Fig 1.
Section I: General characteristics of a coral reef soundscape and presentation of the recordings made on the island of Bora-Bora that serve as example data in the present study. The base map was reproduced from [59]. Section II: Presentation of CoralSoundExplorer software (analysis workflow, graphical output and measurable metrics). Section III: Results obtained with CoralSoundExplorer from the sample recordings made on Bora-Bora. Section IV: Parametric study of CoralSoundExplorer settings using the Bora-Bora dataset. Supplementary sections: S1 Text: Detailed Methodology of CoralSoundExplorer, S2 Text: Parametric study of CoralSoundExplorer, S3 Text: Software installation procedure and instructions.
Fig 2.
Soundscapes diversity in a coral reef.
(A) Spectrogram of a reef soundscape recorded during the day. Fish sounds (yellow square) are mostly concentrated at low frequencies (< 1 kHz). (B) Spectrogram of a reef soundscape recorded at night showing an overall increase of sounds levels mostly due to an increase in the number of snapping shrimp sounds. An unidentified fish sound is also present (yellow square). (C) Spectrogram showing the broadband noise of a passing boat, with higher intensity during the first 5 seconds. (D) Spectrogram showing broadband masking effect caused by rain. Spectrograms were plotted with CoralSoundExplorer, using FFT window size = 2048 samples, sampling frequency = 44100 Hz.
Fig 3.
Locations of French Polynesia and the recording sites on the Bora Baro Island.
The base map was reproduced from [59]. (A) World map showing the location of the French Polynesia archipelago (red square) and the world distribution of coral reefs area (blue square). (B) Map of Bora-Bora showing the three recording sites (boat site, undisturbed site, and tourist site). In dark grey: land areas; in light grey: reef area.
Fig 4.
(A) From sound recordings to graphical representations and distance matrices. Raw sound recordings, associated with predefined discrete (categorical) or continuous labels (e.g., recording locations and times), are cut into one-second extracts and projected into an acoustic space (using mel-spectrum, mel-spectrogram or VGGish embedding). These one-per-second samples are then aggregated according to the chosen integration time (15 seconds for Bora-Bora’s study case) and their acoustic descriptors are averaged over this duration. Dimensions containing no information (i.e., descriptors equal to 0 or constant for the whole dataset) are eliminated. This step is followed by a normalization process on each remaining dimension (robust scaling). The population of aggregated sample descriptors is projected into a two- or three-dimensional space using the UMAP or PCA method for visualization purposes. A UMAP (Xd) operation with a fixed dimensionality (X) is parallelly repeated N times. For each iteration, the distances between each pair of sounds in the acoustic space are calculated to establish distance matrices (Dist Matn with 1 ≤ n ≤ N), which are then averaged (Average Dist Mat). (B) Analyses based on UMAP embeddings. Top panel: Evaluation of the suitability of predefined labels to describe the organization of the acoustic data. This evaluation can be made visually from the 2/3D UMAP or PCA of the initial acoustic space, where each sound is identified by a colored dot according to a predefined label. It is also quantified from the average distance matrix by calculating silhouette indices (Pairwise Silhouette index). Middle panel: Unsupervised sound clustering. The HDBSCAN algorithm identifies sound clusters based on their proximity in the average distance matrix. These clusters are associated with so-called unsupervised labels (Unsupervised cat. labels) that identify sounds in acoustic space. The unsupervised clusters can be visualized in a 2/3D projection (visual UMAP or PCA) and compared with clusters derived from predefined labels (Predefined cat. labels) to assess the correspondence between the two categories of labels (Unsupervised/Predefined comparison matrix). Bottom panel: The temporal trajectories of soundscapes in sound space (Rolling path) are calculated and plotted based on the position of UMAP points relative to the starting position (recording start time). This calculation is averaged over the N UMAPs (see text for details).
Fig 5.
Analysis offered by CoralSoundExplorer: Assessing the suitability of predefined labels for describing data organization (here 3D UMAP view).
Graphical display possibilities: Color points (sounds) according to the values of a categorical variable. Pairwise silhouette indices matrix for one categorical variable. Export possibilities: UMAP/PCA plots (.png,.svg), Pairwise silhouette indices matrix (.csv) and its plot (.png,.svg).
Fig 6.
Analysis offered by CoralSoundExplorer: Unsupervised sound clustering (here 3D UMAP view).
Graphical display possibilities: Color points (sounds) according to the values of unsupervised clusters id (cluster label -1 is related to unclassified sound samples), Contingency matrix (i.e., percentage of each category of one categorical variable within each category of another categorical variable). Export possibilities: UMAP/PCA plots (.png,.svg), contingency matrix (.csv) and its plot (.png,.svg).
Fig 7.
Analysis offered by CoralSoundExplorer: Temporal trajectory of soundscapes (here 3D UMAP view).
