Emergence constraint-based method and device for improving prediction accuracy of future land water storage

Abstract

The application discloses a land water storage future prediction precision improvement method and device based on emergence constraint, and belongs to the technical field of hydrology, water resources and climate change prediction. In view of the problems that the existing multi-mode set structure is strong in heterogeneity, the hydrological mechanism of a basin is mixed, and the structured framework of GRACE observation constraint is insufficient, which leads to insufficient prediction reliability, the application obtains multi-solution center GRACE / GRACE-FO Mascon products and completes multi-dimensional correction, extracts basin scale historical diagnostic quantities, divides three types of dominant basins, namely irrigation, climate and snow, groups the multi-mode hydrological model according to the structure, constructs an emergence constraint equation which is statistically and physically tested on the subset of the structure and mechanism consistent model, substitutes the real observation diagnostic quantity to correct the original prediction, and weightedly integrates and outputs the regional future land water storage change reconstruction field, so that the prediction precision and physical consistency are effectively improved, and the application is applied to water resource management, drought early warning and climate change risk assessment.

Description

Technical Field

[0001] This invention relates to the field of hydrological, water resources and climate change prediction technology, specifically to a method and apparatus for improving the accuracy of future predictions of terrestrial water storage based on emergent constraints. Background Technology

[0002] Global terrestrial water storage (TWS) comprehensively reflects changes in various water reservoirs, including soil water, groundwater, snowmelt water, vegetation intercepted water, and surface water, and is an important indicator for measuring the state of the terrestrial water cycle. The seasonality of TWS is primarily driven by precipitation input and evapotranspiration, and is influenced by the rapid response of soil water, the seasonal phase modulation of snowmelt, and the slow changes in groundwater. On interannual to decadal scales, groundwater storage exhibits significant lag and "hydrological memory" characteristics, making it a key factor controlling medium- and long-term trends in TWS. Furthermore, human activities such as irrigation water withdrawal, groundwater extraction, reservoir management, and inter-basin water transfer are significantly altering the natural regulation patterns of regional TWS.

[0003] Current TWS (Trend Forecasting) predictions primarily rely on a multi-model ensemble of terrestrial hydrology models (such as ISIMIP2b and ISIMIP3b), including land surface process models (LSMs), global hydrological and water resources models (GHWRMs), vegetation-agriculture coupled models, and large-scale integrated nature-human systems models. These models exhibit significant differences in water storage structure, human activity parameterization, and physical process representation, leading to substantial structural uncertainty in the multi-model ensemble. Directly averaging different structural models results in physical signal cancellation, making it difficult to reliably obtain trend estimates for future TWS.

[0004] On the other hand, GRACE / GRACE-FO satellite observations provide global-scale TWS quality change data, capable of identifying large-scale hydrological change characteristics such as groundwater decline, snowmelt, and irrigation impacts. However, existing technologies rarely use GRACE observations in a structured manner to constrain multi-model future predictions. Without distinguishing between the dominant watershed hydrological mechanisms (such as irrigation, precipitation, and snow cover) or model structure categories, confounding diagnostic parameters and spurious correlations can easily arise, thus weakening the effectiveness of emergent constraint techniques.

[0005] Therefore, there is an urgent need for a method that can identify differences in model structure, distinguish watershed dominant mechanisms, and use GRACE observations to construct stable emerging relationships in order to improve the accuracy and reliability of future TWS predictions. Summary of the Invention

[0006] This application provides a method and apparatus for improving the accuracy of future prediction of terrestrial water storage based on emergence constraints, which is used to overcome the problems of large multi-mode structural heterogeneity, significant prediction bias, high uncertainty and lack of reliable observation constraints in the existing prediction of future terrestrial water storage.

[0007] To achieve the above objectives, the present invention provides the following technical solution: A method for improving the accuracy of future predictions of terrestrial water storage based on emergent constraints includes the following steps: S1. Obtain GRACE and / or GRACE-FO Mascon products from multiple solution centers, and successively perform scale factor correction, leakage error recovery, glacier isostatic adjustment (GIA) correction, geocentric term and low-order spherical harmonic coefficient correction and deseasonalization processing to construct a multi-time point terrestrial water storage anomaly sequence for the study area. At the watershed scale, extract the interdecadal trend, seasonal amplitude, and historical diagnostic quantity of terrestrial water storage in response to climate driving and human activities. S2. Based on the historical diagnostic data, combined with reanalysis data, human footprint data and snow cover data, identify the dominant hydrological mechanism of the watershed and divide the watershed into irrigation-dominated, climate-dominated and snow cover-dominated types. S3. The multi-model hydrological models used for future predictions are divided into at least a global hydrological and water resources model group and a land surface process model group according to water storage structure, human activity parameterization method and hydrological process representation method. Within each model group, for each type of watershed, the historical diagnostic quantity of the model and the corresponding future land water storage response quantity are calculated to form a pair of samples. S4. In a subset of models with consistent structure and mechanism, construct a cross-mode emergent regression relationship between historical diagnostic quantities and future land water storage response quantities; perform statistical significance test and hydrophysical consistency test on the regression relationship, and determine the regression relationship that meets the preset significance threshold and has physical rationality as the emergent constraint equation; S5. Substitute the real historical diagnostic quantities obtained from GRACE and / or GRACE-FO observations into the emergence constraint equation to obtain the constrained future land water storage response, thereby correcting the original future prediction results of the multi-model ensemble. S6. The future land water storage response quantities constrained by each watershed are weighted and integrated according to the watershed area to obtain and output the reconstructed field of future land water storage changes at the regional or global scale.

[0008] The GRACE Mascon product itself has inherent biases such as spatial signal leakage, glacier isostatic adjustment tectonic signal confusion, and low-order spherical harmonic coefficient errors. The data set of multiple solution centers can effectively suppress the systematic bias of the algorithm of a single solution center. By performing multi-dimensional correction processing in sequence, the gravity signal components caused by non-water storage changes can be gradually removed. After deseasonalization processing, the seasonal fluctuations of the annual and semi-annual cycles can be removed, thereby obtaining anomaly sequences that can truly reflect the interannual and long-term changes in water storage, providing clean and reliable benchmark data for subsequent trend diagnosis and mechanism identification. Extracting historical diagnostic data at the watershed scale allows for matching natural spatial units of hydrological processes, eliminating the interference of grid-scale random noise on subsequent emergence relation construction. Watersheds are classified according to dominant hydrological mechanisms because the response patterns of terrestrial water storage differ fundamentally under different driving mechanisms. Conflating these mechanisms to construct regression relationships would introduce significant heterogeneity and variance. Classifying watersheds by three dominant mechanisms—irrigation, climate, and snow cover—ensures the inherent consistency of hydrological processes within each sample category, fundamentally reducing the probability of spurious correlations and improving the stability of emergence relations.

[0009] Grouping multi-model hydrological models by physical structure is crucial because models with different structures exhibit fundamental differences in the parameterization of water storage processes. The mapping relationship between historical characteristics and future responses lacks cross-structure transferability. Constructing emergent relationships within models of the same structure effectively avoids the physical signal cancellation problem caused by structural heterogeneity, ensuring that regression relationships have a unified physical basis and improving the reliability of prediction corrections. Constructing cross-model emergent regression relationships within subsets with both structural and mechanistic consistency, utilizing the synergistic changes between historical diagnostic characteristics and future responses across multiple models, and combining statistical significance and hydrophysical consistency tests for screening, can eliminate accidental statistical spurious correlations, ensuring that the final emergent constraint equations possess both statistical robustness and clear hydrophysical meaning.

