Rapid methods to assess biodiversity and ecosystem health at large spatial and temporal scales are paramount for informing conservation planning (Stem et al., 2005). Because sound-producing species are important indicators of environmental health, acoustic surveys offer an important approach for biodiversity monitoring programs (Gregory and Strien, 2010; Blumstein et al., 2011). Advances in digital recording technology allow standardized acoustic surveying of increasingly large landscapes over longer, coordinated time-periods (Merchant et al., 2015). Passive acoustic monitoring would thus be particularly useful in biodiversity hotspots, where understanding the impact of widespread land conversion on numerous species requires rapid surveying at a large scale.
In order to extract meaningful biological information from the enormous datasets that result from large-scale acoustic monitoring, a number of acoustic indices have been developed (Buxton et al., in review). Acoustic indices examine the heterogeneity of the acoustic environment, under the assumption that more species found in a community will produce a greater number of different signals at the same time (Sueur et al., 2008b). By summarizing acoustic energy, this approach foregoes species identification and focuses instead on community level diversity (Towsey et al., 2014). Many types of acoustic indices have been tested and have been found to reflect vocal species diversity and abundance, community composition, vegetation structure, habitat type, human perception of a soundscape, human activity, and ecological condition in terrestrial and aquatic habitats (Sueur et al., 2014). However, the development and standardization of appropriate acoustic indices have occurred predominantly in temperate regions (Buxton et al., in review).
The mountain chains of the Western Ghats in southern India are a biodiversity hotspot –with high endemism and extreme rates of habitat loss (Myers et al., 2000). Between 1920 and 1990, the Western Ghats experienced a 40% decline in forest cover due to land conversion for plantations of timber, tea, coffee, and other agriculture (Chandran, 1997; Menon and Bawa, 1997). The Western Ghats are by far the most densely populated and widely altered of global biodiversity hotspots (Cincotta et al., 2000). Human-modified landscapes in the Western Ghats typically retain a matrix of native forest cover and feature high habitat heterogeneity (Anand et al., 2010). Land-use ranges from coffee grown under the shade of native forest trees to open monocultures of tea (Raman, 2006). This juxtaposition of natural areas and different land-use types warrants a biodiversity monitoring approach which accounts for both local and landscape-scale characteristics.
The Western Ghats have an especially high diversity of amphibians, birds, crickets, and cicadas (Mittermeier et al., 2011), many of which produce sounds for territory defense, mate attraction, predator deterrence, navigation, foraging, and maintaining social groups (Brumm, 2013). This high diversity of sound-producing species suggests that acoustic monitoring could be a particularly valuable method. However, properties of the acoustic environment could render the use of acoustic indices for analysis of recordings challenging. For example, there are high densities of insects which produce extended, constant, and broadband vocalizations (Diwakar and Balakrishnan, 2007; Price et al., 2016). This may confound a high diversity of bird song, which is characterized by an intrinsic variability of intensities (Pieretti et al., 2011). Moreover, heterogeneity of the acoustic environment may be expected to increase with avian vocal diversity and reach a threshold due to the redundancies in acoustic characteristics of species (Sueur et al., 2008b). Thus, in a crowded and complex acoustic space, the relationship between acoustic indices and vocal species diversity is unclear.
With a high diversity of species embedded in a complex mosaic of land-use types, the Western Ghats would benefit from a standardized monitoring approach capable of measuring biodiversity at particular places and times, but which is also predictive and scalable across the landscape. Passive acoustic monitoring would be an ideal technique to capture acoustic diversity at large spatio-temporal scales. However, to extract meaningful biological information from large acoustic datasets, an automated approach is required. Here we tested the ability of acoustic indices to predict the richness, diversity, and total number of avian vocalizations in recordings from 28 sites in the tropical wet evergreen region of the Western Ghats. Additionally, using a combination of model predictions and observations from recordings, we compare acoustic diversity and the richness of avian vocalizations between a number of different land-use types, including tea, coffee, and cardamom.
