Traffic managers need tools capable of providing online traffic flow monitoring and short-term predictions on large-scale networks. Data Assimilation (DA) techniques provide traffic state estimates based on a dynamic model and an observation model. Their performance are very sensitive to the observation model definition. The paper proposes a solution to optimize the number and the location of loop sensors in the context of DA. The optimality is defined based on travel time and travelled distance indicators, and Pareto efficiency is used to determine the optimal sensor allocation. The method is tested on two toy networks to assess the role and level of contribution of each sensor. The results provide a better understanding for deploying additional sensors or selecting sensors to be removed, for optimal traffic state estimation and traffic forecast.