Probabilistic risk assessment of groundwater contamination requires us to incorporate large and diverse datasets at the site into the stochastic modeling of flow and transport for prediction. In quantifying the uncertainty in our predictions, we must not only combine the best estimates of the parameters based on each dataset, but also integrate the uncertainty associated with each dataset caused by measurement errors and limited number of measurements. This dissertation presents a Bayesian geostatistical data assimilation method that integrates various types of field data for characterizing heterogeneous hydrological properties. It quantifies the parameter uncertainty as a posterior distribution conditioned on all the datasets, which can be directly used in stochastic simulations to compute possible outcomes of flow and transport processes. The goal of this framework is to remove the discontinuity between data analysis and prediction. Such a direct connection between data and prediction also makes it possible to evaluate the worth of each dataset or combined worth of multiple datasets. The synthetic studies described here confirm that the data assimilation method introduced in this dissertation successfully captures the true parameter values and predicted values within the posterior distribution. The shape of the inferred posterior distributions from the method indicates the importance of estimating the entire distribution in fully accounting for parameter uncertainty. The method is then applied to integrate multiple types of datasets at the Hanford 300 Area for characterizing a three-dimensional heterogeneous hydraulic conductivity field. Comparing the results based on the different numbers or combinations of datasets shows that increasing data do not always contribute in a straightforward way to improving the posterior distribution: increasing numbers of the same data type would not necessarily be beneficial above a certain number, and also the combined effect of multiple datasets could be very different from the individual effects. This research thus shows that any site characterization should consider not only the worth of individual datasets, but also the combined effects of multiple datasets and the sequence in which they are assimilated.