Arctic diel and circadian acoustic pattern of Orcas, Fin, and Humpback whales revealed by deep learning from two months of continuous recordings

In the northern Norwegian fjords, orcas, fin and humpback whales gather each winter to feed on herring, overlapping with intense human activities such as fishing and whale watching. To assess how anthropophony and geophony influence their acoustic behavior, we conducted two months of continuous passive acoustic monitoring in Kvænangen fjord during winter 2022–2023. Whale vocalizations were automatically detected using a deep learning framework based on YOLOv5, enabling quantification of species-specific acoustic presence and activity. Ambient noise was estimated from power spectral density. Low- and high-noise conditions were identified for geophony and anthropophony using a three-step filtering procedure. Model performances were evaluated under various noise conditions to ensure robust and consistent detection accuracy. Analyzes were then performed to characterize diel, circadian and daily patterns of acoustic activity. All three species were detected nearly continuously, with orcas activity peaking in November. Acoustic patterns were strongly influenced by noise: orcas and fin whales were less vocally active with increasing anthropophony (ρ< −0.31, p < 0.05), while humpback whales showed a time-dependent response, increasing vocal activity on short timescales (p < 0.01) but decreasing over longer periods (ρ = −0.33, p = 0.008). Geophony was associated with reduced acoustic presence for all three species on a daily basis (ρ< −0.34, p < 0.01), suggesting changes in spatial distribution or vocal behavior. Positive correlations between orcas and humpback whales vocal behavior indicated potential concurrent feeding. These findings revealed species- and timescale-specific acoustic responses to noise and illustrate how deep-learning can enhance ecoacoustic monitoring.