Predicting random fields with increasing dense observations