Currently, wind farms typically rely on greedy control, in which the individual turbine's structural loading and power are optimized. However, this often appears suboptimal for the whole wind farm. A promising solution is closed-loop wind farm control using state feedback algorithms employing a dynamic model of the flow. This control method is a novelty in wind farms, and has potential to provide a temporally optimal control policy accounting for time-varying inflow conditions and unmodeled dynamics, both often neglected in current methods. An essential building block for state feedback control is a state estimator (observer) that reconstructs the system states for the dynamic model using a small number of measurements. As computational efficiency is critical in real-time control, lower-fidelity models are proposed to be used. In this work, WindFarmObserver (WFObs) is introduced, which is a state estimator relying on the WindFarmSimulator (WFSim) model and an Ensemble Kalman Filter (EnKF). The states of WFSim form the two-dimensional flow field in a wind farm at hub height. WFObs is tested in a two-turbine setup using a high-fidelity simulation model. With a realistic sensor setup where only 1.1% of the to-be-estimated states are measured, WFObs reduces the RMS error by 21% compared to open-loop simulation of WFSim, at a low computational cost of 0.76 s per timestep, a factor 102 faster than the common Extended Kalman Filter.