Listening to ecosystems: data-rich acoustic monitoring through landscape-scale sensor networks

Recording sites

We have installed acoustic monitors at all 24 OKEON field sites on Okinawa-jima (Fig. 1), and have been collecting data at all sites since February 15, 2017, with pilot studies beginning earlier at a few sites. These field sites are distributed along an approximate North–South gradient at locations entailing different degrees of urbanization, disturbance, and species endemism. Permission for acoustic monitoring was obtained from local land owners, park officials, and community leaders as part of the OKEON project. Here we present data from a pilot study conducted in the period between July 30 and August 15 2016. This period was chosen because it includes most of the adverse conditions for field operations and automated species detection that are experienced throughout the year on Okinawa-jima: in particular high temperatures, typhoons, and noise from cicadas. This pilot study focuses on five field sites, which reflect a variety of closed forest habitats across the subtropical island of Okinawa-jima. One site was located at the Yambaru Discovery Forest (hereafter “Manabi no Mori”) within the newly designated Yambaru National Park, an important region with many endemic flora and fauna (Itô et al. 2000). Three sites were located in parks of varying urban density: in Nago Central Park (hereafter “Nago”), Nakagusuku Park (hereafter “Nakagusuku”), and Sueyoshi park in Naha city (hereafter “Sueyoshi”), the island’s largest city. One site was located in a forest in the suburbs outside Okinawa City (hereafter “Takiyamabaru”; see Fig. 1). To compare the efficacy of our methods across a longer timescale in what is arguably a less complex acoustic environment (i.e., without cicadas), we conducted additional recordings following the same protocols as described above between October 19, 2016 and February 10, 2017, at Tamagusuku Youth and Children’s Center (hereafter “Tamagusuku”; see Fig. 1).

Open image in new window

Acoustic monitoring

We used Song Meter recording devices (Model SM4, Wildlife Acoustics Inc., Concord, MA, USA), which made digital recordings in stereo.wav format from two omnidirectional microphones, saved to an SD card. We kept gain settings at their default parameters (+ 16 dB), and programmed a recording schedule to activate for 10 min at the beginning of every hour and half-hour (i.e., 10 min on, 20 min off), recording at a sampling rate of 24 kHz. We attached recorders to trees at each field site at approximately breast height (~ 1.3 m). Recording was conducted for approximately one month at each site, constrained by battery life, with recorders allowed to run on a regular schedule until their batteries ran out of power (D 1.2 V NiMH 9500 mAh—average depletion time on this schedule was 227.4 h). At the end of the study period, we recovered 1136.3 h of recordings, resulting in a total of 365 GB of data, which we have archived with the Okinawa Institute of Science and Technology’s SANGO high-performance computing center. Wind speed and rainfall values for each recording were approximated using data from the site’s nearest Japan Meteorological Agency weather station (http://www.jma.go.jp/; JMA 2016). To assess the replicability of our measurements, we include data from an additional field trial conducted on OIST campus in April 2016 (see Online Supplement for details).

Geographic analysis

Using GPS coordinates collected at each of our sites we created circular buffers around each site at 25 m intervals from 25 to 2000 m. We observed that a majority of change occurred up to 1000 m before leveling off and so have constrained our discussion and analysis up to 1000 m. Using the area within each circular buffer we calculated the proportion of land cover surrounding each site including building footprints, road surface area, and land cover. Polygons for building footprints and roads were generated using polyline data from a 2008 to 2009 survey obtained from the Geospatial Information Authority of Japan (http://www.gsi.go.jp). Land cover classification was carried out on a Jan 4, 2015 Landsat 8 Operational Land Imager image obtained from the US Geological Survey (Scene: LC81130422015004LGN00). For additional details on the classification see the online supplement. Land classes were defined as: Agriculture, dominated by sugarcane and a mix of fields at various phenologies; Shrub, medium to low growth coastal and disturbed vegetation; Forest, dense forest stands with closed canopy; Grass, short trimmed vegetation found in golf courses and airfields; Urban, a complex mix of man-made surface materials such as concrete and asphalt with limited vegetation; Other, remaining surfaces, bare earth, rocks, and sand (this class also includes unclassified pixels); Water, classified but not used in analysis (see Tables S1–S3 for more details).

