Deep Learning for Biodiversity Detection and Classification


Monitoring biodiversity is a fundamental challenge in protecting the natural world. Human-based biodiversity surveys are expensive, logistically difficult, and time-consuming. To increase the scale of data collection, ecology has increasingly turned to remote image capture. However, the challenge of turning enormous pools of imagery into information on species abundance and behaviour has severely restricted the growth of automated monitoring. Bringing the recent progress in image-based artificial intelligence and computer vision to the wider ecological community depends on close collaborations between domain and data scientists. This collaboration will broaden the use of computer vision in ecology, as well as provide opportunities for ground- breaking research in computer vision. Building from existing datasets in tropical ecosystems and Swiss aquatic environments, we will develop, build, and distribute computer vision tools for biodiversity detection and identification. The data and algorithms developed for this proposal will be openly shared with the ecology and computer vision communities to foster growth between the disciplines.

Starting Date

October 2020

PI / Partners

Biodiversity and Conservation Biology, Spatial Evolutionary Ecology (WSL):

Fish Ecology and Evolution, Eco-Evolutionary Dynamics (EMPA):



  • The project aims at developing and adapting robust computer vision methods for biodiversity monitoring. The methods developed in the framework of this project will be tested on two different case studies: hummingbird detection and recognition in camera trap videos and inference of species traits from images of macroinvertebrates in Swiss rivers. For both case studies results from computer vision models will be used to formulate and support hypotheses about interaction of species with their environment (e.g. hummingbirds and plant species).

Proposed approach:

  • We will deal with methods able to perform fine-grained classification of species by integrating domain-knowledge, such as information about the ecosystems and geographical settings. In parallel, we will work on methods able to infer species traits from image data, thus taking one step further from plain fine-grained classification.


  • The impact of the project will reach beyond computer vision, but potentially enabling accurate biodiversity monitoring based on video and image data. The project will make the contributions openly available so that any research in animal biodiversity could profit.


  • Pegoraro, L.; Volpi, M.; Matthews, B.; Graham, C.; BioDetect: automated image processing for ecological monitoring. 50th Annual Meeting of the Ecological Society of Germany, Austria and Switzerland, Virtual Conference, 30 August – 1 September, 2021.


  • Graham, C.H. and Weinstein, B. G. Towards a predictive model of species interaction betadiversity. Ecology Letters, 2018, 21: 1299-1310.

  • Weinstein, B.G. A computer vision for animal ecology. J. Anim. Ecol. 2018, 87, 533–545.