Conventional techniques for resource estimation generally involve the laborious analysis of variograms and the specification of parameters for covariance functions. In large deposits, construction of resource models and their update require significant amount of manual work, expert knowledge and time. In this paper we take a machine learning view to the problem of resource estimation. This is a non-parametric Bayesian method, leading to automatic parameter estimation (learning). Through Bayesian learning, the parameters of the model can be obtained by optimising a quantity known as the marginal likelihood. The main advantage of this approach over conventional maximum likelihood estimators is that the objective function incorporates the Occam's razor principle. This allows the comparison of different statistical models - for example, models designed with different covariance functions - and naturally penalises complexity, thus avoiding overfitting. Our technique presents significant flexibility for the definition of mean and covariance functions and permits the usage of more sophisticated models.The second contribution of this work is the formulation of a data fusion mechanism that continuously updates the resource model. As mining progresses, more information becomes available. For example, the hardness of the rocks can be measured while drilling blast holes, grades of ore and contaminants are obtained from cone logging. This information is in general available at a finer spatial resolution but can be more uncertain. The data fusion mechanism can seamlessly estimate the uncertainties of the different types of information and fuse them as new measurements are acquired. This procedure can significantly improve the quality of the resource model, defining new grounds for stochastic mine planning. We demonstrate the effectiveness of our approach in estimating iron ore deposits by integrating information about different grades.