Fupeng Li
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Previous Organizations / Institutes
The organizations / institutes from the scholar's other articles in the past ten years
Research Subjects
Research Interests
Fupeng Li has a total of 116 co-authors. Here are the top co-authors by number of shared publications:
Jürgen Kusche
Institute of Geodesy and Geoinformation, University of Bonn, Bonn, Germany
21 shared publications- Zhengtao Wang
School of Geodesy and Geomatics Wuhan University Wuhan China
11 shared publications - Helena Gerdener
University of Bonn, Bonn, 53115, Germany
10 shared publications - Anne Springer
University of Bonn, Bonn, Germany
8 shared publications - Nengfang Chao
College of Marine Science and Technology, China University of Geosciences, Wuhan 430074, China
7 shared publications Bailing Li
6 shared publications
Latest Publications
- Open Access
- 9 November 2026
- Published by Copernicus GmbH
- Vol. 2026, 1-23
Reply on RC2
Fresh water availability is critical for ecosystems, agriculture, industry, and human communities. Anticipating drought conditions benefits from forecasting changes in terrestrial water storage (TWS), the total water stored on land across all compartments, including groundwater, rivers, glaciers, and soil moisture. While individual compartments, such as groundwater, are difficult to observe directly at large scales, TWS integrates their combined changes and can be measured globally through satellite gravimetry. Since 2002, the GRACE and GRACE Follow-On (GRACE-FO) missions have delivered monthly, global estimates of terrestrial water storage anomalies (TWSA), deviations from a long-term mean, making TWSA the most accessible large-scale indicator of hydrological change. Predicting TWSA is nonetheless challenging as it reflects processes operating at vastly different temporal and spatial scales. We present HiGNN-LSTM, a hierarchical graph neural network that represents the Earth across two spatial scales, coupled with an LSTM module to forecast global TWSA for up to six months ahead. As a proof of concept, we show that the hierarchical graph neural network can automatically generate meaningful input features for TWSA forecasting from ERA5 climate variables, without requiring manual predictor selection or lag-correlation analysis. Trained on a GRACE-like reconstruction of TWSA extending from 1979 to 2020, the model substantially reduces the one-month-lead RMSE relative to a seasonal climatology baseline (1.83 cm vs 3.70 cm) and consistently outperforms a ConvLSTM across the full six-month horizon. Skill over climatology shrinks at longer leads and is lost by six months, indicating that most of the gain concentrates at short leads. Evaluation against GRACE- and GRACE-FO-derived TWSA highlights the difficulty of transferring a model trained on reconstructed TWSA to satellite-derived observations.
- 9 November 2026
- Published by Copernicus GmbH
- Vol. 2026, 1-23
Open AccessReply on RC2
Fresh water availability is critical for ecosystems, agriculture, industry, and human communities. Anticipating drought conditions benefits from forecasting changes in terrestrial water storage (TWS), the total water stored on land across all compartments, including groundwater, rivers, glaciers, and soil moisture. While individual compartments, such as groundwater, are difficult to observe directly at large scales, TWS integrates their combined changes and can be measured globally through satellite gravimetry. Since 2002, the GRACE and GRACE Follow-On (GRACE-FO) missions have delivered monthly, global estimates of terrestrial water storage anomalies (TWSA), deviations from a long-term mean, making TWSA the most accessible large-scale indicator of hydrological change. Predicting TWSA is nonetheless challenging as it reflects processes operating at vastly different temporal and spatial scales. We present HiGNN-LSTM, a hierarchical graph neural network that represents the Earth across two spatial scales, coupled with an LSTM module to forecast global TWSA for up to six months ahead. As a proof of concept, we show that the hierarchical graph neural network can automatically generate meaningful input features for TWSA forecasting from ERA5 climate variables, without requiring manual predictor selection or lag-correlation analysis. Trained on a GRACE-like reconstruction of TWSA extending from 1979 to 2020, the model substantially reduces the one-month-lead RMSE relative to a seasonal climatology baseline (1.83 cm vs 3.70 cm) and consistently outperforms a ConvLSTM across the full six-month horizon. Skill over climatology shrinks at longer leads and is lost by six months, indicating that most of the gain concentrates at short leads. Evaluation against GRACE- and GRACE-FO-derived TWSA highlights the difficulty of transferring a model trained on reconstructed TWSA to satellite-derived observations.
