Attention-based Convolutional Autoencoders for 3D-Variational Data Assimilation

TL;DR

This paper introduces a convolutional autoencoder-based approach for 3D variational data assimilation that reduces computational costs significantly while maintaining or improving accuracy, demonstrated on a pollution model in London.

Contribution

The paper presents a novel 'Bi-Reduced Space' method that lowers computational complexity in 3D variational data assimilation using autoencoders, with proven equivalence to existing methods.

Findings

Reduced background covariance matrix size by approximately 1000 times

Achieved higher data assimilation accuracy compared to existing reduced space methods

Validated on real-world pollution data from London

Abstract

We propose a new 'Bi-Reduced Space' approach to solving 3D Variational Data Assimilation using Convolutional Autoencoders. We prove that our approach has the same solution as previous methods but has significantly lower computational complexity; in other words, we reduce the computational cost without affecting the data assimilation accuracy. We tested the new method with data from a real-world application: a pollution model of a site in Elephant and Castle, London and found that we could reduce the size of the background covariance matrix representation by O(10^3) and, at the same time, increase our data assimilation accuracy with respect to existing reduced space methods.