Machine-learned dimethyl sulphide (DMS) for the North Atlantic (2002-2024) to support movement studies

Area of Science:

• Marine ecology
• Biogeochemistry
• Machine learning applications

Background:

• Dimethyl sulphide (DMS) is a crucial olfactory cue for seabird navigation.
• Existing DMS data have coarse spatiotemporal resolutions, inadequate for individual movement scales.

Purpose of the Study:

• To develop high-resolution DMS estimates using machine learning for the North Atlantic.
• To create a tool for integrating DMS data with animal tracking for ecological interpretation.

Main Methods:

• Developed an ensemble machine learning model using in-situ DMS observations and satellite-derived environmental predictors.
• Utilized SHAP analysis to identify key drivers of DMS prediction (mixed layer depth, nitrate, chlorophyll).
• Generated a daily 4km DMS dataset for the North Atlantic (2002-2024) and an open-source Python package (AniDMS).

Main Results:

• The machine learning model achieved high accuracy (test R² = 0.88) for DMS prediction.
• Identified key environmental drivers influencing DMS concentrations, consistent with biogeochemical understanding.
• The new dataset revealed fine-scale DMS gradients and hotspots previously uncaptured.

Conclusions:

• High-resolution DMS mapping is feasible with machine learning, improving upon coarse-resolution products.
• The AniDMS package facilitates the integration of environmental data with animal tracking.
• This framework supports advanced research on seabird olfactory navigation and movement ecology.