PROJECTING GRACE-DERIVED TERRESTRIAL WATER STORAGE (TWS) DATA OVER THE AFRICAN WATERSHEDS

The currently available GRACE-derived Terrestrial Water Storage (TWS) time series suffers from: (1) gaps in GRACE temporal records due to battery performance issues, and (2) relatively large latency; the available GRACE data lags month(s) behind the real time measurements. We developed and implemented a statistical approach that fills the gap in the available GRACE time series and forecasts GRACE data six months in advance. Our approach could potentially advance GRACE’s applications and benefits to the GRACE user community at large. The nonlinear autoregressive network with exogenous inputs (NARX) of the Artificial Neural Networks (ANNs) were used to derive relationships between the six month-shifted GRACE-derived TWS data (e.g., target) and the controlling factors (precipitation, temperature, evapotranspiration, and NDVI) over the nine African watersheds. The ANN technique was selected for pattern recognition for the following reasons: (1) no prior knowledge of underlying physical phenomena is required; (2) handles large databases of spatial and temporal records; (3) complex and nonlinear relationships can be derived; and (4) could be used as a forecasting tool. The ANNs models were trained, tested, and used to project the GRACE-derived TWS data over fifty runs and the average of the model runs were then used to measure the performance of the developed model. The trained ANNs model (training period: 04/2002 to 09/2014) successfully projected the GRACE-derived TWS data for the period from 10/2015 through 03/2016. The average Pearson correlation coefficient (r) throughout the testing period (10/2014 to 09/2015) was found to be 0.98, 0.96, 0.91, 0.79, 0.98, 0.97, 0.90, 0.70, and 0.70 for the Niger, Zambezi, Okavango, Limpopo, Lake Chad, Volta, East Central Coast, Nile, and Mozambique Northeast Coast basins, respectively. Ongoing research activities are concentrated on: (1) understanding the capabilities and the limitations of the advanced methodology, (2) incorporating additional relevant inputs (e.g., climatic indices), and (3) investigating the potential of the developed methodologies for forecasting drought and flooding events over the African watersheds.