Soundscape indices

We used the packages soundecology (Villanueva-Rivera and Pijanowski 2016) and seewave (Sueur et al. 2008) in R Version 3.3.2 (R Development Core Team 2016) to estimate five soundscape indices for each recording. The Bioacoustic Index (BioA) is calculated as the area under the power spectrum curve across the range of frequencies typically occupied by bird species (2–8 kHz), and is intended to describe the proportion of this acoustic space that is occupied in a recording (Boelman et al. 2007). Acoustic Diversity (ADI) is calculated by dividing the frequency range into 10 bins of 1 kHz width each, and returning the Shannon index of the occupancy of these bins for a recording (Villanueva-Rivera et al. 2011). The Acoustic Evenness index (AEI) also uses 1 kHz frequency bins, but returns the evenness (Gini index) of their occupancy. The Acoustic complexity index (ACI) measures the variability of sound intensity within a recording, which can be an accurate index of avian diversity, but is also easily misled by a small number of insects (Farina et al. 2011). Lastly, Normalized Difference Soundscape Index (NDSI) provides a range of values comparing sound intensity between frequency ranges that correspond to biophony (2–12 kHz) and anthropophony (1–2 kHz), such that values approaching − 1 represent dominant anthropogenic noise and approaching + 1 represents a soundscape dominated by biophony (Kasten et al. 2012). Taken together, these indices are intended to describe the soundscape in ecologically relevant ways (Gasc et al. 2016).

Automated recognition of bird vocalizations

We used Kaleidoscope Pro (version 4.1.0; Wildlife Acoustics Inc., Concord, MA, USA) to train software recognizers for five bird species. We selected species that are representative of the full range of vocal repertoires and population sizes from among those on Okinawa-jima, thus making differing contributions to the acoustic soundscape of Okinawa and ultimately to ecosystem functioning. These species included those common species that were widespread across the island (Hypsipetes amaurotis and Corvus macrorhynchos), potentially disturbance-sensitive forest species (Halcyon coromanda and Otus elegans), and the endangered Yambaru-endemic Okinawa Rail (Gallirallus okinawae).

Statistical analysis

We compared soundscape data at a fine scale (30 min blocks) using generalized least-squares (GLS) multiple regression models in the R programming environment (R Development Core Team 2016). We included a temporal auto-regression term (AR1: ~ Time|Site) as a covariance structure in all regression models to correct for time series pseudo-replication. To test general characteristics of soundscape time series, we used a model selection approach to choose from a set of all possible combinations of diurnal vs. static, flat vs. trending, and site-specific vs. site-independent regression models, using the Akaike Information Criterion (AIC) to assess model fit. We assessed the support for each model component by comparing the goodness of fit of a model that included the component to a model which had the component removed. To test for correlations between the number of species detections and land cover components, we calculated the sum of detections for each recording day and fit GLS models with temporal auto-regression terms as above. We used a similar approach to test for the effects of wind and rain on soundscape index values.

Soundscape indices

We generated soundscape indices for five sites during the period July 30–August 15, 2016, and found clear differences in the degree to which each index varied over time. Two out of five indices we measured—acoustic diversity (ADI) and the bioacoustic index (BioA)—showed significant non-zero trends within our initial study period. Both indices had significantly negative slopes (Table 2), indicating a trend in soundscape composition across the study period. Similarly, we found that regression models with a diurnal component were a significantly better fit to the Normalized Difference Soundscape Index (NDSI), ADI and BioA indices (Table 2), and indeed these indices exhibited higher values at night (Fig. 2). For several indices we observed daily peaks at dawn, around 5:00–7:00 for ADI and BioA (Fig. 3).

Table 2

Model selection for soundscape indices

Soundscape index	Model component	dAIC	LRT
NDSI	Diurnal	11.51	15.51	***
	Site	4.01	12.10	*
ACI	Empty
ADI	Diurnal	10.14	14.14	***
	Site	8.11	16.11	**
	Trend	6.86	8.86	**
AEI	Empty
BioA	Diurnal	2.01	6.01	*
	Site	25.44	33.43	***
	Trend	4.43	6.43	*

Best fitting models are shown, with significance tests for each component representing the likelihood ratio test (LRT) of a comparison to the best model with the component removed

* P < 0.05; ** P < 0.01; *** P < 0.001

Open image in new window

Open image in new window

We observed clear differences between sites for each soundscape index, and regression models including separate intercepts for each site were a significantly better fit to our data for NDSI, ADI, and BioA than those without (Table 2). Some variation among sites appeared to follow Okinawa-jima’s urban–rural gradient (Table 2; Figs. 2, 3). NDSI was highest in recordings from Manabi no Mori, the site with the most forest cover, and lowest in recordings from the urban Sueyoshi Park (Fig. 3). This pattern was most apparent in recordings made at night (19:00–5:00; Fig. 2), but was also observable for recordings made during the day (6:00–18:00). However, no significant correlation between daily average NDSI and land cover was found (Table 2). In contrast, ADI was lowest on average in recordings from rural sites, and highest in recordings from urban sites (Fig. 2; Table 2). Conversely, acoustic evenness (AEI) was lowest in Sueyoshi, and BioA was highly variable regardless of site but showed a generally negative association with forest cover (Fig. 2; Table 2). Acoustic complexity (ACI) was largely similar in the most rural sites and Sueyoshi, our most urban site, but Nakagusuku and Nago displayed higher complexity than the other three sites (Fig. 2).

