Morphing Ensemble Kalman Filters

A new type of ensemble filter is proposed, which combines an ensemble Kalman filter (EnKF) with the ideas of morphing and registration from image processing. This results in filters suitable for nonlinear problems whose solutions exhibit moving coherent features, such as thin interfaces in wildfire modeling. The ensemble members are represented as the composition of one common state with a spatial transformation, called registration mapping, plus a residual. A fully automatic registration method is used that requires only gridded data, so the features in the model state do not need to be identified by the user. The morphing EnKF operates on a transformed state consisting of the registration mapping and the residual. Essentially, the morphing EnKF uses intermediate states obtained by morphing instead of linear combinations of the states.

💡 Research Summary

The paper introduces a novel data‑assimilation algorithm called the Morphing Ensemble Kalman Filter (Morphing EnKF), which blends the classical Ensemble Kalman Filter (EnKF) with image‑processing concepts of registration and morphing. Traditional EnKF relies on linear combinations of ensemble members to produce an analysis state; this works well when the underlying dynamics are approximately linear or when errors are small. However, many geophysical and environmental problems involve sharp, coherent structures—thin fire fronts, atmospheric temperature gradients, oceanic fronts—that move rapidly and change shape. In such cases, the linearity assumption leads to substantial filter degradation, and the ensemble may fail to capture the true state even with a modest number of members.

The key innovation is to decompose each ensemble member (x_i) into three components: a common reference field (x_c) that represents the shared “shape” of the state, a spatial transformation (registration mapping) (T_i) that warps the reference field to align with the individual member, and a residual field (r_i) that accounts for differences after warping. Formally, \