We obtained acoustic recordings from four sites within Kadumane tea estate in Hassan district and 24 sites at plantations in Kodagu district of Karnataka state in the Western Ghats (Figure 1). The altitude ranged from 900–1,100 m, where rainforest at this elevation is classified as tropical wet evergreen (Champion and Seth, 1968). Kadumane tea estate occupies over 1,000 acres, with extensive areas cleared for tea (Camellia sinensis; Figure 2) and small interspersed fragments of remnant forest. The plantations in Kodagu district have a variety of shade-grown crops, including coffee (Coffea spp.) and cardamom (Elettaria cardamomum), interspersed with black pepper (Piper nigrum) vines (Figure 2).
At Kadumane tea estate, recordings were collected for 3–12 days in May 2017 using Song Meters (model SM2; Wildlife Acoustics Inc., Concord, MA; Supplementary material, Appendix 1). Audio data were collected in mono, 16 bit WAV format at a sampling rate of 22,050 Hz with a gain of 48 dB. Song Meters were programmed to record in 5 minutes increments for 2 hours after sunrise and for the first 5 minutes each 30 minutes thereafter. Two Song Meters were placed at the edge of tea fields, two were placed >300 m inside forest fragments, and all recording sites were >300 m to ensure independence of acoustic environments.
At plantations in Kodagu, recordings were collected for 2–4 days from April 2010–January 2011 using Song Meters model SM1 (Supplementary material, Appendix 1). Audio data were collected in stereo, 16 bit WAV format at a sampling rate of 22,050 Hz with a gain of 42 dB for 2 minutes each hour from 0600–0800 IST. Half of the recordings were collected in coffee and half were collected in cardamom plantations. Similar to Kadumane, all recording sites were spaced >250 m apart.
Analysis of recordings
In order to calibrate data acquired from different recording devices we converted all audio data to 1-second 1/3 octave band SPLs measurements from 12.5–8,000 Hz using an end-to-end calibration method (Merchant et al., 2015). We then calculated a total of 36 acoustic indices and acoustic descriptors (Table 1) for each 5 minutes (Kadumane) and 2 minutes (Kodagu) sound file. We used indices calculated at a 1-second 1/3 octave band resolution because we found indices calculated at a finer resolution (0.02 sec, 43 Hz) to be of limited utility (Supplementary material, Appendix 2). Because the lowest frequency of a bird vocalization was found at 230 Hz (Greater coucal, Centropus sinensis), we calculated indices intended to capture biological activity using frequency bands from 250–8,000 Hz and those intended to capture background noise (i.e., anthropogenic and weather sounds) using bands from 31.5–200 Hz (Brumm and Slabbekoorn, 2005).
|Index abbreviation||Index name||Frequency band||References|
|L10–L90||Difference between 10th and 90th exceedence levels||Full spectrum||(Hong and Jeon, 2017)|
|Mamp||Median sound level||Full spectrum||(Depraetere et al., 2013)|
|dBA_noise||A-weighted mean sound level||Noise||(Torija et al., 2013)|
|dBA_bird||Biological||(Torija et al., 2013)|
|dB_noise||Mean sound level||Noise||(Torija et al., 2013)|
|dB_bird||Biological||(Torija et al., 2013)|
|Biophony||Biophony||Biological||(Joo et al., 2011)|
|Anthrophony||Anthrophony||Noise||(Joo et al., 2011)|
|avgAMP||Average signal amplitude||Full spectrum||(Desjonquères et al., 2015)|
|L10AMP||SPL exceeded 10% of the time||Full spectrum||(Torija et al., 2013)|
|Peakfreq||Peak frequency||Biological||(Sueur et al., 2008a)|
|AA_noise||Acoustic activity||Noise||(Towsey et al., 2014)|
|AA||Biological||(Towsey et al., 2014)|
|AAc_noise||Count of acoustic events||Noise||(Towsey et al., 2014)|
|Aac||Biological||(Towsey et al., 2014)|
|AAdur_noise||Duration of acoustic events||Noise||(Towsey et al., 2014)|
|AAdur||Biological||(Towsey et al., 2014)|
|Kurtosis||Spectral kurtosis||Biological||(Bormpoudakis et al., 2013)|