[0010] By substituting real historical diagnostic data obtained from GRACE satellite observations into validated emergent constraint equations, the overall shift of the multi-model ensemble can be anchored using observed ground truth values. This enables a paradigm shift in future land water storage prediction from traditional empirical ensemble averaging to observation-driven constraints, effectively narrowing the uncertainty range of multi-model predictions and significantly improving the reliability of future trend estimates. Spatial integration using watersheds as units and weighted by area ensures the consistency of water balance relationships during spatial aggregation, generating a continuous spatial distribution field that facilitates subsequent regional water resource assessment and management applications. The entire method possesses good scalability and can adapt to prediction needs at different spatial scales and for different watershed types.

[0011] Preferably, step S1 includes: Select GRACE / GRACE-FO RL06 Mascon data released by at least three solution centers: CSR, JPL, and GSFC; We perform unified grid, masking and missing measurement interpolation on data from different solution centers, and calculate the ensemble mean as the center estimate and the cross-center standard deviation as the observation uncertainty. The GIA correction field of the pre-set GIA model is introduced to perform GIA correction on high latitude and areas with significant post-glacial rebound, and the leakage error is recovered using the official scale factor field. By subtracting the multi-year average seasonal cycle from the corrected terrestrial water storage series and smoothing it, a terrestrial water storage anomaly series suitable for trend identification and climate response analysis is obtained.

[0012] CSR, JPL, and GSFC are the mainstream official solvers of GRACE / GRACE-FO data. The three institutions have different assumptions and error sources in their solver algorithms. Using a multi-center ensemble approach can effectively reduce the systematic bias of a single institution's solution. At the same time, the uncertainty level at the observation end can be quantified by cross-center standard deviation. Performing unified grid, masking, and missing interpolation processing on multi-source data can eliminate spatial mismatch problems caused by differences in spatial resolution and inconsistencies in land masking, ensuring the spatiotemporal continuity of the time series and providing a complete input basis for subsequent trend calculation and response analysis. Introducing the GIA model correction field can effectively remove the gravity signal of crustal uplift caused by glacial isostatic adjustment, avoiding misinterpretation of tectonic movement signals as changes in terrestrial water storage. Using the official scale factor field for leakage error recovery can restore the signal attenuation caused by spatial smoothing of Mascon products and accurately restore the true water storage variation at the watershed scale. Subtracting the multi-year average seasonal cycle can remove regular seasonal fluctuation signals, highlighting the long-term trend and interannual anomaly characteristics of water storage. Combined with smoothing processing, it can further suppress high-frequency random noise, effectively improving the stability and accuracy of subsequent trend identification and climate response analysis.

[0013] Preferably, step S2 includes the following: First, the Mann-Kendall trend significance test and Sen's slope estimation were performed on the watershed-scale land water storage anomaly sequence to identify watersheds that showed a significant long-term decline or a significant long-term increase in the historical period, thereby obtaining target watersheds with significant trend characteristics. Subsequently, precipitation and potential evapotranspiration sequences from ERA5-Land and / or other reanalysis datasets were read and their trends were analyzed. After meteorological attribution, influencing factors caused by natural drought or precipitation anomalies were removed so that the residual trend of terrestrial water storage could reflect the effects of non-meteorological factors such as human activities. Then, the residual trend was matched with the aforementioned significant trend watersheds to screen out areas that may be affected by human intervention. Next, the residual trend is spatially overlaid with the proportion of irrigated farmland, population density, and other indicators reflecting the intensity of human activities in the HYDE data; if a watershed shows significant signs of groundwater extraction, agricultural irrigation, or other human intervention, then the watershed is identified as an irrigation-dominated watershed. Then, read the snowmelt water equivalent SWE data, calculate the ratio of the seasonal amplitude of SWE to the seasonal amplitude of land water storage, and further calculate the lag correlation coefficient between SWE anomalies and land water storage anomalies; if the correlation coefficient exceeds the preset threshold during the snowmelt period and the contribution rate of SWE to seasonal water volume changes is significant, then the watershed is identified as a snowmelt-dominated watershed. For the remaining watersheds that were not classified as irrigation-dominated or snow-dominated, partial correlation analysis and lag regression were used to determine the watershed based on the standardized regression coefficients of precipitation and potential evapotranspiration on land water storage. If precipitation was the dominant factor, the watershed was classified as a water-limited climate-dominated watershed. If evapotranspiration or energy was the dominant factor, the watershed was classified as an energy-limited climate-dominated watershed.

[0014] By employing nonparametric Mann-Kendall trend tests and Sen's slope estimation, without assuming that the time series follows a normal distribution, we can robustly identify long-term significant trends in watershed water storage and eliminate spurious trends caused by random fluctuations. By eliminating the influence of natural fluctuations in precipitation and potential evapotranspiration through meteorological attribution, we can separate the residual trends in water storage caused by non-meteorological factors such as human activities, avoiding misjudging natural climate fluctuations as the impact of human activities. By spatially overlaying the residual trends with human activity intensity indicators, we can verify the degree of human intervention at the spatial distribution level, ensuring the objectivity and accuracy of the results for identifying irrigation-dominated watersheds.

[0015] By calculating the amplitude proportion of snowmelt equivalent and performing lag correlation analysis, the regulatory effect of snowmelt on the seasonal cycle of terrestrial water storage can be accurately captured, and watersheds with snowmelt as the main source of replenishment can be accurately identified, ensuring the physical basis for the classification of snow-dominated watersheds. Partial correlation analysis and lag regression methods are used for the remaining watersheds to quantify the respective contributions of precipitation and evapotranspiration to changes in water storage, further distinguishing between water-limited and energy-limited climate driving mechanisms, ensuring the physical integrity and full coverage of the watershed classification system. The entire watershed mechanism classification process is driven by quantitative indicators, enabling the objective and automatic classification of watershed-dominant hydrological mechanisms, avoiding biases caused by subjective classification, ensuring the inherent consistency of hydrological response patterns within the same category, and providing a reasonable spatial stratification basis for subsequent grouping and the construction of emergent constraint relationships.

[0016] Preferably, the model structure grouping in step S3 includes: For each hydrological model, read its metadata and configuration file to identify whether it has a groundwater module, a river water reservoir module, a reservoir scheduling module, a human water intake and return module, and a multi-layer soil water reservoir structure. Models that simultaneously possess the capabilities to simulate groundwater, river confluence, and human water use are classified into the Global Hydrology and Water Resources Model Group. Models that focus on land surface energy balance, soil water and heat transfer, and vegetation evapotranspiration processes, but lack complete surface water and groundwater modules, are classified as land surface process model groups. Models with natural-human system coupling modules are classified into the global hydrological and water resources model group in irrigation-dominated watersheds based on their ability to simulate human activities, and are separately labeled or grouped as needed in other watersheds.

[0017] The design of the water storage structure and the parameterization scheme of the process in a hydrological model fundamentally determine its ability to simulate changes in water storage and its response characteristics. The presence or absence of core modules such as groundwater, reservoir scheduling, and human water intake directly determines whether the model can accurately reproduce the impact of human activities such as irrigation and extraction on water storage. Global hydrological and water resource models have complete surface water-groundwater cycle processes and human water use parameterization capabilities, and can better simulate the water storage change process in areas with strong human activity. Grouping them into the same group can ensure the consistency of process expression within the group. Land surface process models focus on land surface energy balance and soil water and heat transfer, emphasizing the simulation of climate response in natural watersheds. They lack a complete structure of surface water and groundwater storage. Grouping them separately can avoid structural heterogeneity caused by mixing with complex structural models. A differentiated grouping strategy is adopted for models with natural-human system coupling capabilities. This strategy can take into account the multi-process coupling characteristics of the models, match the corresponding model groups according to the watershed type, avoid the structural heterogeneity residue caused by a one-size-fits-all grouping, further improve the homogeneity of models within the group, and ensure the physical consistency and robustness of subsequent emergent regression relationships.

