Machine Learning Methods for the Detection of Antarctic Minke Whales ( <scp> <i>Balaenoptera bonaerensis</i> </scp> ) in East Antarctica and Western Australia

ABSTRACT Passive acoustic monitoring is a cost‐effective means of studying marine mammals that inhabit remote and poorly accessible habitats. Since the 1970s, the mysterious “bio‐duck” sound has been reported throughout the Southern Ocean. In 2014, this was attributed to the Antarctic minke whale and has since been retrospectively categorized into different variants of bio‐duck calls by multiple studies across a wide geographic range. To date, more than 20 different bio‐duck variants have been identified, with intra‐ and inter‐regional variation. Our study presents a bespoke convolutional neural network (CNN) detector trained to identify bio‐duck call variants across sites in East Antarctica and Western Australia. The detector achieved high recognition performance across nine geographically distinct datasets, demonstrating strong generalization. Detector performance differed among sites, with the highest performance reported for the Antarctic sites and poorer performance in the Pilbara region of the Australian Northwest Shelf. These differences were explored, comparing the target‐signal (bio‐duck) levels to ambient noise levels. Variation in performance was likely driven by variable signal‐to‐noise ratios across testing datasets. This work presents an advancement in the acoustic monitoring of Antarctic minke whales, providing a tool for assessing their acoustic presence across diverse marine soundscapes.