Passive acoustic data yields insights into bird vocalization behavior associated with invasive shrub removal

Monitoring biodiversity changes associated with ecological restoration is crucial in the current UN Decade on Restoration. Although several studies highlight the impacts of restoration on ecosystems, it is also important to understand how restoration alters species' behavior, including vocalization. Advances in conservation technology, such as passive acoustic monitoring, facilitate rapid and noninvasive monitoring. In this study, we quantified changes in the vocal behavior of a bird species in response to small‐scale restoration carried out by removing the invasive shrub, Lantana camara (lantana), in a Central Indian tropical dry forest. We examined associations between sites of varying lantana densities and the vocalization of one of its primary dispersers, red‐vented bulbul (RVBU). We found statistically significant differences in note‐length and bandwidth of RVBU vocalizations across sites. A random forest classification model showed that Lantana density was not an important predictor of RVBU vocalizations. Apart from the percentage of forest area and farms in our sites, total human population was the most important predictor for RVBU vocalizations suggesting RVBU's use of human habitations as possible resource hubs. Our findings suggest that lantana removal is not associated with significant changes in RVBU vocalization. This study provides empirical evidence that ecological restoration may not alter species' vocalization in the short term and highlights the importance of moving beyond species presence and understanding the impact of restoration on species behavior.