- Open Access
- 9 November 2026
- Published by Copernicus GmbH
- Vol. 2026, 1-23
Reply on RC1
Fresh water availability is critical for ecosystems, agriculture, industry, and human communities. Anticipating drought conditions benefits from forecasting changes in terrestrial water storage (TWS), the total water stored on land across all compartments, including groundwater, rivers, glaciers, and soil moisture. While individual compartments, such as groundwater, are difficult to observe directly at large scales, TWS integrates their combined changes and can be measured globally through satellite gravimetry. Since 2002, the GRACE and GRACE Follow-On (GRACE-FO) missions have delivered monthly, global estimates of terrestrial water storage anomalies (TWSA), deviations from a long-term mean, making TWSA the most accessible large-scale indicator of hydrological change. Predicting TWSA is nonetheless challenging as it reflects processes operating at vastly different temporal and spatial scales. We present HiGNN-LSTM, a hierarchical graph neural network that represents the Earth across two spatial scales, coupled with an LSTM module to forecast global TWSA for up to six months ahead. As a proof of concept, we show that the hierarchical graph neural network can automatically generate meaningful input features for TWSA forecasting from ERA5 climate variables, without requiring manual predictor selection or lag-correlation analysis. Trained on a GRACE-like reconstruction of TWSA extending from 1979 to 2020, the model substantially reduces the one-month-lead RMSE relative to a seasonal climatology baseline (1.83 cm vs 3.70 cm) and consistently outperforms a ConvLSTM across the full six-month horizon. Skill over climatology shrinks at longer leads and is lost by six months, indicating that most of the gain concentrates at short leads. Evaluation against GRACE- and GRACE-FO-derived TWSA highlights the difficulty of transferring a model trained on reconstructed TWSA to satellite-derived observations.
- 9 November 2026
- Published by Copernicus GmbH
- Vol. 2026, 1-23
Open AccessReply on RC1
Fresh water availability is critical for ecosystems, agriculture, industry, and human communities. Anticipating drought conditions benefits from forecasting changes in terrestrial water storage (TWS), the total water stored on land across all compartments, including groundwater, rivers, glaciers, and soil moisture. While individual compartments, such as groundwater, are difficult to observe directly at large scales, TWS integrates their combined changes and can be measured globally through satellite gravimetry. Since 2002, the GRACE and GRACE Follow-On (GRACE-FO) missions have delivered monthly, global estimates of terrestrial water storage anomalies (TWSA), deviations from a long-term mean, making TWSA the most accessible large-scale indicator of hydrological change. Predicting TWSA is nonetheless challenging as it reflects processes operating at vastly different temporal and spatial scales. We present HiGNN-LSTM, a hierarchical graph neural network that represents the Earth across two spatial scales, coupled with an LSTM module to forecast global TWSA for up to six months ahead. As a proof of concept, we show that the hierarchical graph neural network can automatically generate meaningful input features for TWSA forecasting from ERA5 climate variables, without requiring manual predictor selection or lag-correlation analysis. Trained on a GRACE-like reconstruction of TWSA extending from 1979 to 2020, the model substantially reduces the one-month-lead RMSE relative to a seasonal climatology baseline (1.83 cm vs 3.70 cm) and consistently outperforms a ConvLSTM across the full six-month horizon. Skill over climatology shrinks at longer leads and is lost by six months, indicating that most of the gain concentrates at short leads. Evaluation against GRACE- and GRACE-FO-derived TWSA highlights the difficulty of transferring a model trained on reconstructed TWSA to satellite-derived observations.
- Open Access
- 15 October 2026
- Published by Copernicus GmbH
GRACE/-FO Data Assimilation for Hydrological Applications: From Model-Specific Solutions Towards Broader Use
For 20 years now, GRACE and GRACE-FO data have been integrated into hydrological and land-surface models using data assimilation (DA) approaches. Numerous studies have demonstrated the value of GRACE/-FO DA systems for improving estimates of groundwater evolution, enhancing drought monitoring, and providing more accurate flood potential estimates. However, unlike other observation types such as remotely sensed soil moisture, GRACE/-FO observations are rarely assimilated into hydrological and land-surface models on an operational basis. With the prospect of near-real-time observations from upcoming satellite gravimetry missions, the question of how to make GRACE/-FO DA more broadly applicable is becoming increasingly relevant.Currently, there is no single established method for assimilating GRACE/-FO data into hydrological and land-surface models. While some settings have already reached a certain degree of consensus, for example suitable ensemble-based DA methods and the treatment of GRACE/-FO error covariance matrices, other aspects remain less well established. A key challenge is that GRACE/-FO observes integrated terrestrial water storage, whereas hydrological and land-surface models represent this storage through different components, such as soil moisture, snow, groundwater, and surface water. Assimilating GRACE/-FO data therefore requires determining how the observed TWS information is distributed among the model states. The formulation of the state vector and observation operator plays an important role in this process: the state vector determines which model variables are included in the assimilation, while the observation operator determines how these states are combined to represent the integrated TWS observed by GRACE/-FO. Different formulations of the state vector and observation operator can therefore change how information from the observations is distributed among storage components. Additionally, the representation of model uncertainty can strongly influence the assimilation results. We illustrate these challenges using two different DA systems: one based on the global model WaterGap Global Hydrological Model (WGHM) and one based on the regional high-resolution European land-surface model enCore Community Land Model (eCLM).Our results from eCLM-DA show that the choice of the state vector and the representation of model uncertainty strongly influence the assimilation results. Experiences with WGHM further indicate that some of these challenges are common across modelling systems, while others depend on model structure and resolution. We therefore distinguish between general methodological challenges, model-specific requirements, and computational and operational constraints, and identify areas where further methodological development is needed, including for coupled and multivariate DA systems. These considerations highlight challenges that must be overcome for the broader application of GRACE/-FO DA across modelling systems and regions, including its integration into regional monitoring and early warning systems.
