The data presented in this study was collected in an ecologically relevant area named Xixova´-Japu´ı State Park (XJSP), located in the southwest of the Santos Estuarine System (São Paulo State, Brazil). The park encompasses a marine area inside the Santos Bay and an adjacent inland region of tropical rainforest. Situated in the vicinity of Santos Harbour, the largest port in Brazil, the conservation unit is within a region severely affected by environmental impacts and port activities Araujo et al. (2013). Specifically, the data presented in our study was collected underwater, in the marine area of the conservation unit, as illustrated in Figure 1.

A custom-made autonomous passive monitoring system Caldas-Morgan et al. (2015) was deployed at the monitoring site. The autonomous system continuously recorded underwater sound from February 4 to 4 March 2015. The recording was carried out at a 11.025 kHz sampling rate (fs), with 16-bit resolution, and was continuously stored in a local SD card, in WAV files of 15-min durations.

In order to obtain a time-spectral representation of the entire collected dataset, daily PSD matrices were computed for each monitored day. It is worth noting that this procedure entails a significant volume of data reduction, enabling quick visualization of large datasets while retaining relevant information about the monitored site, which will be discussed in the following sections with more details.

The procedure used to calculate the daily PSD matrices is explained as follows: for each WAV file recorded on a specific day (from 04/02/2015 to 04/03/2015), a respective PSD matrix was computed using the Welch method Welch (1967) with a 1-s Hamming window, 1,025 frequency points, 50% overlap, and 60-s temporal signal segments. Subsequently, all PSD matrices pertaining to the same day were concatenated, thus producing daily PSD matrices Pulgar-Pantaleon et al. (2023).

Daily Power Spectral Density (PSD) matrices were utilized both for the calculation of Sound Pressure Levels (SPL) for each respective day and for facilitating the visualization of 24-h spectrograms Sánchez- Gendriz and Padovese (2017b). SPL, indicative of a sound signal’s energy, can be deduced from the PSD matrices based on Parseval’s theorem. This theorem posits that the total energy of a signal can be represented equivalently in the time or frequency domain. Consequently, by aggregating the daily PSD matrices over specific frequency bands, a time series representing the energy within those bands is generated. This series serves as a proxy for identifying events localized within certain spectral ranges, such as fish choruses.

Previous investigations have identified four distinct fish choruses (Ch1, Ch2, Ch3, and Ch4) characterized by regular daily occurrences, primarily at dawn and dusk periods Sánchez-Gendriz and Padovese (2016), Sánchez-Gendriz and Padovese (2017b). To ascertain the onset, peak, and cessation of each chorus type, manual labeling was conducted through the spectrogram’s visual examination and the SPL time series for designated frequency bands. Specifically, the 0.3–1 kHz band was applied for Ch1 annotations; the 0.3–1.5 kHz band for Ch2 and Ch4; and the 1.5–2.0 kHz band for Ch3 annotations. These temporal markers were meticulously noted on the SPL plots Sánchez-Gendriz and Padovese (2017b).

The selection of different frequency bands for manual annotation was based on the distinct and specific spectral characteristics of each type of chorus. These differences in spectral concentration allowed for the discrimination of chorus types, as previously noted in the literature and illustrated in the long-term spectrograms presented in this study. The meticulous annotation of fish chorus timings serves as a reference for subsequent analyses, enabling a comparative evaluation of the outcomes derived from the PCA methodology delineated in this study.

To achieve efficient data representation while maintaining relevant temporal and frequency information, spectral dimensionality reduction was applied to the analyzed data.

Firstly, spectral downsampling was performed at a ratio of 1:5, reducing the original 1,025 frequency bins to 205 points. Next were selected the frequency bins within the 200–2,000 Hz range, which represent chorus activities as illustrated in Section 3.1, resulting in 67 spectral points. PCA was then applied to the spectrally subsampled data within this frequency band. The number of principal components was chosen to capture 95% of the variance.

PCA was utilized to compress the spectral dimensions of the 24-h spectrograms, facilitating subsequent analysis and visualization. This reduction was achieved by concatenating daily spectrograms and treating the frequency dimensions as features projected onto the principal component space, with time bins serving as individual instances. The spectrograms were constructed with 1,025 frequency points and 1,440 time points per day, corresponding to a 1-min resolution.

This methodological approach effectively reduces the dimensionality of the spectrogram data while preserving the essential spectral characteristics that define the temporal patterns of fish choruses. The effectiveness of this approach will be demonstrated in the subsequent section.

To reconstruct the spectrograms, we performed an inverse PCA transformation, projecting the data back from the principal component space to the original frequency space. Given the spectral downsampling performed initially, linear interpolation was employed during reconstruction to restore the frequency resolution to its original state. This interpolation step ensures that the visual and analytical resolution of approximately 5 Hz is maintained, thereby preserving the fidelity of the spectral information. The reconstructed spectrograms retain the key features of the original data, enabling a detailed analysis of the temporal and spectral patterns of fish choruses.

Acoustic heatmaps, also known as Acoustic Maps, were developed utilizing the PCA derived from the 24-h spectrograms. These maps are intended to augment the visual representation and aid in the identification of dynamic acoustic patterns within the analyzed soundscape.

