A hot topic in ice sheet modelling is to run prognostic simulations over the next 100 years to investigate the impact of Antarctica and Greenland ice sheets on sea level change. Such simulations require an initial state of ice sheets which must be as close as possible to what is currently observed. Large scale ice sheet dynamical models are mostly governed by the following input parameters and variables: basal dragging coefficient, bedrock topography, surface elevation, temperature field. But we do not have satisfying initial states for simulations. Fortunately, some observations are available such as surface and (sparse) bedrock topography, surface velocities, surface elevation trend. The use of advanced inverse methods appears to be the adequate tool to produce satisfying initial states. We develop ensemble methods based on Ensemble Kalman filter to infer optimal initial states for ice sheet model initialisation thanks to available observations. As we first want to assess the validity of the method we begin with twin experiments with a simple flow-line large scale model, Winnie, as a first step toward data assimilation for a full 3D ice sheet model, GRISLI. Despite its simplicity, Winnie flow line model is strongly non-linear and is a good prototype to validate our methods. We also run several diagnostics to assess the quality of the recovered parameters.