A Machine Learning Approach for Animal Trajectory Classification

Abstract

The ocean is the largest ecosystem on Earth where diverse human activities threaten marine life. Thus, knowing how, when, where and why animals move is important for their conservation. As a result of the study of marine animal movement through tracking devices during the past decades, we have collected a large database of around 13000 individual trajectories from more than 100 species, which can be analyzed via data-driven methods. Since its potential remains generally unexplored under these novel techniques, our goal will be to assess their performance and adequateness through the classification of species associated with spatio-temporal points (latitude, longitude, time). When shifting the trajectories to a common origin, we find that the initial accuracy of 88% falls to 66%, indicating that while the initial location is a useful feature, the algorithms are able to extract information from the shape of the trajectory. Furthermore, performance is robust to noise (artificially generated trajectories) and through the error analysis we are able to provide insight for identifying corrupted or inaccurate data, which can be useful for determining potential flaws in the data collection.