- 15 October 2026
- Published by Copernicus GmbH
Open AccessGRACE/-FO Data Assimilation for Hydrological Applications: From Model-Specific Solutions Towards Broader Use
For 20 years now, GRACE and GRACE-FO data have been integrated into hydrological and land-surface models using data assimilation (DA) approaches. Numerous studies have demonstrated the value of GRACE/-FO DA systems for improving estimates of groundwater evolution, enhancing drought monitoring, and providing more accurate flood potential estimates. However, unlike other observation types such as remotely sensed soil moisture, GRACE/-FO observations are rarely assimilated into hydrological and land-surface models on an operational basis. With the prospect of near-real-time observations from upcoming satellite gravimetry missions, the question of how to make GRACE/-FO DA more broadly applicable is becoming increasingly relevant.Currently, there is no single established method for assimilating GRACE/-FO data into hydrological and land-surface models. While some settings have already reached a certain degree of consensus, for example suitable ensemble-based DA methods and the treatment of GRACE/-FO error covariance matrices, other aspects remain less well established. A key challenge is that GRACE/-FO observes integrated terrestrial water storage, whereas hydrological and land-surface models represent this storage through different components, such as soil moisture, snow, groundwater, and surface water. Assimilating GRACE/-FO data therefore requires determining how the observed TWS information is distributed among the model states. The formulation of the state vector and observation operator plays an important role in this process: the state vector determines which model variables are included in the assimilation, while the observation operator determines how these states are combined to represent the integrated TWS observed by GRACE/-FO. Different formulations of the state vector and observation operator can therefore change how information from the observations is distributed among storage components. Additionally, the representation of model uncertainty can strongly influence the assimilation results. We illustrate these challenges using two different DA systems: one based on the global model WaterGap Global Hydrological Model (WGHM) and one based on the regional high-resolution European land-surface model enCore Community Land Model (eCLM).Our results from eCLM-DA show that the choice of the state vector and the representation of model uncertainty strongly influence the assimilation results. Experiences with WGHM further indicate that some of these challenges are common across modelling systems, while others depend on model structure and resolution. We therefore distinguish between general methodological challenges, model-specific requirements, and computational and operational constraints, and identify areas where further methodological development is needed, including for coupled and multivariate DA systems. These considerations highlight challenges that must be overcome for the broader application of GRACE/-FO DA across modelling systems and regions, including its integration into regional monitoring and early warning systems.
- Open Access
- 11 September 2026
- Published by Copernicus GmbH
- Vol. 2026, 1-23
Comment on egusphere-2026-2312
Fresh water availability is critical for ecosystems, agriculture, industry, and human communities. Anticipating drought conditions benefits from forecasting changes in terrestrial water storage (TWS), the total water stored on land across all compartments, including groundwater, rivers, glaciers, and soil moisture. While individual compartments, such as groundwater, are difficult to observe directly at large scales, TWS integrates their combined changes and can be measured globally through satellite gravimetry. Since 2002, the GRACE and GRACE Follow-On (GRACE-FO) missions have delivered monthly, global estimates of terrestrial water storage anomalies (TWSA), deviations from a long-term mean, making TWSA the most accessible large-scale indicator of hydrological change. Predicting TWSA is nonetheless challenging as it reflects processes operating at vastly different temporal and spatial scales. We present HiGNN-LSTM, a hierarchical graph neural network that represents the Earth across two spatial scales, coupled with an LSTM module to forecast global TWSA for up to six months ahead. As a proof of concept, we show that the hierarchical graph neural network can automatically generate meaningful input features for TWSA forecasting from ERA5 climate variables, without requiring manual predictor selection or lag-correlation analysis. Trained on a GRACE-like reconstruction of TWSA extending from 1979 to 2020, the model substantially reduces the one-month-lead RMSE relative to a seasonal climatology baseline (1.83 cm vs 3.70 cm) and consistently outperforms a ConvLSTM across the full six-month horizon. Skill over climatology shrinks at longer leads and is lost by six months, indicating that most of the gain concentrates at short leads. Evaluation against GRACE- and GRACE-FO-derived TWSA highlights the difficulty of transferring a model trained on reconstructed TWSA to satellite-derived observations.
