On generating a flexible class of non-stationary spatial models using Gaussian predictive processes