Evaluating models for classifying movement of whale-watching vessels

Abstract

A continued rise in global ocean vessel activity has led to growing concerns for the health of whales around the world. Of particular interest is the increase in recreation vessels, including those related to whale-watching activities. However, there is an absence of established procedures to identify vessels engaged in whale-watching, thus limiting the ability to quantify whale-watching impacts on whales. This study evaluates three computational classification models and their ability to utilize Automatic Identification System (AIS) data to describe wildlife-viewing vessel behaviour. These models include a density-based spatial clustering application with noise (DBSCAN), a hidden Markov model (HMM), and logistic regression (LR), all of which have been previously used to classify vessel behaviour in industries, such as fishing, shipping, and marine security. The results of each model's classification were validated against observed whale sighting data using statistical performance and accuracy metrics. The findings suggest that all three classification models sufficiently detect wildlife-viewing behaviour, but the HMM and LR had preferable performance metrics compared to DBSCAN. Further, although LR provides an informative glance at which AIS variables are most important to detecting wildlife-viewing events, the HMM has comparable performance metrics and requires less data processing. Therefore, this study recommends the use of HMM due to its computational efficiency and because it provides an accurate classification of wildlife-viewing behaviour for whale-watching vessels. The results of this study can be used to support policy decisions, monitor regulation compliance, and inform marine conservation initiatives.