Tom Ratsakatika, Researcher in AI and Environmental Risk at the University of Cambridge, reflects on developing an AI-based alert system for bears and wild boars in Romania’s Carpathian Mountains.
Artificial Intelligence (AI) has revolutionised how we analyse wildlife camera trap photos. For example, it can automatically label common species in thousands of photos within hours – a task that would take months if done manually. With the rise of 4G-enabled camera traps, AI is set to transform real-time wildlife monitoring. In partnership with Fundația Conservation Carpathia, we adapted these technologies to build an automated alert system for bears and wild boars in Romania’s Carpathian Mountains.
The Romanian Carpathian Mountains are home to one of Europe’s largest forest ecosystems and most significant populations of bears, wolves and lynx. The forests also host thousands of wild boars, which are an important food source for bears and wolves. However, in recent years, logging, shifting land-use patterns, and climate change have degraded this habitat, forcing animals to approach mountain villages in search of food. Bears and wild boars pose particular risks to rural communities as they can damage crops, grassland, and livestock, considerably impacting people’s livelihoods and wellbeing.
The rural village of Rucăr is nested in a valley surrounded by dense forests, making it easy for bears and wild boars to reach farms and orchards undetected. Photo: Tom Ratsakatika.
As part of the Carpathian Mountains restoration landscape, Fundația Conservation Carpathia is working with rural communities to promote the peaceful coexistence between people and wildlife. This involves deploying Rapid Intervention Teams who, if alerted swiftly, can address wildlife intrusions before damage occurs. However, given the vast area each team must cover and the fact that intrusions mostly happen at night, they are frequently notified too late. To help optimise and scale the nightly patrols, Fundația Conservation Carpathia approached our team at the University of Cambridge to develop an AI-based alert system for bears and wild boars.
A major consideration was selecting the right hardware for the alert system. Many existing AI-based alert systems rely on satellite networks since, historically, cellular network coverage has been unreliable in rural areas. The prohibitive cost of satellite communications means that camera traps are modified to analyse photos directly on site and only simple text alerts are sent. However, 4G access is expanding rapidly, including in rural Romania. Therefore, we opted to deploy off-the-shelf 4G-enabled camera traps configured to send every photo to a remote server for AI analysis. If the AI identifies an animal of interest, an instant message is sent to the Rapid Intervention Team via a group chat.
The AI processes the photos in two stages. First, a generic object detection model called MegaDetector locates any animals in the photo sequence. While this model is very good at drawing boxes around animals, it cannot tell what species the animal is. To achieve this, a second AI model specialised in detecting bears and wild boars is used. This model was adapted from research by the DeepFaune Initiative by using 26,000 photos labelled manually by Fundația Conservation Carpathia.
The alert system has now been deployed for three months at two locations near the village of Albești. So far, it has sent correct alerts for nearly 100% of the bears and wild boars captured by the camera traps (a couple of particularly large wild boars have been misclassified as bears!). The Rapid Intervention Teams are pleased with the system’s speed, which is a major improvement over directly speaking with farmers. Looking ahead, our next goal is to deploy more cameras across additional villages. Only by observing how the technology integrates with the Rapid Intervention Team’s work at scale will we know its true impact on protecting villages and conserving wildlife.
Camera trap image from the study showing labels created by the AI system identifying the species present.
The research report detailing the system’s development can be downloaded here. The alert system software can be downloaded here (requires some Python coding knowledge). More information about the Carpathian Mountains restoration landscape can be found on their project page.
This article was written by Tom Ratsakatika, Researcher in AI and Environmental Risk at the University of Cambridge. He would like to thank Professor Srinivasan Keshav from the Department of Computer Science and Technology at the University of Cambridge, and Dr Ruben Iosif, Zsolt Miholcea, Laviniu Terciu and Gabi Dudu from Fundația Conservation Carpathia for contributing their expertise and support to this research.