- 11 September 2026
- Published by Copernicus GmbH
- Vol. 2026, 1-23
Open AccessComment on egusphere-2026-2312
Fresh water availability is critical for ecosystems, agriculture, industry, and human communities. Anticipating drought conditions benefits from forecasting changes in terrestrial water storage (TWS), the total water stored on land across all compartments, including groundwater, rivers, glaciers, and soil moisture. While individual compartments, such as groundwater, are difficult to observe directly at large scales, TWS integrates their combined changes and can be measured globally through satellite gravimetry. Since 2002, the GRACE and GRACE Follow-On (GRACE-FO) missions have delivered monthly, global estimates of terrestrial water storage anomalies (TWSA), deviations from a long-term mean, making TWSA the most accessible large-scale indicator of hydrological change. Predicting TWSA is nonetheless challenging as it reflects processes operating at vastly different temporal and spatial scales. We present HiGNN-LSTM, a hierarchical graph neural network that represents the Earth across two spatial scales, coupled with an LSTM module to forecast global TWSA for up to six months ahead. As a proof of concept, we show that the hierarchical graph neural network can automatically generate meaningful input features for TWSA forecasting from ERA5 climate variables, without requiring manual predictor selection or lag-correlation analysis. Trained on a GRACE-like reconstruction of TWSA extending from 1979 to 2020, the model substantially reduces the one-month-lead RMSE relative to a seasonal climatology baseline (1.83 cm vs 3.70 cm) and consistently outperforms a ConvLSTM across the full six-month horizon. Skill over climatology shrinks at longer leads and is lost by six months, indicating that most of the gain concentrates at short leads. Evaluation against GRACE- and GRACE-FO-derived TWSA highlights the difficulty of transferring a model trained on reconstructed TWSA to satellite-derived observations.
- Open Access
- 4 August 2026
- Published by Copernicus GmbH
- p. 1-23
Comment on egusphere-2026-2312
Fresh water availability is critical for ecosystems, agriculture, industry, and human communities. Anticipating drought conditions benefits from forecasting changes in terrestrial water storage (TWS), the total water stored on land across all compartments, including groundwater, rivers, glaciers, and soil moisture. While individual compartments, such as groundwater, are difficult to observe directly at large scales, TWS integrates their combined changes and can be measured globally through satellite gravimetry. Since 2002, the GRACE and GRACE Follow-On (GRACE-FO) missions have delivered monthly, global estimates of terrestrial water storage anomalies (TWSA), deviations from a long-term mean, making TWSA the most accessible large-scale indicator of hydrological change. Predicting TWSA is nonetheless challenging as it reflects processes operating at vastly different temporal and spatial scales. We present HiGNN-LSTM, a hierarchical graph neural network that represents the Earth across two spatial scales, coupled with an LSTM module to forecast global TWSA for up to six months ahead. As a proof of concept, we show that the hierarchical graph neural network can automatically generate meaningful input features for TWSA forecasting from ERA5 climate variables, without requiring manual predictor selection or lag-correlation analysis. Trained on a GRACE-like reconstruction of TWSA extending from 1979 to 2020, the model substantially reduces the one-month-lead RMSE relative to a seasonal climatology baseline (1.83 cm vs 3.70 cm) and consistently outperforms a ConvLSTM across the full six-month horizon. Skill over climatology shrinks at longer leads and is lost by six months, indicating that most of the gain concentrates at short leads. Evaluation against GRACE- and GRACE-FO-derived TWSA highlights the difficulty of transferring a model trained on reconstructed TWSA to satellite-derived observations.
