Place-based artificial intelligence for wildlife monitoring in Hawaiʻi

(A wild pig (puaʻa) captured by a camera trap in Hawaiʻi and classified using AI Puaʻa. The outlined box marks the detected animal, while the label displays its predicted identity and model confidence score of 0.98. U.S. Forest Service photo by Kealohanuiopuna M. Kinney, in collaboration with Addax Data Science.

Introduced animals are among the most persistent threats to Hawaiʻi’s forests. Ungulates such as feral pigs, goats, sheep, cattle, and deer can damage native vegetation, reducing forest regeneration through overgrazing, soil disturbance, increased erosion, and by spreading invasive plants and diseases. Rapid ʻŌhiʻa Death, commonly known as ROD, is one such disease that can be spread by feral ungulates, with feral pigs being some of the most common culprits. As the pigs move throughout a landscape, they rub their hides against bark leaving small “wounds” on trees. It’s through these openings that the Ceratocystis fungus can infect trees, leading to ʻōhiʻa tree mortality.

To mitigate the damage done by introduced animals, effective forest management depends on knowing where these animals occur, how they use the landscape, and whether control efforts are reducing their impact. Motion-activated cameras provide an important way to monitor invasive animals, but a single project can generate millions of photographs. Reviewing those images manually is slow and can delay information needed for research and management. Although artificial intelligence can accelerate this work, available services often require costly subscriptions, cloud-based processing, or reliable internet access that may not be available in remote areas where this information is needed most.

To help address this gap, Kealohanuiopuna Kinney, a research ecologist with the Forest Service’s Institute of Pacific Islands Forestry led the development of a wildlife image-recognition model called AI Puaʻa. This was a collaborative effort that brought together Forest Service scientists, University of Hawaiʻi at Mānoa researchers and Cooperative Extension specialists, state wildlife managers, conservation partners, and artificial-intelligence specialists.

AI, place-based imagery, and local expertise

Researchers assembled approximately 1.64 million images from 5,417 camera locations across the state and adapted freely available software (AddaxAI) to recognize 15 animal categories based on the U.S Register of Introduced and Invasive Species of mammals in Hawaii. Evaluations across 173,431 held-out images produced 97.4% overall accuracy. For wild pigs, the model achieved 98.3% precision (how often the model was correct in identifying a pig as a pig) and 98.7% recall (when a pig present in the area, how often did the model find it). In a reference comparison using 27,726 locally sourced images, the Hawaiʻi-fine-tuned model achieved 97.3% species-level accuracy, compared with 66.3% for the SpeciesNET global model. The result demonstrates the practical value of combining broadly trained models with place-based imagery and local expertise.

Examples of AI Puaʻa classifications across different animals and field conditions in Hawaiʻi: cattle (upper left), goats (upper right), mongoose (lower left), and sheep (lower right). Colored boxes mark detected animals, with labels showing predicted identities and associated confidence scores. These examples represent four of the model’s 15 animal categories. U.S. Forest Service photo by Kealohanuiopuna M. Kinney, in collaboration with Addax Data Science.

The model and associated open-source tools can process images locally after installation, without Wi-Fi or cloud uploads, making them useful for managers working with limited resources or connectivity. Where communications infrastructure is available, the model can also be paired with cellular cameras and edge-computing platforms.

AI Puaʻa demonstrates how collaborative, place-based science can produce a practical and cost-effective system that extends personnel capacity by building on local expertise and keeping the people who hold that knowledge actively involved throughout the process. The model provides a rapid, consistent first review and directs uncertain results to human experts. This allows limited personnel to spend less time sorting routine images and more time interpreting ecological patterns, evaluating management outcomes, and focusing action where it is needed most. By making image review more manageable, AI Puaʻa gives scientists, field personnel, and managers more time to interpret results, assess what is working, and decide where to act next.