[0018] Preferably, the historical diagnostic data includes at least one of the following or a combination thereof: Linear trends and significance of historical terrestrial water storage; Seasonal amplitude of terrestrial water storage and its seasonal phase difference with precipitation, snow cover, or temperature; The elasticity coefficient and response time lag structure of terrestrial water storage to precipitation anomalies; Contribution rate of terrestrial water storage to snowmelt equivalent anomaly and its lag correlation coefficient; The model's human activity module includes parameters such as water intake sensitivity, human water return rate, groundwater pumping gradient, or soil depth and water storage capacity.

[0019] Linear trend indicators can reflect the long-term surplus and deficit status and rate of change of water storage in the basin; seasonal amplitude and phase difference can characterize the regulation mechanism and time matching characteristics of the seasonal cycle of water storage; precipitation elasticity coefficient and response lag can quantify the dynamic response law of water storage to climate-driven anomalies; snow accumulation contribution rate and lag correlation can reflect the replenishment effect of snowmelt on water storage and time lag characteristics; and human activity-related parameters can reflect the intensity and mode of human intervention. By selecting historical diagnostic parameters from multiple dimensions, we can comprehensively characterize the changing features and driving mechanisms of watershed water storage, avoid the information gaps caused by single indicators, provide rich feature variables for constructing emergent constraint relationships, effectively improve the explanatory power of regression relationships for future responses, and ensure that the final constraint relationships correspond to clear hydrophysical processes, thus avoiding the problems of uninterpretability and insufficient generalization ability caused by pure statistical correlation.

[0020] Preferably, the emergent regression relationship in step S4 is a linear or nonlinear regression relationship, and the emergent regression relationship satisfies: The absolute value of the regression correlation coefficient is not less than a preset first threshold; The variance of the regression residuals in the cross-validation results is lower than the preset proportion of the variance of future responses between models; The signs of the regression parameters are consistent with the expected hydrophysical mechanism, and the residuals exhibit an approximately normal distribution.

[0021] Setting a threshold for the absolute value of the correlation coefficient ensures that the regression relationship has a basic statistical correlation strength, eliminating invalid relationships with weak correlations. Cross-validation is used to examine the ratio between the residual variance and the variance of future responses between models, verifying the constraint's ability to narrow multi-mode uncertainty and ensuring that the emergence constraint can effectively compress uncertainty—this is the core value of the emergence constraint technique. Requiring the regression parameter signs to be consistent with the expected hydrological physical mechanism is a key constraint to ensure the regression relationship has real hydrological significance, effectively eliminating statistically significant but physically contradictory spurious correlations. The approximate normal distribution of the residuals satisfies the statistical assumptions of classical regression analysis, ensuring the reliability and rationality of subsequent confidence interval estimates. The emergence constraint equations obtained through dual statistical and physical criteria possess both statistical robustness and hydrological physical rationality, effectively avoiding overfitting and spurious correlations, and ensuring the credibility and physical consistency of subsequent prediction correction results.

[0022] Preferably, step S5 further includes: The observation uncertainties of GRACE and / or GRACE-FO are propagated through the emergent regression relationship to obtain the confidence interval of the constrained future terrestrial water storage response. Based on the confidence interval, the upper limit, lower limit and optimal estimate of the future land water storage trend are constructed, generating uncertainty products that match drought risk assessment and groundwater depletion assessment.

[0023] GRACE satellite observations inherently involve computational uncertainties. By employing emergent regression relationships for error propagation calculations, these uncertainties at the observational end can be fully transmitted to the prediction results. This quantifies and constrains the credibility range of future water storage responses, avoiding information gaps caused by outputting a single estimate. Based on confidence intervals, upper and lower bounds of trends and optimal estimates for multiple scenarios can be constructed, adapting to application scenarios with different risk preferences. This provides probabilistic input support for risk-oriented applications such as drought risk assessment and groundwater depletion assessment. The generation of standardized uncertainty products can directly address the operational needs of water resource risk assessment, enhancing the practicality and application value of prediction results and providing a more comprehensive reference for risk response decisions.

[0024] Preferably, step S6 further includes: Spatial interpolation and smoothing are performed on the constrained future terrestrial water storage response to improve boundary continuity while ensuring that the trend sign and overall amplitude remain unchanged. Based on the reconstructed field of future changes in terrestrial water storage, the regional drought risk index, long-term groundwater deficit index, and water resource carrying capacity pressure index are calculated and output as map products or raster data.

[0025] Constraints obtained by basin-based analysis naturally exhibit abrupt boundary changes. Spatial interpolation and smoothing can effectively improve the boundary continuity and visualization of the spatial field without altering the trend direction or overall magnitude of change, making it more compatible with the usage habits of spatial data products. Based on the reconstruction of future land water storage changes, various water resource assessment indicators can be derived and calculated, directly outputting assessment results for business applications, avoiding secondary calculations and processing by users, and expanding the application scenarios of the method. Standardized map products and raster data are output, adaptable to mainstream GIS analysis platforms and water resource management systems, enhancing the practical application value of the results and providing direct and usable data support for large-scale water resource management, drought early warning, and climate change risk research.

[0026] A device for improving the accuracy of future predictions of terrestrial water storage based on emergent constraints, the device comprising: The observation benchmark construction module is used to obtain GRACE and / or GRACE-FOMascon products from multiple solution centers. After scale factor correction, leakage error recovery, GIA correction, geocentric term and low-order spherical harmonic coefficient correction and deseasonalization processing, it constructs a historical benchmark for terrestrial water storage and calculates historical diagnostic quantities. The watershed mechanism diagnostic module is used to identify the dominant hydrological mechanism of the watershed by combining historical diagnostic data, reanalysis data, human footprint data, and snow cover data, and to classify the watershed into irrigation-dominated, climate-dominated, and snow cover-dominated types. The model grouping and feature construction module is used to group multi-model hydrological models according to their structure and calculate the historical diagnostic values ​​and future terrestrial water storage response values ​​of the models in each model group and in various watersheds. The Emergent Relationship Construction Module is used to construct an emergent regression relationship between historical diagnostic quantities and future response quantities in a subset of models with consistent structure and mechanism, and to perform statistical significance and physical consistency tests. The constrained prediction module is used to substitute the real historical diagnostic values ​​of GRACE and / or GRACE-FO into the emergent constraint equation to obtain the constrained future terrestrial water storage response and correct the original multi-model prediction results. The results integration module is used to integrate the constrained response quantities of each watershed by weighting the watershed area to generate a reconstruction field of future land water storage changes at the regional or global scale.

[0027] The device adopts a modular architecture design, with each module corresponding to a core step of the methodology, clearly defining responsibilities and interfaces, facilitating phased development, independent upgrades, and routine maintenance. The observation benchmark construction module corresponds to the observation data preprocessing and diagnostic quantity calculation stages, responsible for outputting standardized water storage benchmark sequences. The watershed mechanism diagnosis module corresponds to the watershed dominant mechanism identification stage, responsible for automated watershed classification. The model grouping and feature construction module corresponds to the model structure grouping and paired sample preparation stage, responsible for generating compliant model sample sets. The emergence relation construction module corresponds to the constraint equation construction and verification stage, responsible for outputting robust and reliable emergence constraint equations. The constraint prediction module corresponds to the observation substitution and prediction correction stage, responsible for generating constrained prediction results. The result integration module corresponds to the spatial integration and product output stage, responsible for generating the final spatial data product. This modular architecture enables fully automated processing to improve the accuracy of terrestrial water storage prediction. Each module can be independently replaced and optimized, possessing good scalability and maintainability, and supporting batch and routine prediction update operations.

