Big whales and big data: A convolutional neural network approach to investigating long-term spatial-temporal patterns of blue and fin whales in the Southern California Bight

Blue (Balaenoptera musculus) and fin whales (Balaenoptera physalus) produce low-frequency calls associated with foraging, social interactions, and reproduction. Sonobuoys deployed during California Cooperative Oceanic Fisheries Investigation (CalCOFI) surveys contain recordings of distinct calls produced by these species; namely blue whale A, B, and D calls, and fin whale 20 and 40 Hz calls. Here, we fine-tuned the object detection model Faster-R-CNN with an initial dataset of 6598 labeled blue and fin whale calls and refined the model through an iterative human-in-the-loop labeling approach. The model performed well on blue whale A, B, and D, and fin whale 20 Hz calls (average precision 0.80–0.93), while fin whale 40 Hz calls proved more challenging (0.71). The best performing model was deployed on long-term CalCOFI recordings. Seasonal peaks were consistent with previous studies and long-term trends suggested an increase in the number of blue whale calls detected per survey over the study period (2004–2025). During the 2014–2016 marine heatwave, calling activity declined for both species, with blue whale calls largely restricted to the shelf break, suggesting habitat compression. This work illustrates the value of integrating passive acoustic monitoring and deep learning to investigate ecological patterns and track species’ responses to environmental change.