Research

Our knowledge of whales at sea is contingent upon our ability to find them: whether visually from an airlane or vessel or acoustically from hydrophones. Both of those observation types are shaped as much by the survey as by the animal’s behavior and environmental conditions.

My work as a marine ecologist focused on conservation considers what we can infer from the data we are able to collect. I build models that separate the observation process from the ecological process, so that what comes out the other end is an estimate of what is actually there rather than an estimate of where we happened to look. That matters because these models are not academic exercises: they feed directly into the decision support tools used to manage the critically endangered North Atlantic right whale.

Where the food is

Right whales eat copepods — primarily Calanus finmarchicus — and they are extremely particular about it. A whale does not respond to average copepod abundance; it responds to whether a patch is dense enough that feeding on it returns more energy than it costs. That distinction turns out to matter enormously for modelling.

So rather than model copepod abundance and hope it stands in for prey, I model the environmental conditions associated with aggregations that exceed a feeding threshold — what we call τ-patches. These behave differently from abundance-based models: they favour a warmer temperature range and show much higher spatial variability. Extending the approach past C. finmarchicus to Centropages typicus and Pseudocalanus spp. showed that the prey base is an assemblage that shifts across space and time, not a single species.

Key papers: Estimating right whale prey based on Calanus finmarchicus thresholds · Beyond Calanus: changes to the copepod community · Calanus prey and foraging habitat in Canadian waters

Code: taupatch implements the threshold patch models; datamatch and derivoce build the environmental covariates.

Where the whales are

Density surface models turn line-transect survey data into maps of animal density. They are only as good as their treatment of the observation process: how detectability falls off with distance, how effort is distributed, how much of the variance in a map is survey design rather than biology.

The question I have spent the most time on is what goes in as a prey covariate. The convention is chlorophyll concentration — cheap, satellite-derived, and several steps removed from anything a whale eats. Substituting tailored prey fields for that proxy improves predicted distributions: the best-performing model combined all three copepod prey fields and put density where right whales actually are, concentrated in the deep basins of the Gulf of Maine and the Great South Channel, rising through the summer.

The corollary is a claim about monitoring, not just modelling. Prey fields only exist if somebody keeps sampling zooplankton.

Key papers: Incorporating prey fields into right whale density surface models · Impacts of an oceanographic regime shift on U.S. right whale density estimates (in review)

Code: narwcr and distsamp handle survey ingest and segmentation; dsfit fits and compares detection functions on effective strip half-width rather than the Akaike information criterion alone.

Where both are going

The Gulf of Maine is one of the fastest-warming pieces of ocean on the planet, and right whales have already redistributed in ways that broke the assumptions behind existing management areas. Projecting habitat suitability forward under climate scenarios suggests the trend continues: declining suitability across much of the Gulf of Maine from mid-summer into autumn, with a north-eastward shift toward the Scotian Shelf, the Bay of Fundy, and the Newfoundland and Labrador shelves — much of it outside where conservation effort is currently concentrated.

Change of that kind is not only gradual. The Gulf of Maine has a documented history of oceanographic surprises, and they have been arriving more often than the long-term trends alone would predict. That raises a harder question than “what will the mean look like in 2050”: what happens to a density estimate built across a regime shift, when the relationship the model learned no longer holds? That question is the current work.

Key papers: Projecting right whale habitat suitability for 2050 · Foraging habitat under future climate scenarios · The surprising oceanography of the Gulf of Maine

Methods and tools

• Observation process — detectability modelling from line-transect data, density surface models with explicit treatment of survey effort, propagating observation uncertainty into management-facing estimates
• Statistical learning — generalized additive models, boosted regression trees, random forests, artificial neural networks (including CNNs trained from scratch)
• Time series — spectral analysis of multi-decadal oceanographic and biological series: Fourier transforms and periodograms, coherence, filtering, autocorrelation structure
• Data sources — Copernicus Marine and NASA EarthData products, ocean physics model output (HYCOM, GLORYS, FVCOM, DFO internal models), aerial and vessel survey archives, passive acoustic monitoring
• Practice — R, Python, git, bash, Linux; reproducible pipelines; R Shiny and Leaflet