Automatic identification of rainfall in acoustic recordings
Introduction
Among climatic processes, the rainfall is one of the main abiotic factors influencing the temporal patterns of reproductive activity in tropical and temperate animal species (Birkett et al., 2012, Gottsberger and Gruber, 2004, Keast and Marshall, 1954). The rain modifies the physical properties of the breeding sites, by adding water, and increases the relative humidity of the environment (Busby and Brecheisen, 1997, Ladányi et al., 2015). As a result, it stimulates or discourages reproductive behavior, which could produce modifications in the animal communication (Amézquita and Hödl, 2004, Lack, 1950, Zina and Haddad, 2005), changes in population dynamics (Georgiadis et al., 2003, Mondet et al., 2005, Ogutu and Owen-Smith, 2003), and shifts in phenology (Primack et al., 2009, Saenz et al., 2006, Yoo and Jang, 2012, van der Kolk et al., 2016).
Developing studies to estimate instant rainfall effects in animal sound production is a burdensome task. In general, weather stations are not necessarily located near the study area (Carey and Alexander, 2003, Mendelsohn et al., 2007), i.e., they are far from the site to be representative of the local weather conditions experienced by the community (Carey and Alexander, 2003). This implies that the correlation of rainfall data with the sounds produced by populations with narrow distribution is not always accurate. Additionally, the rainfall is normally measured in daily and hourly rates by weather stations (Ban and Schmidli, 2015, Hidalgo et al., 2014, Meek and Hatfield, 1994, Du et al., 2016), despite rain can disrupt or encourage animal auditory communication in shorter temporal spans (Lack, 1950, Zina and Haddad, 2005).
Automatic acoustic recorders have been extensively used in passive monitoring of animal species (Bedoya et al., 2014, Busby and Brecheisen, 1997, Depraetere et al., 2012, Farina, 2014, Gregory and van Strien, 2010, Kalan et al., 2015, Laiolo, 2010, Pace et al., 2010, Pieretti et al., 2011, Towsey et al., 2014, Sueur et al., 2008). These devices acquire the sounds of the landscape regardless the nature of their source; therefore, the collected recordings could be used to identify sounds of abiotic sources (e.g. rainfall) in terrestrial and aquatic ecosystems. The rainfall produces one of the most recognizable, and variable, sounds in nature; however, the detection of rainfall events in ecoacoustic studies (Sueur and Farina, 2015) has been barely studied.
Most of the current acoustic methods for rainfall identification have been proposed for signals collected with hydrophones in marine ecosystems (Amitai and Nystuen, 2008, Ma and Nystuen, 2005, Medwin et al., 1992). Although the sound of rainfall in aquatic and terrestrial ecosystems presents several similarities, they are generated under different circumstances. The underwater ambient sound generated by rainfall is a consequence of raindrops colliding with the water surface and trapping bubbles (Amitai and Nystuen, 2008). In a forest, or other landscapes, the sound of the rainfall is produced by the impact of raindrops on the vegetation layers (mainly in the canopy). In aquatic and terrestrial environments the rainfall sound is present throughout the audible spectrum with heterogeneous power in specific frequency bands. Nonetheless, sensors, soundscapes, and power distribution of the rainfall sound across the spectrum differ from both environments. For these reasons, methodologies developed for analyzing underwater rain sound cannot be directly extrapolated to the terrestrial case.
Few studies have been focused on rainfall identification in acoustic recordings of terrestrial ecosystems. To the best of our knowledge, the method proposed by Ferroudj et al. (2014) was the first and only known approach to solve this issue. They proposed a pattern recognition approach with the use of 5 features (temporal entropy, spectral entropy, acoustic complexity index, background noise, and spectral cover) and a decision tree for a bi-class classification problem (heavy rain or non-rain), obtaining an overall accuracy of 93%.
The aim of our study is to provide a tool for detecting, quantifying, and analyzing rainfall events in acoustic recordings obtained from evergreen forests. Thereby, short- and long-term effects of rainfalls in the animal sound production can be analyzed. This approach illustrates the potential for studies related with changes in phenology, auditory communication, or population dynamics as a result of changing rainfall conditions, by solely using acoustic signals provided by automatic recorders.
Section snippets
Rainfall detection
The rainfall produces spectral components throughout the audible spectrum (the sound produced by the impact of each raindrop depends on its size and the material of the receptor). For this reason, the rainfall can be seen in the spectrogram as a background sound (similar to the static television noise). However, the power of this sound is not uniformly distributed across the spectrum. This effect is less visible in the spectrogram, but it is evident in the power spectral density (PSD).