We estimated NDSI values for a late-year monitoring period at our Tamagusuku field site. These estimates showed a general decline in biophony throughout the recording period. To examine changes in the timing of the dawn chorus, we tested the fit of seasonal models with a year-long period to hourly subsets of this data. We found that seasonal models were a better fit than flat or trending models for recordings at 5:30 and 6:00 am (Likelihood Ratio Test: 5.6, P < 0.05; LRT: 6.1, P < 0.05). Seasonal and trending models showed similar fit at 6:30am (LRT: 1.8, P = 0.17), and no models showed significant improvement over an empty model for 7:00 am (LRT: 0.8, P = 0.36), which exhibited a consistently high NDSI value throughout the study period (see Fig. S1).

Automated species detection

Automated detection results revealed considerable differences in vocalization densities between species and across sites (Fig. 4). Species differed in their detection frequency (Table 1), with C. macrorhynchos being detected most often (total 19698 vocalizations across sites), followed by Hypsipetes amaurotis (13302 vocalizations), Halcyon coromanda (5288 vocalizations), O. elegans (4496 vocalizations) and finally G. okinawae (319 vocalizations). For all species with the exception of H. amaurotis, we observed a substantial decline in the number of detected vocalizations from north to south, with Manabi No Mori, Nago and Takiyamabaru displaying higher densities than Nakagusuku and Sueyoshi (see Fig. 4, Tables 1, 4). All species except H. amaurotis exhibited their lowest recorded detection frequencies at the two southernmost sites, Nakagusuku and Sueyoshi park (see Fig. 1). Likewise, all species (with the exception of G. okinawae) exhibited their highest vocalization density at the Nago field site in northern Okinawa-jima. The Okinawa Rail (G. okinawae) is endemic to the Yambaru forest in northern Okinawa-jima, and we did not detect any vocalizations from this species outside of the Manabi No Mori field site.

Open image in new window

Table 4

GLS regression models comparing daily call rates of species to land cover proportion

Species	Forest				Urban				Wind				Rain
	phi	ES (std. B)	Std. err		phi	ES (std. B)	Std. err		phi	ES (std. B)	Std. err		phi	ES (std. B)	Std. err
Gallirallus okinawae	0.28	0.41	0.10	***	0.34	−0.30	0.11	*	0.01	−0.02	0.09	NS	0.00	−0.16	0.07	*
Halcyon coromanda	0.49	0.37	0.13	**	0.55	−0.16	0.15	NS	0.26	−0.04	0.10	NS	0.27	−0.03	0.09	NS
Hypsipetes amaurotis	0.75	−0.07	0.20	NS	0.75	0.09	0.20	NS	0.70	−0.15	0.10	NS	0.71	−0.02	0.06	NS
Corvus macrorhynchos	0.90	0.28	0.26	NS	0.91	0.05	0.28	NS	0.64	0.00	0.06	NS	0.66	−0.09	0.04	*
Otus elegans	0.63	0.28	0.16	–	0.64	−0.23	0.17	NS	0.60	−0.09	0.12	NS	0.60	−0.05	0.08	NS

Details and formatting follow from Table 3

We observed significant site-specific effects for all species except Hypsipetes amaurotis and O. elegans (Table 4). The number of H. amaurotis detections differed little between sites, but was lowest in Takiyamabaru. In contrast, despite showing no significant site effect, O. elegans was entirely absent from Sueyoshi to Nakagusuku. We tested whether this geographic trend might be explained by the proportion of forest or urban land cover near each site. We found significant positive relationships between forest cover and daily detection rate for G. okinawae and Halcyon coromanda, indicating that they are more often detected in forested areas (Table 4). A similar relationship was uncovered for O. elegans but was not significantly supported by our data.

Our data also showed temporal patterns in species detections across the study period. Halcyon coromanda showed a significantly decreasing trend in daily detection rate (P < 0.001; std. B −0.21, B err. 0.15). For all other species, we found no evidence of a trend in detection rate over the study period. Whilst most species showed perceptible decreases in detection rate during a period of rough weather, no correlations were recovered between daily detection rate and either wind speed or rainfall (Table 4). Halcyon coromanda vocalized almost exclusively in the morning during our focal period, and Hypsipetes amaurotis began vocalizing one or two hours later across most days at Sueyoshi and Takiyamabaru than at the other sites. C. macrorhynchos vocalized throughout the day, typically persisting from sunrise until sunset. O. elegans was detected most often at night but occasionally during the day, and G. okinawae was detected sparsely throughout the day but only rarely between 10 pm and pre-dawn (Fig. 4).

