Big data insights for wildlife.
An introduction to Wildlife Insights.
Deep in the Congolian forest, a small box silently keeps watch as the sun sets, rises, and sets again. Despite its unobtrusive appearance, the box has not gone unnoticed. A Mountain Gorilla is peering at it curiously.
Inside the box is a camera capturing the movement of every animal that passes by. Soon a scientist will come to collect the camera’s memory card and begin the time-consuming task of reviewing the thousands of images their camera traps have taken.
But, thanks to Wildlife Insights, the task of processing camera trap data is now faster and easier. Wildlife Insights is an open access platform that uses artificial intelligence to automatically identify species captured in camera trap images. It’s the first global platform of its kind — and it already contains millions of images that are ready to be analysed with the inbuilt suite of tools.
With so much data in one place, Wildlife Insights has the potential to change everything we know about species and how to conserve them. By bringing together data from all over the world, Wildlife Insights is connecting isolated pockets of knowledge and encouraging a truly collaborative approach to global wildlife conservation.
The founding and core members behind Wildlife Insights includes Conservation International, Wildlife Conservation Society, Google, North Carolina Museum of Natural Sciences, Smithsonian Institution, Zoological Society of London, WWF, and Map of Life. Our role as a project partner was to build the platform from scratch, ensuring it would remain stable and scalable as it grows. Given that thousands, if not millions of images could be uploaded to Wildlife Insights in the coming years, we knew we’d have to build a solid foundation on which the platform could grow.
The backend is totally custom build, using NestJS for the main API, and it’s deployed to a Kubernetes cluster, all of it running on Google Cloud. We’ve gone cloud native this time, and in addition to Kubernetes we’re using a lot of other cloud technology: Airflow runs the data analysis pipelines (running Rocker-based images), computer vision is performed in Google’s AI Platform and many processes are delegated to cloud functions. We’ve ended up with a system that is performant while dealing with what is already a massive dataset, and that can be scaled for its future needs.
Studying animals in the wild has traditionally been a labour-intensive endeavour. Ecologists spend months in the field collecting data, often in difficult or uncomfortable conditions. Camera traps have become an essential tool for researchers, providing repeated samples of the presence and absence of a species that can be used to inform population counts, behaviour analysis, and conservation planning. Thousands of images can be collected each field trip, and reviewing them takes time. By letting a computer do the tagging, some early users of Wildlife Insights have estimated that they have reduced the amount of time they spend processing and reviewing images by 80%.
The Wildlife Insights automatic species identification tool was trained on data from the TEAM Network, Snapshot Serengeti, Caltech Camera Traps, North American Camera Trap Images, and One Tam, which includes 614 species from around the world. Currently, common classes of species can be identified with an accuracy of between 80% and 98.6%, and accuracy is predicted to increase as more images are added to the training dataset.
Knowing where species are and how their populations are faring is an essential part of conservation. A quarter of all assessed species are threatened with extinction, and we need to act quickly if we want to prevent any more species from going extinct. Wildlife Insights will provide the data that enables the decisions on which places and species most urgently need protection. Wildlife Insights is the perfect opportunity for researchers to collaborate and share their knowledge with others who are working towards the same aims.