Passive acoustic monitoring offers a distinct advantage over traditional survey methods especially in challenging habitats (Marques et al., 2013;Stowell & Sueur, 2020) where surveys can be prone to biases in terms of identifying species, for example.Data collection under passive acoustic monitoring allows the use of automated recording devices that collect ecological information, continuously or in a regular manner, encrypted in a "soundscape", a term used to define the collection of all sounds in an environment (Stowell & Sueur, 2020).These soundscapes are then available for multi-scalar analysis at various ecological levels, including assessing the broader or ecosystem-level analyses (e.g.Burivalova et al., 2018;Choksi et al., 2023) and targeted species-specific analyses (e.g., Digby et al., 2013).
Passive acoustic data is rich in ecological information on species, communities, and entire ecosystems (Sueur & Farina, 2015).This data, thus, has a lot of utility in assessing the species-and ecosystem-level impacts of restoration efforts, beyond the presence or absence of a species, which is commonly the indicator used in impact assessments (Campos-Cerqueira & Aide, 2016;Gibb et al., 2019).Other restoration impact studies focus on the larger ecosystem through the examination of soundscapes (Burivalova et al., 2018;Choksi et al., 2023;Ramesh et al., 2023).Although these indicators are important to measure, there remains a gap in our understanding of the narrower, species-level impacts of restoration, for example, on vocalizing behavior.Studying the impact of restoration efforts on species' vocalization behavior holds value, as vocalizations play a crucial role in sexual and social communication in vocalizing taxa (Garcia & Favaro, 2017;Kumar, 2003), which thereby benefit their survival and existence.Consequently, since restoration changes habitat structure, species are expected to adapt their vocal signals to optimize signal transmission (Acoustic Adaptation Hypothesis; Morton, 1975).Since closed habitats tend to cause signal reverberation and attenuation, species have been shown to produce long-distance, low-frequency signals with longer notes and lower frequency modulations (Bradbury, 2011;Ey & Fischer, 2009;Morton, 1975;Wiley & Richards, 1978).In comparison, species in open habitats vocalize with shorter, more repetitive notes (Bradbury, 2011;Ey & Fischer, 2009;Morton, 1975;Wiley & Richards, 1978).Open and closed habitats are also highly contrasting in light penetrability and availability, which might affect the temporal structure (signalling at different times of the day) of the vocalizations.Thus, the influence of habitat structure on acoustic structure warrants studies on adaptations of vocalizing species to ecological changes associated with restoration.
The red-vented bulbul (Pycnonotus cafer), henceforth RVBU, is a tropical generalist frugivore with a wide distribution across the Indian subcontinent.Studies examining the mutualistic link between the common invasive shrub Lantana camara (lantana) and avifauna found that bulbul species, especially RVBU, were among the primary avian frugivores responsible for lantana dispersal (Ramaswami et al., 2016(Ramaswami et al., , 2017)).Lantana, which is native to Central America, was introduced in India by the British in the 1800s (Mungi et al., 2020).Due to its high ecological resilience, lantana tends to predominate the forest understory (Negi et al., 2019), thereby changing the forest structure and bird assemblages (Aravind et al., 2010).Previous studies also reported a higher density of RVBU associated with a higher density of lantana compared to other frugivore species (Aravind et al., 2010), suggesting that RVBU could be used as an indicator of ecological changes associated with lantana.Bhatt and Kumar (2001) found that ripe lantana fruits were among the top three food preferences for RVBU among nine varieties of indigenous plants and eight varieties of exotic plants.Lantana-invaded forests thus provide a preferred habitat for RVBU.RVBU is also known to be vocally active (Brooks, 2013) making it an ideal target species for assessing the species-level outcomes of restoration through lantana removal in a tropical dry forest in Central India using acoustic technology.
Despite providing a multitude of ecosystem services such as climate regulation and flood control (Maass et al., 2005), tropical dry forests, particularly compared to other biomes, remain a lesserstudied biome (Dirzo et al., 2011;Powers et al., 2018;Schröder et al., 2021).Tropical dry forests cover 42% of the tropics, and 97% of the current tropical dry forest area stands threatened (Miles et al., 2006).These forests are crucial biodiversity reserves with high levels of endemism (Morales et al., 2015).In several parts of the world, they support local livelihoods (Choksi, 2020) making them susceptible to land use change and other human modifications.
Therefore, tropical dry forests are an important biome to restore and monitor (Morales et al., 2015).In this study, we examined the impact of restoration through the removal of lantana in Central India on RVBU's vocalization behavior.The restoration effort was carried out primarily for the benefit of the local people (Choksi et al., 2023).
We performed a passive acoustic survey originally with the goal of examining the differences in the entire forest soundscapes associated with restoration (Choksi et al., 2023).We now extend this work through this study with the objective of examining the differences in vocal behavior of RVBU, which we use to indicate the overall response of this species to restoration through lantana removal in a tropical dry forest in Central India.Specifically, we ask the following questions: (i) Are there significant differences in the total number of detections and the proportion of RVBU vocalization types detected across sites?
We expected a higher number of total RVBU detections in lantana-invaded unrestored sites (sites with high lantana density in the understory) compared to restored sites (sites where lantana has been removed) and sites with naturally occurring low density of lantana in the understory, or low lantana density (LLD) sites, due to RVBU's association with lantana.We also expected a significant difference in the proportion of RVBU songs and contact calls across our sites as we hypothesized that RVBU, being one of the primary dispersers on lantana, utilize unrestored sites more than restored and LLD sites.
(ii) Are there significant spectral differences in RVBU vocalizations across sites?
We expected that the sites, which vary in terms of the density of lantana in their understory have an impact on how RVBUs vocalize.Following the Acoustic Adaptation Hypothesis (Ey & Fischer, 2009;Morton, 1975), differences in the vegetation structure could make RVBUs adapt to these differences by adjusting their spectral features for efficient signal transmission.
Therefore, we specifically hypothesize that the spectral parameters of RVBU vocalizations in restored sites with no understory, lantana density would be significantly different than RVBU vocalizations in unrestored and LLD sites.
(iii) Are there significant temporal differences in RVBU vocalizations across sites?
Because of the existing mutualistic link between lantana and RVBU, we expected RVBUs to adjust their time of vocal activity in the sites where lantana has been removed from the understory.Since vegetation structure affects light penetrability and, in turn, avifauna activity, we hypothesized that with differences in lantana density in the understory, we would also find significant differences in the vocal activity of RVBU across dawn hours in our sites.
(iv) Is there a significant association between lantana density and RVBU based on acoustic data?
We expected RVBU vocalization to be associated with lantana density in our sites.Because RVBU is one of lantana's known primary dispersers, we hypothesized that removing lantana from the understory would have an impact on RVBU's visitation rate in restored sites, post restoration.We hypothesized that changes in the visitation rate of RVBU would also change RVBU's vocalization behavior, And thus, we expected an association between lantana density and the proportion of vocalizations.We expected acoustic data to reflect the mutualistic link between lantana and RVBU that has previously been proven by studies using visual surveys.

