Using accelerometer‐based behavioural classification to enhance scavenger conservation

Abstract Human activities are endangering animal species globally, and implementing effective conservation strategies requires understanding animal behaviour and ecology. Advancements in GPS tracking technology, accelerometry and machine learning algorithms are allowing the in situ study of animal movement and behaviour remotely. However, the challenge of building supervised machine learning algorithms and collecting the large datasets required to train them is hampering the widespread use of these tools. Additionally, the reliability of these models in classifying unobserved behaviours is rarely validated, resulting in possible classification errors. We built a supervised accelerometer‐based behavioural classification model for griffon vultures ( Gyps fulvus ). Similarly to most other avian scavenger populations worldwide, griffons are critically endangered in Israel and neighbouring countries, mostly due to feeding on poisoned carcasses. Thus, identifying this scavenger's feeding behaviour and foraging areas is crucial for their conservation. We trained a Random Forest model on acceleration data of 14 captive and 17 free‐roaming griffons. We classified 5783 behavioural observations into 6 classes: feeding, lying, standing, other ground behaviours, flapping and soaring flight. The model performed well (0.96 accuracy, 0.89 precision and 0.82 recall) and, importantly, feeding behaviours were accurately classified (0.87 precision, 0.92 recall). We calculated an observation‐specific confidence score and demonstrated its effectiveness in identifying true‐ and false‐positive classifications, in both captive and free‐roaming individuals. Finally, we used our model to reliably identify feeding hotspots, where vultures can be at higher risk of poisoning. Synthesis and applications. We provide a tool to help identify vulture feeding hotspots, supporting carcass management efforts to prevent poisoning. Integrated with near real‐time tracking, our model can support global efforts to combat scavenger poisoning. The training dataset, model and codes are provided in a user‐friendly platform, along with a conceptual framework, to encourage use by ecologists and conservation practitioners.