|Skew||Spectral skewness||Biological||(Bormpoudakis et al., 2013)|
|Hf||Spectral entropy||Biological||(Sueur et al., 2008b)|
|Ht||Temporal entropy||Biological||(Sueur et al., 2008b)|
|H||Total entropy||Biological||(Sueur et al., 2008b)|
|Hm||Entropy of spectral maxima||Biological||(Towsey et al., 2014)|
|HvPres||Entropy of spectral variance||Biological||(Towsey et al., 2014)|
|HvSPL||Entropy of spectral variance||Biological||(Towsey et al., 2014)|
|Roughness||Roughness||Biological||(Rychtáriková and Vermeir, 2013)|
|ACI||Acoustic complexity||Biological||(Pieretti et al., 2011)|
|ADI||Acoustic diversity||Biological||(Pekin et al., 2012)|
|Evenness||Acoustic evenness||Biological||(Villanueva-Rivera et al., 2011)|
|AR||Acoustic richness||Biological||(Depraetere et al., 2013)|
|NDSI||Normalized difference soundscape index||Full spectrum||(Kasten et al., 2012)|
|Bio_anth||Ratio of biophony to anthrophony||Full spectrum||(Kasten et al., 2012)|
|SpecPers||Spectral persistence||Biological||(Towsey et al., 2014)|
|CLdurDiff||Difference between cluster duration||Biological||(Towsey et al., 2014)|
|CLpkDiff||Difference between cluster peak frequency||Biological||(Towsey et al., 2014)|
|CLLeqDiff||Difference between cluster Leq||Biological||(Towsey et al., 2014)|
|NumCL||Spectral diversity||Biological||(Towsey et al., 2014)|
To assess the diversity and richness of bird species vocalizations for comparison with acoustic indices, a trained technician (for Kadumane) and a bird song expert (SA, for Kodagu) identified bird vocalizations in a subset of recordings. The observers used software Raven Pro 1.5 (Cornell University, Ithaca, NY) to visualize spectrograms, listen to audio files, and tag vocalizations. Before analyzing the final dataset, to ensure the technician was able to correctly identify most (>70%) bird sounds, they were asked to identify 65 randomly chosen bird species songs from the Karnataka region of the Western Ghats (bird song accessed from xeno-canto—http://www.xeno-canto.org). Although we were interested in classifying insect vocalizations, they have little frequency modulation; thus, identifying species would require more time-consuming observation of oscillograms or power spectra (Metrani and Balakrishnan, 2005). Instead, we identified all insect sounds as “insect.” Other non-avian sounds were classified as “anthropogenic,” “rain,” “frog,” “squirrel,” or “langur.” Sounds that observers were unable to identify to species (<2%) were labeled as “unknown.”
In recordings from Kodagu, the number of unique avian species vocalizations was noted for each 2-minute clips (i.e., vocal species richness). In recordings from Kadumane, the technician measured the duration, counted, and classified each vocalization (i.e., vocal species richness and number of each vocalization) in three randomly selected 5-minute clips from each day at each site (58 clips total). Because we were unable to isolate individual animals, the total number of individual sounds from each species was counted. When it was difficult to distinguish individual sounds, the total duration of a series of vocalizations was calculated and divided by the average duration of a single sound.
Because we collected longer recording samples at Kadumane, to standardize vocal species richness among Kadumane and Kodagu we divided the total number of unique avian species heard by the length of the recording sample (in min). For recordings from Kadumane, we calculated the Shannon index of diversity of avian vocalizations (Equations 1, vegan package; (Oksanen et al., 2013), the total number of avian vocalizations, and the total number of the most common avian species vocalizations for each 5 minutes recording sample. The Shannon index was calculated as:
where S was the number of avian species observed in the recording sample and was proportion of unique vocalizations belonging to the ith species (Gotelli and Chao, 2013).
