Extending species-area relationships into the realm of ecoacoustics: The soundscape-area relationship

The rise in species richness with area is one of the few ironclad ecological relationships. Yet, little is known about the spatial scaling of alternative dimensions of diversity. Here, we provide empirical evidence for a relationship between the richness of acoustic traits emanating from a landscape, or soundscape richness, and island area, which we term the SoundScape-Area Relationship (SSAR). We show a positive relationship between the gamma soundscape richness and island area. This relationship breaks down at the smallest spatial scales, indicating a small-island effect. Moreover, we demonstrate a positive spatial scaling of the plot-scale alpha soundscape richness, but not the beta soundscape turnover, suggesting disproportionate effects combined with acoustic niche partitioning as an underlying mechanism. We conclude that the general scaling of biodiversity can be extended into the realm of ecoacoustics, implying soundscape metrics are sensitive to fundamental ecological patterns and useful in disentangling their complex mechanistic drivers.

Methods
Raw acoustic data:

Acoustic data were collected at the Balbina Hydroelectric Reservoir in Brazilian Amazonia. We used a standardized spatial sampling design, scaling the number of plots per island with island size. Long-duration acoustic surveys were conducted at Balbina between July-December 2015, sampling 74 forest islands (Bueno et al. 2020). The number of sampling plots per island ranged from 1-7 and increased with island size. At each plot, a passive acoustic sensor was deployed on a tree trunk at 1.5 m height with the microphone pointing downward. Each sensor consisted of an LG smartphone in a waterproof case linked to an omnidirectional microphone, set to record 1 in every 5 min at a sampling rate of 44.1 kHz for 4-10 days using the ARBIMON Touch application ( arbimon.rfcx.org). Due to poor recording quality, and to retain proportional sampling, some sites were excluded from the study, retaining 69 sampling plots (1-4 plots per island) on 49 islands (0.45 - 668.03 ha). 

Due to the size of the raw sound file dataset, recordings could not be uploaded onto Dryad. The recordings are stored on an external repository specialized in large ecoacoustic datasets and can be accessed here: https://arbimon.org/p/balbina/insights. 

Soundscape richness data:

To quantify the diversity of acoustic traits emanating from the landscape, we followed the analytical pipeline outlined in Luypaert et al. (2022) to calculate the soundscape richness index. This acoustic index was previously shown to correlate positively with soniferous species richness at Balbina (Luypaert et al. 2022). To capture ecological patterns without identifying species, the pipeline quantifies the richness of Operational Sound Units (OSUs), a unit of measurement that groups sounds by their shared spectro-temporal coordinates in the 24-hour acoustic space in which species produce sound. 

To ensure consistency in temporal sampling efforts, we designated a 5-day period for sampling across all study plots. Using a sampling rate of 44,100 Hz and a window length of 256 frames, we calculated the Acoustic Cover (CVR) spectral index for each 1-minute sound file at each plot. To do this, we used the 'ss_index_calc' function from the 'soundscapeR' R-package, which is under development on GitHub. The CVR index consists of a series of values, each corresponding to a specific frequency bin within a 1-minute spectrogram. Each bin's CVR value represents the percentage of cells surpassing a 3-dB threshold, ranging from 0 to 1 (Fig. 1 in Luypaert et al. 2022). We merged these CVR-index files for each plot chronologically, resulting in a frequency-by-time dataframe that contains the CVR-index values. By employing the ‘IsoData’ binarization algorithm, we converted raw CVR-index values into a binary variable. Doing so, we determined whether OSUs were detected within each 24-hour sample of the soundscape (Fig. 2 in Luypaert et al. 2022). Subsequently, an incidence matrix was constructed for each plot, providing information on the detection or non-detection of OSUs in each 24-hour soundscape sample throughout the 5-day acoustic survey. These incidence matrices serve as the foundation for all subsequent computations related to soundscape richness.

The resulting incidence matrices are stored in the repository under ".~/Data/RData/inc_mat.rds". All code to perform soundscape richness computations can be found in RNotebook 2: "2_Soundscape_metrics.Rmd".

Predictor variables - island isolation:

As islands at Balbina may contain non-forest patches lacking an arborescent cover, we focused on soundscapes produced by forest-dwelling species. Thus, island size was calculated as the total forest area per island, omitting areas of non-forest vegetation or bare soil. We downloaded a classified image from MapBiomas (30m resolution; Souza et al. 2020) and calculated the amount of ‘dense forest’ per island (pixel value 3), as other pixel values contained either heavily degraded or non-forest cover types (Bueno et al. 2020).

Many different definitions for island isolation exist, and the most appropriate metric likely varies between ecosystems, island type (i.e. oceanic vs land-bridge), taxonomic groups, and more (Itescu et al. 2020). To assess whether different metrics influence the perceived impact of isolation on the soundscape richness, we calculated three metrics: (i) distance to nearest mainland (DNM); (ii) distance to nearest island (DNI); and (iii) proportion of water (PW) within a buffer around the island perimeter.

The optimal scale-of-effect for the ’proportion of water in surrounding matrix’ isolation variable (see Jackson & Fahrig 2015) was determined by trialing 40 different buffer sizes (50-2000m at 50m intervals), choosing the spatial scale at which isolation attained the strongest relationship with soundscape richness.

The required GIS files to calculate the isolation variables can be found under ".~/Data/QGIS_data". The code to extract isolation variables and determine the scale-of-effect is found in RNotebook 1: "1_Metadata_preparation.Rmd" and RNotebook 5: "5_Isolation_scale_of_effect.Rmd".

Subsequent analyses:

All other required files are produced by following the RNotebook files provided in this repository in sequence from RNotebook 1 - 10.

Note for downloading the dataset:

Some users may experience difficulties downloading or unzipping the data in this repository. If you are experiencing issues with the data download, be patient and only click the download button once to avoid overwhelming the Dryad download prompt. This dataset contains many thousands of files, corresponding to the individual CVR-index value files for each 1-minute file in the study, which can lead to longer download times. If you experience difficulties unzipping the downloaded data file, particularly on Mac computers, consider using the following data unpacking software: https://theunarchiver.com. If you still experience issues after following these steps, you can reach out at [e-mail].