| Study area and selection of sites
We conducted this study in the buffer region of Kanha National Park (KNP), a critical tiger conservation landscape in the Central Indian Highlands.The study sites are in the subdistrict Bichhiya, in the district of Mandla, Madhya Pradesh, India.The region contains predominantly tropical deciduous vegetation (Agarwala et al., 2016), which supports several threatened species such as the Bengal tiger (Panthera tigris) and sloth bear (Melursus ursinus) (Jhala et al., 2008;Thatte et al., 2018).The sites also support the livelihoods of legally established populations of Scheduled Caste and Tribe members who depend on forest resources for sustenance (Choksi, 2020;DeFries et al., 2022).These forests lie adjacent to farmlands and human habitations (see Figure S1).We specified 20 sampling sites (minimum 20 ha, mean area: 58.32 ± 30.93 ha) inside the buffer area of KNP that are utilized by the local communities to gather and use forest resources for livelihood.We identified eight restored sites that were restored by removing lantana from the understory and matched them to eight unrestored sites with high lantana density and four reference (LLD) sites with naturally occurring LLD using a propensity score.We carried out the statistical matching of sites based on a propensity score derived from geographic, socio-economic and site-level vegetation factors using the R package matchIt (Stuart et al., 2011).Further statistical matching details are provided in Choksi et al. (2023).

| Acoustic data collection
At every recorder location, we used Audiomoth 1.0.0 acoustic recorders (Hill et al., 2019) with a 48 kHz sampling rate and medium gain.We placed at least two recorders (± 1) within the core of the sampling sites (core was created by "buffering in" the sampling site polygon by 70 m to ensure no sounds from outside the sampling site were recorded) at locations determined by the random point generator in QGIS (46 total recorder locations).The distance between any two recorders within each sampling site was 380-540 m to ensure there was no pseudo-replication in the data (Alcocer et al., 2022).We tied the recorders on the trunks of trees at a height of approximately 1.5-2 m above the ground.The recorders were put in a Ziploc bag to avoid any damage from water.We programmed the recorders to automatically collect 1-min recordings every 5 min for 24 h a day over

| Estimating RVBU presence in acoustic data
We manually annotated the presence of RVBU in the sound recordings.To do this, we randomly sampled a subset of a minimum of 45 min of data during the dawn hours between 0530 h and 0930 h.
We then split the 1-min recordings into 10-s clips to help make the manual annotation process easier.Two regional eBird reviewers with expert knowledge of regional avifauna used Raven Pro 1.5 (Cornell Lab of Ornithology, 2014) to manually investigate the 10-s clips.This resulted in a presence/absence matrix of all avifauna species present in each 10-s clip.This matrix was built to estimate the overall species richness for Choksi et al. (2023).We then used a filter on this matrix to only extract a random sample of fifty 10-s clips, per recorder location, that had the presence of RVBU vocalizations to further extract temporal and spectral features of RVBU vocalizations.We carried out a spectral analysis of each of the 50 10-s clips in Raven Pro 1.5.We referred to Kumar (2004) to identify the type of RVBU vocalization (Song, contact call or alarm call).In the event of a misannotation, that is, if we came across a segment with no RVBU calls, we removed the segment from the analysis.We analyzed a total of 659 10-s clips across all sites.For this section of the analysis, we considered four outcome variables which we explain, in detail, in the sub-sections below.

Total RVBU detections
For this outcome variable, we used the presence/absence matrix subset with RVBU detections only (Refer to section 2.3.1).We used this outcome variable to study the differences in RVBU detection in the three site categories-restored, unrestored and LLD.First, we randomly selected a thousand 10-s clips for each site category from the presence/absence matrix.We then computed the cumulative total of all RVBU detections (all clips where RVBU were marked present).This was repeated for a thousand iterations using a simple bootstrap algorithm.

Proportion of all vocalizations
For this outcome variable, we considered all RVBU vocalizations (songs, alarm calls and contact calls) detected in the 10-s clips for each of the three site categories Equation 1 explains the computation of the proportion of all vocalizations.

