Predicting species distributions in the open ocean using satellite-derived environmental data and convolutional neural networks

While policymakers are committed to a 30% global protection target by 2030, including the ocean, our knowledge of marine species distributions remains limited compared to terrestrial species. This gap is a major barrier to science-based decision-making in the field of marine conservation. The high cost of data acquisition is partly responsible for the situation. But a significant lever would be to use the sparse data we have more efficiently. Indeed, current species distribution models (SDMs), though robust, are simplistic in terms of environment-species interactions: they often rely only on long-term climate averages. This seriously hinders the discovery of dynamic processes, such as seasonal migrations or adaptation to environmental change, and therefore limits our knowledge of how marine species use space. Recently, computational ecologists have successfully designed and tested new types of SDMs based on deep learning to model plant species distributions in the terrestrial realm. They treat the environmental landscape as a multi-layered image and use convolutional neural networks to extract relevant geographical features and predict associated species, with promising results on data-poor species using knowledge transfer. Our work adapts this approach to the open ocean, in particular by taking into account the high variability of environmental conditions. In an initial trial, we predicted relative presence probabilities for 38 marine taxa at a global scale using 18 satellite-derived environmental variables, achieving 89% Top-3 accuracy. This work provided valuable insights on data curation, variable importance and hyperparameter fine-tuning. This approach provides ways to enhance our knowledge of data-poor marine regions or species, and to deepen our understanding of the dynamic impact of environmental conditions. It paves the way for a diversity of use cases ranging from other marine environments to predictions of future species distributions. This allows more comprehensive biodiversity mapping, which is a significant step towards well-designed ecosystem protection measures.