Uncertainty in the initial conditions is one of the main sources of uncertainty in numerical weather prediction. Operational forecast Centres represent uncertainty in the initial conditions by perturbing the best estimation of the initial state or analysis. The nonlinear dynamics of numerical models makes perturbations grow exponentially in time and localize in space. There is not a unique method to perturb the initial conditions and the spatio-temporal growth of the perturbations differs from one technique to the other. The characterization of perturbations growth contributes to determine the predictability barrier of the system. Different initial perturbation techniques (singular vectors, breeding vectors and random perturbations) have been applied to the same numerical model, the ECMWF EPS. A comparison of the spatio-temporal dynamics of the initial perturbations will be provided by using theoretical results obtained in non-equilibrium surface growth. The recently introduced MVL diagram is found to distinguish the linear and non linear regimes, outstanding the differences in the spatio-temporal growth of the perturbations. Northern hemisphere, Southern hemisphere and tropics will be analyzed separately.