Classifying accuracy of fin whale range estimates from single seismic sensors

Ocean-bottom seismometers (OBSs) are used increasingly often to track baleen whale signals, employing single-station ranging techniques such as the three-component (3C) method. By using the orientation of ground motion from OBS components, the 3C method provides robust range estimates of direct-path signals within a validity range that relates to instrument depth. Consequently, the method requires a classification process to determine whether a signal falls within the validity range. Fin whale tracks, composed of 20-Hz notes from six locations, were used to develop and evaluate three classification models: decision trees (DTs), generalized additive models, and neural networks. Models were trained using different data combinations and incorporated a comprehensive set of variables related to channel amplitude, signal quality, polarization, and estimated signal angles. The DT achieved the highest performance, reaching an accuracy of 0.94 on the test data. Key variables for predicting the validity of the 3C ranges included the difference between observed horizontal-to-vertical amplitude ratios and its theoretical value, polarization metrics, and the amplitude of one horizontally oriented OBS component (Y-channel). The resulting framework contributes to improving the utility of seismic data for biological studies, which are critical for marine mammal monitoring and conservation strategies.