Measuring biodiversity with sound: How effective are acoustic indices for quantifying biodiversity in a tropical dry forest?

Large‐scale biodiversity conservation and management necessitate rapid assessment tools and technologies. Indices derived from passive acoustic data offer a novel solution for rapid biodiversity monitoring. Although these indices quantify vocalizing biodiversity at a site, previous studies indicate inconsistencies in the indices' performance across different biomes. We tested the efficacy of seven acoustic indices—acoustic complexity index, acoustic diversity index, bioacoustic index (BI), acoustic entropy index, total entropy (H), normalized difference soundscape index, and number of peaks in an understudied biome, tropical dry forests, in central India. We measured the relationship between every index and a combination of these indices with a biodiversity metric, avian species richness. We found a weak correlation between individual indices and species richness (0.00 ≤ R ≤ 0.35), while a combination of indices was comparatively better at predicting species richness (R2 = 0.54). Although BI performed better than all other indices, our results indicate that acoustic indices do not accurately quantify avian species richness in this forest in central India. However, combining multiple indices increases their efficacy, limitedly. We recommend evaluating the efficiency of acoustic indices, especially in underrepresented habitats, before their application in avifauna‐based rapid acoustic surveys.


| INTRODUCTION
As humans continue to alter landscapes at a fast pace (Sala et al., 2000), it is crucial to monitor and manage the impacts on biodiversity with efficient and scalable monitoring tools (Gardner, 2010;Siddig et al., 2016).For biodiversity conservation and management practice, technology-based tools can help provide detailed information on the status of biodiversity and determine species responses to ecological changes in a timely manner (Lahoz-Monfort & Magrath, 2021).Recent advances in conservation technologies using sound, such as passive acoustic monitoring (PAM), have paved the way for rapid biodiversity assessments at large spatio-temporal scales (Stowell & Sueur, 2020).PAM provides a faster, noninvasive monitoring method, which also reduces possible biases in species identification and recording, by providing a permanent record of the raw data (Sueur et al., 2008;Sugai et al., 2019).
The analysis of soundscapes, which are collections of sounds emerging from a landscape, has helped answer crucial ecological questions such as detecting the presence of a species (Brodie et al., 2020;Digby et al., 2013), estimating the arrival times of migratory species (Oliver et al., 2018), estimating population densities (Marques et al., 2013;Pérez-Granados & Traba, 2021), and monitoring species assemblages using acoustic surveys (Ulloa et al., 2019).Acoustic surveys can help reduce the cost of biodiversity surveys substantially through the use of open-source acoustic hardware and software (Hill et al., 2019;Kahl et al., 2021;Pérez-Granados, 2023;Whytock & Christie, 2017) and offer a method to collect large acoustic datasets in a less invasive technique for longer intervals (Bardeli et al., 2010).
One of the many approaches to soundscape analysis focuses on broader ecosystem-level questions and falls under the umbrella of ecoacoustics (Stowell & Sueur, 2020;Sueur & Farina, 2015).Ecoacoustics is the study of sounds and their ecological role within soundscapes, comprising biophony (sounds originating from biological sources), anthrophony (sounds originating from human sources), and geophony (sounds originating from geographical or climatic sources) (Pijanowski et al., 2011;Stowell & Sueur, 2020;Sueur & Farina, 2015).Ecoacoustic analyses aim to study the vocalizing biodiversity in soundscapes by quantifying the variation in acoustic features, such as amplitude and frequency, via the use of various acoustic indices (Buxton et al., 2018;Sueur et al., 2008).
Some acoustic indices (AI) are based on the assumption that high species diversity is associated with higher acoustic complexity (Sueur et al., 2014) and thus a richer soundscape, signifying a "healthier" ecosystem (Eldridge et al., 2018).Sueur et al. (2014) 1).Commonly used acoustic indices such as the acoustic evenness index (AEI), acoustic diversity index (ADI), and total entropy (H) quantify the evenness of acoustic energy across the soundscape (Villanueva-Rivera et al., 2011).Whereas, indices such as the bioacoustic index (BI) and the acoustic complexity index (ACI) quantify the variation in the intensity of sounds present in the soundscape (Boelman et al., 2007;Pieretti et al., 2011) which are characteristic of avian vocalizations.Number of peaks (NP) measures the complexity of the soundscape by counting the number of frequency peaks in a mean spectrum (Gasc et al., 2013), and the normalized difference soundscape index (NDSI) quantifies anthropogenic disturbance in the soundscape by providing a theoretical ratio of anthrophony to biophony (Kasten et al., 2012).
Researchers have used acoustic indices to quantify the impact of human disturbance in various ecological regions.For example, Burivalova et al. (2018) used an acoustic index to test whether land management performed by communities affects the vocal biodiversity in Papua New Guinea.The authors of the study found significantly higher levels of soundscape saturation in areas where forest cover was preserved.In another study from Colombia, G omez et al. ( 2018) used acoustic indices to assess the quality of habitat and found that acoustic indices helped achieve higher accuracy while classifying habitats based on varying degrees of disturbance.Rajan et al. (2019) used acoustic indices to quantify soundscape characteristics by measuring the temporal changes in biophony in three habitat types in Kerala, India.These studies, thus, highlight the potential use of acoustic indices as proxies for indicators of biological diversity and for quantifying ecological changes such as ecological restoration as well as degradation.
Despite these encouraging studies, many indices have shown inconsistencies across various types of biomes, degrees of species diversity, vocalizing communities, and environmental complexity, especially in tropical forests (Buxton et al., 2018;Jorge et al., 2018;Mammides et al., 2017;Sethi et al., 2023).For example, Eldridge et al. (2018) found significant correlations between ACI and bird species richness in a temperate biome in the United Kingdom, but the index weakly correlated with bird species richness in a neotropical biome in Ecuador.Although some indices show great individual potential for acoustic surveys in some biomes, Towsey, Parsons, et al. (2014) and Towsey, Wimmer, et al. (2014) suggest using a combination of multiple indices to increase their efficiency.However, the dissimilarity in the performance of acoustic indices within similar biomes and across biomes could be associated with factors such as differences in the structural complexity of the biome and level of disturbance, all of which affect the propagation of sound in these biomes.For example, the type of habitat influences the degree of attenuation of sound signals and thus affects how far sound travels (Marten & Marler, 1977;Wiley & Richards, 1978).Further, Alcocer et al. (2022) reported a decrease in the number of studies testing the relationship between acoustic indices and biodiversity metrics, such as species richness, before using them for quantifying biodiversity in underexamined biomes.The authors also emphasize that the use of these indices without establishing the strength of their relationship with biodiversity may lead to false findings.This underscores the need for more research on the potential and limitations of the indices as rapid biodiversity monitoring tools in different kinds of ecosystems.
Most of the literature evaluating the performance of acoustic indices focuses on temperate and tropical humid forests.The few studies on tropical dry forests are from Latin America (e.g., Rendon et al., 2022) and Africa (e.g., Rankin & Axel, 2017).Tropical dry forests, particularly in Asia, remain underrepresented in the field of ecology and ecoacoustics literature, especially concerning acoustic indices (Izaguirre & Ramírez-Al an, 2018;Oliveira et al., 2021).We speculate that the differences between temperate, tropical humid, and tropical dry forests in terms of vegetative differences, diversity of vocalizing communities, and socio-ecological differences could lead to differences in the performance of acoustic indices in these biomes.Unlike tropical humid forests, tropical dry forests are structurally less complex, with more floristic species in the understory than trees (Van Bloem et al., 2004).With their low structural complexity, sound signals could travel farther in tropical dry forests than in tropical humid forests, leading to differences in the way indices perform in this biome.Tropical dry forests provide a plethora of important ecosystem services (Morales et al., 2015), such as climate regulation, bio-regulation, and flood control (Maass et al., 2005).These forests are also often socio-ecological systems that provide livelihood support for forest-dependent local communities (Choksi, 2020) and are thus crucial to monitor.
In our study, we evaluate the performance of seven acoustic indices in a tropical dry forest in the Central Indian Highlands (CIH) for predicting general biodiversity using an acoustically derived metric-avian species richness.Specifically, we ask the following questions: I. Do the acoustic indices correlate with the avian species richness?
We expect all indices to correlate with avian species richness.We expect a stronger correlation for intensity-based indices (ACI and BI), which are expected to capture variation in the intensity of signals in the soundscape caused by avifauna vocalizations.II.Does a combination of acoustic indices increase their efficacy in predicting avian species richness?
We expect that a combination of multiple indices collectively increase their efficacy in predicting avian species richness.

