Detecting and quantifying animal migration from occurrence data using temporally-explicit distribution models

Background/Question/Methods

Identifying the migratory strategies employed by animals is key to understanding patterns of biodiversity and identifying key conservation priorities. Tracking seasonal shifts in geographic distributions is an emerging method to describe migration from the population to the macroecological scale. Although the rapidly-increasing availability of occurrence records hold much promise for increasing understanding of animal migration, models relying on such records are subject to bias introduced by uneven and variable data collection methods. To overcome these challenges, we developed a methodological toolkit to efficiently characterize population-level migratory strategies from presence-only occurrence data. The process extracts a suite of migration metrics from modeled distributions, which quantify measures of migration distance, geographic range size shift, and geographic seasonality. We evaluated the following research questions: 1) How to detect and classify shifting population distributions driven by animal migration? and 2) Do our methods compensate for opportunistic sampling regimes? To demonstrate and test these migration metrics against a suite of benchmarking metrics, we tested the comparative performance of our mapping approach versus one relying on occurrence records alone.

Results/Conclusions

We extracted migration metrics from the distributions of 43 species of well-studied bird in the Americas that were selected to represent a broad suite of migratory characteristics and patterns of geographic distribution. Benchmarking metrics (migration distance, range size shift, and seasonality) generated from season-specific range maps suggested that these species represent diverse migratory strategies, including long-distance migration, residency, and partial migration. Our approach compensates for opportunistic sampling in two ways: first, data cleaning spatially thins data, compensating for uneven sampling; second, the machine learning approach detects correlations between environmental covariates and species distribution, predicting distributions more accurately than from opportunistic sampling alone. We identified the metrics that perform most reliably at predicting benchmarking metrics, and note that in most cases the novel metrics outperform a geography-only model which relies on occurrence records only. These metrics will present a useful suite of tools for classifying and summarizing the patterns of animal migration, as well as distribution shifts generally.