[0028] Preferably, the device further includes a data management and update module, used for unified storage, version management and timed updates of GRACE / GRACE-FO data, reanalysis data, human footprint data, land use data and multi-model outputs, and automatically triggering the re-run of the observation benchmark construction module, watershed mechanism diagnosis module, model structure grouping and diagnostic feature construction module and emergence relationship construction module when data is updated.

[0029] GRACE satellite data, reanalysis data, and multi-model outputs all undergo regular version updates and data supplementation. A unified data management and update module is set up to achieve full lifecycle management of multi-source input data, ensuring the consistency, traceability, and timeliness of input data. An automatic trigger rerun mechanism is set up to automatically complete the entire process iteration from baseline construction to constraint equation generation when input data is updated, without manual intervention, ensuring that the prediction results can be promptly integrated with the latest observation and simulation data. This module can effectively reduce manual maintenance costs, continuously ensure the timeliness and accuracy of prediction results, and adapt to the needs of long-term operational use.

[0030] Compared with existing technologies, this invention provides a method and apparatus for improving the accuracy of future prediction of terrestrial water storage based on emergent constraints, which has the following beneficial effects: First, by grouping multi-model hydrological models according to characteristics such as water storage structure and human activity parameterization capabilities, physical signal cancellation caused by structural heterogeneity can be effectively avoided, and emergent relationships can be constructed only within structurally consistent models, thereby improving the reliability of prediction corrections.

[0031] Secondly, by classifying the watershed into three mechanisms—irrigation-dominated, climate-dominated, and snow-dominated—we can avoid the mixing of hydrological signals from different sources, such as human water diversion, precipitation anomalies, and snow accumulation processes, thereby improving the stability of emergence relationships.

[0032] Furthermore, by introducing real observational diagnostics from GRACE to replace model diagnostics, the prediction of future terrestrial water storage shifts from traditional empirical averaging to observation-driven approaches, significantly improving the reliability of future trend estimates. The future terrestrial water storage trends corrected by the method described in this application show significant convergence, and are reasonably corrected by physical mechanisms in areas of strong human intervention, climate-driven areas, and snow-dominated areas.

[0033] Furthermore, the method proposed in this application has good scalability and is applicable to scenarios such as large-scale water resource management, drought early warning, groundwater depletion assessment, and climate change risk research, achieving deep integration of observational information and multi-model ensembles. Attached Figure Description

[0034] Figure 1 This is a schematic diagram of the overall process of the terrestrial water storage future prediction method based on emergence constraints of the present invention; Figure 2 This is a schematic diagram showing the detailed process of constructing the error-corrected GRACE observation benchmark provided in an embodiment of the present invention; Figure 3 A detailed schematic diagram of the watershed mechanism diagnostic system driven by multi-source data provided in an embodiment of the present invention; Figure 4This is a flowchart of the multi-mode structural grouping and historical diagnostic feature construction provided in the embodiments of the present invention; Figure 5 A schematic diagram illustrating the process of identifying emerging constraint relationships and correcting future predictions provided in an embodiment of the present invention; Figure 6 A schematic diagram of the processing flow involving observation constraints, trend estimation, and uncertainty quantification provided for embodiments of the present invention; Figure 7 This is a schematic diagram of the software and hardware combination system structure with complete functions in an embodiment of the present invention; Figure 8 This is a schematic diagram of the electronic device structure of the present invention.

[0035] In the diagram: 21. Memory; 22. Processor; 23. Communication unit; 24. Cloud platform. Detailed Implementation

[0036] This embodiment addresses key technical bottlenecks in existing terrestrial water storage prediction systems, such as significant differences in multi-model structures, complex climate-driven mechanisms, difficulty in quantifying the impact of human activities, and lack of reliable observational constraints. It proposes a method that integrates multi-source observational data, reanalysis data, land use information, and multi-model simulation results. Through model structure partitioning, watershed mechanism classification, diagnostic feature construction, and emergence relationship identification, it aims to significantly improve the accuracy of future terrestrial water storage trend prediction.

[0037] Current land water storage simulations typically rely on dozens of global hydrological models and land surface models. However, due to significant differences in their water storage structures, human activity parameterization methods, soil depth settings, groundwater representations, and snowmelt mechanisms, these models exhibit significant biases in their TWS responses to the same climate drivers. Traditional simple averaging methods often mask these differences in physical mechanisms and can even lead to a "strong model suppressing a weak model" phenomenon, making future trend predictions unreliable. The inventors, through in-depth research, believe that only by classifying models according to their physical structures and water storage systems, and dividing watersheds according to their dominant mechanisms, can a reliable emergent relationship between historical diagnostic quantities and future response quantities be established within structurally consistent subsets. This allows for future corrections using the observational benchmarks provided by GRACE.

[0038] To overcome the above-mentioned technical defects, the technical solution proposed in this embodiment is as follows: Figure 1 As shown, the method includes the following steps: constructing an error-corrected GRACE observation benchmark; constructing a watershed mechanism diagnostic system driven by multi-source data; structurally grouping multiple models and constructing historical diagnostic features; identifying emergent relationships within subsets with consistent structure and mechanisms; constraining future predictions using GRACE real diagnostic data; and integrating the constrained prediction results globally.

[0039] Figure 1 The flowchart illustrates the overall process of the method proposed in this application. However, those skilled in the art should understand that this flowchart is merely an illustrative example and should not be construed as a sole limitation on the order of steps. In practical applications, each step can be adjusted, extended, or simplified according to system requirements; some steps may be executed in parallel or combined in certain scenarios.

[0040] To more clearly illustrate the technical solution of this application, the following will focus on five core modules.

[0041] The first core module is the construction of the GRACE terrestrial water storage observation benchmark. This embodiment uses GRACE / GRACE-FORL06 Mascon data provided by CSR, JPL, and GSFC, and fuses these observations using ensemble averaging. Due to the low spatial resolution, significant leakage effect, and non-negligible GIA influence of the GRACE data, this embodiment first performs data cleaning, scale factor correction, GIA correction, geocentric term correction, and deseasonalization to obtain a high-precision TWSA sequence. By ensemble averaging the data from the three calculation centers, a central estimate of terrestrial water storage changes can be obtained; and observational uncertainty can be calculated based on the standard deviations between different calculation centers. Based on these processing steps, a high-confidence global terrestrial water storage change reference sequence can be constructed, thus providing a reliable observation benchmark for subsequent emergent constraint analysis.

[0042] The second core module is the construction of a multi-source data-driven watershed mechanism diagnostic system. This embodiment argues that the dominant mechanisms of TWS (True Water Soundness) changes differ significantly across watersheds. Some regions are primarily influenced by human activities (water resource extraction, irrigation water demand, reservoir scheduling, etc.), some are mainly controlled by snow accumulation processes (seasonal variations in snowmelt equivalent and snowmelt replenishment lags), and some are primarily driven by climate variability (precipitation and evapotranspiration processes). Without identifying and classifying these mechanisms, any operation using globally unified emergent relationships will generate severe confounding, leading to the invalidation of constraints. Therefore, this embodiment establishes a three-tiered watershed classification system—human activity-driven, snow accumulation-driven, and climate-driven—based on GRACE trends, human footprint data, ERA5-Land precipitation trends, SWE variance contribution, and lag correlation analysis.

