Revising precipitation – water storages – vegetation signatures with GRACE-based data assimilation

Abstract

• Assessing modeled water – vegetation dynamics by observation-based data sets. • Observation-based data derived from GRACE assimilation and remote sensing. • Duration for precipitation refilling water storages is shorter modeled than observed. • Model overestimates water storage amount contributing to vegetation growth. • Insights will help modelers improve model structures; e.g. climate events like ENSO. The availability of freshwater is highly influenced by climate change, extreme climate events and by anthropogenic use. Countries where a large part of the population depends on the agricultural sector, such as South Africa, are strongly affected by changes in climate, which emphasizes that water is an essential source for food production and drinking water. To analyze changes in surface and subsurface water, model simulations and in situ data are commonly used. However, both have limitations, for example, the models rely on potentially erroneous forcing data and insufficient process representations and the in situ data do not represent the larger-scale weather and climate due to spatial and temporal heterogeneity. This can be mitigated by assimilating remote-sensed satellite data into models. In this research, we build a realistic picture of the water and its propagation (measured between peak times) from fluxes as precipitation, to its way through the storages and its impact on vegetation at the 50 km scale by using observation-based data. The observations are derived from MODIS remote-sensing and integrating GRACE total water storage anomaly (TWSA) observations into a hydrological model via data assimilation. Our objective is to identify shortcomings in model simulations by confronting them with the (synthesized) observations. Moreover, we demonstrate the importance of integrating observations into the models. We base these comparisons on signatures or sub-signals (e.g., temporal lags and annual amplitudes) that we derive via regression analysis, principal component analysis, sensitivity analysis and correlation analysis from the synthetic data and the model output. Our main results show that correlations and signatures in real observations are found weaker as compared to what is simulated in the model, e.g. for the contribution of precipitation to groundwater. Lag times between precipitation and surface and groundwater storage peaks are observed to be longer than in the model. The observed propagation of soil water from storages to vegetation is often shorter than in the model, while for groundwater it is longer. We believe our findings will be highly relevant for modelers; the gained knowledge can be used to improve models. In addition, we feel our study underlines the potential of GRACE assimilation into hydrological models.