Non-stereotypy (to species) in mysticete downsweeps

Paul Nguyen Hong Duc ^1* Christine Erbe ¹ Shyam Madhusudhana ¹ Daniel Wilkes ¹ Lachlan Gill ¹ Cristina DS Tollefsen ¹ Narissa de Bruin ¹ Aiyana Erbeking ¹ Curt Jenner ² Micheline-Nicole Jenner ² Angela Recalde-Salas ³ Chandra Paulina Salgado Kent ^4,5 Kautilya Srivastava ^1,6 Chong Wei ¹ Robert McCauley ¹

• ¹ Centre for Marine Science and Technology, Curtin University, Perth, Australia
• ² Centre for Whale Research WA, Fremantle, Western Australia, Australia
• ³ Bush Heritage Australia, Melbourne, Victoria, Australia
• ⁴ Edith Cowan University, Joondalup, Western Australia, Australia
• ⁵ Oceans Blueprint, Cogee, Australia
• ⁶ University of Western Australia, Perth, Western Australia, Australia

The Australian EEZ provides habitat for ten species of mysticete whales seasonally supporting critical life functions ranging from feeding to breeding. All of these species produce downsweeping calls, which may confound passive acoustic monitoring efforts. In an attempt to optimize a detector for Eastern Indian Ocean pygmy blue whale (EIOPBW) downsweeps, we tried a spectrogram correlator based on confirmed templates and a neural network trained on general blue whale D-calls followed by clustering algorithms. Outputs were manually validated by bioacousticians. We found that downsweeps exhibit significant variability and form a graded continuum of acoustic features, as opposed to clusters. Comparative analysis demonstrated parallels between EIOPBW call variants and downsweeps of other mysticete species, raising concerns about the reliability of assigning calls to species based solely on spectrographic features. Geographical and seasonal patterns of downsweeps were more conclusive for EIOPBW when aligned with known migratory routes and timings. Challenges in automated detection, variability in environmental noise, and human biases in manual classification were acknowledged. To improve species identification, we suggest integrating soft labeling, advanced acoustic transforms, sound propagation corrections, and cross-referenced databases. Until automated methods achieve higher reliability, passive acoustic monitoring will require a multidisciplinary approach incorporating regional ecological insights and manual validation.