Methodology for click and short duration pulses recognition as a dual factor in the classification of marine species

Passive acoustic monitoring is a non-invasive way to observe marine biodiversity where visual surveys are limited, but continuous recordings mix biological clicks with impulsive noise from anthropogenic and non-target biological sources, and their volume makes manual analysis infeasible; methods that do not resolve this mixture produce biased biodiversity estimates. This work presents an automated two-stage methodology: unsupervised detection of impulsive events with Isolation Forest, using normalized spectro-temporal features and requiring neither labeled training data nor absolute hydrophone calibration, followed by classification of each detected source into one of four acoustic categories (biological echolocation, stereotyped biological, anthropogenic, or ambiguous), from which biodiversity indicators are separated from non-biological noise. It was evaluated on five case studies comprising seven acoustic sources: snapping shrimp, bottlenose dolphins, orcas, Gobius cobitis , stingrays, Mustelus lenticulatus and vessels. Detection sensitivity ranged from 0.67 for snapping shrimp to 1.00 for Gobius cobitis , stingrays and Mustelus lenticulatus , with a window-level false-positive rate below 0.06; the most demanding case was an orca sequence overlapping vessel noise. Against energy, STA/LTA and Teager–Kaiser baselines, Isolation Forest achieved the highest mean F 1 (0.72) and a comparable mean AUC (0.91). The classification resolved four of the seven sources; the remaining three stayed ambiguous owing to scarce reference data. A ± 20 % perturbation of the decision thresholds left every assignment stable except the vessel, which required an entropy- and kurtosis-weighted composite score. The methodology is a viable, calibration-independent tool for acoustic source characterization supporting marine conservation monitoring.