Abstract
Data assimilation (DA) aims to achieve consistent atmospheric analyses with observations and numerical model forecasts. However, the increasing trend in forecast resolution and observation richness places an increasing computational burden on DA. To address this challenge, we develop a novel latent space data assimilation (LSDA) framework that performs efficient DA in a reduced-dimensional latent space learned by an Autoencoder from numerical atmospheric states. Distinct from previously reported LSDA methods, our approach introduces an extra neural network, O2Lnet, trained on simulated observations derived from model states, to map real observations onto the AE latent space. The observation-derived latent state obtained by O2Lnet can then be directly decoded to obtain the analysis in model space using the decoder component of the Autoencoder. In Part I, we aim to demonstrate the feasibility of this observation-only analysis method, denoted as LSDA-OOA, by inferring 2-meter temperature (T2) analyses on 1km-grids with both idealized and real T2 observations.
The idealized experiments demonstrate that given sufficient observations, LSDA-OOA can yield high-quality analyses while exhibiting a favorable resiliency to random observation errors. When applied to analyze real T2 observations, LSDA-OOA produced T2 analyses with an accuracy comparable to the WRF (Weather Research and Forecast) four-dimensional data assimilation (FDDA) method. In particular, it greatly outperforms WRF-FDDA for the cases containing larger errors in forecasts (background fields). Finally, we replace the training data from WRF-FDDA analyses with the forecasts instead and find that this only results in a small increase of the error in the LSDA-OOA analyses.
© 2025 American Meteorological Society. This is an Author Accepted Manuscript distributed under the terms of the default AMS reuse license. For information regarding reuse and general copyright information, consult the AMS Copyright Policy (www.ametsoc.org/PUBSReuseLicenses).
