Machine Learning Techniques for Quantifying Geographic Variation in Leach’s Storm-Petrel (Hydrobates leucorhous) Vocalizations

Leach’s storm-petrel (Hydrobates leucorhous), a nocturnal seabird that breeds in the Northern hemisphere, has geographically separated populations in the North Atlantic and North Pacific. Although some mixing occurs during the non-breeding season, genetic evidence demonstrates that these populations are diverging. However, genetic information for the study of phylogenetics can be costly and time-consuming to obtain. Vocalizations could offer a more cost-effective way of obtaining similar information (or could be used in conjunction with it). In this chapter, we examine if the chatter call of the Leach’s storm-petrel can be used to classify the Atlantic and Pacific populations. We used a machine learning context by testing several implementations of random forests and boosted regression trees in the R and Python programming languages. We discuss the implementations with respect to accuracy, speed, and memory handling. We found that random forests from the h2o and ‘randomForest’ packages in R performed best with regards to accuracy, ‘randomForest’ and ‘gbm’ performing best with regards to speed, and ‘tensor forest’ and ‘h2o’ implementations performing best with regards to memory. Furthermore, we were able to classify the Atlantic versus Pacific populations of Leach’s storm-petrel with AUC values >0.8 (generally considered ‘good’ in ecology). We expect that this could be adapted on a larger scale to assist taxonomic classification without having to perform invasive DNA sampling on many individuals of sensitive populations.