Acoustic indices have been used as proxies to monitor biodiversity and quickly analyze increasing amounts of large audio datasets. One of the main assumptions of the developing field of ecoacoustics is that greater vegetation structural complexity will lead to greater acoustic activity, but this has received limited empirical validation. Our goals were to determine the existence of habitat-specific soundscapes and to test how the vegetation structure from these habitats affects acoustic indices. We conducted this research in the Viruá National Park, north of the Brazilian Amazon. Soundscape and vegetation data were collected in 143 sites spanning eight natural and anthropogenic habitat types, which we grouped into three categories: open habitats, flooded-forests, and non-flooded forests. Thirteen acoustic indices were retrieved from 92,283 one-minute recordings to summarize the soundscapes. We used this soundscape data to build random forest (RF) models to test if each habitat displayed a unique soundscape pattern that could be identified by the acoustic indices. To determine how vegetation structure affected acoustic indices, we used linear mixed models with acoustic indices as dependent variables, six vegetation variables (canopy cover, canopy height, litter depth, shrub cover, large trees, and small trees) as independent fixed effects, and sites as a random effect.
Results/Conclusions
We found that each habitat type had predictable soundscapes. Accuracy of RF models to predict the testing dataset was 74% and 87%. We found that acoustic indices that relied on statistical features of recordings were better at identifying habitat-specific soundscapes than acoustic indices based on signal complexity. For example, variable importance of acoustic indices in distinguishing the habitats was higher for third quartile (TQ) and centroid (CENT). Canopy cover significantly affected eleven of 13 acoustic indices, while other vegetation variables (e.g. shrub cover and number of trees) were less important in explaining differences in acoustic indices. The acoustic evenness index (AEI) and skewness (SKEW) were the best indices connecting soundscapes to vegetation structure, with canopy cover explaining 81% and 52% of the variance in these indices, respectively. Acoustically rich soundscapes were linked to high canopy cover in forested habitats, while acoustically poor soundscapes were linked to low canopy cover in open habitats. This is similar to the effects of canopy cover on species richness across different taxa. Our results suggest that acoustic indices can be used together with other scalable vegetation remote sensing methods for multi-taxa animal surveys at policy-relevant extents.