Thu, Aug 18, 2022: 3:45 PM-4:00 PM
518B
Background/Question/MethodsTechnological advances in tagging and tracking animals allows us to collect high-resolution data in space and time on species that migrate large distances. Many of these species are however partial migrants, where only a subset of the population migrates in any given year. Distinguishing between migratory and non-migratory behaviors based on consistent criteria and across large datasets (automation) remains a challenge. To address this issue, we developed a binary classifier to delineate migrants from non-migrants within a given year and across years. The classifier is designed to process large datasets with minimal human supervision, and to output, for each individual in each season, the probability that this individual was migrating. Our classifier is based on the Earth Mover’s Distance (EMD) between the distributions of positions in either geographical or environmental space, occupied by the focal animal during two user-defined seasons, and is calibrated using a training set of manually classified trajectories. This classifier provides an efficient method to categorize individuals accurately and precisely across a migratory continuum, allowing for automated and fully reproducible migratory classification.
Results/ConclusionsOur model proved reliable across multiple species and systems, with temporally dynamic environmental covariates (e.g., NDVI or snow depth) providing a great deal of predictive power. Based on the area under the Receiver Operating Characteristic (ROC) curve, our best geographical-space-only model delineated migrants and residents with an overall accuracy of 98.67%, whereas our best environmental-space-only model delineated migrants and residents with an accuracy of 93.3%. Snow depth and NDVI are often considered key environmental covariates that contribute to migration, and our classifier identifies these covariates as important contributors to delineating migrants from residents. This suggests that our classifier may allow users who are not familiar with key components of migration for their study species to identify potential environmental covariates. Our results suggest that our classifier provides the means to categorize individuals accurately and precisely across a migratory continuum, based on utilization distribution shifts in both geographical and environmental space.
Results/ConclusionsOur model proved reliable across multiple species and systems, with temporally dynamic environmental covariates (e.g., NDVI or snow depth) providing a great deal of predictive power. Based on the area under the Receiver Operating Characteristic (ROC) curve, our best geographical-space-only model delineated migrants and residents with an overall accuracy of 98.67%, whereas our best environmental-space-only model delineated migrants and residents with an accuracy of 93.3%. Snow depth and NDVI are often considered key environmental covariates that contribute to migration, and our classifier identifies these covariates as important contributors to delineating migrants from residents. This suggests that our classifier may allow users who are not familiar with key components of migration for their study species to identify potential environmental covariates. Our results suggest that our classifier provides the means to categorize individuals accurately and precisely across a migratory continuum, based on utilization distribution shifts in both geographical and environmental space.