Abstract
Rice’s whales (Balaenoptera ricei) are a critically endangered species endemic to the Gulf of America, with recent acoustic detections suggesting a broader distribution than previously documented. This study evaluates whether a region-based convolutional neural network (rCNN) detector, trained on vocalizations from the eastern Gulf, can effectively identify Rice’s whale calls in the western Gulf. Autonomous acoustic recordings from four sites were analyzed, and detector performance was assessed by calculating precision and miss rates for each site based on validation of the automatic detections and a manual review of a 2% random subsample of the data. Results were compared to a spectrogram cross-correlation detector to evaluate detection reliability across varying acoustic conditions. The rCNN-based detector showed higher precision at most sites (EB: 0.78 vs. 0.53, CCF: 0.79 vs. 0.771), while the spectrogram cross-correlation detector detected more true calls overall (CCE: 2345 vs. 966). The spectrogram cross-correlation detector also produced higher false positive detections, especially in high-noise environments (EB: 3061 vs. 269). Neither detector demonstrated consistently strong performance, highlighting the need for a dedicated western long-moan detector. These results highlight a trade-off between accuracy and sensitivity and underscore the importance of region-specific training when using automated detectors in acoustically variable environments.