[0043] In identifying human-dominated watersheds, this embodiment employs a trend significance test combined with meteorological attribution to eliminate the effects of natural drought. Then, the remaining areas with significant negative trends are overlaid with the proportion of irrigated farmland and population density to ultimately identify watersheds globally that are strongly influenced by human activities. In these watersheds, the bias in future predictions mainly stems from insufficient model simulation of water withdrawal; therefore, it is necessary to establish emergent relationships within a subset of GHWRMs.

[0044] In identifying snow-dominated watersheds, this embodiment uses the ratio of seasonal amplitude of SWE to seasonal amplitude of TWS as the main indicator, and combines the hysteresis correlation between SWE and TWSA to determine the control intensity of snowmelt on TWS. If SWE makes a significant contribution to TWS amplitude and there is obvious synchronicity during the snowmelt period, it is considered a snow-dominated watershed.

[0045] For the remaining watersheds, this embodiment employs partial correlation analysis to identify their climate-driving mechanisms. Orthogonal decomposition of precipitation and potential evapotranspiration is used to distinguish between water-limited and energy-limited watersheds. Emergent relationships in these watersheds typically reflect model storage capacity bias and the intensity of TWS response to P and PET drives.

[0046] The third core module is model structure grouping and diagnostic feature construction. This embodiment categorizes multi-models into two types based on structure: GHWRMs and LSMs. The core reason for this is that GHWRMs typically explicitly simulate complex processes such as groundwater, rivers, reservoirs, and human water intake, with a complete water storage structure; while LSMs emphasize land surface energy flux, soil thermal processes, and vegetation evapotranspiration, resulting in a significantly simplified water storage structure. Diagnostic features are calculated for both types of models, including TWS trends, seasonal amplitude, sensitivity to P or SWE, human activity simulation capabilities, and phase deviation analysis.

[0047] The fourth core module is the construction of emergent relationships. The key idea of ​​this application is that within a subset of models with consistent structure and mechanisms, there exists a stable relationship between historical diagnostic features and future TWS responses. For example, a model that better reproduces historical GRACE trends may have more reliable future trends; a model that responds weakly to precipitation anomalies may have a lower future climate-driven trend. This cross-model relationship is identified using linear or nonlinear regression methods, and its statistical significance and physical mechanism consistency are verified.

[0048] The fifth core module is the constraint and global integration of future forecasts. The core improvement in this embodiment is to substitute the actual diagnostic data of GRACE into the emergent relation equation to calculate the constraint value for future trends, thereby correcting the model's future forecast bias. By integrating the constraint results of all watersheds by area, a future global TWS trend map and regional risk distribution map are generated.

[0049] The above five core modules constitute the overall framework of the future terrestrial water storage prediction method based on emergence constraints proposed in this application. The specific steps, data flow, calculation principles, equipment composition, and technical characteristics of each module in this embodiment will be described in more detail below.

[0050] See also Figure 1To more clearly illustrate the overall technical solution, this embodiment provides a more detailed description of the five core modules in terms of data flow, computational path, and logical relationships, so that those skilled in the art can accurately reproduce the technical method of this application in actual implementation. It should also be emphasized that the execution order of the following steps can be appropriately adjusted without departing from the inventive concept, and some steps can be implemented through parallel computing. This application aims to describe a repeatable, interpretable, scalable, and highly reliable system for predicting future terrestrial water storage.

[0051] First, we will further explain the process of building the GRACE data benchmark. See [link / reference] Figure 2 Because the spatial resolution of GRACE data is approximately 300 kilometers, its use at the watershed scale can lead to issues such as boundary leakage, grid overlap, and increased noise. Therefore, this embodiment employs the officially provided scale factor correction to convert the low-resolution modular data into an estimate that more closely approximates the actual changes in water storage. Furthermore, for regions significantly affected by Glacial Isotropic Adjustment (GIA), such as northern Canada, Scandinavia, and the vicinity of Antarctica, this embodiment introduces the GIA correction field from the ICE-6G_C model for inversion correction on the GRACE TWS benchmark, thereby avoiding interference from the slow crustal rebound effect on TWS trend identification. During processing, the geocentric term and low-order spherical harmonic coefficients in the GRACE data can also be replaced to ensure that the correlation coefficients are not contaminated by orbital gravity field inversion biases, thus preventing contamination of the overall signal.

[0052] To obtain a more robust benchmark, this embodiment subtracts the multi-year average seasonal cycle from the original GRACE time series to obtain the TWSA series. In data processing, local regression smoothing and bandpass filtering methods can be used to reduce high-frequency noise, thereby obtaining a stationary series more suitable for trend identification and climate relationship analysis. It should be noted that the time window used for trend analysis can be flexibly adjusted according to the characteristics of different regions; for example, short time series are not suitable for areas strongly influenced by monsoon systems, while longer time windows are more suitable for areas with long-term over-exploitation of groundwater to improve the stability of trend calculations. The watershed mechanism identification system will be further explained next. Figure 3 This embodiment demonstrates the five-step process for constructing a watershed mechanism classification on a global scale. The accuracy of the watershed mechanism classification directly determines whether emergent relationships can be stably identified and whether they have physical meaning. Therefore, this embodiment employs rigorous error propagation analysis in data fusion, feature calculation, significance testing, and spatial overlay, and introduces an adaptive threshold determination method to drive the mechanism judgment.

[0053] In the identification of human-dominated watersheds, this embodiment first performs the Mann-Kendall trend test and Sen's slope estimation on the GRACE TWSA sequence to assess the significance and gradient of the trend. If the trend is significantly negative and the rate of decline exceeds an empirical threshold, the next step, meteorological decoupling, is initiated. In subsequent processing, precipitation sequences from ERA5-Land are read and precipitation trend curves are constructed. If the precipitation trend does not show a significant decline, or its decline is significantly less than the decline in land water storage (TWS), it can be determined that there may be a TWS decline driven by non-meteorological factors in the region. Subsequently, human activity-related indicators such as the proportion of irrigated farmland area and population density from HYDE data are introduced, and the aforementioned regions with potential human-dominated characteristics are spatially overlaid with this type of data. If a watershed is located in an area with a significantly high irrigation ratio or a densely populated area, the watershed can be identified as a human-dominated watershed.

[0054] In human-dominated watersheds, the TWS trend is typically associated with groundwater extraction or large-scale irrigation. Therefore, the emergence relation primarily focuses on the model's ability to simulate human activities; that is, models with strong human activity parameterization tend to better reproduce the GRACE trend. Thus, this embodiment sets specific diagnostic features in the emergence constraints of these watersheds, such as the model's sensitivity to water withdrawal simulation, the strength of human activity module parameters, and groundwater reservoir depth settings, and constructs corresponding emergence relations based on these features.

[0055] Next, let's examine the process of identifying snow-dominated watersheds. During data processing, the SWE data from ERA5-Land can be read, and its seasonal amplitude can be converted into equivalent water thickness. In high-latitude or high-altitude regions, since SWE has a significant dominant role in the seasonal variation of terrestrial water storage (TWS), the contribution rate of SWE can be further calculated, and the lag correlation coefficient between the SWE anomaly (SWEA) and the terrestrial water storage anomaly (TWSA) can be calculated. Simultaneously, the maximum correlation can be searched within multiple lag windows to determine the degree of snow cover dominance over TWS changes. If the maximum correlation coefficient exceeds a threshold and occurs during the snowmelt period, it can be considered that the TWS changes in this region are mainly controlled by snowmelt.