Proportion of songs
For this outcome variable, we considered all RVBU songs detected in the 10-s clips for each site category.We defined the proportion of RVBU songs as the total number of songs detected in a day over the total number of vocalizations detected in a day.Equation 2 explains the computation of the proportion of songs.

Proportion of contact calls
For this outcome variable, we considered all RVBU contact calls detected in the 10-s clips.We defined the proportion of RVBU contact calls as the total number of contact calls detected in a day over the total number of vocalizations detected in each day for each site category.Equation 3 explains the computation of the proportion of contact calls.

| Spectral analysis
For this section of the analysis, we broke down each vocalization detected into separate notes.We manually built bounding boxes around each note detected within each vocalization occurring in each 10-s clip.We considered each note of a song or call as an individual unit of analysis because of the absence of any amplitude breaks within them.By minimizing the presence of background noise in the white spaces between our units of analysis, notes, we aimed to get the most accurate spectral measures in Raven Pro 1.5.We identified notes which were visually distinct from their subsequent notes within a song or call (Figure S3).We visually classified notes, separating them by the presence of a silent gap of less than 1-s.
Table 1 provides information about the five commonly tested spectral outcome variables that were considered in Raven Pro during the spectral analysis with their respective range.

| Statistical analysis
We used the response variables-proportion of all vocalizations, the proportion of songs and the proportion of contact calls to assess the association between lantana density in each type of site (restored, unrestored or LLD) (Table S4 provides details about categorizing lantana by density) and the type of RVBU vocalization.We do not show results for RVBU alarm calls since we had under 10 detections for this type of call in two site types (Table S5) and do not consider that a large enough sample for meaningful analysis.
For the temporal analysis, we took the average of the proportion of all vocalizations, the proportion of songs and the proportion of contact calls for every 15-min time bin within the dawn hours 0530 h-0930 h.We then used the Wilcoxon test to determine if the proportions in each time bin for each site type were significantly different.For spectral analysis, we used violin plots to visualize the distribution of the spectral parameters and used the function geom_signif (), which carries out a Wilcoxon test, from the R package ggsignif (Ahlmann-Eltze & Patil, 2021) to display the statistical significance of differences within the three site groups.
We also used the random forest classification algorithm using the R package randomForest (Liaw & Wiener, 2001) to measure the importance of six geographic and vegetation predictor variables as used in Choksi et al., 2023 (Table S3 provides details about all six geographic variables) for explaining our response variables, especially lantana density (Table S4).We used the results of this analysis to comment on the association between lantana and RVBU using acoustic data.All statistical analyses were performed in the R programming environment (R Core Team, 2021).

| Differences in RVBU vocalization types
We observed a higher number of RVBU detections in unrestored Proportion of all vocalizations = Total number of RVBU vocalizations detected in a day Total number of clips analysed in a day (2) Proportion of songs = Total number of RVBU songs in a day Total number of all RVBU vocalizations in a day (3)

Proportion of contact calls =
Total number of RVBU contact calls in a day Total number of all RVBU vocalizations in a day Table S6) suggested a significant difference in the total number of detections across the three site categories (Figure 2).S7).
We also found a significantly lower median bandwidth in LLD  S8, S9).We found no significant difference between all site categories for the other spectral parameters considered in this study (Figure S4; Tables S8 and S9).We provide the results and fig- ures for PFC Min frequency and the number of notes in Figure S4, Tables S8 and S9 in the supplementary information.

| Differences in temporal patterns of RVBU vocalizations
Temporally, we found no significant differences between the global peak values which represent the average for the variables-the proportion of all vocalizations (LLD-restored: z = −2.17S10) in each 15-min time bin for each site type (Figure 6).