- 4 August 2026
- Published by Copernicus GmbH
- p. 1-23
Open AccessComment on egusphere-2026-2312
Fresh water availability is critical for ecosystems, agriculture, industry, and human communities. Anticipating drought conditions benefits from forecasting changes in terrestrial water storage (TWS), the total water stored on land across all compartments, including groundwater, rivers, glaciers, and soil moisture. While individual compartments, such as groundwater, are difficult to observe directly at large scales, TWS integrates their combined changes and can be measured globally through satellite gravimetry. Since 2002, the GRACE and GRACE Follow-On (GRACE-FO) missions have delivered monthly, global estimates of terrestrial water storage anomalies (TWSA), deviations from a long-term mean, making TWSA the most accessible large-scale indicator of hydrological change. Predicting TWSA is nonetheless challenging as it reflects processes operating at vastly different temporal and spatial scales. We present HiGNN-LSTM, a hierarchical graph neural network that represents the Earth across two spatial scales, coupled with an LSTM module to forecast global TWSA for up to six months ahead. As a proof of concept, we show that the hierarchical graph neural network can automatically generate meaningful input features for TWSA forecasting from ERA5 climate variables, without requiring manual predictor selection or lag-correlation analysis. Trained on a GRACE-like reconstruction of TWSA extending from 1979 to 2020, the model substantially reduces the one-month-lead RMSE relative to a seasonal climatology baseline (1.83 cm vs 3.70 cm) and consistently outperforms a ConvLSTM across the full six-month horizon. Skill over climatology shrinks at longer leads and is lost by six months, indicating that most of the gain concentrates at short leads. Evaluation against GRACE- and GRACE-FO-derived TWSA highlights the difficulty of transferring a model trained on reconstructed TWSA to satellite-derived observations.
- Open Access
- 23 June 2026
- Published by Copernicus GmbH in EGUsphere
- Vol. 2026, 1-23
Hierarchical Graph Networks for Seasonal Forecasts of Terrestrial Water Storage Anomalies
Fresh water availability is critical for ecosystems, agriculture, industry, and human communities. Anticipating drought conditions benefits from forecasting changes in terrestrial water storage (TWS), the total water stored on land across all compartments, including groundwater, rivers, glaciers, and soil moisture. While individual compartments, such as groundwater, are difficult to observe directly at large scales, TWS integrates their combined changes and can be measured globally through satellite gravimetry. Since 2002, the GRACE and GRACE Follow-On (GRACE-FO) missions have delivered monthly, global estimates of terrestrial water storage anomalies (TWSA), deviations from a long-term mean, making TWSA the most accessible large-scale indicator of hydrological change. Predicting TWSA is nonetheless challenging as it reflects processes operating at vastly different temporal and spatial scales. We present HiGNN-LSTM, a hierarchical graph neural network that represents the Earth across two spatial scales, coupled with an LSTM module to forecast global TWSA for up to six months ahead. As a proof of concept, we show that the hierarchical graph neural network can automatically generate meaningful input features for TWSA forecasting from ERA5 climate variables, without requiring manual predictor selection or lag-correlation analysis. Trained on a GRACE-like reconstruction of TWSA extending from 1979 to 2020, the model substantially reduces the one-month-lead RMSE relative to a seasonal climatology baseline (1.83 cm vs 3.70 cm) and consistently outperforms a ConvLSTM across the full six-month horizon. Skill over climatology shrinks at longer leads and is lost by six months, indicating that most of the gain concentrates at short leads. Evaluation against GRACE- and GRACE-FO-derived TWSA highlights the difficulty of transferring a model trained on reconstructed TWSA to satellite-derived observations.
- 23 June 2026
- Published by Copernicus GmbH in EGUsphere
- Vol. 2026, 1-23
Open AccessHierarchical Graph Networks for Seasonal Forecasts of Terrestrial Water Storage Anomalies
Fresh water availability is critical for ecosystems, agriculture, industry, and human communities. Anticipating drought conditions benefits from forecasting changes in terrestrial water storage (TWS), the total water stored on land across all compartments, including groundwater, rivers, glaciers, and soil moisture. While individual compartments, such as groundwater, are difficult to observe directly at large scales, TWS integrates their combined changes and can be measured globally through satellite gravimetry. Since 2002, the GRACE and GRACE Follow-On (GRACE-FO) missions have delivered monthly, global estimates of terrestrial water storage anomalies (TWSA), deviations from a long-term mean, making TWSA the most accessible large-scale indicator of hydrological change. Predicting TWSA is nonetheless challenging as it reflects processes operating at vastly different temporal and spatial scales. We present HiGNN-LSTM, a hierarchical graph neural network that represents the Earth across two spatial scales, coupled with an LSTM module to forecast global TWSA for up to six months ahead. As a proof of concept, we show that the hierarchical graph neural network can automatically generate meaningful input features for TWSA forecasting from ERA5 climate variables, without requiring manual predictor selection or lag-correlation analysis. Trained on a GRACE-like reconstruction of TWSA extending from 1979 to 2020, the model substantially reduces the one-month-lead RMSE relative to a seasonal climatology baseline (1.83 cm vs 3.70 cm) and consistently outperforms a ConvLSTM across the full six-month horizon. Skill over climatology shrinks at longer leads and is lost by six months, indicating that most of the gain concentrates at short leads. Evaluation against GRACE- and GRACE-FO-derived TWSA highlights the difficulty of transferring a model trained on reconstructed TWSA to satellite-derived observations.
