Discovering Biotic Interactions from Citizen Science Imagery using Deep Learning

Interactions between species are the main contributor for many ecosystem services, including pollination. Knowing which species interact with each other, and where, allows us to assess productivity like crop yield, and take measures to protect relevant ecosystem hotspots. Interactions are traditionally observed through field experiments; however, this is tedious to do and difficult to scale. 

A recently emerging alternative is to identify interactions in photographs automatically with machine learning. Thanks to crowdsourcing initiatives like iNaturalist and eBird, we now have billions of such images available, with many of them depicting interactions (e.g., a bee visiting a flower). 

This project explores identifying multi-species interactions with machine learning. Deep learning in particular has gained strong traction for visual species classification, but interaction discovery is less studied with these tools. You will work on tasks from the detection of individuals to automated classification of interaction types, including state-of-the-art deep learning methods like foundation models. Once established, such models can then be used on the wealth of environmental, geolocated images to automatically detect and map biotic interactions at scale. This research has the potential to jumpstart future ecological studies about interactions and help uncover interaction hotspots and protect biodiversity at scale.