Research on Atlantic surface pCO2 reconstruction based on machine learning

Jiaming Liu, Jie Wang, Xun Wang, Yixuan Zhou, Runbin Hu, Haiyang Zhang

TL;DR

This study leverages machine learning, specifically the XGBoost model, to create a highly accurate reconstruction of sea surface pCO2 in the Atlantic Ocean. By integrating satellite data with key environmental variables identified through spatial analysis, the model achieves an impressive 94% accuracy, offering a crucial tool for understanding ocean acidification and its impacts.

What Problem Does This Paper Address?

Ocean acidification, driven by increased atmospheric CO2 absorption, poses a significant threat to marine ecosystems and global climate. Accurately measuring and predicting sea surface partial pressure of carbon dioxide (pCO2) is vital for monitoring this process. However, traditional ship-based measurements are sparse and lack the spatial resolution needed for comprehensive analysis. Existing global pCO2 datasets have limitations, and understanding the complex interplay of environmental factors influencing sea surface pCO2 is challenging. This paper aims to address these issues by developing a robust and accurate method for reconstructing Atlantic sea surface pCO2, providing a more detailed picture of this critical oceanographic parameter.

Methods

The researchers employed a multi-pronged approach:

  • Geographical Detector: This spatial statistical method was used to quantitatively assess the influence of various environmental factors (longitude, latitude, wind speed, precipitation, evaporation, wave height) on sea surface pCO2. This helped identify the most significant predictors for the reconstruction models.
  • Machine Learning Models: Six different machine learning algorithms were tested: Convolutional Neural Network (CNN), Back Propagation Neural Network (BP), Long Short-Term Memory Network (LSTM), Extreme Learning Machine (ELM), Support Vector Regression (SVR), and Extreme Gradient Boosting Tree (XGBoost). These models were trained and validated to reconstruct monthly sea surface pCO2 data.
  • Data Integration: The study combined data from two major global pCO2 products (Copernicus Marine Service and Chinese Academy of Sciences Ocean Science Research Center) with environmental data.
  • Stability Testing: The robustness of the best-performing model was evaluated by introducing artificial uncertainties into the input variables to assess its sensitivity to data errors.

Datasets

  • SOCAT Cruise Survey Data: Used as the ground truth for validating the reconstructed pCO2 data.
  • Copernicus Marine Service pCO2 Data: A global sea surface pCO2 product used as input for the reconstruction models.
  • Chinese Academy of Sciences Ocean Science Research Center (CODC-GOSD) pCO2 Data: Another global sea surface pCO2 product used as input.
  • ERA5 Reanalysis Data (ECMWF): Provided environmental variables including:
  • Sea surface 10 m wind speed (U10)
    • Total precipitation (TP)
    • Evaporation (E)
    • Significant height of combined wind waves and swell (SHWW)

Key Insights and Ideas

  • Environmental Factors Matter: The geographical detector revealed that while satellite-derived pCO2 products (Copernicus and CODC-GOSD) are strong predictors, environmental factors like total precipitation (TP), latitude, and longitude also play significant roles in sea surface pCO2 reconstruction.
  • XGBoost Dominance: Among the tested machine learning models, XGBoost consistently outperformed others, achieving the highest accuracy and lowest error rates, especially when combining multiple environmental variables.
  • Synergy is Key: Combining the Copernicus and CODC-GOSD pCO2 products with key environmental variables (TP, U10, SHWW) proved to be the most effective strategy for high-precision reconstruction.
  • Robustness to Uncertainty: The XGBoost model demonstrated remarkable stability, showing low sensitivity to uncertainties in its input variables, making it reliable even with noisy or incomplete data.
  • Localized Accuracy: The model's high performance extended to regions with complex environmental conditions and data anomalies, suggesting its broad applicability.

Key Results

  • Overall Accuracy: The XGBoost model achieved an impressive overall reconstruction accuracy of 94% for the Atlantic Ocean, significantly outperforming the standalone use of Copernicus or CODC-GOSD products.
  • Regional Performance: In regions with numerous outliers, the XGBoost model maintained a reconstruction accuracy of ≥95%.
  • Environmental Variable Importance: The geographical detector identified Copernicus pCO2 and CODC-GOSD pCO2 as the most influential factors (~0.72 Q value), followed by TP, latitude, longitude, SHWW, U10, and E. Evaporation (E) showed a very low contribution and was excluded from the final model.
  • Model Stability: When input variables were altered by ±20%, the XGBoost model's R² remained high (0.94), and RMSD values stayed within acceptable limits, demonstrating its robustness.

