Skills in sub-seasonal to seasonal terrestrial water storage forecasting: insights from the FEWS NET land data assimilation system

Abstract. Accurate prediction of terrestrial water storage (TWS), the sum of soil moisture, groundwater, snow/ice, and surface water, is critical for informing water resource management and disaster responses. In this study, we evaluated subseasonal to seasonal (S2S) TWS forecasts, produced by the FEWS NET land data assimilation system (FLDAS), over Africa using observations from the Gravity Recover and Climate Experiment (GRACE) and its Follow-On (GRACE/FO) mission. FLDAS consists of two advanced land surface models, Noah-MP and the NASA Catchment Land Surface Model (CLSM), both of which simulate key TWS components including groundwater. Results show that CLSM is more skillful in forecasting TWS anomalies at S2S scales than Noah-MP, with >0.6 relative operating characteristics (ROC) scores over more than half of the study domain across the 1–6 months lead times. CLSM forecasts also maintain stronger correlations with GRACE/FO data than Noah-MP, particularly at longer lead times, owing to more skillful reanalysis-based initial conditions and stronger persistence in simulated TWS. In contrast, Noah-MP forecasts show weaker skill, especially in central Africa where the skill also declines rapidly with lead time.

Evaluation results show that accuracy of TWS forecasts is strongly influenced by precipitation interannual variability: forecasts driven by precipitation products with lower precipitation interannual variability are generally more accurate than those driven by higher precipitation variability. The performance gap between Noah-MP and CLSM is also more pronounced in regions with higher precipitation variability such as central Africa. This sensitivity arises because TWS often exhibits strong multi-year variability in responses to interannual precipitation, making realistic simulation of long-term variability critical for skillful TWS forecasts. The superior performance of CLSM is attributed to its strong representation of upward groundwater movement, especially during prolonged droughts, which enhances TWS interannual variability. In contrast, the weak representation of capillary rise in Noah-MP limits its ability to capture effects of long-term precipitation variability on TWS. Both models exhibit lower correlation and higher RMSEs when evaluated against GRACE/FO data than relative to reanalysis, further underscoring substantial uncertainty in model physics.

Autocorrelation analyses show that TWS persistence is closely linked to groundwater persistence. CLSM groundwater exhibits stronger persistence than that of Noah-MP, owing to its ability to simulate groundwater responses to long-term precipitation variability. While persistence provides an important source of predictability, our results also show that inaccurate persistence, such as that associated with anthropogenically induced trends and changes in precipitation that are often inadequately captured by land surface models, can degrade forecast skill. These findings underscore the importance of using independent datasets such as GRACE/FO observations to evaluate TWS forecasts.