Based on the "Back and Forth Nudging" (BFN) and Ensemble Kalman Filter (EnKF) methods, a Hybrid BFN EnKF (HBFNEnKF) data assimilation method was designed and tested using a channel shallow water model and a global shallow water model, separately. Furthermore, the performances of the HBFNEnKF, Hybrid Nudging EnKF (HNEnKF), and Ensemble Square-Root Filter (EnSRF) methods are discussed, with model error considered. The results showed that the HBFNEnKF method retains the continuity and smoothness of HNEnKF, avoids the discontinuity and unbalance problem of EnSRF, and has the highest convergence speed. Through a single variable observation experiment, the advantage of HBFNEnKF was clear; that is, HBFNEnKF can maintain the balance between different model variables. A scale investigation on the increment field showed that, compared with EnSRF, HBFNEnKF produces a better assimilation result at larger scales, and avoids a number of spurious increments at medium and smaller scales.