| 29 Jun 2026
Towards accurate daily 3D temperature and salinity reconstruction from remote sensing enhanced by explainable AI
Abstract. Using remote sensing data to reconstruct three-dimensional (3D) temperature and salinity, a field called Deep Ocean Remote Sensing (DORS), is essential for the study of ocean dynamics and climate change. However, existing DORS studies predominantly focus on monthly scales and lack explainability, leaving the mechanisms governing reconstruction errors poorly understood and hindering daily-scale operational applications. Here we report a transformer-based framework (i.e., EarthFormer) to reconstruct daily 3D temperature and salinity fields from multi-source remote sensing inputs, across 19 standard depth levels in the Northwest Pacific (NWP) (105–160° E, 0–40° N) at 0.25° resolution using reanalysis model product as the labeled data. Validated against Argo observations, the reconstruction achieves an root-mean-squared error of 0.893 °C (R2 = 0.989) for temperature and 0.141 PSU (R2 = 0.827) for salinity, approaching reanalysis accuracy while offering near-real-time timeliness and lightweight computation. Notably, explainable AI analysis reveals that the counter-intuitively low contribution of satellite-derived sea surface salinity (SSS) stems not from weak physical relevance but from data-quality limitations; substituting high-accuracy SSS shifts the salinity error profile from a monotonic depth decrease to a V-shaped structure, with SSS contribution rising from ~10 % to ~50 %. Overall, this study demonstrates the feasibility of explainable-AI-enhanced daily 3D thermohaline reconstruction, providing a new technical pathway for real-time ocean monitoring and underscoring that improving satellite SSS retrieval and data quality is as critical as advancing model architectures for reliable DORS applications. The data are publicly available at https://doi.org/10.5281/zenodo.20602639 (Fang, 2026).
I had the opportunity to review another paper that appears to be a “sister” paper to this one. The use of AI makes it possible to extend the production of “intermediate datasets” that, according to the authors, can be used for final products in modeling or other applications. In the case of article ESSD-2026-261, the intermediate product was a weekly dataset; here, it is a daily dataset.
Here, the different datasets are normalized and correlated. The method is not innovative, but it is well established and provides highly acceptable results. The extension of data from the surface to deeper layers has previously been carried out (e.g.,) using historical profile data, whereas here it is done using GLORYS data.
My first comment is that the resulting data are certainly smoothed, and therefore their applicability may also be limited (is this appropriate for a ‘real-time’ data set?)
My second comment concerns the use of GLORYS both as input data for the method and as validation data. I do not consider this procedure appropriate. The key issue is defining independent data.
Normally, reanalyses make use of all available data up to the time they are performed. In my opinion, the only meaningful way to verify the validity of the method is to compare the results with data collected after the reanalysis exercise. This would give me greater confidence in the quality of the methodology developed by the authors.
The article may be published after the authors have addressed my comments and following the editor’s final evaluation.