There are four known cetacean species that produce NBHF echolocation clicks in the California Current Ecosystem, including harbor, Dall’s porpoise, as well as dwarf and pygmy sperm whales (Kogia sima and Kogia breviceps, respectively). These species all produce NBHF clicks with similar acoustic features (peak frequency greater than 100 kHz and 3 dB bandwidth less than 10 kHz), and to date, the species cannot be distinguished acoustically. Their presence in acoustic surveys is generally reported within a “NBHF” category; however, each species has distinct habitat preferences (Carretta 2023), and likely responds differently to anthropogenic impacts and environmental stressors. We build upon unsupervised clustering methods developed by (Griffiths et al. 2020) by adding visually-verified species assignments to train an event level classification model in a supervised approach. This work also expands on a San Francisco State University master’s thesis by Jackson VanFleet-Brown (thesis will be publicly at SFSU Thesis website).
Visually verified acoustic recordings for Dall’s and harbor porpoises from PASCAL, CCES, and Adrift surveys and NBHF clicks in the offshore waters of Baja California (Kogia spp. are the only NBHF here) were used as a training dataset (see table, right) to train a 2-stage BANTER (BioAcoustic EveNT ClassifiER) model (Rankin et al. 2017). Click detections were assigned to a detector category based on the presence of a peak frequency below 125 kHz (lo-range) and greater than 125 kHz (hi-range). A suite of features was calculated for each click detection using the R package PAMpal, and the median inter-click interval for each event was included as an event-level feature. The model was trained in an iterative way to achieve high classification accuracy and stability. The classification model was then used to predict labels on the Adrift survey data.
gt<-gt(head(TrainTable)) %>%
cols_label(Species='Species', Survey='Survey', NEvents='N Events', MedClicks='Event Clicks', TotClicks='Total Clicks')%>%
tableTheme()
gt| Species | Survey | N Events | Event Clicks | Total Clicks |
|---|---|---|---|---|
| Kspp. | CCES-drifter | 13 | 7 (3-8) | 106 |
| Pd | BC-array | 4 | 7 (4-13) | 40 |
| Pd | CalCURSeas-array | 6 | 14 (10-19) | 84 |
| Pd | PASCAL-array | 5 | 10 (5-12) | 44 |
| Pp | ADRIFT-drifter | 40 | 34 (9-116) | 2954 |
| Pp | CalCURSeas-array | 7 | 6 (5-64) | 278 |
The classification accuracy of the BANTER model was 83% overall (Figure 1), ranging from 77% for harbor porpoise to 93% for Dall’s porpoise. All classification results were greater than expected by chance (see priors in Figure 1).
Dall’s porpoise were the dominant species found in all study areas and seasons, accounting for 91% (n=2,836 of 3093 events) of NBHF detections overall. Harbor porpoises were detected in all study areas, although 54% (n=105 of 192) of events were detected during the upwelling season in Oregon. Only 2% (n=65 of 3093) of all NBHF events were attributed to Kogia spp., and 77% (n=50) of these events occurred within the San Francisco and Morro Bay study areas.
This NBHF acoustic classifier can then be used to predict on archived Adrift NBHF detections to better resolve the three separate taxa in the California Current, including Kogia spp., Dall’s and harbor porpoises (Figure 1). The overall classification accuracy of the model (83%) is acceptable, however there are several avenues to improve the model. Recently, (Zahn et al. 2024) reported significant gains in BANTER model performance by considering the ratios of third-octave levels at specific frequencies. The mean spectra of each class within our training data indicate distinct distributions of spectral energy in each class, and the inclusion of a third octave level ratio (or other similar metric) may improve model performance. Additionally, the use of an iterative training approach merits consideration.
Acoustic events that are labeled with high probabilities can be included when re-training a new model (Figure 2).
This iterative approach would be biased toward acoustic events most similar to the original training dataset, however the gains from including additional variation in an increased sample size should be evaluated. The development of a more robust classification model should be investigated, but the model we report here has sufficient classification performance to apply to Kogia-specific species habitat models, investigations of species-specific responses to disturbance, and the potential development of acoustic density estimates.