Four-Dimensional Super-Resolution Data Assimilation (4D-SRDA) for Prediction of Three-Dimensional Quasi-Geostrophic Flows

Abstract

Super-resolution (SR) in deep learning is a technique to generate high-resolution (HR) outputs from low-resolution (LR) inputs. Recently, combining SR with data assimilation (DA) has been proposed, leading to the development of super-resolution data assimilation (SRDA). The SRDA method simultaneously performs SR and DA by inputting LR simulation results and observations into a neural network. This study develops a four-dimensional SRDA (4D-SRDA) model to predict temporal evolutions of three-dimensional quasi-geostrophic flows in a baroclinic jet system. To evaluate the performance of 4D-SRDA, we compare it with a Local Ensemble Transform Kalman Filter (LETKF), which uses an HR model. 4D-SRDA successfully reproduces both small- and large-scale structures of potential vorticity, visually similar to those produced by the LETKF. We compare grid-wise and pattern-similarity errors to quantify the accuracy of the analysis and forecast states. Despite using an LR fluid model, 4D-SRDA achieves accuracy comparable to that of the LETKF. Comparing the computational time required for prediction reveals that 4D-SRDA is substantially more efficient than the LETKF. These results suggest that 4D-SRDA is a promising approach for predicting HR atmospheric flows.