delphinID: deep learning models for automatic identification of Northeast Atlantic delphinid species from acoustic recordings of whistles and clicks

delphinID models are deep convolutional neural networks (CNNs) trained in Python Tensorflow 2.18.0 to automatically classify detections of delphinid whistles and clicks to species based on the distributions acoustic frequencies in groups of whistles or clicks detected using PAMGuard’s Whistle & Moan Detector (WMD) and Click Detector, respectively. delphinID click and whistle models for the Northeast Atlantic were trained using groups of detections within 4-second time windows, while the output of these models is used as a feature vector for additional Random Forest layer which provides a recording-level species classification based on acoustic information from both whistles and clicks. As the detection-based input for classification is computationally inexpensive to compute, delphinID can easily run in PAMGuard at up speeds up to 64x real-time. Both click and whistle classifiers were trained and evaluated exclusively on visually confirmed single species recordings of these species made in the Northeast Atlantic and North Sea and only evaluated on recordings spatiotemporally independent from those used to train the models. Cross-validated testing estimates delphinID to classify recordings with an average accuracy of 86%, and ranges from 80% for Delphinus delphis to 92% for Lagenorhynchus albirostris. Species that can be classified by the Northeast Atlantic delphinID models are:

• Atlantic white-sided dolphin (Lagenorhynchus acutus)
• Common bottlenose dolphin (Tursiops truncatus)
• Killer whale (Orcinus orca)
• Long-finned pilot whale (Globicephala melas)
• Risso's dolphin (Grampus griseus)
• Short-beaked common dolphin (Delphinus delphis)
• White-beaked dolphin (Lagenorhynchus albirostris)

Use delphinID easily within the widely used open-source software PAMGuard. delphinID classifiers are stored in a compressed format compatible for use within PAMGuard's Deep Learning module. Using the configuration file provided, users can import the classifiers to classify whistle detections from the Whistle & Moan Detector module and click detections from the Click Detector module. delphinID classifications exported from PAMGuard should then be processed with Collator-delphinID, a Quarto notebook designed for intuitive summary and analysis of delphinID classifications. Here, classifications from whistle and click classifiers are integrated to produce final recording-level classifications. Please refer to the README file in the zipped folder for more information and tutorials.