To obtain and fully use long series of high-precision observations of water balance components, it is necessary to quantify and reduce random and systematic data errors. This thesis applies and evaluates a previously developed data fusion methodology for bias-correcting and noise-filtering remote sensing observations of water balance variables that results in a consistent set of estimates that close the water balance. The method combines monthly water balance constraints and probabilistic data models for each water balance variable (precipitation, evaporation, river discharge and water storage), and uses Markov chain Monte Carlo sampling and iterative smoothing to estimate data errors and water balance variables without the need for any ground-truth. The methodology is evaluated here by application to three river basins in China located in different climate zones (humid to semi-arid). The evaluation assesses (i) how the method performs in the case study basins and to what extent data error assumptions in the probabilistic data models are satisfied, (ii) how sensitive the results are to changes in the datasets, and (iii) whether the use of dataset ensembles rather than dataset pairs (as in the original method) changes the results and further improves the overall data fit. The findings for these three research questions are as follows. First, the posterior water balance estimate for humid and arid basins with average standard error of 6-10 mm/month for precipitation, 4-6 mm/month for evaporation, 8-14 mm/month for water storage. Significant increase in both precipitation and evaporation uncertainty during wet summer period and the data error assumption is violated for precipitation in Wuding basin. Second, the results of replacing precipitation are more sensitive for Baihe humid basin, with both precipitation and evaporation increased, while for Wuding basin, replacing evaporation datasets is more sensitive.Third, significant increases in likelihood value by fusing all precipitation and evaporation datasets, with posterior uncertainty of water balance components increased. And the data error assumption is satisfied for Wuding basin.
Therefore, the use of dataset ensembles as opposed to dataset pairs is recommended in further applications of the data fusion methodology. Additional applications may focus on applying the methodology across a wider range of river basins, using a wider range of ensemble datasets (including different GRACE solutions), as well as comparison of different data error models.