Ensemble-based data assimilation techniques are being explored as possible alternatives to current operational analysis techniques such as 3-or 4-dimensional variational assimilation. Ensemblebased assimilation techniques utilize an ensemble of parallel data assimilation and forecast cycles. The background-error covariances are estimated using the forecast ensemble and are used to produce an ensemble of analyses. The background-error covariances are flow dependent and often have very complicated structure, providing a very different adjustment to the observations than are seen from methods such as 3-dimensional variational assimilation. Though computationally expensive, ensemble-based techniques are relatively easy to code, since no adjoint nor tangent-linear models are required, and previous tests in simple models suggest that dramatic improvements over existing operational methods may be possible.A review of the ensemble-based assimilation is provided here, starting from the basic concepts of Bayesian assimilation. Without some simplification, full Bayesian assimilation is computationally impossible for model states of large dimension. Assuming normality of error statistics and linearity of error growth, the state and its error covariance may be predicted optimally using Kalman filter (KF) techniques. The ensemble Kalman filter (EnKF) is then described. The EnKF is an approximation to the KF in that background-error covariances are estimated from a finite ensemble of forecasts. However, no assumptions about linearity of error growth are made.Recent algorithmic variants on the standard EnKF are also described, as well as methods for simplifying the computations and increasing the accuracy. Examples of ensemble-based assimilations are provided in simple and more realistic dynamical systems.2