprocessed to ensure that the first training data and the second training data have the same spatial resolution, the same temporal resolution, and the same equivalent water level in the same units.

7. The method of reconstructing global terrestrial water storage changes according to any one of claims 1 to 6, characterized in that, The Unet deep learning model uses the GELU Gaussian error linear unit as the activation function.

8. A device for reconstructing global terrestrial water storage changes, characterized by, include: The data acquisition module is used to acquire first training data and second training data. The first training data represents grid data of global land water storage changes over a past time span. The global land water storage change grid data includes multiple temporal data, multiple spatial data, and multiple land water storage change data. Among the multiple temporal data, the spatial data and land water storage change data corresponding to at least one temporal data are missing. The second training data includes global land evapotranspiration data, global land precipitation data, global runoff data, and land surface model TWS data corresponding to the multiple temporal data. The target model training module is used to train the target model based on the first training data, the second training data, and the Unet deep learning model. In the Unet deep learning model, the last layer of the network uses a 1×1 convolutional layer without any activation function, directly mapping the multi-channel feature map to the single-channel TWSC prediction value. The grid data output module is used to determine complete global land water storage change grid data within the time span based on the at least one time data, the second training data corresponding to the at least one time data, and the target model.

9. An electronic device, comprising: The electronic device includes a processor and memory; The memory is used to store code and related data; The processor is configured to execute code in the memory to implement the method according to any one of claims 1 to 7.

10. A storage medium having a computer program stored thereon, which, when executed by a processor, implements the method as described in any one of claims 1 to 7.

Images