All statistical analyses were performed in R version 3.4.1 (R Core Team, 2017). To predict the richness, diversity, and the total number of avian vocalizations in recordings from acoustic indices, we used a random forest (RF) machine learning procedure (Breiman, 2001; randomForest package; Liaw and Wiener, 2002). First we removed multivariate redundant acoustic indices using QR decomposition (threshold = 0.05; (Golub and Van Loan, 1996). We tested global RF models with all remaining acoustic indices as predictor variables. Using the 58 sub-setted clips from Kadumane and data from Kodagu combined we tested a global RF model with the richness of avian species vocalizations as the response variable. Using the 58 clips from Kadumane only we tested global RF models with one of three response variables: Shannon diversity of avian vocalizations, total number of avian vocalizations, and total number of the most common species vocalizations (Indian scimitar babbler, Pomatorhinus horsfieldii, see results). We selected the model with the lowest mean squared error (MSE) and highest R2 and used a model selection procedure to fit a final model with the fewest number of variables (rfUtilities package; (Evans and Murphy, 2017). We used this final model to predict the diversity of avian vocalizations in all recordings. To examine the importance of each acoustic index we used the mean percent increase in MSE, which is the loss of predictive accuracy due to the permutation of each variable. The more important a parameter, the larger the percent increase in MSE (i.e., the larger the effect on the model; (Breiman, 2001).
Continuous sounds are thought to confound the ability of acoustic indices to detect variation in bioacoustic signals (Pieretti and Farina, 2013; Parks et al., 2014). Thus, we tested whether continuous sounds affected the predictive ability of top acoustic index RF model by examining the relationship between model residuals and the duration and count of different sound types. We used a linear mixed model with absolute values of model residuals as a response variable and the following sound types as predictor variables: the duration (in sec) of anthropogenic (e.g., voices, motorcycles, vehicles) and constant biological sound (sounds >8 sec; e.g., insects, some birds); the presence or absence of rain; and the number insect and frog sounds. We used these sound types as covariates as they are relatively continuous over time and use a wide range of frequencies.
For recordings collected at Kadumane, 290 minutes (5% of total) were manually analyzed to count and categorize all sounds. For recordings from Kodagu, 328 minutes (95% of total) were manually analyzed to count the total number of avian species vocalizations. We found 39 avian species in recordings from Kadumane and 67 in recordings from Kodagu. At both locations, we observed the Large-billed crow (Corvus macrorhynchos) in the largest percentage of recording clips (29% and 39% of Kadumane and Kodagu recording clips, respectively). In recordings from Kadumane, the greatest number of total vocalizations was produced by the Indian scimitar babbler (Supplementary material, Appendix 3). Thus, we further analyzed whether acoustic indices could predict the total number of Indian scimitar babbler vocalizations, which were the most common species sounds.
Acoustic indices vs avian vocal activity and diversity
We found that acoustic indices weakly predicted the richness of avian species vocalizations in Kodagu recordings, Kadumane recordings, and recordings from both sites combined (R2 < 0.51, MSE >0.25, Table 2). Similarly, acoustic indices weakly predicted the total number of avian vocalizations and the number of Indian scimitar babbler vocalizations in Kadumane recordings (R2 < 0.47, MSE > 276.35, Table 2). Conversely, acoustic indices were strongly related to Shannon diversity of avian vocalizations in Kadumane recordings (R2 = 0.64, MSE = 0.17, Figure 3, Table 2); thus, we use this model in all further analyses. The top RF model predicting Shannon diversity of vocalizations included the following acoustic indices as predictors: acoustic complexity index (ACI) (Pieretti et al., 2011); spectral and temporal entropy (Sueur et al., 2008b); biophony and mean sound levels in the avian frequency bands (Joo et al., 2011; Tucker et al., 2014); acoustic activity, duration of acoustic events in the avian and noise bands, entropy of spectral maxima, and spectral persistence (Towsey et al., 2014), acoustic evenness index and acoustic diversity index (Villanueva-Rivera et al., 2011); spectral kurtosis (Bormpoudakis et al., 2013); and roughness and the difference between the 10th and 90th exceedance levels (Rychtáriková and Vermeir, 2013; Hong and Jeon, 2017). The most important acoustic indices in this top model were the ACI, roughness, and temporal entropy (Figure 4).
