How AI Is Revolutionizing Wildlife Conservation and Ecosystem Monitoring

As we stand face-to-face with accelerated biodiversity loss, experts are harnessing the powerful potential of artificial intelligence to track and safeguard our vanishing wildlife. According to a recent study by Oregon State University, habitat destruction, overexploitation of resources and climate change have left more than 3,500 animal species marching towards become extinct.

High-tech solutions for pressing environmental problems

Within the world of academia, bright minds like Justin Kay, a PhD student at MIT and member of the Computer Science and Artificial Intelligence Laboratory (CSAIL), are diving into this challenge headfirst. Kay, under the guidance of CSAIL principal investigator Sara Beery, develops innovative computer vision algorithms to monitor wildlife – his current focus is on tracking the crucial salmon species in the Pacific Northwest. These fish are the linchpins of their ecosystem; they feed predators like birds and bears, and manage insect populations.

Choosing the appropriate AI model for a specific dataset can sometimes resemble finding a needle in a haystack, given the explosion of AI tools readily available. Today, there are over 1.9 million pre-trained models on platforms like HuggingFace. Kay, in collaboration with his team at MIT and the University of Massachusetts Amherst, proposed a solution to this predicament: “consensus-driven active model selection,” or CODA. This novel method enables researchers to swiftly pinpoint the most effective AI model for their data, eliminating the need for lengthy annotation and testing.

CODA reshapes the traditional route of using AI, which required building a model from the ground up, an exercise that necessitates technical expertise and a representative dataset. With CODA, users can make the most of existing pre-trained models by annotating only a few salient examples to pick the best-suited model for their data. This method banks on the “wisdom of the crowd” approach, analyzing the consensus among numerous AI models to identify the model likely to perform best across the entire dataset.

Efficacy and Future Applications of CODA

CODA has already exhibited remarkable results in classifying wildlife in images, an exercise that is paramount for ecologists handling huge datasets from field cameras. For instance, CODA can help a researcher swiftly figure out which AI model will most accurately classify species in a cache of hundreds of thousands of wildlife images, even with minimal labeled data.

At Beerylab, where Kay is currently working, there is a flurry of AI applications in ecology being explored. These include using drones to monitor coral reefs, identifying individual elephants over time and integrating satellite and ground-based data to fathom environmental changes. The team also works on models that deal with data bottlenecks using scalable computer vision and machine learning tools.

Rethinking AI Evaluations in Ecology with Human Touch

Kay emphasizes the importance of framing the output of vision models—such as detecting animals in images— within the context of broader analyses aligned with answering ecological questions, like species distributions or tracking population changes over time. His team is developing ways to evaluate AI performance that integrates human expertise and multi-stage prediction pipelines to enhance the relevance and trustworthiness of AI tools in ecology. “The natural world is changing at unprecedented rates,” Kay rightly notes, “and being able to move quickly from scientific questions to data-driven answers is more important than ever.”

For a deeper dive into this topic, you can read the full interview and article at MIT News.