Foraging behaviour of humpback whales (Megaptera novaeangliae): Automatic detection of feeding lunges from two-dimensional data

Abstract

Rorqual whales (Family Balaenopteridae) forage almost exclusively by performing so-called feeding lunges. It is difficult to study rorqual foraging behaviour through direct observation because most of the lunges are carried out deep in the water column. The introduction of high-resolution digital tags recording three-dimensional (3D) acceleration has allowed for the collection of complex movement data, increasing our understanding of their foraging behaviour. Lunges can be detected from specific movement signals in the 3D acceleration data. However, there are still datasets obtained using simpler tags, such as time-depth recorders (TDR) that have yet to be analysed, and there is a lack of automated methods for analysis. In this study, an algorithm allowing for automatic detection of lunges from these two-dimensional (2D) depth-time diving profiles is developed. This detector was applied on 16 humpback whales (Megaptera novaengliae) tagged with high-resolution multisensory tags. The data was subset into a simpler 2D format and then validated against lunge detections on the same data using 3D detectors. Optimisation of the 2D detector was done by manually changing the algorithm parameter settings, and then using ROC and AUC to find the best possible settings. The optimisation found much individual variation, with optimised settings resulting in AUCs ranging from 0.499 to 0.805. The detector was then run on data from 36 TDR tagged whales. The detector performed relatively well and have the potential to help with decreasing time and increasing standardisation of dive data analyses. Also, this method can be useful for adding knowledge in relation to rorqual whale foraging behaviour, especially when combined with additional data such as ecological information.