- Top Cited PapersOpen Access
- 19 February 2026
- Published by Copernicus GmbH in Hydrology and Earth System Sciences
- Vol. 30 (4) , 985-1022
A review of current best practices and future directions in assimilating GRACE/-FO terrestrial water storage data into numerical models
Water cycle reanalyses, generated by integrating observations into hydrological and land surface models, provide long-term and consistent estimates of key water cycle components. Reanalyses are essential to understand hydrological variability, extreme events such as droughts and floods, and to improve water resource management. Over the past two decades, the assimilation of terrestrial water storage anomaly data from the GRACE and GRACE Follow-On (GRACE/-FO) missions has significantly enhanced these reanalyses, as GRACE/-FO observations uniquely constrain total water storage variability across all terrestrial compartments. Incorporating GRACE/-FO data has led to major advances in representing trends in key hydrological variables, climate-driven changes in the water cycle, and anthropogenic influences such as irrigation-induced groundwater depletion – factors often poorly captured in models. With processing pipelines now being developed for low-latency short-term data products from the upcoming next-generation gravity missions, we expect that low-latency periodically updated reanalyses and analyses from assimilation will become more relevant. However, challenges remain, particularly in resolving mismatches in spatial and temporal resolution between GRACE/-FO observations and high-resolution models, and there is no consensus yet on the optimal approach for assimilating GRACE/-FO data. In light of the upcoming launches of next-generation gravity missions and the development of increasingly sophisticated Earth system modeling frameworks, this review synthesizes insights from approximately 60 GRACE/-FO data assimilation studies in an attempt to converge to best practices. The review reveals that the most effective assimilation strategies leverage (robust modifications of) the classical ensemble Kalman filter and localization techniques, explicitly account for correlated observation errors, and address biases contained in the observations as well as those arising from model perturbations. Unmodeled processes must be carefully handled through signal separation, multi-source assimilation, or removal prior to assimilation. Future directions include developing low-latency products for near-real-time assimilation, integrating enhanced and combined satellite observations, and employing machine-learning approaches for downscaling and hybrid assimilation. Collectively, these strategies provide a pathway toward more accurate, physically consistent, and operationally useful water cycle reanalyses.
- 19 February 2026
- Published by Copernicus GmbH in Hydrology and Earth System Sciences
- Vol. 30 (4) , 985-1022
Top Cited PapersOpen AccessA review of current best practices and future directions in assimilating GRACE/-FO terrestrial water storage data into numerical models
Water cycle reanalyses, generated by integrating observations into hydrological and land surface models, provide long-term and consistent estimates of key water cycle components. Reanalyses are essential to understand hydrological variability, extreme events such as droughts and floods, and to improve water resource management. Over the past two decades, the assimilation of terrestrial water storage anomaly data from the GRACE and GRACE Follow-On (GRACE/-FO) missions has significantly enhanced these reanalyses, as GRACE/-FO observations uniquely constrain total water storage variability across all terrestrial compartments. Incorporating GRACE/-FO data has led to major advances in representing trends in key hydrological variables, climate-driven changes in the water cycle, and anthropogenic influences such as irrigation-induced groundwater depletion – factors often poorly captured in models. With processing pipelines now being developed for low-latency short-term data products from the upcoming next-generation gravity missions, we expect that low-latency periodically updated reanalyses and analyses from assimilation will become more relevant. However, challenges remain, particularly in resolving mismatches in spatial and temporal resolution between GRACE/-FO observations and high-resolution models, and there is no consensus yet on the optimal approach for assimilating GRACE/-FO data. In light of the upcoming launches of next-generation gravity missions and the development of increasingly sophisticated Earth system modeling frameworks, this review synthesizes insights from approximately 60 GRACE/-FO data assimilation studies in an attempt to converge to best practices. The review reveals that the most effective assimilation strategies leverage (robust modifications of) the classical ensemble Kalman filter and localization techniques, explicitly account for correlated observation errors, and address biases contained in the observations as well as those arising from model perturbations. Unmodeled processes must be carefully handled through signal separation, multi-source assimilation, or removal prior to assimilation. Future directions include developing low-latency products for near-real-time assimilation, integrating enhanced and combined satellite observations, and employing machine-learning approaches for downscaling and hybrid assimilation. Collectively, these strategies provide a pathway toward more accurate, physically consistent, and operationally useful water cycle reanalyses.
