I am looking for insight on the feasibility of identifying seabirds from photos taken at sea (such as these https://www.flickr.com/photos/[email protected]/albums/72157708853397813) to different levels of precision:
1) species level. This can already be tested, I think, as it is already implemented in camera trap pictures.
2) phenotype level within species. In some species (for example, the Northern Fulmar https://www.flickr.com/search/?text=northern%20fulmar), a continuous phenotypic gradient exists (for example, from darkest to lightest plumage). I suspect that the difficulty would be in assigning a score - though one end of the gradient could be considered as the typical plumage (100% ID score) and the other end as the aberrant plumage (0% ID score), with scores in between giving a proxy for phenotype.
3) individual level. Like the BearID Project is able to identify individual bears using face recognition techniques, could deep learning identify individual birds using not only face geometry but also plumage patterns (for example, face, under- and upper-wing patterns as in the Black-capped Petrel https://www.flickr.com/search/?text=black-capped%20petrel)?
Using geo-referenced images, this kind of analyses would help look into species/phenotype distribution. Individual identification could also be used to estimate population sizes (a conservation metric that is missing in several endangered pelagic seabirds).