|Response variable||Recording sites||R2||MSE|
|Total avian sounds||Kadumane||0.47||4,265.40|
|Total sounds of Indian scimitar babbler||Kadumane||0.33||276.35|
|Species richness||All sites||0.51||1.12|
Predicting avian vocal diversity among sites
We found that RF models best predicted high and low Shannon diversity of avian vocalizations (Figure 5). Shannon diversity between 0.5–1.5 had the highest model residuals indicating that acoustic indices were less reliable when predicting mid-levels of avian vocalization diversity.
Linear models examining the effect of sounds on residuals from RF models showed that residuals were higher in recordings with longer duration biological sounds (Figure 6). These sounds ranged from 8 to 238 seconds (almost the entire duration of the 300 sec sound clip) and were produced by insects, birds, and frogs (n = 12, n = 23, and n = 5 of 58 respectively). Anthropogenic and rain sounds had little effect on the predictive ability of models, although they were present in very few of the recording clips analyzed (n = 6 and n = 2 of 58, respectively).
RF models predicted the highest Shannon diversity of avian vocalizations between sunrise and two hours after sunrise (0600–0800 Indian Standard Time, IST; Figure 7). At some sites, diversity of avian vocalizations also peaked just before sunset (1700–1800 IST). The predicted Shannon diversity of avian vocalizations from acoustic index models were weakly correlated with avian species richness observed in recordings from each site at Kodagu (Table 3). Avian species richness was lowest at the recording sites at Kadumane placed within forest remnants and 1 m from an open tea plantation (Kadumane South remnant forest and Kadumane North tea plantation, Table 3). Generally, we found that avian species richness was higher at sites in coffee and cardamom plantations versus tea plantations or remnant forest stands on tea plantations (Table 3, Figure 8).
|Site||Crop||Bird species richness||Predicted diversity|
|Mojo plantation 3||Cardamom||0.833 ± 0.255||0.761 ± 0.158|
|Faisal||Coffee||0.556 ± 0.032||0.278 ± 0.009|
|Subramani||Coffee||0.528 ± 0.052||0.329 ± 0.041|
|Rainforest R||Cardamom||0.514 ± 0.028||0.602 ± 0.084|
|Naveen||Coffee||0.5 ± 0.025||0.373 ± 0.054|
|Faisal||Cardamom||0.5 ± 0.012||0.497 ± 0.043|
|Shivappa||Cardamom||0.479 ± 0.036||0.206 ± 0.017|
|Rajeev||Cardamom||0.472 ± 0.064||0.282 ± 0.025|
|Subaiya||Coffee||0.444 ± 0.059||0.417 ± 0.022|
|Cloudbowl||Cardamom||0.417 ± 0.05||0.495 ± 0.046|
|Rajeev||Coffee||0.403 ± 0.056||0.311 ± 0.024|
|School estate||Coffee||0.403 ± 0.048||0.625 ± 0.043|
|Mojo plantation 2||Coffee||0.4 ± 0.074||0.494 ± 0.023|
|School estate||Cardamom||0.375 ± 0.049||0.566 ± 0.028|
|Uthappa||Cardamom||0.36 ± 0.054||0.383 ± 0.031|
|Bharat||Coffee||0.35 ± 0.027||0.383 ± 0.036|
|Golden mist||Coffee||0.347 ± 0.034||0.629 ± 0.04|
|Kushalappa||Coffee||0.306 ± 0.042||0.267 ± 0.026|
|Mojo plantation 1||Cardamom||0.278 ± 0.043||0.188 ± 0.013|
|Uthappa 2||Cardamom||0.253 ± 0.008||0.581 ± 0.036|
|Vivek||Coffee||0.191 ± 0.011||0.251 ± 0.013|
|Emmanuel||Cardamom||0.181 ± 0.012||0.346 ± 0.006|
|Brookview||Coffee||0.139 ± 0.008||0.453 ± 0.027|
|Brookview||Cardamom||0.095 ± 0.008||0.563 ± 0.023|
|Kadumane North||Tea||0.089 ± 0.051||0.472 ± 0.007|
|Kadumane South||Forest (remnant on tea plantation)||0.054 ± 0.005||0.578 ± 0.002|
|Kadumane South||Tea||0.043 ± 0.002||0.734 ± 0.001|
|Kadumane North||Forest (remnant on tea plantation)||0.02 ± 0.001||0.62 ± 0.001|
