Estimating two-decadal variations of global oceanic particulate organic carbon using satellite observations and machine learning approaches

Abstract

Particulate organic carbon (POC) is fundamental to the marine carbon cycle, yet accurately estimating its concentration from satellite data remains challenging. In this study, we developed a novel machine learning framework that incorporates multiple data streams, covering apparent and inherent optical properties, biological indicators, and environmental variables, to improve global POC retrieval. Our model achieved high accuracy, with a Spearman’s correlation coefficient of 0.92, root-mean-square error of 68.46 mg m⁻³, and median absolute percentage error of 25.01 %, outperforming conventional algorithms that rely solely on remote sensing reflectance. Applying our approach to a long-term satellite dataset (1997–2023), we identified four major seasonal variation patterns across different oceanic regions: In high-latitude regions (Type 1), POC peaks in summer due to increased light availability, while mid-latitudes (Type 2) exhibit a stable pattern with a spring peak driven by water mixing and favorable sea surface temperatures. In the Equatorial Atlantic and Indian Oceans (Type 3), a spring trough and autumn peak suggest potential significant wind-driven nutrient inputs, whereas the Equatorial Pacific (Type 4) maintains high POC levels year-round, likely influenced by persistent upwelling and nutrient dynamics. These findings highlight the advantages of integrating machine learning for improved POC estimations and provides deeper insights into global POC dynamics.