Predicting the biodiversity of reef fishes – FISH-PREDICT

Biodiversity and species’ habitats are deteriorating in most ecosystems worldwide, but perhaps nowhere more rapidly than on shallow-water reefs that provide well-being and socioeconomic benefits to over 1.3 billion people globally. These coastal reefs host over 30,000 fish species, which are the mainstay of ecosystem functioning, food security and incomes but are threatened by human activities and climate change. Their conservation and sustainable management raise major challenges that scientists, nations, non-governmental organizations, and managers are facing today.

Surprisingly we still lack a set of relevant ecological indicators and predictive models that could help better understand the dynamics of reef social-ecological systems to better anticipate their futures within a context of global change (climate and over-exploitation). The FISH-PREDICT project will help fill this gap by (i) taking advantage of massive but unstructured datasets on coastal ecosystems in the Mediterranean Sea and the Pacific Ocean, (ii) proposing a set of 14 complementary ecological indicators related to reef fish biodiversity and representing the main dimensions of conservation priority (pathways to extinction, ecosystem functioning, contribution to people, etc.), and (iii) developing new algorithms and methods in Artificial Intelligence (AI) integrating multiple sources of information across scales and mixing symbolic approaches based on textual knowledge and subsymbolic approaches based on data in a new generation of hybrid models.

Meeting this ambition requires to tackle four AI-Biodiv challenges:

1. Accurately and rapidly assigning short Environmental DNA barcodes to a given fish taxa in order to increase the number and quality of fish occurrence data over space and time.
2. Automatically and accurately detecting fish species in underwater videos and photos to increase the number and quality of fish occurrence and abundance data over space and time.
3. Extracting oceanographic, habitat and anthropogenic features from satellite images and seabed data where fish surveys have been carried out in order to learn low-dimensional representation of high-dimensional information and feed predictive models with parsimonious, synthetic and relevant information.
4. Accurately predicting fish species and biodiversity distribution using hybrid AI models to ultimately select the most relevant indicators of human pressure and protection effectiveness.

Beyond methodological advances, the novelty of the project lies in its purpose to challenge the classic and over-simplified view that human development is merely a systematic source of ecosystem degradation, and instead to uncover smart solutions addressing the long-term sustainability of reef social-ecological systems. We thus expect to propose novel strategies sustaining both biodiversity and people.

These societal and scientific ambitions fit with the main goals of the IA-Biodiv Challenge and the French AI Plan. To increase our impact, we plan to interact with 7 stakeholder organizations (e.g., OFB, Agence de l’Eau, Secretariat of the Pacific Community etc..). We also plan to closely interact with the COpé and other consortia by (i) co-building and co-managing the virtual research environment « IA-BiodivNet », (ii) proposing additional data (Environmental DNa, underwater videos) to train the models, and (iii) testing all models using new data collected between 2023 and 2025.

FISH-PREDICT is also very timely since the International Union for Conservation of Nature and other organizations have committed to make 30% of the world’s coastal and marine areas protected from fishing and other forms of exploitation by 2030. Our findings may serve as catalysts to boost the creation of marineprotected areas in regions (Mediterranean Sea and the Pacific Ocean) lagging behind conservation objectives.