Fig. 1
Experimental results and discussion
In the identification of rainfall – non-rainfall recordings of all levels of intensity (light, moderate, heavy, and violent), the method attained an overall accuracy of 92.9% (Tmean = 0.9 × 10−6, Tsnr = 1.8) (see Table 1). This is because light intensity events also contained drizzles. These drizzles were slightly audible, and they could not be measured by the pluviometer (i.e. the amount of rainfall was inferior to the measurement error, ± 0.04 mm). For this reason, the accuracy reached a value of
Conclusions
In this paper, we propose a method to infer the rainfall intensity levels directly from acoustic recordings. This avoids the use of additional devices for acquisition of rainfall information (e.g. pluviometer) by using an already existent infrastructure of passive monitoring acoustic recorders. Additionally, this procedure allows directly observing local weather conditions experienced by the community that is being acoustically monitored.
The power spectral density (PSD) is a good estimator of
Conflict of interests
There is not any relationship between the manufacturer of the recorders used for this research and the authors of this paper. The corresponding author was affiliated to Universidad de Antioquia during the development of this research.
Acknowledgements
This project was supported by “Fondo de Sostenibilidad Universidad de Antioquia – Estrategia de sostenibilidad”, and “Isagen – contract 47/574”.
References (44)
- et al.
Automatic recognition of anuran species based on syllable identification
Ecol. Inf.
(2014) - et al.
Monitoring animal diversity using acoustic indices: implementation in a temperate woodland
Ecol. Indic.
(2012) - et al.
Assessment of wind and water erosion risk in the watershed of the N ingxia-Inner Mongolia Reach of the Yellow River, China
Ecol. Indic.
(2016) - et al.
Towards the automated detection and occupancy estimation of primates using passive acoustic monitoring
Ecol. Indic.
(2015) - et al.
Multi-indicator sensitivity analysis of climate change effects on landscapes in the Kiskunság National Park, Hungary
Ecol. Indic.
(2015) The emerging significance of bioacoustics in animal species conservation
Biol. Conserv.
(2010)- et al.
Data quality checking for single station meteorological databases
Agric. For. Meteorol.
(1994) - et al.
Subunit definition and analysis for humpback whale call classification
Appl. Acoust.
(2010) - et al.
A new methodology to infer the singing activity of an avian community: the Acoustic Complexity Index (ACI)
Ecol. Indic.
(2011) - et al.
Spatial and interspecific variability in phenological responses to warming temperatures
Biol. Conserv.
(2009)
The use of acoustic indices to determine avian species richness in audio-recordings of the environment
Ecol. Inf.
Using a phenological network to assess weather influences on first appearance of butterflies in the Netherlands
Ecol. Indic.
How, when, and where to perform visual displays: the case of the Amazonian frog Hyla parviceps
Herpetologica
Underwater acoustic measurements of rainfall
Precipitation: Advances in Measurement, Estimation and Prediction
Heavy precipitation in a changing climate: Does short-term summer precipitation increase faster?
Geophys. Res. Lett.
Animal perception of seasonal thresholds: changes in elephant movement in relation to rainfall patterns
PLoS ONE
Chorusing phenology and habitat associations of the crawfish frog, Rana areolata (Anura: Ranidae), in Kansas
Southwest. Nat.
The Essential Physics of Medical Imaging
Climate change and amphibian declines: is there a link?
Divers. Distrib.
Soundscape Ecology: Principles, Patterns, Methods and application
Detection of rain in acoustic recordings of the environment. PRICAI 2014: Trends in Artificial Intelligence
Lect. Notes Comput. Sci.
The influence of rainfall on zebra population dynamics: implications for management
J. Appl. Ecol.
Cited by (43)
Towards the development of a citizens’ science-based acoustic rainfall sensing system
2024, Journal of HydrologyAutomatic acoustic heterogeneity identification in transformed landscapes from Colombian tropical dry forests
2022, Ecological IndicatorsCitation Excerpt :In consequence, we decided to automatically detect and exclude noisy recordings. The Power Spectral Density (PSD)-based method proposed by Bedoya et al., 2017 is an adequate estimator for detecting recordings with geophonic and anthropogenic elements. PSD indicates how the signal power is distributed across frequencies.