Predicting Shape of Dives of Southern Elephant Seals Using Regression Tree Models

In recent years, the study of animal movements in the ocean has been revolutionized with massive use of miniature measuring devices providing access to complex behavioral data and associated environmental data sampled at very high frequency. However, improvements in collecting original large data sets have exceeded our ability to process them.

This work purposely presents statistical methods adapted to high frequency data that can be handled as curves. The objectives of this study are to highlight the relationships between dives of Southern elephant seals, Mirounga leonina, and the physical environnement in which elephant seals operate. Starting from a huge data set of elephant seal dives, we first show how to construct dive profiles from point-wise samples. We then propose a generalized regression tree method where the predictive variable is a curve. Regression trees models are built to predict the shape of dives using discret environmental variables (i.e. temperature at a depth of 250m) and environmental profils (i.e. temperature, salinity) as predictors. We will discuss the connection between shapes of dives and shapes of environmental profiles, and we will also discuss tree capabilities for predictor selection.