Combining Physically Based Modeling and Deep Learning for Fusing GRACE Satellite Data: Can We Learn From Mismatch?

Abstract

Global hydrological and land surface models are increasingly used for tracking terrestrial total water storage (TWS) dynamics, but the utility of existing models is hampered by conceptual and/or data uncertainties related to various underrepresented and unrepresented processes, such as groundwater storage. The gravity recovery and climate experiment (GRACE) satellite mission provided a valuable independent data source for tracking TWS at regional and continental scales. Strong interests exist in fusing GRACE data into global hydrological models to improve their predictive performance. Here we develop and apply deep convolutional neural network (CNN) models to learn the spatiotemporal patterns of mismatch between TWS anomalies (TWSA) derived from GRACE and those simulated by NOAH, a widely used land surface model. Once trained, our CNN models can be used to correct the NOAH-simulated TWSA without requiring GRACE data, potentially filling the data gap between GRACE and its follow-on mission, GRACE-FO. Our methodology is demonstrated over India, which has experienced significant groundwater depletion in recent decades that is nevertheless not being captured by the NOAH model. Results show that the CNN models significantly improve the match with GRACE TWSA, achieving a country-average correlation coefficient of 0.94 and Nash-Sutcliff efficient of 0.87, or 14% and 52% improvement, respectively, over the original NOAH TWSA. At the local scale, the learned mismatch pattern correlates well with the observed in situ groundwater storage anomaly data for most parts of India, suggesting that deep learning models effectively compensate for the missing groundwater component in NOAH for this study region.Plain Language Summary Global hydrological models are increasingly being used to assess water availability and sea level rise. Deficiencies in the conceptualization and parameterization in these models may introduce significant uncertainty in model predictions. GRACE satellite senses total water storage at the regional/continental scales. In this study, we applied deep learning to learn the spatial and temporal patterns of mismatch or residual between model simulation and GRACE observations. This hybrid learning approach leverages strengths of data science and hypothesis-driven physical modeling. We show, through three different types of convolution neural network-based deep learning models, that deep learning is a viable approach for improving model-GRACE match. The method can also be used to fill in data gaps between GRACE missions. Key Points: • Strong interests exist in fusing GRACE satellite TWS data into global hydrological models to improve their predictive performance • We train CNN deep learning models to learn the mismatch between TWS anomalies simulated by a land surface model and that observed by GRACE • Results show deep learning models significantly improved the predictive skills of land surface model by compensating for missing components Supporting Information: • Supporting Information S1

Multi-decadal hydrologic change and variability in the Amazon River basin: understanding terrestrial water storage variations and drought characteristics

“…S3); however, this is a general pattern observed in many hydrological models as reported in a recent study (Scanlon et al, 2018). Scanlon et al (2018) indicated a low correlation between GRACE and models, which they attributed to the (i) lack of surface water and groundwater storage components in most of the models, (ii) uncertainty in climate forcing, and (iii) poor representation of human intervention in the models (Scanlon et al, 2018;Sun et al, 2019). Here, we shed more light on the disagreement issue by investigating the contributions from the explicitly simulated surface and sub-surface storage components and their latitudinal patterns, addressing the first concern noted above which is the most critical among the three in the Amazon because of the varying contribution of different stores across scales (Pokhrel et al, 2013).…”

Reconstruction of GRACE Data on Changes in Total Water Storage Over the Global Land Surface and 60 Basins

“…To further corroborate the learning-based models, we compare the simulated TWSA to in situ GWSAs in the Haihe River basin (including the major portion of the North China Plain, NCP) ( Figure S11). We assigned groundwater wells to the nearest grid cells and then calculated CC values between the modeled TWSA and in situ GWSA, referring to the study in Sun et al (2019). Results can be found in Figures S12-S14 in the supporting information.…”

Groundwater Storage Change in the Jinsha River Basin from GRACE, Hydrologic Models, and In Situ Data

“…The successful launch of the GRACE satellites has provided considerable support for monitoring GWSC, but many limitations still need to be overcome. The most serious shortage is the low temporalspatial resolution of the GRACE mission (Rodell et al 2018), revealing that (1) the GRACE data have northsouth stripe noise (Wahr et al 1998), which requires filter and signal attenuation (Landerer and Swenson 2012); (2) LWSC from GRACE data are the total water storage change, including snow, surface water, soil moisture, and groundwater storage, which cannot be isolated using GRACE data alone (Sun et al, 2019). There are errors in the background model and hydrological data, which can also affect the results of GWSC.…”