MEGO-C: A Novel Interpretable Machine Learning Architecture for Fusing Satellite and In Situ Data Into a Submesoscale-Resolving 3-D Chlorophyll Atlas

View at publisher Summarise with VIRAT
Authors: Yanfeng Wen, Cédric Jamet, Peng Chen, Kun� Shi, Delu Pan
Cited by: 0
References: 61

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

Observing the ocean's biological carbon pump at its critical meso- and submesoscales remains a significant challenge, leading to major uncertainties in global carbon cycle estimates. This challenge stems from a fundamental data disparity: while satellite ocean color sensors provide unparalleled global coverage of the sea surface, they cannot penetrate the ocean's interior. Conversely, in-situ profilers like BGC-Argo measure the vertical dimension but are too sparse to resolve fine-scale horizontal variability. Existing 3D chlorophyll-a(Chl-a) products, whether based on statistical models or numerical reanalysis, are often too coarse to resolve these key dynamics. To address this long-standing scale-mismatch problem, we introduce MEGO-C, a novel machine learning architecture specifically engineered for Earth science data fusion. It employs an interpretable, heterogeneous stacked ensemble to synergistically fuse 144,936 BGC-Argo profiles with a suite of multi-sensor satellite observations. This framework produces the first-ever global, daily, submesoscale-resolving (~4 km) 3D Chl-aatlas. Validation demonstrates its breakthrough performance, with prediction errors less than half of those from benchmark CMEMS products and an order-of-magnitude finer spatial resolution (~4 km vs. ~25 km). Crucially, our eddy case studies based on this high-resolution dataset show that coarse-resolution products can substantially underestimate area-integrated Chl-a (ΣChl·dA; biomass proxy) within Chl-a–based eddy-like coherent features by ~25% and miss peak concentrations by over 50%. This finding suggests the biogeochemical impact of eddies is substantially undervalued in global assessments, a result with important implications for revising carbon cycle estimates. SHAP analysis confirms the model’s physical consistency, moving beyond a "black-box" approach, while scale-selective filtering verifies its ability to resolve submesoscale filaments and sharpen eddy variability. This work provides a crucial new observational benchmark for validating the biogeochemical components of Earth system models and presents a globally applicable, effective, and transparent approach for synergistic data fusion in remote sensing and oceanography.

Fields of Study & Subjects

Marine Plankton and Biogeochemistry Oceanographic and Atmospheric Processes Aquatic Ecosystems and Phytoplankton Dynamics

Publication details

Journal / source
IEEE Transactions on Geoscience and Remote Sensing
Volume
64
Pages
1–23
Publication date
1 January 2026
Language
EN