23 September 2023 in Journal of Geophysical Research: Oceans
22 May 2025 in Journal of Geophysical Research: Atmospheres
16 February 2024 in Journal of Geophysical Research: Atmospheres
24 December 2024 in Journal of Geophysical Research: Atmospheres
1 May 2026 in Journal of Atmospheric and Oceanic Technology
In this study, we developed a U-Net-based mixed layer depth (MLD) inversion method to generate daily 0.08° MLD distributions in the northwest Pacific, inverted from satellite-observed sea surface temperature (SST), sea surface height (SSH), and wind fields. The model is trained on MLD datasets diagnosed from HYCOM reanalysis (2016–18 and 2020–22) and evaluated in 2019 and 2023. Inverted MLD performs well in the northwest Pacific tropical cyclone formation region (TCFR), with overall root-mean-square errors (RMSEs) (approximately 25% of the typical regional MLD) of 10.27 m (2019) and 11.64 m (2023). It captures MLD deepening after tropical cyclones, reducing absolute errors against climatology by over 40% during tropical cyclone Khanun (2023). The impacts of the training data properties were thoroughly checked based on the proposed U-Net inversion method. Experiments show the following three results: 1) Moderately reducing the resolution significantly reduces the training duration while preserving pixelwise metric-evaluated inversion accuracy. 2) Retaining data for summer improves training stability and inversion performance compared to full-year data training. 3) Strengthening physical constraints between input and output provides the most comprehensive training and inversion improvements. Our study demonstrates the potential of deep learning for ocean MLD inversion and offers guidance on training data selection and processing for future applications.
28 July 2026 in Journal of Advances in Modeling Earth Systems
Two of the challenges for paleoclimate data assimilation (PDA) are the extreme sparsity of proxy observations and the non‐Gaussian nature of background errors in the climate system. In this study, we develop a latent‐space variational data assimilation framework based on variational autoencoders (VAEs) and latent‐space minimization (VAE‐Var), and apply it to assimilating sparse air and sea temperature proxies. Employing a VAE‐based encoder–decoder to learn a data‐driven latent prior that renders non‐Gaussian background errors, VAE‐Var addresses the challenge of non‐Gaussian background errors through an adjustable Gaussian constraint, relaxing the Gaussian‐error assumption underpinning traditional variational schemes. Using surface air temperature and sea surface temperature as targets, we systematically evaluate “reconstruction” skills across varying levels of observational sparsity, comparing VAE‐Var with conventional approaches. Diagnostics of analysis increments demonstrate that VAE‐Var yields improved “reconstruction” skill and more physically plausible analyses under sparse observations, highlighting its merits in non‐Gaussian PDA for high‐dimensional reconstruction applications.