Data assimilation or filtering of nonlinear dynamical systems combines numerical models and observational data to provide the best statistical estimates of the systems. Ensemble-based methods have proved to be indispensable filtering tools in atmosphere and ocean systems that are typically high dimensional turbulent systems. In operational applications, due to the limited computing power in solving the high dimensional systems, it is desirable to use cheap and robust reduced-order forecast models to increase the number of ensemble for accuracy and reliability. This talk describes a multiscale data assimilation framework to incorporate reduced-order multiscale forecast methods for filtering high dimensional complex systems. A reduced-order model for two-layer quasi-geostrophic equations, which uses stochastic modeling for unresolved scales, will be discussed and applied for filtering to capture important features of geophysical flows such as zonal jets. If time permits, a generalization of the ensemble-based methods, multiscale clustered particle filters, will be discussed, which can capture strongly non-Gaussian statistics using relatively few particles.