Soundscape indices

The soundscape indices reported here displayed systematic variation across sites, particularly along an urban–rural gradient. The Normalized difference soundscape index (NDSI) has been shown to accurately reflect anthropogenic noise when values are low (Gage and Axel 2014), and avian species richness when values are high (Fuller et al. 2015). We found that, on average, our urban site in Sueyoshi park had higher values than our suburban sites in Nago and Nakagusuku. However, the typical baseline NDSI at Sueyoshi was the lowest of all sites, indicating both a weak biophony and a persistent and constant noise from the city that was easily observable in spectrograms (NRF pers. obs.). In contrast, Nago and Nakagusuku had high baseline NDSI values with frequent incidents of high anthropophony, likely due to passing trucks or aircraft. This suggests that in a dynamic acoustic environment, complex temporal models may be necessary to better disentangle the effects of temporary and persistent acoustic disturbance (Botteldooren et al. 2006).

Acoustic diversity (ADI), which describes the diversity of 1 kHz frequency bands in a recording (Villanueva-Rivera et al. 2011), showed a strongly negative association with forest land cover and a strongly positive one with urban land cover. Given that bird diversity is highest at in the northern Yambaru forest, both in prior wildlife surveys (Itô et al. 2000; Biodiversity Center of Japan 2004) and in our species detections, we do not expect this ADI gradient to reflect species diversity. One explanation for this pattern is the constant broadband sound emitted by cicadas, which were recorded during most of the day throughout the study period. However, this pattern may also be explained by the prevalence of broadband anthropophony events in urban and suburban Okinawa. This result highlights the need for cross-validation of soundscape index values with estimates of local biodiversity, particularly birds and stridulating insects.

It is important to note some of the limitations inherent in drawing spatial conclusions from a small number of sites. Previous studies have shown that while soundscapes tend to be spatially auto-correlated (nearby areas sound similar), they can show important variation in sound amplitude, propagation, and diversity indices (Pekin et al. 2012; Rodriguez et al. 2014). Thus, further sampling is necessary to draw solid conclusions regarding the relationship between Okinawa’s soundscape and geography, and is indeed underway.

We identified several soundscape trends which were driven either solely or in part by weather conditions. The identified bands of lower acoustic diversity and complexity, and higher evenness roughly midway through our study correspond to a period of poor weather. During this time, high wind speed reduced acoustic diversity but increased evenness, whereas higher daily rainfall was associated with increased ADI and acoustic complexity (ACI). This seems intuitive, as it is widely accepted that geophony influences soundscape indices either directly through the impact of wind or rain on the soundscape (Farina 2014), or indirectly through the impact of geophony on the signaling behavior of sound-producing animal species (Davidson et al. 2017).

Automated species detection

Using an automated recognition approach, we were able to consistently detect vocalizations from all five focal species. The degree of accuracy varied for these detections between sites and species, with an average accuracy of 76.8% in locations where species were confirmed as present (Table 1). This accuracy level was generally sufficient to allow human confirmation of detection events. Furthermore, cluster assignment score cutoffs can be applied to enable conservative scoring of species presence or absence with minimal human interference. Ultimately, we were able to document a clear effect of geographic location—and potentially urbanization—on the focal species in our study. The two most common species in our study (Corvus macrorhynchos and Hypsipetes amaurotis; see Biodiversity Center of Japan 2004) were present at all five sites, and vocalization density of these species was least affected by our urban–rural gradient. We found that forest species (Halcyon coromanda and Otus elegans) showed a clear decline in vocalization density towards the more urban southern sites, suggesting greater sensitivity to disturbance. At the Yambaru forest site, Manabi no Mori, we were able to consistently detect the endangered Okinawa Rail, Gallirallus okinawae, whilst false positives for this species at other sites were minimal and exhibited poor clustering scores. Our finding that the detection rate of both G. okinawae and H. coromanda was affected by forest cover, with these culturally important species both being detected more frequently in forest sites, further highlights the value of Yambaru forest in supporting endemic, endangered and culturally significant biota (Itô et al. 2000).