Graphical display possibilities: plot predefined time trajectories, plot “Relative distances [from average starting point]”. Export possibilities: UMAP/PCA plots (.png,.svg), “Relative distances [from average starting point]” (.csv) and their plots (.png,.svg).
Fig 8.
Other display facilities and data exploration tools provided by CoralSoundExplorer 1 (3D UMAP view).
(A) Display features: color points (sounds) according to the values of a categorical variable, highlight points according to the values of a categorical variable (the same or another). Export possibilities: UMAP/PCA plot (.png,.svg). (B) Display features: sounds’ information and their spectrograms after clicking on the corresponding point. An audio player is provided to listen to the selected sound. Export possibilities: sound (.wav).
Fig 9.
Other display facilities and data exploration tools provided by CoralSoundExplorer 2 (3D UMAP view).
(A) Display features: color points (sounds) using a continuous scale according to the starting time of each point (starting time of the recording). Export possibilities: UMAP/PCA plot (.png,.svg). (B) color only points corresponding to sounds recorded in a specified time interval. Export possibilities: UMAP/PCA plot (.png,.svg).
Fig 10.
Three-dimensional visualization of the acoustic space of coral reef soundscapes using CoralSoundExplorer software.
Each dot corresponds to 15 seconds of sound recording. Recordings were made at three sites on the Bora-Bora reef (undisturbed site, tourist site and boat site), over 24-hour periods, defining a day period and a night period, repeated over 3 non-consecutive 24-hour days (replicates 1, 2 and 3). In panels (A), (B) and (C), each sound is colored according to one of these predefined labels. Panel (D) combines the three labels using the following nomenclature site/replica/period (site: undisturbed as Und., tourist as Tour., and boat as Boat, replicate number 1, 2, or 3, day and night as D or N) to form 18 predefined clusters (3*2*3 = 18). These graphical representations make it easy to explore soundscapes and qualitatively grasp their correspondence with predefined labels. CoralSoundExplorer’s interface is interactive, allowing the 2D or 3D representation to be oriented as desired and zoomed in on areas of interest. Each dot is associated with the corresponding sound recording, which can be listened to and exported with a click.
Fig 11.
Quantification of acoustic similarity (silhouette indices) between reef sounds recorded at Bora-Bora.
Silhouette indices are calculated from 100 3D UMAPs. The continuous color scale represents the index value (0: the two groups are similar, signifying homogeneous soundscapes; 1: the two groups are completely dissimilar). (A) Recordings are labeled by site only (tour.: tourist, boat: boat site, und.: undisturbed). (B) Recordings are labeled by site, day/night period (D: day, N: night) and replicate number (3 replicates, corresponding to 3 non-consecutive 24-hour recording periods).
Fig 12.
Unsupervised clustering of soundscapes recorded on the coral reefs of Bora-Bora.
(A) UMAP visualization using HDBSCAN with the Excess of Mass (EOM) clustering method, which identifies two clusters separating the boat recording site from the other two sites (tourist and undisturbed). (B) Contingency matrix of the different categories of the composite label for each of the two unsupervised acoustic clusters obtained with the excess of mass method (EOM). (C) UMAP visualization using HDBSCAN with the Leaf clustering method. This method identifies eight clusters. These clusters can then be linked to specific sound sources (e.g., boat noise) by manually exploring the recordings (by clicking on the dots, inspecting the spectrograms and listening to the sounds). (D) Contingency matrix of the different categories of the composite label for each of the eight unsupervised acoustic clusters obtained with the Leaf clustering method.
Fig 13.
Time path of soundscapes recorded at the Bora-Bora undisturbed site.
The color scale of the trajectories is related to time. (A) shows the 24-hour trajectory for three different recording days (the three replicates). (B), (C), and (D) show the trajectory for each replicate. Acoustic clusters were generated unsupervised using the Leaf method.
Fig 14.
Time path of soundscapes recorded at the Bora-Bora tourist site.
The color scale of the trajectories is related to time. (A) shows the 24-hour trajectory for three different recording days (the three replicates). (B), (C), and (D) show the trajectory for each replicate. Acoustic clusters were generated unsupervised using the Leaf method.
Fig 15.
Time path of soundscapes recorded at the Bora-Bora boat site.
The color scale of the trajectories is related to time. (A) shows the 24-hour trajectory for three different recording days (the three replicates). (B), (C), and (D) show the trajectory for each replicate. Acoustic clusters were generated unsupervised using the Leaf method.
Fig 16.
Time trajectories of the three recording sites.
(A) Relative trajectories of the three replicates of the boat site over time. (B) Relative trajectories of the three replicates of the undisturbed site over time. (C) Relative trajectories of the three replicates of the tourist site over time. (D) Relative trajectories of replicate 3 of the three sites over time.