High-resolution terrestrial water storage dynamics in Central Asia: Evaluating hydrological forcing datasets for GRACE downscaling

Abstract. The Gravity Recovery and Climate Experiment (GRACE) and GRACE Follow-On (GRACE-FO) missions provide highly valuable large-scale observations of terrestrial water storage (TWS), but their coarse spatial (~200 km) and temporal (~monthly) resolutions limit their direct use in regional applications. In this study, we implement and refine a three-step downscaling framework to downscale GRACE-based TWS changes (TWSCs) to daily, 1 km resolution over the Naryn – Kara Darya basins and Fergana valley in Central Asia by integrating GRACE data with high-resolution hydrological forcing datasets, including precipitation, evapotranspiration, and runoff from Global Land Data Assimilation System (GLDAS), Famine Early Warning Systems Network Land Data Assimilation System Central Asia (FLDAS-CA), the land component of the Fifth Generation European ReAnalysis (ERA5-Land), and a mixed combination (Mix) comprising the Multi-Source Weighted-Ensemble Precipitation (MSWEP), the Global Land Evaporation Amsterdam Model (GLEAM), and the Global Flood Awareness System (GloFAS). Temporal downscaling corrects daily water-balance-derived storage changes using spline interpolation constrained by monthly GRACE observations. Spatial downscaling employs a Partial Least Squares regression to map coarse GRACE anomalies onto fine-scale predictors, and a post-bias correction ensures consistency with the original GRACE signal. Given the lack of in situ data in the study region, we implement three validation strategies: comparison with the ITSG-Grace2018 daily solution, an “upscaling-back” consistency test, and event-based analysis. The results show that all forcing scenarios capture the broad seasonal and interannual variability of GRACE, but their performance differs substantially. GLDAS retains grid-like artefacts, FLDAS-CA systematically underestimates long-term declines and seasonal amplitudes, and ERA5-Land introduces high-frequency noise in the daily TWSCs. In contrast, the Mix forcing achieves the best overall performance, yielding the highest correlation coefficients (up to 0.8) with ITSG-Grace2018, the most satisfactory Nash-Sutcliffe Efficiency (NSE) distribution (mean = 0.65) relative to GRACE signals in the upscaling-back test, and a realistic negative long-term trend (−5.7 mm yr⁻¹) compared with −8.3 mm yr⁻¹ from GRACE. The Mix-based downscaled product also captures short-term hydrological events, such as the significant January 2006 snow event, and human-induced impacts associated with potential return flows from surface water irrigation. These results highlight the importance of carefully selecting input hydrological datasets in downscaling applications. Additionally, the presented framework is computationally flexible and transferable, allowing specification of target resolutions and adaptation of input hydrological datasets accordingly for applications in other regions.