Despite being in close proximity to roads and having large proportional surrounding urban land-use, the Nago forest site appears to be a local refuge for several of our focal species. Our common focal species, Hypsipetes amaurotis and C. macrorhynchos, were highly vocal at this site, as were our disturbance-sensitive focal species, Halcyon coromanda and O. elegans. This may suggest that the forest near Nago is an important site for biodiversity despite its urban setting, which could potentially support some of Okinawa-jima’s endemic species, were their ranges to expand further southwards. The Nakagusuku field site displayed high values of the bioacoustic index (BioA) and ACI, low NDSI and was a local hotspot for Hypsipetes amaurotis vocalizations, and to a lesser degree also Halcyon coromanda, despite the site’s southern location and mixed land-use cover. Along with Takiyamabaru, this site also exhibited a clear effect of diurnal cycle on soundscape indices, mainly for ADI and AEI. But unlike Takiyamabaru, at Nakagusuku we also observed high intensities of both anthropophony—mostly trucks, aircraft, and construction—and biophony—mostly cicadas and birds (SRP-JR and NRF, pers. obs.). This combination of intense biophony and anthropophony may also have contributed to the lower accuracy of our species detection models for Nago and Nakagusuku, particularly for C. macrorhynchos, H. coromanda and O. elegans (see Table 1).

Combining approaches

We identified several patterns in our soundscape indices that appear to be driven by stridulating insects. Further, we found a clear effect of weather conditions on soundscape indices and in some cases species detection. Lastly, we observed a ‘dawn chorus’ effect in several indices, which we attributed to bird species vocalizing in the morning.

We found that Acoustic complexity (ACI) was highest at our suburban sites in Nago and Nakagusuku, likely due to the abundance and activity of cicadas (Graptopsaltria bimaculata, Platypleura kaempferi and P. kuroiwae) during the day and a repetitive orthopteran (Hexacentrus unicolor) during the night. Previous studies have found that ACI increases drastically in the presence of cicadas and other orthopterans that produce a broadband signal (Farina et al. 2011; Fuller et al. 2015). BioA was also highest at night in Nago and Nakagusuku, likely due to the persistent broadband signaling of several orthopteran species, especially Hexacentrus unicolor. However, the degree of diurnal differentiation in soundscape index values (particularly NDSI) appeared to be greater at suburban sites, potentially due to the presence of vehicular traffic noises during the day (dark blue lines in Fig. 3).

We also observed several notable temporal patterns that were concordant between our soundscape index and species detection results. Firstly, there was a general decrease in all species’ vocal activity during the period of inclement weather on Okinawa-jima. These low levels of vocalization detection across all species were concurrent with low soundscape values for NDSI and high values for ADI and ACI. This suggests an impact of inclement weather on species behavior, and also serves as a proof of concept by demonstrating the capability of both methods to detect it (Keast 1994). Secondly, we observed a general increase in diurnal species vocal activity during the “dawn chorus” at around 5:30 am. Three of our most abundant species: Hypsipetes amaurotis, C. macrorhynchos and Halcyon coromanda, began vocalizing around this time, and most detections of Halcyon coromanda were observed during the dawn chorus period (5:30–7:30). We found this effect to be correlated with several soundscape indices, with some indication of a dawn chorus being visible for ACI and ADI, but most noticeably for peaks of the Bioacoustic Index, which were a good temporal match with the onset of Halcyon coromanda and Hypsipetes amaurotis vocalizations. Our further recording of soundscapes from the winter supports the detection of a later dawn chorus in October through February (see Fig. 5). Despite the dawn chorus effect being well documented in studies of bioacoustics and soundscape ecology (Kacelnik and Krebs 1983; Wimmer et al. 2013; Towsey et al. 2014), future acoustic monitoring should explore the phenology of this event, and role of seasonal, weather, and biotic interference in determining its timing (Farina et al. 2015; Hart et al. 2015).

Open image in new window

Taken together, our results indicate a weather-related decline in both soundscape indices and species detections, a detectable dawn chorus of individual species being visible in our soundscape results, and that several of our soundscape indices were saturated by insect activity for periods of our study. Overall, this is an encouraging sign that the soundscape approach succeeds in reflecting species’ behavioral phenology on Okinawa-jima (Farina et al. 2011; Davidson et al. 2017). Further, our findings suggest that future soundscape studies conducted across the summer season should recognize and treat differently periods of inclement weather and high cicada activity. However, it should be noted that perhaps poor weather conditions and broadband insect acoustic activity are equivalent insofar as they both reduce the signal efficacy of soniferous species due to a ‘filling up’ of acoustic niche space (Hart et al. 2015). We find it encouraging to note that this issue appears to be more manageable when comparing soundscape indices across longer duration recordings, as was the case for our winter recording period.