| Association between lantana and RVBU vocalization behavior
Our random forest classification model demonstrated that the most important explanatory variable for the proportion of all RVBU vocalizations is the percentage of forest area within a 3 km buffer area (Figure 7).We found that lantana density ranked lowest in variable importance for all vocalizations.However, we found that lantana density ranked low (if not lowest) in variable importance for the proportion of songs and contact calls, individually (Figure S5a,b).
Apart from the percentage of forest area, among the top three predictors for the proportion of all vocalizations, the proportion of songs and contact calls were the percentage of farms and the total human population within a 3 km buffer.Lantana density was among the top three predictors for the spectral features-number of notes (for songs only), minimum frequency and maximum frequency (for both songs and contact calls) (Figures S6 and S7).However, overall, we did not find lantana to be a significant predictor of RVBU vocal behavior.

| DISCUSS ION
Our study examines the association between lantana and RVBU vocalization behavior in sites with varying densities of lantana in the understory, post-restoration.Overall, we found that lantana density was not significantly associated with the proportion of RVBU vocalizations.We also did not find an impact of the restoration carried out by removing lantana on the temporal and spectral characteristics of RVBU vocalizations.

| Association between RVBU vocalization and lantana density
Previous studies report a greater number of visits by RVBU on lantana shrubs compared to other frugivores, especially for foraging, and suggest RVBU as one of lantana's prime dispersers (Aravind et al., 2010;Bhatt & Kumar, 2001;Ramaswami et al., 2016).
Removing lantana from the understory could therefore have an impact on RVBU vocalization which could indicate changes in the visitation rate and overall activity of RVBU in our sites.We therefore, expected lantana density to be associated with different aspects of RVBU vocalization.
Our results show that the median total detections were significantly higher in unrestored sites with high lantana density compared to the restored sites with no lantana in the understory.
However, our Random Forest classification model showed that lantana density was not the most important predictor of the acoustically derived proportion of all vocalizations.Additionally, we found that the most important predictor of the proportion of contact calls was the total human population within 3 km of our sampling sites.
RVBU is known to frequent parks, gardens and farmlands, suggesting an association with human habitations (Islam & Williams, 2020), especially for congregations.We postulate that because RVBU is a generalist frugivore that sometimes also feeds on insects and plant material (Islam & Williams, 2020), the presence of farmlands and home gardens in human-inhabited areas across all our sites may act as potential resource hubs for RVBU in sites with no or less lantana.We also speculate that the observed differences in the total number of RVBU vocalizations detected in our sites could be the result of factors other than lantana density such as differences in the structural complexity of the forest and fluctuating microclimate that affect the propagation and detection of sound signals (Wiley & Richards, 1978).

| Temporal and spectral responses of RVBU to lantana removal
Sound propagation in forested areas is largely affected by the environment (Marten & Marler, 1977).In LLD sites, the vocalizations had notes with a narrow bandwidth and longer note length, whereas, in unrestored sites with high lantana density, the vocalizations had significantly shorter notes and a broader bandwidth.Variables such as vegetative elements comprising stems and leaves and climatic variables such as wind and temperature could affect the way sound signals propagate in a habitat.Consequently, animals often adjust their vocalization to maximize transmission (Forrest, 1994).Longer notes with a narrow bandwidth travel farther in dense overall vegetation and shorter, more frequent notes ensure effective transmission in relatively open habitats (Bradbury, 2011).We speculate that RVBUs adjust the bandwidth and note length of their songs and calls according to the varying levels of lantana density in the understory in our sites.Morton (1975) found that tropical birds that prefer undergrowth avoid frequency modulation to reduce scattering and ensure effective long-range communication.We speculate that the difference in vegetation structure due to varying lantana density could lead to the observed differences in note length of songs and contact calls with shorter notes in high lantana density sites and longer notes in LLD sites.Therefore, we speculate that a higher density of lantana in the understory of unrestored sites could be associated with shorter note lengths of songs and calls.
Although our results suggest that lantana removal is not associated with changes in RVBU vocalization pattern and behavior, we believe there may be other behavioral aspects that could show an association, for example, habitat use.We invite researchers to further examine the less-explored influence of restoration efforts, especially through invasive species removal, on the behavior of species known to be associated with the species being removed.Further, this study demonstrates an effective method to understand how species respond to landscape changes and how acoustic data could be used to reveal species-specific restoration impacts.

