Skip to main content

Data from: Low frequency sampling rates are effective to record bottlenose dolphins

Cite this dataset

Romeu, Bianca et al. (2021). Data from: Low frequency sampling rates are effective to record bottlenose dolphins [Dataset]. Dryad.


Acoustic monitoring in cetacean studies is an effective but expensive approach. This is partly because of the high sampling rate required by acoustic devices when recording high-frequency echolocation clicks. However, the proportion of recording echolocation clicks at different frequencies is unknown for many species, including bottlenose dolphins. Here, we investigated the echolocation clicks for two subspecies of bottlenose dolphins in the western South Atlantic Ocean. The possibility of record echolocation clicks at 24 and 48 kHz was assessed by two approaches. First, we considered the clicks in the frequency range up to 96 kHz. We found a loss of 0.95-13.90% of echolocation clicks in the frequency range below 24 kHz, and 0.01-0.42% below 48 kHz, to each subspecies. Then, we evaluated these recordings downsampled at 48 and 96 kHz and confirmed that echolocation clicks are recorded at these lower frequencies, with some loss. Therefore, despite reaching high frequencies, the clicks can also be recorded at lower frequencies because echolocation clicks from bottlenose dolphins are broadband. We concluded that ecological studies based on presence-absence data are still effective for bottlenose dolphins when acoustic devices with a limited sampling rate are used.


Data sampling

Echolocation clicks of two bottlenose dolphin subspecies that occur across different environments in the western South Atlantic Ocean (wSAO) were analyzed. These were Tursiops truncatus truncatus, which are found in open waters, and T. t. gephyreus, which are found in coastal areas. We collected T. t. gephyreus data from the lagoon system adjacent to Laguna (28º20’S, 48º50’ W), southern Brazil. This lagoon system has depths of between 0.4 and 13 m, with an average depth of 1.8 m. The data were collected from December 4th to 12th, 2017, using a 4.4 m inflatable research boat with a 30 hp outboard engine. The recordings were made with the engine off at sites with depths of 2.5 to 5.0 m. The hydrophone was positioned 1.5 m from the surface and 15 groups were sampled using recordings that lasted around 4 to 34 min. Each sampled group contained two to seven dolphins that were displaying travel or foraging behavior.

The T. t. truncatus were opportunistically recorded in open waters by the Cetaceans Project Monitoring in the Santos Basin area, south and southeastern Brazil. The data were collected on February 14th and 24th, 2017, from three sites near the coast, at depths of 21, 28, and 35 m. Photographic records confirmed the subspecies T. t. truncatus. The group sizes ranged from 12 to 30 individuals, and two groups showed travelling behavior, and one was travelling/foraging. The recordings lasted approximately 7 to16 min. Two types of vessels were used: a 23.7 m mini supply vessel with two 325 hp engines and a 5 m inflatable boat with a 50 hp outboard engine. The hydrophone was positioned 5 m from the surface and the engines were off during the recordings.

All the recordings from both subspecies were made using a Reson TC 4032 hydrophone (0,005-120 kHz) connected to a Sony PCM-D100 recorder with a sampling rate of 192 kHz/24 bit (maximum frequency of 96 kHz – Nyquist frequency).


Sample processing

Two different approaches were used to analyze if echolocation clicks recorded in a frequency range up to 96 kHz can be recorded in frequency ranges below 24 and 48 kHz. First, the recordings were assessed visually through spectrograms to quantify the proportion of echolocation clicks occurring in each frequency range of interest. That is, each click observed in the spectrograms was verified if it occurred below and above frequency thresholds of 24 and 48 kHz. Second, the recordings were downsampled at 48 and 96 kHz (Nyquist frequency of 24 and 48 kHz, respectively) and processed by an automatic signal detector, to test if echolocation clicks are recorded when recordings are made at a lower sampling rate.


Proportion of echolocation clicks in each frequency range

The recordings from each dolphin group were fractionated into 1-minute samples to standardize the sample units. However, some recording durations were longer than multiples of 1 minute. Therefore, to use all records, we included samples of less than 1 minute in the analysis, that is those sections beyond the 1-minute limit of the last sample. We analyzed the samples in spectrograms using Raven Pro 1.6.1 software, with a sampling rate of 192 kHz/ 24 bit, Hann window, 512 points in size, and overlap of 50%. Raven uses Fourier transform to create spectrograms as a frequency domain representation of the signal.

