Artificial intelligence tools to assess different levels of activity performed by semi-wild horses in grassland ecosystems.

In order to understand the role of horses in ecosystems and to effectively use their grazing in the protection of grasslands, it is important to assess where they primarily stay, followed by whether these habitats are used for grazing or resting. The main goal of the study was the model development based on artificial intelligence tools which allow to distinguish the basic levels of activity performed by horses using data from an accelerometer mounted in a collar worn by animals. The model calibration was based on direct observations of five randomly selected Polish primitive horse mares. In order to create a model that allows for classification into three groups of behaviours: grazing, resting, and moving, an approach based on machine learning, one of the basic technologies of artificial intelligence, was used. The carried out analyses allowed for the determination of the most important features, among the fourteen determined from raw X, Y, and Z axis acceleration values across 5-s measurements. The recommended method for the classification of behaviours of primitive Konik horses based on the selection of variables observed from the accelerometer is the CART method, whereas the most accurate tool for its construction is learning neural networks. Our research indicates the usefulness of the accelerometer and proposed artificial intelligence methods in distinguishing the main activities performed by horses.