Session 8: Machine learning in the marine sciences: opportunities and challenges

Hosts: David Greenberg and Moritz Mathis

Recently the availability of ocean and climate data has grown immensely due to significant advances in high-performance computing and observational monitoring. This opens new possibilities for understanding the dynamics and susceptibility of the Earth system, improving the predictability of future climate states and developing management strategies to ensure sustainable food and energy usage. 

At the forefront of new information technologies that can benefit marine science and resource management is machine learning (ML). ML systems are applicable to a wide range of data analysis problems, such as supervised detection of data features (e.g. storms) or automatic discovery of patterns (e.g. leading variability modes in the ocean and atmosphere). ML can also be used to improve existing computational workflows, such as model parameter tuning, causal inference, sequential data assimilation and quantification of predictability and uncertainty.

However, the systems, models, and data used in marine science pose major conceptual and technical challenges for ML. Data is often irregularly structured in space or time and may have missing values, requiring standard ML techniques (such as convolutional networks) to be modified. Image-based data in the Earth sciences can reach a size and resolution seldom considered in ML, pushing existing software and hardware to their limits. When the physical systems studied and modeled exhibit both chaotic dynamic, ML algorithms based on automatic differentiation and floating point arithmetic can exhibit numerical instabilities. Overcoming these challenges is critical for realizing the full potential of ML in marine science.

This session aims at the transfer and exchange of recent ML developments, highlighting opportunities and challenges for an effective adoption of this technology across the marine sciences. We hope to bring together ML specialists interested in marine applications and marine practitioners using ML or curious about its possibilities, promote collaborations and identify common challenges and solutions. We call for contributions that demonstrate or propose various applications of ML algorithms in the scientific analysis of marine observational products and modeling data.