Revealing 100 years of global ocean deoxygenation, Shanghai Jiao Tong University uses artificial intelligence to reconstruct the 'suffocating ocean', ICML has included

Author | Lu Bin, Han Luyu

Marine dissolved oxygen is a key factor in maintaining the function of marine ecosystems. With the impact of global warming and human activities, the ocean has shown a trend of deoxygenation in recent years. The increasingly suffocating ocean has serious consequences for fishery development, climate regulation and other aspects.

Recently, the team of Professors Wang Xinbing and Gan Xiaoying from the School of Electronic Information and Electrical Engineering of Shanghai Jiao Tong University, together with Academician Zhang Jing, Professor Zhou Lei and Associate Professor Zhou Yuntao from the School of Oceanography of Shanghai Jiao Tong University, jointly proposed a sparse ocean observation data driver The deep learning model OxyGenerator. For the first time, the century-old global ocean dissolved oxygen data from 1920 to 2023 was reconstructed, and the reconstruction performance significantly exceeded the results of the CMIP6 series of numerical models dominated by expert experience.

The research result "OxyGenerator: Reconstructing Global Ocean Deoxygenation Over a Century with Deep Learning" has been selected by the China Computer Society Class A Conference International Conference on Machine Learning (ICML)Recruitment provides strong data support for the analysis of complex oxygen cycles and climate regulation, and is an active attempt to integrate artificial intelligence and oceanography.

Over the past century, the issue of declining ocean oxygen levels caused by climate change has attracted widespread attention. Among the tools for understanding long-term changes in the OMZ, the rapid expansion of OMZ30 (oxygen minimum zone) is considered a key indicator. By 2023, the ocean area in 1920 has tripled. This finding is important for understanding long-term changes in the OMZ and will help better ocean monitoring and protection in the future.

Paper link: http://arxiv.org/abs/2405.07233

In order to comprehensively and deeply understand ocean deoxygenation, from the effective To explore the oxygen cycle and its changing patterns in the data, in 2017 Schmidtko and other researchers published the article "Decline in global oceanic oxygen content during the past five decades" in "Nature", which was the first time to use space The interpolation method enables the reconstruction and quantitative analysis of global ocean dissolved oxygen data since 1960.

Assessing the specific impacts of long-term human activities since the Industrial Revolution, reconstructing the dissolution climate record of the past fifty years is far from sufficient. Highly sparse historical observations and spatial interpolation methods with limited accuracy have become important bottlenecks in solving the problem.

Observation data from five major public dissolved oxygen databases including World Ocean Database 2018, CLIVAR and Carbon Hydrographic Database, Argo, Global Ocean Data Analysis Project version2.2022, and Geotraces IDP. amount, and the missing rate of ocean observation data exceeds 90%

To this end, the research team of Shanghai Jiao Tong University has brought together ocean survey data since 1900, including scientific research vessel voyage survey data, Argo buoy observation data, and real-time observations of deep-sea submersible buoys. There are a total of 6 billion pieces of dissolved oxygen related data (the amount of stored data is about 2TB), and unified quality control is carried out.

Taking into account the irregular edges of ocean water bodies and the non-uniform characteristics of highly sparse observation data, a four-dimensional spatio-temporal graph network was established through the idea of ​​graph modeling, fully considering the spatial correlation and high values ​​​​in geography. Measurement samples realize the cross-temporal information transfer between observation data and missing data.

In view of the fact that the concentration changes of ocean dissolved oxygen are affected by both ocean physical and biochemical variables, a multi-layer perceptron is first used to extract non-linear features of multi-element data, and is dissolved through a two-way long short-term memory network Mining the temporal variation characteristics of oxygen observations.

Secondly, since the global ocean presents heterogeneous spatiotemporal correlations in different historical periods and regions, inspired by the idea of ​​oceanographic zoning, a graph message passing mechanism with adaptive variable zoning (Zoning-Varying Message-Passing) is proposed. ), through the super network parameter generation algorithm, perform affine transformation on graph messages in different partitions to achieve variable partition graph information transmission.

Finally, the fusion of oceanographic domain knowledge helps calibrate the uncertainty of neural networks. This study takes the ideal balance ratio of nitrogen, phosphorus, and oxygen in the ocean (Redfield Ratio) and designs a gradient regularization method embedded with chemical knowledge to eliminate signal anomalies in the reconstruction results as much as possible.

OxyGenerator Ocean Deoxygenation Reconstruction Framework

After multi-fold cross-validation with observed variables, and comparison with the results of three sets of CMIP6 numerical models led by experts, this study The proposed OxyGenerator achieved the best performance in all four reconstruction performance evaluation indicators, with a MAPE reduction of 38.77%, greatly reducing the reconstruction error in the open ocean.

Comparison of experimental performance between OxyGenerator and CMIP6

In areas such as the Western Pacific with sufficient observation data and the Black Sea affected by special environmental conditions, OxyGenerator performs particularly well, with model performance in remained stable for hundreds of years. At the same time, the results well reconstruct the disturbance of dissolved oxygen distribution caused by special climate events such as El Niño/La Niña in historical periods, and also accurately reflect the large-time-scale water movement characteristics such as thermohaline circulation.

This research is the result of deep intersection and close cooperation between artificial intelligence and marine science, and has opened up new ideas for addressing global climate challenges. In the future, the team will continue to promote in-depth cooperative data-driven geoscience discovery research, and actively develop research in the field of advanced technology empowering scientific intelligence (AI for Science).

The above is the detailed content of Revealing 100 years of global ocean deoxygenation, Shanghai Jiao Tong University uses artificial intelligence to reconstruct the 'suffocating ocean', ICML has included. For more information, please follow other related articles on the PHP Chinese website!