[0056] It is worth noting that this embodiment also sets dynamic SWE thresholds for different climate zones. For example, in the monsoon-controlled Qinghai-Tibet Plateau region, SWE and precipitation jointly affect TWS amplitude, while in high-latitude regions, SWE is the absolutely dominant factor. Therefore, the SWE determination threshold can be adaptively adjusted according to factors such as annual average temperature, altitude, and precipitation type to ensure the accuracy of watershed classification.

[0057] For climate-dominated watersheds, this embodiment employs partial correlation and lag regression methods to identify their hydrological control mechanisms. In the processing, the terrestrial water storage anomaly (TWSA), precipitation (P), and potential evapotranspiration (PET) are first detrended to avoid spurious correlations caused by long-term trend interference. Then, a lag regression model is constructed, using a multi-month lag window to fit the TWSA response to precipitation anomalies or evapotranspiration-driven events. If the standardized regression coefficient corresponding to precipitation is significantly higher than that of evapotranspiration in the model, it is classified as water-limited; if the evapotranspiration coefficient is dominant, it is classified as energy-limited.

[0058] This embodiment suggests that the emergence relationship in climate-dominant regions reflects the differences in water storage capacity among models and their sensitivity to climate drivers. Therefore, diagnostic features typically include the elasticity coefficient of historical TWS to precipitation anomalies, model soil depth settings, and the degree of groundwater-soil coupling.

[0059] After watershed mechanism identification is completed, global watersheds can be divided into three categories, and the model structure grouping process begins. This embodiment classifies the models of ISIMIP2b and ISIMIP3b into two categories: GHWRMs and LSMs. See [link to relevant documentation]. Figure 4 GHWRMs typically feature explicit surface water reservoirs, groundwater reservoirs, and multi-layered soil water storage structures, and include a human water use module; LSMs, on the other hand, primarily focus on surface-atmosphere energy fluxes and generally lack comprehensive surface and groundwater simulation capabilities. This embodiment employs automated model label recognition, reading the metadata of each model to determine whether it includes river channel reservoirs, reservoir scheduling, human water use modules, groundwater modules, etc., thereby enabling clear classification.

[0060] In some complex models, such as MIROC-INTEG-LAND with natural-human coupling modules, this embodiment classifies it as a model with a hybrid structure by examining its simulation structure and reading its human intervention sub-modules. However, in the emergent relationship construction stage, it is usually only involved in the analysis in watersheds with human activity relationships.

[0061] After grouping the models according to their structures, diagnostic features are calculated. In this embodiment, the diagnostic features include historical trends, seasonal amplitude, phase synchronicity, precipitation sensitivity, SWE sensitivity, human activity simulation response capability, and water storage capacity simulation capability. These features include both univariate statistical indicators and intervariate interaction features. For example, the lagged correlation between the land water storage anomaly (TWSA) and the snow cover water equivalent anomaly (SWEA) can be calculated; simultaneously, the response characteristics of precipitation (P) and potential evapotranspiration (PET) to land water storage (TWS) can be analyzed based on the temporal differences between the model inputs.

[0062] After the diagnostic features are constructed, the emerging relationship identification stage begins. See [link to relevant documentation]. Figure 5 Within a subset of models with consistent structure and mechanisms, regression equations describing the relationship between historical diagnostic values ​​and future land water storage responses can be constructed. For example, in a common scenario, the historical TWS trend D and future trend R of the model have the following relationship: R = aD + b. This embodiment uses least squares regression to estimate parameters a and b, and cross-validation to assess the stability of the relationship. Physical consistency is also used; for example, if the model's sensitivity to precipitation significantly deviates from its actual climate sensitivity, it will not participate in the construction of emergent relationships.

[0063] To ensure that the emergent relationships are statistically significant, this embodiment sets significance thresholds, such as requiring the regression correlation coefficient to exceed a certain standard, the residuals to follow an approximately normal distribution, and the standard error to be less than a certain proportion of the differences between models.

[0064] After identifying the emergent relationships, the true diagnostic value of GRACE (Dobs) is substituted into the regression equation to obtain the constrained estimate of the future trend (REC). See [link to relevant documentation]. Figure 6 This process can significantly correct structural biases in future simulations of the original model. For example, in irrigation-dominated areas, if the lack of human activity modules in LSMs leads to a significant underestimation of trends, the constrained results will be closer to reality.

[0065] The constrained prediction results of all watersheds are then integrated to form a global prediction product of future land water storage, which can be further expanded into trend maps, drought risk maps, and regional water resource assessment maps.

[0066] The above content constitutes the overall technical flow of this embodiment. The following Part III will provide a more detailed description of the apparatus, device, storage medium and specific examples to further improve the reproducibility of the implementation of this application.

[0067] See also Figures 1 to 6 To ensure the efficient execution of the emergence-constraint-based terrestrial water storage future prediction method provided in this application within a real-world system, this embodiment further describes the possible hardware and software environment, module relationships, data transmission methods, and model training and computation processes that the method may rely on, thereby guaranteeing the engineering feasibility of the aforementioned technical solution. This embodiment also further illustrates the optional extension methods, algorithm adaptation strategies, and system optimization processes for different scenario requirements, making this application not only theoretically feasible but also operable in practical business systems, research platforms, and even national-level water resources monitoring systems.

[0068] In the future forecasting process proposed in this application, a large amount of multi-source climate, hydrological, and socioeconomic data with different spatial resolutions will be used. This data will be scaled, quality-controlled, physically analyzed, and its mechanisms identified to create a highly consistent dataset suitable for impact analysis. Due to the high complexity of the data in terms of spatial resolution, temporal scale, and variable categories, automated data management capabilities are required to support unified data reading, updating, and organization. This embodiment achieves automatic retrieval and loading of the GRACE dataset, ERA5 series reanalysis data, HYDE human activity data, LUH2 land use change data, SWE data, and ISIMIP multi-model outputs through a pre-defined data interface. The system has a scheduled update function, automatically triggering data updates and retraining processes after new versions of data are released, ensuring the timeliness and accuracy of the forecast results.

[0069] Taking the GRACE data benchmark construction module as an example, the program reads Mascon products from three solution centers and automatically identifies differences in latitude and longitude grids, time series lengths, and variable naming conventions during loading. Then, it performs grid alignment, unified masking, and missing value interpolation on the three datasets. In the GIA correction and leakage error recovery stages, pre-stored high-resolution terrain data and geophysical parameters are used to recover the signals based on the scale factor information provided by the Mascon data. It is worth noting that for years with severe data gaps, adaptive temporal interpolation technology is used to avoid introducing non-physical trend biases, thereby ensuring the stability of subsequent trend analysis.

[0070] In the process of watershed-scale standardization, this embodiment uses unified global watershed boundary data, such as STN-30p or HydroBASINS, and standardizes various data types in terms of resolution and projection method, so that all variables can be directly compared at the "watershed-month" or "watershed-year" scale. An efficient regional aggregation algorithm is used to achieve the area-weighted averaging process, significantly reducing computation time, especially in scenarios requiring monthly calculations for more than 180 watersheds worldwide.

