iiinova — where your digital self keeps evolving

Clément Violet

@clementviolet.bsky.social

PhD Marine Ecology | Postdoc on Marine Non Indigenous Species | IFREMER 🐬🇫🇷

Clément Violet @clementviolet.bsky.social · 13/02/2025

🌍 Understanding habitat states & transitions is key for conservation! 🏝️ Helps track ecosystem shifts from climate change 🛑 Detects early warning signs of degradation (e.g., kelp loss → urchin barrens).

Change in the cluster composition of the Beware Reef. This site shiffted into an urchin barren in 2013 as show by the disaperence of large canopy algae (dark blue line).

112

Clément Violet @clementviolet.bsky.social · 13/02/2025

🌍 With our data-driven standardized way to classify & monitor marine habitats, what did we find? 🏝️ 17 distinct habitat states, including kelp forests, coral reefs & transitional zones 📉 Our pipeline captures fine-scale ecological changes over time

Description of the identified clusters along with photoquadrats examples. Arrows indicate transitions between clusters.

132

Clément Violet @clementviolet.bsky.social · 13/02/2025

This study wouldn’t be possible without citizen scientists! 🌍🌊 👩‍🔬 Reef Life Survey divers collected 6554 transects worldwide 📸 Standardized photoquadrat sampling ensured consistency 📊 It enabled us to track habitat transitions over time

Map of the number of RLS transect sampled per ecoregion.

141

Clément Violet @clementviolet.bsky.social · 13/02/2025

Machine learning models like UMAP & HDBSCAN are powerful, but they’re often hard to interpret. So, we used SHAP to: 📌 Identify which features drive cluster formation 📌 Uncover non-linear interactions in benthic ecosystems 📌 Improve ecological interpretability of results

Using SHAP to Interpreting this cluster as a Foliose brown algae cluster.

111

Clément Violet @clementviolet.bsky.social · 13/02/2025

We used better methods: UMAP (for dimension reduction) & HDBSCAN (for clustering). ✅ Captures complex ecological structures ✅ Finds clusters of varying shapes & sizes ✅ Filters noise instead of forcing data into clusters

UMAP ordination of the benthic cover data. Each point is coloured accordingly to its cluster. Black points represent unclustered data.

111