| Limitations
This study is not without limitations.We used passive acoustic data to study the vocalization behavior of a single species.We hypothesize that actively collected acoustic data using directional microphones with minimal interferences of sounds from other sources may provide fewer but more accurate data points in terms of the spectral parameters relative to passive acoustic data.Active acoustic data comprising focal recordings also allows researchers to account for individual-level differences and quantify species abundances, which is often challenging with passive acoustic data.In this study, we could not account for the individual-level differences, however, the approach of using passive acoustic data allowed us to capture data over larger spatial and temporal scales.Given the scale of restoration efforts being taken worldwide, tools such as passive acoustic monitoring pose as a more favorable option for biodiversity or species-specific monitoring across large areas in a short time.

| CON CLUS ION
Restoring forests changes the ecological and structural complexity of the forests (Bullock et al., 2022;Camarretta et al., 2020) and thus, potentially, affects the way species communicate.However, we found that efforts to restore degraded forests invaded by lantana over a short term (the 3-year period following restoration) do not show an association with major significant changes in the vocalization behavior of a generalist frugivore, the RVBU, in this Central Indian tropical dry forest.We found that lantana density was not

ACK N OWLED G M ENTS
We are grateful to the Madhya Pradesh Forest department for the permission to carry out this study and community members in villages where data was collected.PC thanks the Marawi family for their hospitality while on field to collect data.PC received funding from the National Geographic Society, Center for Environmental Economics and Policy (CEEP) and the Advanced Consortium on Cooperation, Conflict and Complexity (AC4) to carry out research on the impact of ecological restoration on biodiversity and people, for which the data used in this study was originally collected.We thank Devendra Korche, Siddharth Biniwale and Pravar Mourya, who were involved in the collection and analysis of the soundscape data in Choksi et al. (2023).We thank Dr. Geetha Ramaswami for a friendly review that helped us greatly improve this manuscript.

F
I G U R E 1 (a) Map showing the recorder locations (blue, green and red symbols) within each of the 20 sampling sites in the buffer area of Kanha National Park in Bicchiya subdistrict, Mandla district, India.Each sampling location within sampling sites is approximately 400 m from any other sampling location to avoid any overlap in acoustic data collection.Photo from sampling location: (b) MH_2, (c) PT_3 and (d) SB_2.TA B L E 1 Spectral parameters considered with definition, the unit of measurement and mean ± standard deviation.

F
Violin plots displaying differences in the spectral parameters (on the Y-axis) of RVBU songs.(a) PFC Max frequency, (b) bandwidth 90%, and (c) note length.Outliers for all parameters greater than 97.5 percentile were removed.F I G U R E 5 Violin plots displaying differences in the spectral parameters (on the Y-axis) of RVBU contact calls.(a) PFC Max frequency, (b) bandwidth 90% and (c) note length.Outliers for all parameters greater than 97.5 percentile were removed.

F
I G U R E 6 The proportion of (a) All RVBU vocalizations, (b) songs, and (c) contact calls plotted over the dawn hours between 0530 h and 0930 h.Peaks highlighted by red dots represent the global peaks in proportion for each of the three site types.F I G U R E 7Variable importance plot from the random forest model with the proportion of all vocalizations as the response variable.'IncNodePurity' is a measure of how much the model error increases when a particular variable is randomly permuted or shuffled.
associated with changes in RVBU spectral and temporal vocalization patterns across our sites.It is possible for significant changes in vocalization behavior to show over the long term as RVBU adapts to new niches formed post-restoration.Long-term monitoring could, therefore, help better understand the association between lantana removal and RVBU.This study makes a timely contribution to the field of exploring narrower, species-level impacts of restoration practices.AUTH O R CO NTR I B UTI O N S Mayuri Kotian: Investigation, Methodology, Formal analysis, Visualization, Writing-original draft.Pavithra Sundar: Investigation, Data Curation, Writing-review & editing.Taksh Sangwan: Investigation, Methodology.Pooja Choksi: Conceptualization, Methodology, Supervision, Writing-review & editing.
Table S2 provides the exact number of clips ana- lyzed per site.We considered nine outcome variables (four outcome variables in section 2.3.2 and five outcome variables explained in Table1) for analysis that was broadly classified as spectral analysis and species presence analysis.