The creation of Acoustic Maps entailed several steps: Initially, for each principal component (PC) extracted at a 1-min temporal resolution, a corresponding 24-h time series was constructed. To achieve temporal smoothness and enhance interpretability, a moving average filter was subsequently applied to these time series, adhering to the methodology described in Sanchez-Gendriz (2021).

These filtered time series were then arranged into a two-dimensional (2D) array to form a heatmap matrix, where each row represented a 1-min interval within the 24-h cycle, and each column corresponded to a day within the study period. To further improve the readability of the Acoustic Maps, horizontal interpolation was performed on the heatmap matrix. This step is especially beneficial for datasets encompassing a relatively short duration, as it significantly enhances the visual continuity of the data representation. In the context of this study’s 29-day dataset, horizontal interpolation was instrumental in refining the visual quality of the maps, a point that will be detailed in the following sections.

We used the annotated timings for each of the four chorus types across individual days to create a map of the acoustic activity. For ease of identification, choruses were labeled sequentially based on their occurrence in time. Therefore, we used the labels Ch1, Ch2, Ch3, and Ch4 to denote these continuous sound events. Figure 2 provides a twofold representation: a) An Acoustic Map showcasing the activity of the choruses. Notably, the Acoustic Map in Figure 2A has a unique characteristic: each time series, representing a specific chorus, spans a 24-h duration with a 1-min time resolution. If chorus i occurs, it adopts a value of i, and during non-occurrences, the value is 0. By using different colors for each chorus and a separate colormap for periods with no chorus activity, we can superimpose the time occurrence of all four chorus types onto the same Acoustic Map. b) In Figure 2B, a depiction of a 24-h spectrogram with annotations for the four chorus types, is shown. A keen observation from the sample days in Figure 2B shows that chorus Ch1 and Ch2 overlap in their time and frequency band concentration, but they vary significantly in their spectral shape. In contrast, while Ch2 and Ch4 exhibit similar spectral patterns, they differ in their time of occurrence. Ch3 stands apart both in terms of its spectral content and its temporal proximity to the end of Ch2.

Long-term spectrograms have been instrumental in characterizing soundscapes in both terrestrial and marine environments Guan et al. (2015); Sankupellay et al. (2015), especially when monitoring fish communities Carric¸o et al. (2020). Even though spectrograms transform audible information into a visual domain, the inherent data reduction is worth noting. Aural analysis of a 24-h sound file is a lengthy process, while 24-h spectrograms allow for rapid exploration, often with minimal training required for the recognition of prolonged sound events like fish choruses.

For instance, Figure 3 illustrates a 7-day spectrogram derived from our dataset. A recurring pattern of fish choruses is evident. In the top panel, the spectrogram covers a frequency band from 0 to 5 kHz. The bottom panel presents annotations of notable fish choruses demarcated by rectangular boxes. Here we illustrate the long-term spectrogram, showing the annotations of choruses on the spectrograms. For that we will illustrate as representation 7 days of the data. Extended long-term spectrograms covering 28 complete days are included as (Supplementary Figure S1–Original spectrograms, Supplementary Figure S2–PCA reconstructed spectrograms).

As part of the visualization pipeline, for representing values in spectrograms, heatmaps, and spectral information, we have transformed the magnitude numerically using A _dB = 20 log₁₀(A). This logarithmic transformation is a standard practice in acoustics to effectively represent the wide dynamic range of sound pressure levels, making it easier to visualize and interpret variations in acoustic signals.

In this section, we present the results obtained from applying PCA to the spectral dimension of long-term spectrograms. Following spectral downsampling at a ratio of 1:5, we retained 205 frequency components from the original 1,025 frequency bins. Based on prior analysis (Section 3.1), we focused on the frequency range from 200 to 2,000 Hz, resulting in 67 spectral components, as this range contains relevant acoustic activity.

PCA was then applied to the spectrally subsampled spectrograms. As illustrated in Figure 4, seven principal components were selected based on the cumulative variance criterion, accounting for 95% of the variance in the original data. This approach effectively reduces the dimensionality of the spectrogram data while preserving the essential spectral characteristics that define the temporal patterns of fish choruses, as demonstrated in Section 3.2.2. The spectral contribution of each component is detailed in the Supplementary Material.

The principal components are initially illustrated as Acoustic Maps. Subsequently, we provide a visualization of 7 days of reconstructed spectrograms obtained from a representative period and compare them with the original spectrograms.

Additionally, we have included Supplementary Figure S3, which illustrates the contribution of each individual PC. This figure was obtained by performing an inverse PCA using each specific PC represented in the figure. The figure also includes the contributions of the combined seven Principal Components (1:7 PCs). It is important to note that while the spectral contribution of individual PCs can be challenging to interpret, they provide valuable insights when combined. The individual components show patterns that are a linear combination of the energy concentrated in the frequency bins of the original spectrograms. These patterns, while not corresponding to contiguous frequency ranges, capture different aspects of the spectral characteristics. The individual temporal patterns of each PC can be representative of the dynamics of the soundscape, as previously illustrated in Figure 5 and further complemented by the results presented in Supplementary Figure S3.