The spectrograms were visually inspected to identify echolocation clicks. These clicks were defined and identified as those belonging to click trains with interclick intervals longer than 10 ms. The total number of echolocation clicks recorded up to 96 kHz was counted manually. Each echolocation click was visually inspected to verify its occurrence below and/or above frequency thresholds of 24 kHz and 48 kHz given the full frequency range of the recordings. Then clicks were counted in each frequency threshold to estimate the proportion of clicks that appear in each threshold. The total number of echolocation clicks recorded up to 96 kHz was paired with the total number of echolocation clicks that occurred in 24 kHz and/or 48 kHz thresholds. This is because the same click, when visualized below 24 kHz, was counted as “up to 24 kHz” and “up to 48 kHz”, while the clicks with frequencies above 24 kHz, but below 48 kHz, were only counted as “up to 48 kHz”.


Downsampling and automatic detections

A sampled high-frequency signal can change when it is sampled in different, lower sampling rates. Because of this, in our second approach, the recordings were downsampled at 48 and 96 kHz to analyze if echolocation clicks are recorded at 24 and 48 kHz, using an automatic signal detector. The downsample and the following automatic  clicks detection were made using R 3.6.0. First, a fourth-order Butterworth 15 kHz high-pass filter and an anti-aliasing Finite Impulse Response (FIR) low-pass filter to the 24 or 48 kHz were applied in the original recordings, for each corresponding downsample frequency. Then, the downsample was made using the “downsample” function of the “tuneR” R package. Next, we used the “auto_detec” function of the “warbleR" R package to detect the echolocation clicks automatically in recordings at 96, 48, and 24 kHz—in other words, the original and downsampled frequencies. The parameters to configure the “auto_detec” were defined through the “optimize_auto_detec” function from the “warbleR" R package.

The “optimize_auto_detec” function takes a selection table containing the times of each signal (echolocation clicks, in this case), and then run the automatic detection with multiple parameters to find the ones that maximizes sensitivity and specificity of signal detections compared to the selection table. Subsamples of 5-seconds from different recordings and different dolphins’ groups were used to select the echolocation clicks and validate automatic detections. The selection tables contained one subsample of each T. t. truncatus group (total of three groups), and one subsample of two different T. t. gephyreus groups. The signals were manually selected in the subsamples of each sampling rate, to define the best-adjusted parameters to detect clicks in each frequency (96, 48, and 24 kHz). The least number of subsamples from T. t. gephyreus used was due to these dolphins are from the same population and environment (lagoon system).

Even though we have selected the optimal parameters to configure the signal detector based on a priori manual detections, the sensibility and specificity of the detector were not the same to all recordings. Then, only the presence/absence of signal detections were considered, since we cannot guarantee that all detections are true positive. Furthermore, additional filters were applied to reduce the number of false-positive detections. First, all detections were filtered based on the interval of signals detected, similar to interclick intervals, excluding detections with an interval longer than 0.2 s, since echolocation clicks occur in click trains, not isolated. Second, the detection localization (in time) in the original recording was compared with detection localization in downsampled recordings, assuming the detector performance was better in recordings at 96 kHz. Since the echolocation clicks in the downsampled recordings tend to be longer than in the original record, a buffer was created around each detection of the original recordings by expanding the start and end time of detections by the length of each signal (i.e. click duration) to guarantee the match between the same detections at each frequency. Next, recordings were divided into bins of one second and matching bins from the division of the 1-minute samples from the first approach.

The bins from the original recordings (96 kHz) were binarized. Only 1-second bins with at least one signal detected present were kept and compared with the same respective bins in 24 and 48 kHz. Finally, the matches between bins with the presence of signals in downsampled recordings (24 and 48 kHz) and in original recordings (96 kHz) were quantified. Finally, this yields a proportion of seconds with signals detected in downsampled recordings given the total number of seconds with detections in the original files. These proportions in each 1-minute sample (or fractions) were used to calculate the probability to detect echolocation clicks at 24 or 48 kHz.