[0071] In the watershed mechanism diagnosis module, envelope analysis is introduced to improve the robustness of SWE-dominant relationship identification. This involves assessing not only the contribution of SWE to TWS amplitude but also the stability of SWEA's explanatory power for TWSA across years. If a watershed exhibits SWE-dominant characteristics in multiple years, the confidence level of the mechanism classification will significantly increase. In the anthropogenic identification stage, this embodiment also uses the cumulative water scarcity index and groundwater depletion index to further verify the source of the trend, thereby avoiding misclassification of watersheds affected by natural climate variability as anthropogenic.

[0072] During the model structure grouping process, the simulation configuration files accompanying each model can be read to identify whether the model contains structural information such as human activity simulation modules, groundwater modules, river water reservoir modules, and reservoir scheduling modules. Under the ISIMIP3b framework, different models have different structures. If they are directly used in ensemble averaging without differentiation, non-physical trend mixing will occur in future predictions. This embodiment splits the models according to their structure, allowing models with the same structure to participate in the construction of emergent relationships within the same mechanism watershed, thereby ensuring the physical interpretability of the constraints.

[0073] In constructing diagnostic features, this embodiment not only calculates simple statistical indicators but also constructs composite features that deeply reflect the physical mechanisms of the model. For example, in climate-dominated regions, it calculates the elasticity coefficient of TWS to precipitation anomalies and the response time lag structure, and seeks the optimal regression window among different lag monthly distances. In snow-dominated regions, it not only calculates the synchronization relationship between SWEA and TWSA but also analyzes the interaction between snowmelt amount and soil liquid water content, further confirming the dominant effect of snow accumulation. In human-dominated regions, it constructs feature vectors based on water withdrawal parameterization, such as the agricultural irrigation water demand curve, human activity return rate, and groundwater extraction gradient given by the model. These features can more fully distinguish the strength of the model's ability to simulate human activities, providing a reliable foundation for constructing emergent relationships.

[0074] In the emergent relationship construction module, statistical-physical relationships between historical diagnostic quantities and future response quantities are sought within a subset of models with consistent structure and mechanisms. The stability of these relationships is verified through model-by-model residual analysis. When a relationship is stable and has a clear physical meaning, it is determined to be an interpretable emergent relationship. This embodiment also considers the possibility of nonlinearity during the relationship construction process. For example, in snow-dominated areas, the snowmelt process may exhibit a nonlinear response relationship; therefore, a linear or nonlinear fitting method is selected based on the regional characteristics.

[0075] The actual diagnostic values ​​of GRACE are then substituted into the regression equation to obtain future trend constraint values. In this process, the observation error of GRACE can be propagated through the regression model, and the confidence interval of the constraint results can be calculated accordingly, thus ensuring that the final prediction results have quantitative interpretability.

[0076] In the global integrated processing, the constraint results of future trends in each watershed can be weighted according to the watershed area to generate the future land water storage (TWS) change pattern of the global landmass (excluding ice sheets). To enhance the spatial continuity of the results, a spatial smoothing algorithm can be used to process the regional boundaries without changing the sign and overall magnitude of the trend, thereby ensuring that the prediction results maintain physical rationality in space.

[0077] At the application level, the method proposed in this application can support various scenarios. For example, national water resources management departments can use this method to construct water storage evolution trends for the next ten to twenty years and identify areas at risk of groundwater depletion in advance. In the agricultural production sector, this method can be used to predict the sustainability of future irrigation water demand. In the field of climate research, this method can provide a scientific basis for risk assessment of extreme drought events. In the model comparison task of the Global Climate Change Assessment Report, this method can serve as an important tool for the structured correction of multi-model results.

[0078] Regarding devices based on the above method, for example Figure 7 As shown in the embodiments of this application, a software and hardware combination system with complete functionality is provided. The diagnostic construction module in this device performs GRACE correction and benchmark construction; the mechanism identification module performs SWE diagnosis, trend analysis, and human activity discrimination; the model structure grouping module reads model structure information and performs classification; the relationship construction module establishes emergent relationships; the constraint prediction module corrects future outputs; and the result integration module is responsible for generating the final product. These modules communicate with each other through data interfaces, possessing scalability and pluggability.

[0079] Regarding storage media, the software system of this embodiment can be stored in a read-only memory, flash memory, hard disk, optical disk or other readable media, and can run on any compatible electronic device to perform all the above steps.

[0080] In an electronic device embodiment, such as Figure 8 As shown, the electronic device includes a processor 22, a memory 21, and a communication unit 23. The processor 22 is responsible for executing the entire calculation process of the above method, the memory 21 is used to store model parameters, datasets, and program instructions, and the communication unit 23 is used to obtain observation data from external servers or databases, such as updating GRACE or ERA5 data in real time, to ensure that the system can operate independently or collaboratively with the outside world.

[0081] In scenarios requiring deployment in a high-performance computing environment, processor 22 can be a combination of multi-core CPU and GPU to accelerate computations such as regression analysis, data resampling, and feature construction. This embodiment also supports deploying the method to cloud platform 24, enabling the method to serve not only single-machine computing but also to be provided as an API to various business systems.

[0082] To illustrate the applicability of the method in this application, the following application demonstration is provided. In a certain type of irrigation area, long-term groundwater extraction has led to a significant decline in the TWS trend, and some land surface models do not simulate human activities, resulting in predictions that significantly deviate from GRACE observations. Through the mechanism identification method of this embodiment, this area is classified as human-dominated; subsequently, an emergent relationship is constructed within the GHWRMs structure set, and the actual GRACE diagnostic value is substituted into the model to obtain future trend constraint values. The results show that this method can significantly improve the accuracy of future predictions, and the trend estimation is more consistent with independent observation data.

[0083] In snow-dominant regions, discrepancies in model simulations of SWE changes often lead to TWS amplitude deviations. This embodiment identifies the snow-dominant mechanism through SWE contribution rate, phase bias, and hysteresis correlation analysis, and establishes emergent relationships. After constraints, the model's estimation of TWS trends caused by future snow cover reduction becomes more reliable.

[0084] In summary, this embodiment combines GRACE observations, climate reanalysis data, multi-model outputs, and socioeconomic data. Through mechanism partitioning, structural grouping, and emergent constraints, it constructs an interpretable, scalable, and reproducible system for predicting future terrestrial water storage, which fully compensates for the shortcomings of existing technologies in handling model heterogeneity, identifying mechanism uncertainties, and fusion of observations.

Claims

1. A method for improving the accuracy of future predictions of terrestrial water storage based on emergent constraints, characterized in that, Includes the following steps: S1. Obtain GRACE and / or GRACE-FO Mascon products from multiple solution centers, and after scale factor correction, leakage error recovery, glacier isostatic adjustment (GIA) correction, geocentric term and low-order spherical harmonic coefficient correction and deseasonalization, construct a multi-time point terrestrial water storage anomaly sequence in the study area, and extract the interdecadal trend, seasonal amplitude and historical diagnostic quantity of terrestrial water storage at the watershed scale for climate-driven and human-activity-responsive data. S2. Based on historical diagnostic data, combined with reanalysis data, human footprint data and snow cover data, the dominant hydrological mechanisms in the watershed are identified, and the watershed is divided into irrigation-dominated, climate-dominated and snow cover-dominated types. S3. The multi-model hydrological models used for future predictions are divided into at least a global hydrological and water resources model group and a land surface process model group according to water storage structure, human activity parameterization method and hydrological process representation method. Within each group, for each type of watershed, the historical diagnostic quantity of the model and the corresponding future land water storage response quantity are calculated to form a paired sample. S4. In a subset of models with consistent structure and mechanism, construct cross-mode emergent regression relationships between historical diagnostic quantities and future land water storage response quantities; Statistical significance and hydrophysical consistency tests were performed on the regression relationships, and the regression relationships that met the preset significance threshold and had physical rationality were determined as the emergence constraint equations. S5. Substitute the real historical diagnostic quantities obtained from GRACE and / or GRACE-FO observations into the emergence constraint equation to obtain the constrained future land water storage response, and correct the original future prediction results of the multi-model ensemble; integrate the constrained future land water storage response quantities of each watershed according to the watershed area to obtain the reconstructed field of future land water storage changes at the regional or global scale and output it.

