Reconstruction of dissolved oxygen in the Indian Ocean from 1980 to 2019 based on machine learning techniques

Abstract

Oceanic dissolved oxygen (DO) decline in the Indian Ocean has profound implications for Earth’s climate and human habitation in Eurasia and Africa. Owing to sparse observations, there is little research on DO variations, regional comparisons, and its relationship with marine environmental changes in the entire Indian Ocean. In this study, we applied different machine learning algorithms to fit regression models between measured DO, ocean reanalysis physical variables, and spatiotemporal variables. We utilized the Extremely Randomized Trees (ERT) model with the best performance, inputting complete reanalysis data and spatiotemporal information to reconstruct a four-dimensional DO dataset of the Indian Ocean during 1980–2019. The evaluation results showed that the ERT-based DO dataset was superior to the DO simulations in Earth System Models across different time and space. Furthermore, we assessed the spatiotemporal variations in reconstructed DO dataset. DO decline and oxygen-minimum zone (OMZ) expansion were prominent in the Arabian Sea, Bay of Bengal, and Equatorial Indian Ocean. Through correlation analysis, we found that temperature and salinity changes related to solubility primarily control the oxygen decrease in the middle and deep sea. However, the complicated factors with solubility change, vertical mixing, and circulation govern the oxygen increase in the upper and middle sea. Finally, we conducted a volume integral to estimate the oxygen content in the Indian Ocean. Overall, a deoxygenation trend of −141.5 ± 15.1 Tmol dec −1 was estimated over four decades, with a slowdown trend of −68.9 ± 31.3 Tmol dec −1 after 2000. Under global warming and climate change, OMZ expanding and deoxygenation in the Indian Ocean are gradually mitigating. This study enhances our understanding of DO dynamics of the Indian Ocean in response to deoxygenation.

Topics & Concepts