Elsevier

Ecological Informatics

Volume 59, September 2020, 101120
Ecological Informatics

Application of a sample space for the characterization of shaded coffee plantation soundscape

https://doi.org/10.1016/j.ecoinf.2020.101120Get rights and content

Highlights

  • It is established that there are base, temporary, and sporadic frequency-intensity combinations.

  • Discretization of the power spectrum allowed to model occurrence of frequency-intensity combinations.

  • Frequency-intensity combinations and samples are related by a logarithmic function.

Abstract

The plot of the frequencies that build soundscapes (power spectrum) has been used to show some of their distinctive characteristics. The application of a discretized sample space is proposed to describe, in a concise manner, the contribution of different frequency signals in a soundscape on a yearly scale. A sample space composed of frequency and intensity combinations was used to describe the shapes of power spectrums and their recurrence. The number of elements of the sample space was set so that each element had a low probability of being found in random samples. In this way, the statistical confidence in the occurrence and recurrence of the elements is increased, making it possible to validate traits in the soundscape. According to their occurrence, it was established that there are “base”, “temporary” and “sporadic” combinations of frequency-intensity. When the relative frequencies of the frequency-intensity combinations were plotted versus the number of analyzed power spectrums, it was observed that few combinations predominate in the soundscape. Further, the distribution of the relative frequencies of the combinations tends to stabilize with time; a potential function described the relative frequency of the combinations. It was found that a logarithmic accumulation curve described the relationship between the analyzed power spectrums and the detected combinations in each of the proposed frequency sections; such a function was useful to validate the sampling method.

Introduction

The research development on ecoacoustics could represent an advantage regarding time and sensitivity in a variety of ecology investigations that include population, communities, ecosystems, landscapes, and biomes (Farina and Gage, 2017). Other potential applications may include more complex issues as the evaluation of the effects of climate change on ecosystems (Krause and Farina, 2016). The concept of the acoustic community allows the study of ecosystems in a holistic and non-intrusive way, aided by technologies for processing large amounts of data (Luther, 2009).

Following the hypothesis of the “acoustic niche” (Krause, 1987) it has been tried to find order and structure in soundscapes. It is stated that the structure in the production of sound might be related to the ecological organization present in an equilibrated ecosystem (Klaus, 1993).

Descriptions of soundscapes are useful as sound is a fundamental property of nature, and they can be related to environmental quality (Pijanowski et al., 2011b). Disturbances in a landscape could be assessed by examining how acoustic variables change (Bellisario and Pijanowski, 2019; Parks et al., 2014). The description of soundscapes is intended to concisely represent all the sounds produced in a location over a period (Sueur and Farina, 2015). For this purpose, a profusion of indices has been proposed. Most acoustic indices are developed to assess the richness or complexity of soundscapes, while few others seek for a level of acoustic disparity between soundscapes (Gasc et al., 2015). A more effective application of such indices is often acquired by subdivisions of the frequency spectrum, though there is not a standardized division of the spectrum, rather the division is commonly done according to the researcher's criteria and observations. Each of the indices show particular traits of the soundscape and mostly indicate correlative information (Bellisario and Pijanowski, 2019). Indices can be related to disturbance in the environment (Gómez et al., 2018), time of the day (Sánchez-Gendriz and Padovese, 2017), season (Putland and Mensinger, 2020) number of vocalizing animals (Frommolt and Tauchert, 2014), biodiversity (Denes et al., 2014), ecosystem fragmentation and condition (Tucker et al., 2014) and some other factors.

Soundscapes had been described in a variety of approaches. For example, it was observed that acoustic communities could be defined according to habitat, the frequency of acoustic signals that oscillate throughout the day and with seasons (Farina and James, 2016). It was reported distinctive values and variations of entropy along the year and the day, using the Shannon's diversity index, at a corn plantation, soybean plantations and a forest landscapes (Pijanowski et al., 2011a). In Gitga'at Territory an attempt was made to establish an ecoacoustic baseline with recordings from 8 sites, based on the number of biophony, anthropophony and technophony events and global indexes that measure the intensity of acoustic activity regarding energy (Ritts et al., 2016). At the water-vegetation edge of an uninhabited island on Twin Lakes in Grant Township, in Michigan's northern Lower Peninsula, the relationship between biophony (defined between 2 and 11 kHz) and anthropophony (defined between 1 and 2 kHz), was used to describe soundscapes (Gage and Axel, 2014). From the cited works and several others, it can be seen that soundscapes distinctive features could be perceived using indexes or global measurements, we consider it viable to continue such an approach in further discussion.

In this work, we analyze the power spectrum graph as a descriptor of the soundscape. The graph of the power spectrum shows the portion of the energy per unit of time that falls within each determined frequency band and shows the dominant frequency or frequencies in a sound recording. The distribution of energy at different frequencies was expressed using an average of the intensity for each of the defined frequency bands. By this approach, one could determine the shapes of the power spectrum, assuming that in the soundscape, there would be a finite number of power spectrum shapes.

