Abstract
Abstract
Blue (
Balaenoptera musculus
) and fin whales (
Balaenoptera physalus
) are dominant contributors to low-frequency ocean soundscapes, yet reliably extracting their calls from long-term passive acoustic recordings is methodologically challenging. Here, we train a multi-class deep-learning detector to identify five principal blue and fin whale call types (A, B, D, 20 Hz, and 40 Hz) from low-frequency spectrograms using a Faster R-CNN architecture combined with three rounds of iterative human review and hard-negative mining, progressively expanding and rebalancing the training set using California Cooperative Oceanic Fisheries (henceforth, CalCOFI) sonobuoy and moored hydrophone recordings from the southern California Current Ecosystem. The detector was evaluated on four independent test datasets spanning multiple years, seasons, and recording platforms and then deployed on CalCOFI sonobuoy recordings collected quarterly over two decades (2004–2024). The final model achieved consistently high mean precision, recall and F1 scores for most call types (e.g., A: 0.71/0.71/0.71; B: 0.83/0.59/0.63; D: 0.79/0.84/0.80; 20 Hz: 0.87/0.74/0.78), while 40 Hz calls remained challenging (0.42/0.69/0.51), primarily due to confusion with spectrally overlapping humpback whale downsweeps. Detections were post-processed using call-specific characteristics and received-level thresholds and normalized by recording effort and detection area to derive standardized indices of call density
$$\left( \frac{\text {calls}}{\text {h} \cdot 1000 \text { km}^2}\right) $$
with uncertainty estimates. Densities were aggregated annually and show call-specific differences between inshore and offshore habitats and interannual variability associated with periods of anomalous oceanographic conditions. Inter-call interval analyses suggested seasonal stability in blue whale song, high variability in blue and fin whale social calls, and seasonal and interannual variability in fin whale song repetition rates. This study is among the first to use deep-learning to estimate baleen whale call density from decades of passive acoustic recordings in a complex soundscape.