Temporal dynamics of animal vocal behavior affect semi-automatic identifications

Abstract Background Passive acoustic monitoring (PAM) enables continuous, non-invasive surveys of vocal species but requires careful validation of automated classifications. Critically, intra-specific vocal variation in animals across seasons is often overlooked, even though classifier performance depends on it. Using BirdNET, we analyzed over 20,000 three-second recordings from forest and alpine grassland habitats in Central Italy to assess how temporal changes in bird vocal behavior affect classification accuracy. For 37 species, we built logistic regression models relating manual validation outcomes to BirdNET confidence scores and sampling month. Results Including month as a covariate improved model performance for 32 species, revealing strong temporal variation in detection reliability linked to phenological phases. Species-specific confidence thresholds (CT) corresponding to a 90% probability of correct detection varied widely across months (ΔCT up to 0.9). Average model performance was high (AUC = 0.875; precision = 0.91). Conclusions These results demonstrate that dynamic, time-adjusted thresholds increase the robustness of semi-automatic detection workflows, enhancing the reliability of PAM-derived biodiversity assessments.