Assimilating observational data is essential for improving Arctic sea ice model prediction, yet the high-dimensional nature of such models poses challenges for applying nonlinear particle filtering methods. To address this, we propose a Latent Space Particle Filter (LSPF) approach that leverages a variational autoencoder (VAE) deep neural network to extract low-dimensional representations of sea ice physical fields. This method compresses the high-dimensional data into a latent subspace, enabling efficient statistical sampling and generating a large number of low-dimensional samples for nonlinear particle filtering. We train the VAE using multiple sea ice reanalysis datasets and conduct historical assimilation experiments using the latest ice-ocean coupled model developed by Princeton University’s Geophysical Fluid Dynamics Laboratory. Results indicate that assimilating satellite observations of sea ice concentration and thickness with LSPF during the winter freezing period significantly reduces model errors, particularly for sea ice thickness. All simulations are extended to September without additional assimilation and evaluated with independent satellite observations and mooring data. Findings further demonstrate that wintertime nonlinear particle filter assimilation can improve prediction skill, especially when performed every three days, reducing model errors by approximately 30%–50%. Therefore, the LSPF method proposed in this study provides a promising and effective solution for nonlinear data assimilation in realistic high-dimensional geoscience applications.

Data assimilation,Arctic sea ice seasonal prediction,Nonlinear particle filter,Variational autoencoder