Deep learning algorithms have seen rapid and widespread adoption in ocean science. For many tasks, such as classification and error correction, they now represent the state of the art. However, applying deep learning in the field of oceanography also presents unique challenges, including the various temporal scales of oceanic processes, heterogeneously distributed and noisy observational data, and unresolved processes in numerical models.
In this short course, we aim to present a set of best practices for applying and assessing deep learning methods in oceanographic research. We will also highlight common pitfalls and how to avoid them.
The course will be structured around a series of short presentations and practical examples covering key topics, including:
- Types of oceanographic problems suited for deep learning: reconstruction, prediction, …
- Building datasets appropriate for deep learning applications: constitution of - training/validation/test datasets, effect of non-stationnarity, type/quality/number of data, …
- Training strategies and model selection: normalization, supervised training, generative models, …
- Validation and evaluation of ocean products derived from deep learning: accuracy, realism, …
- Ethical considerations: reproducibility, open science, and the environmental impact of deep learning