A Machine Learning Approach to Simulation of Mallard Movements

Machine learning (ML) is increasingly used in diverse fields, including animal behavior research. However, its application to ambiguous data requires careful consideration to avoid uncritical interpretations. This paper extends prior research on ringed mallards where sensors revealed their movements in southern Sweden, particularly in areas with small lakes. The primary focus is to distinguish the movement patterns of wild and farmed mallards. While well-known statistical methods can capture such differences, ML also provides opportunities to simulate behaviors outside of the core study span. Building on this, this study applies ML techniques to simulate these movements, using the previously collected data. It is crucial to note that unrefined application of ML can lead to incomplete or misleading outcomes. Challenges in the data include disparities in swimming and flying records, farmed mallards' biased data due to feeding points, and extended intervals between data points. This research highlights these data challenges, while identifying discernible patterns, as well as proposing approaches to meet such challenges. The key contribution lies in separating incompatible data and, through different ML models, handle these separately to enhance the reliability of the simulation models. This approach ensures a more credible and nuanced understanding of mallard movements, demonstrating the importance of critical analysis in ML applications in wildlife studies.