- Open Access
- 3 February 2026
- Published by American Geophysical Union (AGU) in Water Resources Research
- Vol. 62 (2)
Observation‐Driven Forecast of Global Terrestrial Water Storage and Evaluation for 2010–2024
Since 2002, the Gravity Recovery and Climate Experiment (GRACE) and GRACE Follow‐On (GRACE/‐FO) satellite missions have provided unprecedented measurements of terrestrial water storage changes (TWSC). These data are essential for monitoring the global water cycle, supporting drought and flood risk management, and informing water‐related decision‐making. However, GRACE products are typically released with a latency of several months, limiting their utility for real‐time and operational forecasting applications. In this study, we use machine learning to forecast GRACE‐like TWSC up to 12 months ahead, relying solely on observational and reanalysis‐based inputs. The observation‐driven forecast approach is evaluated over the period 2010–2024 and benchmarked against seasonal forecasts from the European Centre for Medium‐Range Weather Forecasts (ECMWF)’s new long‐range forecasting system (SEAS5). Our results show that the developed method offers improved accuracy and robustness compared to the ECMWF forecasts, providing a viable data‐driven alternative for operational TWSC forecasting. We generate global forecast data sets at 1° resolution, creating a robust, publicly available resource that extends GRACE‐like insights into the near future. The study addresses the latency of GRACE/‐FO products by offering real‐time TWSC forecasts to support applications such as drought early warning, sea level prediction, hydrological model validation, and geodetic applications such as forecasting Earth orientation parameters via hydrological angular momentum excitation or estimating loading corrections in GNSS and altimetry data analysis. The hindcast data set (2010–2024) evaluated in this study and the regularly updated semi‐operational forecast data set (from 2024 onward) are publicly available at: https://doi.pangaea.de/10.1594/PANGAEA.973113 and https://www.igg.uni‐bonn.de/apmg/de/data‐and‐models/grace‐fo‐forecasting .
- 3 February 2026
- Published by American Geophysical Union (AGU) in Water Resources Research
- Vol. 62 (2)
Open AccessObservation‐Driven Forecast of Global Terrestrial Water Storage and Evaluation for 2010–2024
Since 2002, the Gravity Recovery and Climate Experiment (GRACE) and GRACE Follow‐On (GRACE/‐FO) satellite missions have provided unprecedented measurements of terrestrial water storage changes (TWSC). These data are essential for monitoring the global water cycle, supporting drought and flood risk management, and informing water‐related decision‐making. However, GRACE products are typically released with a latency of several months, limiting their utility for real‐time and operational forecasting applications. In this study, we use machine learning to forecast GRACE‐like TWSC up to 12 months ahead, relying solely on observational and reanalysis‐based inputs. The observation‐driven forecast approach is evaluated over the period 2010–2024 and benchmarked against seasonal forecasts from the European Centre for Medium‐Range Weather Forecasts (ECMWF)’s new long‐range forecasting system (SEAS5). Our results show that the developed method offers improved accuracy and robustness compared to the ECMWF forecasts, providing a viable data‐driven alternative for operational TWSC forecasting. We generate global forecast data sets at 1° resolution, creating a robust, publicly available resource that extends GRACE‐like insights into the near future. The study addresses the latency of GRACE/‐FO products by offering real‐time TWSC forecasts to support applications such as drought early warning, sea level prediction, hydrological model validation, and geodetic applications such as forecasting Earth orientation parameters via hydrological angular momentum excitation or estimating loading corrections in GNSS and altimetry data analysis. The hindcast data set (2010–2024) evaluated in this study and the regularly updated semi‐operational forecast data set (from 2024 onward) are publicly available at: https://doi.pangaea.de/10.1594/PANGAEA.973113 and https://www.igg.uni‐bonn.de/apmg/de/data‐and‐models/grace‐fo‐forecasting .
- Open Access
- 6 January 2026
- Published by American Geophysical Union (AGU) in Geophysical Research Letters
- Vol. 53 (1)
Reproducing GRACE Total Water Storage Change at Finer Spatial Scales
The Gravity Recovery and Climate Experiment (GRACE) satellite mission provides accurate observations of terrestrial water storage anomalies (TWSA), but their coarse spatial resolution (∼3°) limits sub‐regional‐scale applications. Existing downscaling methods rely on high‐resolution hydrometeorological inputs, which often underestimate the full magnitude of GRACE signals. Here, we develop a machine‐learning‐based iterative downscaling method that reproduces TWSA at 0.25° resolution while retaining nearly all of the original GRACE signals, using ERA5 soil moisture, precipitation, and temperature as inputs. We find that the downscaled TWSA has improved agreement with in situ groundwater levels compared to the original GRACE data, with higher correlation at over 63% of wells and reduced RMSE at more than 83% globally. The downscaled TWSA also retains an average correlation of 0.99 with original GRACE data at the basin scale, outperforming a previously released downscaling product. The downscaled TWSA data set is publicly available at https://doi.org/10.5281/zenodo.17265162 .
