26 September 2024 in Peer Community Journal
As biodiversity plummets due to anthropogenic disturbances, the conservation of oceanic species is made harder by limited knowledge of their distributions and migrations. Indeed, tracking species distributions in the open ocean is particularly challenging due to the scarcity of observations and the complex and variable nature of the ocean system. In this study, we propose a new method that leverages deep learning, specifically convolutional neural networks (CNNs), to capture spatial features of environmental variables. This novelty eliminates the need to predefine these features before modelling and creates opportunities to discover unexpected correlations. Our aim is to present the results of the first trial of this method in the open ocean, discuss limitations and provide feedback for future improvements or adjustments. In this case study, we considered 38 taxa comprising pelagic fishes, elasmobranchs, marine mammals, marine turtles and birds. We trained a model to predict probabilities from the environmental conditions at any specific point in space and time, using species occurrence data from the Global Biodiversity Information Facility (GBIF) and environmental data from various sources. These variables included sea surface temperature, chlorophyll concentration, salinity and fifteen others. During the testing phase, the model was applied to environmental data at locations where species occurrences were recorded. The classifier accurately predicted the observed taxon as the most likely taxon in 69% of cases and included the observed taxon among the top three most likely predictions in 89% of cases. These findings show the adequacy of deep learning for species distribution modelling in the open ocean. Additionally, this purely correlative model was then analysed with explicability tools to understand which variables had an influence on the model’s predictions. While variable importance was species-dependent, we identified finite-size Lyapunov exponents (FSLEs), sea surface temperature, pH and salinity as the most influential variables, in that order. These insights can prove valuable for future species-specific ecology studies.
7 September 2022 in Aquatic Conservation: Marine and Freshwater Ecosystems
11 December 2020 in bioRxiv
Coral restoration emerged globally as a form of life support for coral reefs, awaiting urgent mitigation of anthropogenic pressure. Yet its efficiency is difficult to assess, as ambitious transplantation programs handle hundreds of thousands of fragments, with survival rates inherently time-intensive to monitor. Due to limited available data, the influence of most environmental and methodological factors is still unknown. We therefore propose a new method which leverages machine learning to track each colony’s individual health and growth on a large sample size. This is the first time artificial intelligence techniques were used to monitor coral at a colony scale , providing an unprecedented amount of data on coral health and growth. Here we show the influence of genus, depth and initial fragment size, alongside providing an outlook on coral restoration’s efficiency. We show that among 77,574 fragments, individual survival rate was 31% after 2 years (21% after 4 years), which is much lower than most reported results. In the absence of significant anthropogenic pressure, we showed that there was a depth limit below which Pocillopora fragments outperformed Acropora fragments, while the opposite was true past this threshold. During the mid-2019 heatwave, our research indicates that Pocillopora fragments were 37% more likely to survive than Acropora fragments. Overall, the total amount of live coral steadily increased over time, by more than 3,700 liters a year, as growth compensated for mortality. This supports the use of targeted coral restoration to accelerate reef recovery after mass bleaching events.
12 August 2023 in bioRxiv
As biodiversity plummets due to anthropogenic disturbances, the conservation of oceanic species is made harder by limited knowledge of their distributions and migrations. Indeed, tracking species distributions in the open ocean is particularly challenging due to the scarcity of observations and the complex and variable nature of the ocean system. In this study, we propose a new method that leverages deep learning, specifically convolutional neural networks (CNNs), to capture spatial features of environmental variables. This novelty eliminates the need to predefine these features before modelling and creates opportunities to discover unexpected correlations. Our aim is to present the results of the first trial of this method in the open ocean, discuss limitations and provide feedback for future improvements or adjustments. In this case study, we considered 38 taxa comprising pelagic fishes, elasmobranchs, marine mammals, marine turtles and birds. We trained a model to predict probabilities from the environmental conditions at any specific point in space and time, using species occurrence data from the Global Biodiversity Information Facility (GBIF) and environmental data from various sources. These variables included sea surface temperature, chlorophyll concentration, salinity and fifteen others. During the testing phase, the model was applied to environmental data at locations where species occurrences were recorded. The classifier accurately predicted the observed taxon as the most likely taxon in 69% of cases and included the observed taxon among the top three most likely predictions in 89% of cases. These findings show the adequacy of deep learning for species distribution modelling in the open ocean. Additionally, this purely correlative model was then analysed with explicability tools to understand which variables had an influence on the model’s predictions. While variable importance was species-dependent, we identified finite-size Lyapunov exponents (FSLEs), sea surface temperature, pH and salinity as the most influential variables, in that order. These insights can prove valuable for future species-specific ecology studies.
24 September 2026 in SSRN Electronic Journal
Global biodiversity decline and the commitment to protect 30% of land and sea by 2030 increase the need to map and understand the multiple facets of biodiversity. Macroecological models can complement in situ measurements to produce continuous biodiversity maps, but their use by conservation practitioners is constrained by limited interpretability and uncertainty assessment. Here, we analysed an extensive eDNA survey along the Mediterranean coastline to map nine indicators of fish biodiversity: taxonomic, functional and phylogenetic diversity, commercial, Red-list, cryptobenthic, elasmobranch and top-predator richness, and climatic vulnerability. We used a multi-model approach with target-oriented cross-validation, SHAP values, uncertainty and transferability analyses to map indicators across 6,369 km² at 1.2 km resolution. Our results refined previously described biodiversity patterns and identified localized hotspots. Biodiversity facets differed substantially in both their spatial patterns and predictability. Five indicators reached the predictive-performance threshold for mapping: species richness, climatic vulnerability, commercial richness, cryptobenthic richness and functional diversity. After accounting for correlation with species richness, their spatial patterns were markedly distinct. Climatic vulnerability was higher than expected along the Provence-Alpes-Côte d’Azur and Corsican coasts and lower in the Gulf of Lion, whereas commercial species richness showed the opposite pattern. Uncertainty and transferability also varied among indicators and space with less reliable predictions in underrepresented environmental conditions, particularly offshore. Together, our results show that biodiversity indicators differ not only in their spatial distributions, but also in their predictability. Conservation-relevant biodiversity mapping should therefore consider the complementarity among biodiversity facets together with the reliability of their predictions.
30 October 2025 in bioRxiv
In a context of rapidly declining biodiversity, knowing the distribution of endangered species is critical to ensure the protection of the areas they occupy. To achieve this, species distribution models (SDMs) typically use a range of variables to determine the suitability of an area to a given species, which enables scientists to produce species distribution maps. Most SDMs are statistically limited in the number of predictors they can take into account, which leads to using summary variables such as temperature yearly average or bathymetry extrema. Deep-learning-based SDMs have been proposed to tackle this limitation by bringing powerful implicit feature extractors. Here, we describe a new model to advance the adaptation of such Deep-SDMs to marine environments and take advantage of the knowledge of the environmental seascape around surveyed points. We used data from the Reef Life Survey data set to predict the presence of 1,796 fish species around all of Australia, in a wide range of climates. The environmental seascape around each point was encoded as a raster image populated with 15 environmental variables, 4 human activity variables, and completed by a 10-year time series of temperature anomaly. The model performance was highly dependent on the species, with a 0.95 F1 score for the best performing species, Notolabrus parilus , but rapidly decreased, with the 100th best species having a 0.38 F1 score.