Title:Theory of Unsupervised Super-Resolution Data Assimilation with Conditional Variational Autoencoders: Estimating Background Covariances via Super-Resolution

Abstract:This study proposes a theory of unsupervised super-resolution data assimilation (SRDA) using conditional variational autoencoders (CVAEs). The theory is based on an evidence lower bound that leads to an objective function for unsupervised learning. This function can be reduced to the objective function of a traditional data assimilation method, namely the three-dimensional variational (3D-Var) formalism. Similar to 3D-Var, it is essential to use a non-diagonal background error covariance matrix to assimilate distant observations. Instead of using such a non-diagonal matrix, we extend the proposed theory by using the non-locality of super-resolution (SR). For linear SR, SR operators serve as background error covariance matrices. For nonlinear SR, error backpropagation through SR neural networks induces covariance structures in inference. SRDA can be performed using CVAEs because the loss function for CVAEs is generally an evidence lower bound. By incorporating the SR neural network into the CVAE, the encoder estimates the high-resolution (HR) analysis from HR observations and low-resolution (LR) forecasts. The decoder acts as the observation operator and reconstructs the HR observations from the estimated HR analysis. The effectiveness of SRDA was evaluated through numerical experiments using an idealized barotropic ocean jet system. Compared to inference with an ensemble Kalman filter, SRDA demonstrated superior accuracy in HR inference. SRDA was also computationally efficient because it does not require HR numerical integration or ensemble calculations. The findings of this study provide a theoretical basis for integrating SR and DA, which will stimulate further research in this direction.