- 6 January 2026
- Published by American Geophysical Union (AGU) in Geophysical Research Letters
- Vol. 53 (1)
Open AccessReproducing GRACE Total Water Storage Change at Finer Spatial Scales
The Gravity Recovery and Climate Experiment (GRACE) satellite mission provides accurate observations of terrestrial water storage anomalies (TWSA), but their coarse spatial resolution (∼3°) limits sub‐regional‐scale applications. Existing downscaling methods rely on high‐resolution hydrometeorological inputs, which often underestimate the full magnitude of GRACE signals. Here, we develop a machine‐learning‐based iterative downscaling method that reproduces TWSA at 0.25° resolution while retaining nearly all of the original GRACE signals, using ERA5 soil moisture, precipitation, and temperature as inputs. We find that the downscaled TWSA has improved agreement with in situ groundwater levels compared to the original GRACE data, with higher correlation at over 63% of wells and reduced RMSE at more than 83% globally. The downscaled TWSA also retains an average correlation of 0.99 with original GRACE data at the basin scale, outperforming a previously released downscaling product. The downscaled TWSA data set is publicly available at https://doi.org/10.5281/zenodo.17265162 .
- Open Access
- 3 December 2025
- Published by Copernicus GmbH
- Vol. 2025, 1-51
Reply on RC2
Water cycle reanalyses, generated by integrating observations into hydrological and land surface models, provide long-term and consistent estimates of key water cycle components. Reanalyses are essential to understand hydrological variability, extreme events such as droughts and floods, and to improve water resource management. Over the past two decades, the assimilation of terrestrial water storage anomaly data from the GRACE and GRACE Follow-On (GRACE/-FO) missions has significantly enhanced these reanalyses, as GRACE/-FO observations uniquely constrain total water storage variability across all terrestrial compartments. Incorporating GRACE/-FO data has led to major advances in representing trends in key hydrological variables, climate-driven changes in the water cycle, and anthropogenic influences such as irrigation-induced groundwater depletion – factors often poorly captured in models. However, challenges remain, particularly in resolving mismatches in spatial and temporal resolution between GRACE/-FO observations and high-resolution models, and there is no consensus yet on the optimal approach for assimilating GRACE/-FO data. In light of the upcoming launches of next-generation gravity missions and the development of increasingly sophisticated Earth system modeling frameworks, it is an opportune time to compile the recommendations of GRACE/-FO data assimilation studies to date, in an attempt to converge to best practices. This review synthesizes past achievements, critically examines unresolved challenges, and explores future directions for advancing water cycle reanalyses using satellite gravimetry observations through improved assimilation strategies, machine learning, and near-real-time intake of satellite data.
- 3 December 2025
- Published by Copernicus GmbH
- Vol. 2025, 1-51
Open AccessReply on RC2
Water cycle reanalyses, generated by integrating observations into hydrological and land surface models, provide long-term and consistent estimates of key water cycle components. Reanalyses are essential to understand hydrological variability, extreme events such as droughts and floods, and to improve water resource management. Over the past two decades, the assimilation of terrestrial water storage anomaly data from the GRACE and GRACE Follow-On (GRACE/-FO) missions has significantly enhanced these reanalyses, as GRACE/-FO observations uniquely constrain total water storage variability across all terrestrial compartments. Incorporating GRACE/-FO data has led to major advances in representing trends in key hydrological variables, climate-driven changes in the water cycle, and anthropogenic influences such as irrigation-induced groundwater depletion – factors often poorly captured in models. However, challenges remain, particularly in resolving mismatches in spatial and temporal resolution between GRACE/-FO observations and high-resolution models, and there is no consensus yet on the optimal approach for assimilating GRACE/-FO data. In light of the upcoming launches of next-generation gravity missions and the development of increasingly sophisticated Earth system modeling frameworks, it is an opportune time to compile the recommendations of GRACE/-FO data assimilation studies to date, in an attempt to converge to best practices. This review synthesizes past achievements, critically examines unresolved challenges, and explores future directions for advancing water cycle reanalyses using satellite gravimetry observations through improved assimilation strategies, machine learning, and near-real-time intake of satellite data.