Correlative distribution models have become essential tools in conservation, macroecology and ecology more generally. They help turn limited occurrence records into predictive maps that help us get a better sense of where species might be found, which areas might be critical for their protection, how large their range currently is, and how it might change with climate change, urban encroachment or other forms of habitat conversion.
It can be frustrating, however, when species distribution models (and the predictive maps they produce) don’t adequately capture what we already know about the habitat needs of a species. A major challenge to date has been to represent the environmental needs of species that require distinct habitats during different life stages or behavioural states. Rainbow parrotfish (Scarus guacamaia), for example, spend their youth sheltered from predators in mangrove areas before moving onto coral reefs, and European nightjars (Caprimulgus europaeus) breed in heathland but require access to grazed grassland for foraging. Correlative distribution models confronted with occurrence records from both life stages or behavioural modes tend to produce poor predictive maps because they confound these distinct requirements. Continue reading →
Understanding key habitat requirements is critical to the conservation of species at risk. For highly mobile species, discerning what is key habitat as opposed to areas that are simply being traversed (perhaps in the search for key habitats) can be challenging. For seabirds, in particular, it can be difficult to know which areas in the sea represent key foraging grounds. Devices that record birds’ diving behaviour can help shed light on this, but they’re expensive to deploy. In contrast, devices that record the birds’ geographic position are more commonly available and have been around for some time.
In their recent study entitled ‘Predicting animal behaviour using deep learning: GPS data alone accurately predict diving in seabirds,’ Ella Browning and her colleagues made use of a rich dataset on 399 individual birds from three species, some equipped with both global positioning (GPS) and depth recorder devices, others with GPS only. The data allowed them to test whether deep learning methods can identify when the birds are diving (foraging) based on GPS data alone. Results were highly promising, with top models able to distinguish non-diving and diving behaviours with 94% and 80% accuracy. Continue reading →