This study develops a U-Net-based ocean Mixed Layer Depth (MLD) inversion method to generate daily, high-resolution(0.08°) MLD distributions in the Northwest Pacific by integrating satellite-observed Sea Surface Temperature (SST), Sea Surface Height (SSH), and wind fields. The U-Net model was trained on MLD datasets diagnosed from HYCOM reanalysis (covering 2016-2018 and 2020-2022), and evaluated against data from 2019 and 2023.
The inversion method shows high performance within the Northwest Pacific Tropical Cyclone Formation Region (TCFR), with overall RMSEs of 10.27 m (2019) and 11.64 m (2023), which represent approximately 25% of the typical regional MLD. Crucially, the inversion method leverages the neural network’s non-linear fitting capability to capture short-term, non-periodic signals, such as unseasonal deepening of ocean mixed layer caused by Tropical Cyclones (TCs), which cannot be represented by 365-day climatological mean states. For instance, during TC Khanun (2023), the inverted MLD showed deepening along the TC track, reducing absolute errors against climatology by over 40% within cumulative TC impact area.
A central contribution of this work is a systematic evaluation of training data properties on inversion accuracy based on the proposed U-Net inversion method. Comparative experiments were designed to study the impacts of resolution, time domain and physical constraint between input variables and the MLD. Results reveal the following three primary conclusions.
1) Moderately reducing the resolution significantly reduces training time while slightly improving pixel-wise inversion accuracy compared to high-resolution training. Downscaling the target MLD to match input observational resolution allows for the deployment of larger models within the same computational limits, though inversion accuracy does not consistently improve with model scales increasing.
2) Prioritizing reducing within-year temporal range over decreasing total annual coverage improves training stability and inversion performance for specific time windows. The model trained solely on multi-year summer data can have better summer inversion results than models trained on year-round data, which is a very valuable hint for tasks that only need MLD to provide underlying support for atmospheric predictions during the typhoon season.
3) Strengthening physical constraints between input and output datasets by replacing observational inputs with data from a homologous output source provides the most comprehensive improvements. Furthermore, the study highlights the spatiotemporal heterogeneity of performance gains when using a numerical model-based training strategy compared to the reanalysis-based one. While requiring a high level of configuration and tuning, numerical model-based training offers a robust alternative, as the inversion model is trained on consistent, long-term datasets generated without complex assimilation systems.
Ultimately, this research conducted high-resolution daily MLD inversion and the obtained MLD datasets can potentially be used for constraining sea-air coupled operational forecasting. It demonstrates the capability of machine learning for ocean state reconstruction and provides strategic guidance for data selection and processing in future oceanic applications.

Xinyi LI, Xiaolin YU, and Xin LIU, 2026, Impacts of Training Data Properties on U-Net Based Ocean Mixed Layer Depth Inversion in Northwest Pacific, Journal of Atmospheric and Oceanic Technology, in press

Mixed layer depth,U-Net,Neural network,NN training