In this paper, a framework for transferring model parameters from data-rich areas to data-sparse areas is presented. The study shows that there are two critical steps for the success of transferring model parameters in general. The first is to effectively classify basic available data into clusters, and the second is to establish nonlinear relationships between model parameters and the classified basic available data. The framework for transferring model parameters is illustrated with the b-parameter of the Variable Infiltration Capacity (VIC) land surface model. The b-parameter, which controls the shape of spatial distribution of soil moisture capacity for a study area, usually cannot be directly measured. At present, the VIC b-parameter is generally estimated through model calibration, which is not feasible for applications of regional climate and general circulation model studies because of limited observations. Therefore, to transfer the b-parameter from data-rich to data-sparse areas, an effective approach for estimating the b-parameter for the data-rich areas is necessary. In this paper, we present an alternative approach to estimate the b-parameter values using the State Soil Geographic (STATSGO) database. The methodology of transferring the VIC b-parameter from data-rich areas to data-sparse areas combines a self-organizing map neural network with a K-means clustering method to classify soil data based on soil characteristics. A supervised neural network using a Bayesian regularization method is then developed to describe the nonlinear relationships between the VIC b-parameter and the statistics of soil properties. The framework is tested using soil data from Arkansas, California, Oklahoma, and Texas with encouraging results. A case study at the Illinois River watershed near Watts is carried out to evaluate the impact of the presented parameter transferability framework on model performance. The initial result shows that the b-parameter estimated from the proposed transferability framework provides reasonable model-simulated streamflows when compared to the observations.