26-205 AI and Earth Observation to Understand Fish Habitat in a Changing Ocean

Mission

Context and Motivation

Earth Observation (EO) provides an unprecedented view of the dynamic ocean, capturing variables such as sea surface temperature, chlorophyll concentration, sea ice coverage, currents, and wave energy at multiple scales. These observations are essential for understanding marine ecosystems, yet linking them to biological data remains a major challenge.

This PhD project aims to develop new computational and AI-based methods to integrate multi-source satellite observations with biologging datasets, in order to characterize and predict marine habitat dynamics. The approach emphasizes innovation in EO data representation, hierarchical spatial indexing, advanced signal analysis, and explainable machine learning, with applications extending beyond marine ecology to other Earth system sciences.

Scientific and Technological Objectives

The project will advance EO data use for ecosystem analysis through three main directions :

Innovative EO Data Representation — Use Hierarchical Equal Area isoLatitude Pixelation of a sphere (HEALPix) to structure EO data for efficient multi-resolution analysis and direct combination of heterogeneous satellite products (optical, altimetric, etc.).

Advanced Numerical Methods — Implement the Scattering Covariance framework, a multi-scale statistical approach inspired by physics and signal processing, to extract invariant patterns from EO fields. This method mimics deep-learning architectures but requires no large training datasets, making it suitable for sparse biological data.

Integrated Modeling and Explainable AI — Develop interpretable models linking ocean dynamics to biological responses using unsupervised and semi-supervised learning. Explainable AI tools will help identify which physical features (fronts, eddies, sea-ice margins) influence species’ habitats and migration.

These methods will establish a new paradigm for combining EO with biological observations, enhancing our capacity to model ecological processes from space.

Application Case : Atlantic Salmon Migration

The approach will be demonstrated using a unique biologging program in southwest Greenland conducted by the Atlantic Salmon Federation (ASF), Fisheries and Oceans Canada (DFO), and NOAA Fisheries. Atlantic salmon populations have declined in many regions; understanding their marine phase is crucial for conservation.

From 2018–2024, 427 salmon were tagged with acoustic transmitters and pop-off satellite archival tags (PSATs) that record temperature, depth, and light. When the tags detach, summaries are transmitted via satellite; if recovered, high-resolution data can be downloaded. Traditional track reconstruction relies on light-based geolocation, which is error-prone at high latitudes due to cloud cover and daylight variability.

By fusing biologging data with EO (sea-ice concentration, SST, chlorophyll-a, sea-surface height, frontal diagnostics) and ocean model outputs (3-D currents, mixed-layer depth, wave energy, ..), the project will reconstruct probable migration pathways, identify behavioral states, and quantify uncertainty—demonstrating how EO can overcome traditional limitations in biologging analyses.

Methodological Plan

Data integration : Co-locate biologging records with EO and model variables on a HEALPix grid using scalable cloud tools (xarray, dask, Pangeo); store AI-ready data in Zarr format.

Feature extraction : Apply Scattering Covariance to EO / model fields and biologging time series to derive coherent, multi-scale descriptors.

Unsupervised learning : Cluster behavioral states and relate them to environmental patterns using probabilistic models and spatio-temporal cross-validation.

Explainability and visualization : Identify key physical drivers and visualize trajectories, habitats, and uncertainty to support expert interpretation and policy dialogue.

Open science : Release code and data in FAIR formats (HEALPix / Zarr), ensuring reproducibility and reusability within the Pangeo and DestinE ecosystems.

Expected Outcomes and Impact

The project will deliver a Digital Twin–ready EO and model integration framework linking marine animal behavior to dynamic ocean conditions. Outputs include :

(i) Open, FAIR, AI-ready datasets (HEALPix-indexed, cloud-native Zarr),

(ii) Reproducible end-to-end workflows and open-source code, and

(iii) Interpretable predictors and probabilistic habitat / behavior maps derived from EO and ocean-physics models.

Because the data model, tiling, and pipelines are standardized, the same workflow can be applied to other biologging campaigns and species, enabling comparative ecology at scale. Beyond ecology, this work advances EO data-analysis technologies through new methods for multi-source fusion, scalable computation, and numerical diagnostics.

The outcomes align with the vision of Digital Twins for Earth system understanding, supporting adaptive ecosystem management and climate-resilience strategies.

=================