Sampling rate assessment

First, the hypothesis that bottlenose dolphins echolocation clicks occur equally in frequencies ranges of up to 24, 48 and 96 kHz, when recorded from free-living animals, was tested. A generalized linear mixed-effects models (GLMM) was constructed using a binomial error structure with a logit link function. The proportion of echolocation clicks that occurs at each frequency range out of the total number of echolocation clicks recorded up to 96 kHz was the response variable and the frequency range and environment (lagoon system or open water) were fixed effect terms. To account for sample pairing (up to 24 kHz and up to 48 kHz), pairs of samples were treated as random effects and were used nesting within recording identities to account for autocorrelation in group composition and behavioral context during sampling.

            A candidate model was built to test the hypothesis and it was compared to a null model containing only the intercept. Models were compared using the Akaike information criterion corrected (AICc) for small samples and the Akaike weights. The model with the lowest AICc value and the highest Akaike weight was considered to be the most parsimonious and the one that better supported data variation. We simulated 10,000 datasets from the fitted model to validate the model assumption. Then, the Kolmogorov-Smirnov test was used to determine whether the deviations between the observed and expected residual distributions were significant. The random effects were validated by visually comparing a QQ-plot of the random effects quantiles against the standard normal quantiles. The assumptions were valid if most values fell along the line. Model fit was assessed using the theoretical marginal and conditional  R². Marginal R²   represents the proportion of the total variance explained by the predictors (fixed effects) and conditional  R² represents the proportion of the variance explained by both the fixed and random effects.

            Finally, the hypothesis that the echolocation clicks are recorded at frequencies up to 24 and 48 kHz, was tested. A second set of GLMM was constructed using a binomial error structure to test if echolocation clicks can be detected equally across the different sampling rates (the downsampled recordings). Considering only the bins with presence of at least one signal in the original frequency, the response variable was the proportion of bins with detections that occurs in the 1-minute samples at each downsampled frequency out of the total number of bins with detections in the equivalent sample in original frequency. The frequency and the environment (lagoon system or open water) were used as predictors, similar to the first model. The 1-minute samples were treated as random effects, nested within recording identities to account for autocorrelation in group composition and behavioral context during sampling. The next steps in this analysis, which involved candidate model build, models comparison, and the model fit assessment, were conducted as in the first model. All analyses were conducted using R 3.6.0.

Usage notes


The file contains 220 one minute samples (rows) and seven columns: dolphin_group contains a string for each group recorded, with a prefix for coastal (COAST) and offshore (OFF) recordings, and the ID of the group separated by an underline; environment contains information of the environment where the recording was made (coastal or offshore); recording contains the number of the recording file; duration contains the sample duration in seconds (up to 60 s); click_L24 contains the number of clicks only detected in frequencies below 24 kHz; clicks_L48 contains the number of clicks detected in frequencies up to 48 kHz; clicks_total contains all the clicks detected up to 96 kHz.



The file contains 18 locations of recordings, with four columns. environment contains information of whether the recordings were made in the coastal lagoon system (Lagoon_s) or offshore (Open_w); group contains the group ID for each recording in the environment; Lat and Long contains the latitude and longitude in degrees, respectively.



R code to prepare the environment before running analysis.R, figure_seewave.R, map.R. Briefly, setup.R installs and load R packages, load custom functions and define colours to plot figures.



Contains the R code to run the analysis and reproduce the results.



Contains the R code to plot spectrograms from audio samples above_24khz.wav and above_48khz.wav. plotSpectro function is loaded from setup.R.



Contains the R code to generate the maps.



Contains the R code to apply filters and downsample the recordings.



Contains the R code to filter detections based on the interval of signals detected, to analyze the overlapping clicks in each downsampled frequency, and to analyze the results.



Contains the R code to run the automatic detection of clicks, with the parameters used in each frequency and for each environment.



Contains the R code to consider the detections by time bins and bins with detections per minute, and to analyze the results.



Contains the R code to run the 'optimize_auto_detect' funcion.



Sound archive of a sample subset, to plot the spectrogram figure.



 Sound archive of a sample subset, to plot the spectrogram figure.


Coordenação de Aperfeicoamento de Pessoal de Nível Superior, Award: 001