Acoustic monitoring has long been used to capture the diversity and abundance of vocalizing animals. Rapid technological advances in low-cost digital recorders has increased the potential for relatively cheap and effective passive acoustic monitoring over large extents (Merchant et al., 2015; Sueur and Farina, 2015). This new scope of acoustic monitoring has increased the scale of ecological inference, where large networks of acoustic recorders are poised to be the future of wildlife surveying (Sutherland et al., 2016). Given the urgency of addressing the biodiversity crisis, the rapid, standardized, and large-scale monitoring offered by acoustic recordings is imperative. Critical to acoustic methods is finding a way of quickly extracting meaningful information from enormous datasets. We tested the ability of a suite of acoustic indices to capture the diversity and richness of avian species vocalizations in recordings from the Western Ghats, a biodiversity hotspot.
We found that indices were of limited utility to capture the richness of avian species vocalizations, but accurately reflected the Shannon diversity of avian vocalizations in recordings (Table 2). Because indices measure variation in the acoustic environment they depend on both the number of species vocalizing and the abundance of each vocalization. Thus, acoustic indices are more likely to predict the diversity of vocalizations, which takes into account the abundance of each type of vocalization. Other studies found a significant relationship between avian species richness and acoustic indices (Sueur et al., 2008b; Joo et al., 2011; Pieretti et al., 2011). However, these recordings were collected in simpler acoustic environments with lower diversity of vocalizing animals (e.g., temperate forests) or indices were compared with simulated avian communities. In an environment with more complex acoustic characteristics, such as the Western Ghats, indices are more likely to reflect a combination of the abundance and different types of vocalizations.
We found that many acoustic indices were included in the top model predicting the diversity of avian vocalizations in recordings. Different indices have different mathematical properties and reflect different components of a soundscape (Gasc et al., 2015). Thus, combining groups of indices in a flexible modelling approach is most effective to predict the variation in the acoustic environment (Towsey et al., 2014). Indices that most affected the predictive ability of models were those that reflected both the temporal and spectral distribution of acoustic energy. For example, the ACI had the highest variable importance (i.e., permutation caused the highest increase in MSE; Figure 4). The ACI is calculated by summing the absolute difference in sound pressure levels (SPL) between adjacent seconds (i.e., temporal variation) divided by the total SPL in the recording sample over all frequency bins (i.e., frequency variation; (Pieretti et al., 2011). Also important was temporal entropy, which measures the evenness of sound over time (Sueur et al., 2008b). We found that acoustic descriptors, or acoustic indices summarizing sound energy (e.g., mean and median sound levels, Table 1), were less important in the model predicting diversity of avian vocalizations, likely because they are less effective at capturing acoustic heterogeneity. Our results suggest that both the variation and evenness in sound over time and frequency bins contribute to predicting the diversity of avian vocalizations in an acoustic sample.
We found some evidence of difference in predictive ability of models between Kadumane and Kodagu plantation recordings (Table 2). The richness of singing bird species per min of observed recordings was higher in the Kodagu plantation recordings (Table 3). Moreover, the composition of vocalizing bird species was slightly different between the two locations. Although acoustic indices should be insensitive to the identity of species vocalizations, there is some evidence from previous research that indices may change with the presence of particular species (Gasc et al., 2015). Thus, to ensure acoustic indices reflect vocal diversity among locations, future research should investigate the effect of different sets of species and the density of species richness in the acoustic environment.