Limitations and Future Work

  • Data Deficiency in Polar Regions: The authors acknowledge the challenge of data scarcity in polar marine regions and suggest that future high-precision remote sensing data and continued empirical data collection will be crucial for improving global models.
  • Model Combination: While multiple models were tested, the study did not explore combining different machine learning models to see if that would further enhance reconstruction accuracy. This is suggested as a future research direction.
  • Extending to Other Oceans: The success in the Atlantic Ocean suggests potential for extending this methodology to reconstruct sea surface pCO2 in other oceans, which would facilitate the creation of continuous, large-scale data products for broader scientific research.
  • Temporal and Spatial Dimensions: While the study achieved good results, the authors note that the accuracy of such models relies on sufficient temporal and spatial data dimensions, implying that further data collection and refinement are always beneficial.

Summary updated

Abstract

Ocean acidification is transforming marine ecosystems at an unprecedented rate, which in turn requires the estimation of sea surface carbon dioxide partial pressure (pCO 2 ) as a crucial metric to gauge acidification. This has substantial implications for marine resource assessment and management, marine ecosystems, and global climate change research. This study utilizes SOCAT cruise survey data to assess the accuracy of global sea surface pCO 2 products offered by Copernicus Marine Service and the Chinese Academy of Sciences Ocean Science Research Center. Through the application of a geographic information analysis method—geographical detector—the study quantitatively reveals the significance of environmental influencing factors, such as longitude, latitude, sea surface 10 m wind speed (U 10 ), total precipitation (TP), evaporation (E), and significant height of combined wind waves and swell (SHWW), in the reconstruction of sea surface pCO 2 . Subsequently, various machine learning models, which include convolutional neural network (CNN), back propagation neural network (BP), long short-term memory network (LSTM), extreme learning machine (ELM), support vector regression (SVR), and extreme gradient boosting tree (XGBoost), are used to reconstruct the monthly sea surface pCO 2 data for the Atlantic Ocean from 2001 to 2020 to investigate the potential and suitability of high-precision reconstruction of the sea surface pCO 2 dataset for this sea area. The findings indicate that: (1) The geographical detector effectively quantifies the contribution of various environmental factors used in sea surface pCO 2 reconstruction. Notably, the Copernicus pCO 2 and CODC-GOSD pCO 2 contribute the most, with both contributing ~0.72. These are followed by TP, latitude, longitude, SHWW, U 10 , and E. (2) After comprehensive data testing, the six machine learning models select the optimal hyperparameters for reconstruction. Among these, the XGBoost model notably improved the quality of the original dataset when using Copernicus pCO 2 and CODC-GOSD pCO 2 products in conjunction with SHWW, U 10 , and TP environmental variable data. Compared with SOCAT data, the overall reconstruction accuracy in the Atlantic Ocean reached an impressive 94 %, outperforming the standalone use of either Copernicus pCO 2 or CODC-GOSD pCO 2 products. Furthermore, the XGBoost model demonstrated strong applicability in regions with numerous outliers, maintaining a reconstruction accuracy of ≥95 %. (3) Stability test results reveal that the XGBoost model exhibits low sensitivity to uncertainties in all input variables. This indicates that the model can accommodate environmental data errors induced by abrupt changes in marine environments. Such robustness enhances its reliability in sea surface pCO 2 reconstruction. The reconstruction of the Atlantic sea surface pCO 2 is conducive to the assessment of global ocean acidification and provides a theoretical basis for the sustainable development of the marine environment. • The XGBoost machine learning model has been refined, leading to a superior prediction of atmospheric carbon dioxide partial pressure over the Atlantic Ocean. • We introduced the GeoDetector, a spatial statistical analysis method, to quantitatively assess the influence of various factors on the atmospheric carbon dioxide partial pressure over the ocean surface. • Comprehensive validation ensures the robustness and reliability of the XGBoost model in this region. • Investigations into the atmospheric carbon dioxide partial pressure over the Atlantic Ocean offer valuable insights into global marine ecosystems.
BibTeX
@inproceedings{liu2025research,
  author = {Jiaming Liu and Jie Wang and Xun Wang and Yixuan Zhou and Runbin Hu and Haiyang Zhang},
  title = {{Research on Atlantic surface pCO2 reconstruction based on machine learning}},
  booktitle = {Ecological},
  year = {2025},
  doi = {10.1016/j.ecoinf.2025.103094},
}
Figure