The rapid acceleration of the global water cycle caused by changes in global climate trigger complex processes that make conventional machine learning techniques limited in assessing impacts of such changes on terrestrial water storage (TWS). This study introduces an assimilated deep learning neural network to improve the modeling of TWS dynamics. Key predictors and inputs to this deep learning framework include runoff, rainfall, soil moisture, evapotranspiration, global teleconnection patterns and sea surface temperatures (SSTs). Our proposed back propagation model capitalizes on the availability of remotely sensed observations and model datasets to predict monthly TWS, a quantity that is difficult to observe in the field, but important for the estimation of regional energy balance, water resource management and agricultural purposes. By integrating pre-processed outputs from the Principal Component Analysis (PCA) and Independent Component Analysis (ICA) into our network using a deep neural pattern, we synthesized TWS from 2002-2017, and also made future predictions using our trained models. Results from this analysis showed that the ICA-BPNN model has a higher predictive accuracy compared to the PCA-BPNN. These models were used to fit the three dominant temporal patterns of Gravity Recovery and Climate Experiment (GRACE) – observed TWS over Africa. Our simulation results from the testing phase indicate that the fit for the prediction of the first three leading modes of TWS for both models when compared to the observed GRACE-TWS was found to be PCA-BPNN1 (89%), PCA-BPNN2 (82%), PCA-BPNN3 (84%) and ICA-BPNN1 (93%), ICA-BPNN2 (88%), ICA-BPNN3 (82%). The simulation fit of the BPNN corresponding to multi-annual time series, which are captured in the second and third orthogonal modes and localized patterns of TWS were lower than those of annual signals in both the PCA- and ICA-BPNN models. This was attributed to the fact that the multi-annual time series in GRACE-hydrological signals of our test-bed are complex compared to the annual patterns of TWS. On the one hand, this exemplifies the superior performance of our predictive framework in modeling naturalized system (annual changes in TWS driven by only climatic factors). On the other hand, the complexity in modelling multi-annual variations in TWS suggests heavily disturbed naturalized systems evidenced in the presence of human water management operations among other anthropogenic activities.