In a soundscape, given the considerable variability of the signals, the recurrence of a set of sounds with the same frequency and intensity, on a continuous frequency and intensity scale, is nearly impossible. As an alternative, it would be expected that on a discretized frequency and intensity scale, some frequency-intensity combinations may reappear. The concept is analogous with the presented by Keogh et al. (2001): to reduce the amount of information in order to find similarities in large databases. Almeira and Guecha (2018) reported the recurrence of frequency-intensity patterns in a yearly scale observation of a soundscape. According to these results, we considered convenient to establish a sample space with a finite number of intensity-frequency combinations, and use it to define the power spectrum shapes, to detect the recurrent ones, and to make direct comparisons.

Alternative methodologies on describing soundscapes use percentiles of intensity along the frequency spectrum (Richardson et al., 2013), or box and whisker diagrams (Bassett et al., 2012). Probability density has also been used to model changes in sound spectrograms over time (Harris et al., 2016), assuming probabilistic distributions. But an issue with the use of distributions is that they can lead to erroneous results when a single probability distribution does not match the entire analyzed data or if the data fits multimodal probability distributions (Merchant et al., 2015). Uneven probabilistic distribution of intensity along the frequency domain is reported by Haver et al. (2019).

With a continuous sample space, there would be endless combinations to describe an acoustic landscape. But there would be difficulties in making direct comparisons given the number of possibilities and the need of a common probability distribution, valid for several soundscapes, or at different times for a soundscape. Probability distributions will allow determining if the frequency-intensity changes are statistically significant. We did not find a probability distribution that fulfilled such conditions for the observed data. Instead, the probability distributions of the combinations showed different shapes in each frequency band, and at different times of the year. Alternatively, if the problem is solved applying non-parametric statistics (Bayesian inference), it is unnecessary to assume statistical frequency distributions as in the work presented by Kasten et al. (2012), where Gaussian distributions were assumed in a model that described spectrograms.

Ng et al. (2018) applied a chi square test to discriminate frequency spectrums from selected land covers. They applied the chi square distribution to highly scattered power data in 1.0 kHz bands (from 1 to 9 kHz). The problem with this method is that not necessarily the data in all the bands will fit such a distribution, which suggest a single power value as the most probable, and leaves other values as deviations. Rather, the power values could take several values that may be relevant in a soundscape. Being able to detect such dominant values without a rigid probability distribution may improve the description of the soundscape and enhance comparisons.

If we want to set a finite number of shapes of the power spectrum and to examine the shapes that reappear, it is useful to divide the spectrum. One reason is that, given that each frequency band has a distinct behavior, the probability that an entire spectrum shape reappears is much lower than the probability that part of the spectrum shape reappears. Besides, the sounds are rarely synchronized in the same way, making it unlikely to find duplicated entire spectrums. Therefore, we expect to find that portions of the spectrum may consistently reappear. The division of the frequency spectrum was useful in other studies. Gage et al. (2017) used ten frequency intervals (1–11 kHz) to study soundscapes at open Eucalypt and notophyll vine forest, finding that soundscape power at 3–4 kHz was highly correlated with the number of species and total calls. Farina and Pieretti (2014) applied cluster analysis to reveal similarities in the acoustic complexity index (ACI) of 10 frequency classes (1–10 kHz) at a Mediterranean maqui, finding three main clusters.

Adding elements to the frequency-intensity sample space (increasing the sample space size) increases the probability of finding differences between soundscapes or detecting smaller changes in a soundscape. We can have as many levels as we need, but it is convenient to choose a number that allows finding statically significant differences among spectrums, and that also keeps manageable the number of operations needed to describe the power spectrums. Therefore, a discretized sample space was set making the statistical inferences to determine the convenience of its application.

The main goals of this work were:

  • To propose a discrete sample space for the description of power spectrums.

  • To determine a method to detect dominant shapes of the power spectrum using the proposed sample space.

  • To determine a model of the relative frequency of the shapes of the power spectrum over time, using the proposed sample space.

  • To define a model that relates the number of samples and the number of founded frequency-intensity combinations.

  • To determine base, temporary and sporadic combinations of frequency-intensity in the studied soundscape.

Section snippets

Recordings

Recordings were made during the year 2015: one day was recorded in each of 11 months. In each month, one day was randomly selected for a 24-h recording. Portions of the recordings with windy or rainy conditions were included in the analysis without noise reduction or filtering. Table 1 shows the basic recording parameters. The recordings were made in a shaded coffee plantation surrounded by traces of premontane wet forest and pastures. Fig. 1 shows the recording area.

Recordings processing

The power spectra (PSs)

Results

The results show the behavior of sections 1 to 6, which showed more activity; in the remaining sections (7 to 9), a single combination with a very low intensity (−8.3 to 0.3 dBA) predominated most of the time (>95%).

Discussion

Using the model of the accumulated relative frequencies, we could deduce that the relative frequency distribution of the frequency-intensity combinations could be described for the soundscape for the year 2015 based on a several-days sampling: after s lapses, the frequency distribution stabilizes, this may indicate that sufficient lapses have been taken to describe the soundscape. The stability of the measures is necessary when making multi-year comparisons such as those made by Krause and

Conclusions

The method used for the representation and analysis of the acoustic landscape is based on synthesizing the information to a point where regularity can be exposed.

The accumulated relative frequencies of the frequency-intensity combinations are adjusted to a model. The use of such a model allows the estimation of the accumulated relative frequency of the combinations based on several days of sampling; this would reduce the amount of information and effort needed to characterize a soundscape.

The

References (39)

Cited by (0)

View full text