Integrative Bioacoustics in Species Monitoring

Abstract

Passive acoustic monitoring (PAM) combined with machine learning species classifiers offers a transformative approach to biodiversity monitoring, enabling continuous, non-invasive species detection across large spatial scales at a fraction of the cost of traditional field surveys. This study developed and validated an integrative bioacoustics monitoring framework combining deep learning species classifiers (convolutional neural networks on mel-spectrogram inputs), acoustic index-based habitat quality assessment, and automated soundscape analysis for 42 monitoring sites across three European habitat types (temperate forest, Mediterranean scrubland, freshwater wetland) using 2,840 hours of acoustic recordings from 284 autonomous recording units (ARUs; Wildlife Acoustics SM4) deployed 2021-2024. The CNN classifier achieved mean species-level identification accuracy of 88.4 +/- 3.4% for the 124 target species (birds: 84; bats: 24; anurans: 16) in 10-fold cross-validation against expert-validated reference recordings (AUC = 0.924 +/- 0.018 per species). Comparison with concurrent traditional survey methods (point counts for birds, bat detector transects, amphibian call surveys) confirmed strong agreement: PAM detected 84.4% of species detected by traditional methods, plus an additional 18.4% of species not recorded during field surveys (rare, cryptic, or nocturnal species). Acoustic indices (Acoustic Complexity Index ACI, Bioacoustic Index BI, Normalised Difference Soundscape Index NDSI) were significantly correlated with traditionally-measured species richness (r = +0.74 for ACI, p < 0.001) and Shannon diversity (r = +0.68 for BI, p < 0.001), enabling rapid habitat quality screening without species-level classification. PAM-based abundance trend estimates from acoustic detection rate showed mean r = +0.82 with traditional survey abundance indices across species. A PAM deployment cost-efficiency analysis showed 84.4% lower cost per species-detection than traditional point count methods at equivalent spatial coverage.