2. The method according to claim 1, characterized in that, Step S1 includes: Select GRACE / GRACE-FO RL06 Mascon data released by at least three solution centers: CSR, JPL, and GSFC; We perform unified grid, masking and missing measurement interpolation on data from different solution centers, and calculate the ensemble mean as the center estimate and the cross-center standard deviation as the observation uncertainty. The GIA correction field of the pre-set GIA model is introduced to perform GIA correction on high latitude and areas with significant post-glacial rebound, and the leakage error is recovered using the official scale factor field. By subtracting the multi-year average seasonal cycle from the corrected terrestrial water storage series and smoothing it, a terrestrial water storage anomaly series suitable for trend identification and climate response analysis is obtained.

3. The method according to claim 1 or 2, characterized in that, Step S2 includes the following: First, the Mann-Kendall trend significance test and Sen's slope estimation were performed on the watershed-scale land water storage anomaly sequence to identify watersheds that showed a significant long-term decline or a significant long-term increase in the historical period, thereby obtaining target watersheds with significant trend characteristics. Subsequently, precipitation and potential evapotranspiration sequences from ERA5-Land and / or other reanalysis datasets were read and their trends were analyzed. After meteorological attribution, influencing factors caused by natural drought or precipitation anomalies were removed so that the residual trend of terrestrial water storage could reflect the effects of non-meteorological factors such as human activities. Then, the residual trend was matched with the aforementioned significant trend watersheds to screen out areas that may be affected by human intervention. Next, the residual trend is spatially overlaid with the proportion of irrigated farmland, population density, and other indicators reflecting the intensity of human activities in the HYDE data; if a watershed shows significant signs of groundwater extraction, agricultural irrigation, or other human intervention, then the watershed is identified as an irrigation-dominated watershed. Then, read the snowmelt water equivalent SWE data, calculate the ratio of the seasonal amplitude of SWE to the seasonal amplitude of land water storage, and further calculate the lag correlation coefficient between SWE anomalies and land water storage anomalies; if the correlation coefficient exceeds the preset threshold during the snowmelt period and the contribution rate of SWE to seasonal water volume changes is significant, then the watershed is identified as a snowmelt-dominated watershed. For the remaining watersheds that were not classified as irrigation-dominated or snow-dominated, partial correlation analysis and lag regression were used to determine the watershed based on the standardized regression coefficients of precipitation and potential evapotranspiration on land water storage. If precipitation was the dominant factor, the watershed was classified as a water-limited climate-dominated watershed. If evapotranspiration or energy was the dominant factor, the watershed was classified as an energy-limited climate-dominated watershed.

4. The method according to any one of claims 1-3, characterized in that, The model structure grouping in step S3 includes: For each hydrological model, read its metadata and configuration file to identify whether it has a groundwater module, a river water reservoir module, a reservoir scheduling module, a human water intake and return module, and a multi-layer soil water reservoir structure. Models that simultaneously possess the capabilities to simulate groundwater, river confluence, and human water use are classified into the Global Hydrology and Water Resources Model Group. Models that focus on land surface energy balance, soil water and heat transfer, and vegetation evapotranspiration processes, but lack complete surface water and groundwater modules, are classified as land surface process model groups. Models with natural-human system coupling modules are classified into the global hydrological and water resources model group in irrigation-dominated watersheds based on their ability to simulate human activities, and are separately labeled or grouped as needed in other watersheds.

5. The method according to any one of claims 1-4, characterized in that, The historical diagnostic data includes at least one of the following or a combination thereof: Linear trends and significance of historical terrestrial water storage; Seasonal amplitude of terrestrial water storage and its seasonal phase difference with precipitation, snow cover, or temperature; The elasticity coefficient and response time lag structure of terrestrial water storage to precipitation anomalies; Contribution rate of terrestrial water storage to snowmelt equivalent anomaly and its lag correlation coefficient; The model's human activity module includes parameters such as water intake sensitivity, human water return rate, groundwater pumping gradient, or soil depth and water storage capacity.

6. The method according to any one of claims 1-5, characterized in that, In step S4, the emergent regression relationship is either linear or nonlinear, and the emergent regression relationship satisfies: The absolute value of the regression correlation coefficient is not less than a preset first threshold; The variance of the regression residuals in the cross-validation results is lower than the preset proportion of the variance of future responses between models; The signs of the regression parameters are consistent with the expected hydrophysical mechanism, and the residuals exhibit an approximately normal distribution.

7. The method according to any one of claims 1-6, characterized in that, Step S5 also includes: The observation uncertainties of GRACE and / or GRACE-FO are propagated through the emergent regression relationship to obtain the confidence interval of the constrained future terrestrial water storage response. Based on the confidence interval, the upper limit, lower limit and optimal estimate of the future land water storage trend are constructed, generating uncertainty products that match drought risk assessment and groundwater depletion assessment.

8. The method according to any one of claims 1-7, characterized in that, Step S6 further includes: Spatial interpolation and smoothing are performed on the constrained future terrestrial water storage response to improve boundary continuity while ensuring that the trend sign and overall amplitude remain unchanged. Based on the reconstructed field of future changes in terrestrial water storage, the regional drought risk index, long-term groundwater deficit index, and water resource carrying capacity pressure index are calculated and output as map products or raster data.

9. A device for improving the accuracy of future predictions of terrestrial water storage based on emergent constraints, characterized in that, The device includes: The observation benchmark construction module is used to acquire GRACE and / or GRACE-FO Mascon products from multiple solution centers, and to perform scale factor correction, leakage error recovery, GIA correction, geocentric term and low-order spherical harmonic coefficient correction and deseasonalization on the products to construct a historical benchmark for terrestrial water storage and calculate historical diagnostic quantities. The watershed mechanism diagnosis module is used to identify the dominant hydrological mechanism of the watershed based on the historical diagnostic data, reanalysis data and human footprint data, and to classify the watershed into at least irrigation-dominated, climate-dominated and snow-dominated types. The model structure grouping and diagnostic feature construction module is used to group multi-model hydrological models by structure and calculate the historical diagnostic quantity and future land water storage response quantity of the model in each model structure group and each type of watershed. The emergent relationship construction module is used to construct an emergent regression relationship between historical diagnostic quantities and future response quantities in a subset of models with consistent structure and mechanism, and to perform statistical significance and physical consistency tests on the relationship. The constrained prediction module is used to substitute the real historical diagnostic values ​​of GRACE and / or GRACE-FO into the emergent regression relationship to obtain the constrained future terrestrial water storage response, and to correct the original multi-model prediction. The results integration module is used to weight and integrate the future land water storage response quantities constrained by each watershed according to the watershed area, and generate a reconstruction field of future land water storage changes at the regional or global scale.

10. The apparatus according to claim 9, characterized in that, The device also includes a data management and update module, which is used to uniformly store, version manage and update GRACE / GRACE-FO data, reanalysis data, human footprint data, land use data and multi-model outputs, and automatically trigger the re-run of the observation benchmark construction module, watershed mechanism diagnosis module, model structure grouping and diagnostic feature construction module and emergence relationship construction module when the data is updated.

Images