Abstract
This two-part study introduces a novel latent space data assimilation (LSDA) framework comprised of an Autoencoder and an observation-to-latent mapping model, referred to as the AE-O2L network. This network allows observation-only analysis (LSDA-OOA) as demonstrated in Part I. The present work (Part II) extends AE-O2L to incorporate background fields into the data assimilation together with observations, referred to as observation and background assimilation (LSDA-OBA). As in Part I, the 2m temperature (T2) of a 1km-grid NWP system over a complex surface in eastern China is used to train and test the AE-O2L-based LSDA-OBA framework.
The result shows that assimilating backgrounds through the latent space improves LSDA performance. LSDA-OBA also outperforms the variational LSDA method (LSDA-Var), especially when observations are sparse. By assimilating 40 real observations, LSDA-OBA achieves analyses of 933 test cases with an MAE of 0.72 K as verified against the 7 data-withheld stations versus 0.76 K for LSDA-Var. Furthermore, LSDA-OBA runs two orders of magnitude faster than LSDA-Var.
Sensitivity experiments show that the increment of each element of the latent vector corresponds to a mode perturbation in the NWP model space, and this relationship is roughly linear. We also demonstrate that the space spanned by these modes approximates the decoding space of the Autoencoder. When performing an LSDA process, different modes are activated for different weather scenarios. Furthermore, the accumulated effect of the most active modes can approximate the final analysis with proper structures and intensity, which explains how LSDA works with such a small latent space.
© 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).