Acoustic index models predicted high and low vocal species diversity with higher accuracy. Acoustic characteristics are expected to be more readily distinguishable between recordings with low and high diversity relative to sites where acoustic characteristics are only slightly different. Thus, some error can be expected for vocalizing communities with intermediate numbers of species. Moreover, we found that relatively continuous sounds (>8 sec, generally insects) affected the ability of index models to predict vocal diversity. In this sense, the acoustic environment of the Western Ghats makes the use of acoustic indices challenging—particularly, the high density of animals that produce extended sounds covering a broad spectrum of frequencies (e.g., cicadas, (Diwakar and Balakrishnan, 2007; Price et al., 2016). However, the relationship between model residuals and continuous biological sounds was weak (confidence intervals near zero—Figure 6). This suggests that the indices we included in predictive models captured at least some of the variability relating to the presence of continuous sources. The development of new indices may help identify and remove or control for the presence of unwanted continuous sources of broadband sound in long-term recordings. Further research may also consider the presence and effect of anthropogenic sounds, which were rare at our recording sites.
Although we were unable to quantify the richness of insect species, their stridulation sounds are not only an important feature of the acoustic environment, but also reflect a significant component of local biodiversity. Similarly, we did not focus on identifying anuran vocalizations to species. Previous studies have established that richness of insect and anuran species differs across different land-use types (Daniels, 2003; Mone et al., 2014). Thus, future research could focus on developing new indices or examining the predictive ability of a subset of existing acoustic indices for non-avian species diversity.
We found lower avian species richness at recording sites near open tea fields and in remnant forest stands on tea plantations versus sites in coffee and cardamom plantations (Figure 2). These findings are similar to earlier studies, where shade-grown crops are known to maintain higher levels of biodiversity (Sidhu et al., 2010). Because coffee and cardamom are generally grown under the shade of native rainforest trees (Figure 2), habitats are more complex, and some plantations are known to resemble neighboring remnant forest structure (Bhagwat et al., 2008). Conversely, tea plantations are intensively managed, dense monocultures of tea bushes with a sparse canopy of alien tree species, representing the most extreme habitat alteration relative to native rainforest (Raman, 2006). Tea plantations are a major form of land-use in the Western Ghats, having increased by 6,200 Ha (5.5%) between 2000 and 2006 (Mudappa and Raman, 2007). Moreover, recent observation suggests that coffee and cardamom plantations are shifting towards lower canopy cover. Thus, rapid assessments of biodiversity in these altered Western Ghat landscapes could be especially important given the potential for further land-use change.
We caution that recordings were collected at different times of year and in different years between sites at Kadumane and Kodagu. Recordings were collected in May at Kadumane, towards the end of peak breeding season of birds in this part of Western Ghats (Betts, 1952). Recordings were collected between October and March at sites in Kodagu (Table S1.1), which spans the post-monsoon to pre-breeding season, when several migrant species are present in southern India. Many avian species are less vocal outside of the breeding season, likely resulting in less vocal activity during the winter period. Although we would expect species richness of vocalizations to be inflated in Kadumane recordings, we observed the opposite, reinforcing our conclusion that avian species richness is lower in tea plantations. However, we recommend standardizing the season when recordings are collected to examine patterns between land-use types. Moreover, future research could focus on annual patterns in avian vocal diversity, examining the changes in patterns when winter migrants arrive and when resident birds breed. Finally, standardizing recording specifications, such as microphone gain and temporal sampling schemes are important when planning acoustic monitoring strategies (Pieretti et al., 2015).
We found that combining a suite of acoustic indices using a flexible modeling approach accurately predicted high and low avian species vocal diversity in recordings from sites in the Western Ghats, India. We also found that avian species richness was lower in tea versus shade-grown coffee and cardamom plantations.
The Western Ghats are a biodiversity hotspot, with high levels of diversity, endemism, and a complex mosaic of different land-use types. This fragmented landscape faces further changes, with the synergistic impact of continued development, habitat loss, and climate change threatening the persistence of biodiversity (Ponce-Reyes et al., 2013). In this context, acoustic recordings analyzed using acoustic indices represent a promising method of measuring avian diversity at large spatio-temporal scales.