| Study region and site selection
We conducted the study in the officially designated buffer region of Kanha National Park (KNP) in the subdistrict Bichhiya in Mandla district, Madhya Pradesh, which is part of the CIH.The region is an important conservation landscape for threatened species such as the Bengal tiger (Panthera tigris), the Indian leopard (Panthera pardus), and the Sloth bear (Melursus ursinus) (Jhala et al., 2008;Thatte et al., 2018).The landscape is dominated by tropical deciduous vegetation (Agarwala et al., 2016) with mixed forest tree species such as Sal (Shorea robusta), and farmlands nearby.The region holds one of the largest populations of legally established Scheduled Caste and Tribe members in India who depend on timber and non-timber forest products for livelihood (Choksi, 2020;DeFries et al., 2022).We established 15 sampling sites of a minimum area of 20 ha (mean + standard deviation [SD]: 64.72 ± 32.77 ha), which local communities living in the buffer region utilized for subsistence and livelihood needs.We then used the random point generator in QGIS to establish a minimum of two recorders per sampling site (45 total recorder locations, mean = 3 recorders) at a distance of 380-540 m from each other within the core of the 15 sampling sites (mean distance between sampling sites centroids = 2200 m).All 15 sampling sites are statistically comparable (see Table S1 for more information) in terms of vegetation structure and only differ in terms of understory shrub density, with seven sites having practically no understory, four sites having low-density understory and four sites having a high-density understory.The predominant species found in the understory across all sites is an invasive shrub, Lantana camara (Figure 1).

| Data collection
We collected the acoustic data over a period of 7-10 days during the winter season in December-January for the Year 2020 and January-February for the Year 2021.In case of a malfunction of a recorder, we collected data in March 2020 and February 2021, whenever we got the opportunity to travel to sites in between lockdowns for different coronavirus disease of 2019 (COVID-19) variants.At each sampling location, we used the Audiomoth 1.0.0 acoustic recorder (Hill et al., 2019) with a sampling rate of 48 kHz and medium gain.We programmed the recorders to collect 1-min samples every 5 min for 24 h a day to capture the temporal variation in the samples during each year (Bradfer-Lawrence et al., 2019).In total, for each sampling location, we had 71.37 ± 1.78 h of data collected.

| Estimation of avian species richness
For manual annotation of all the avian species detected in the acoustic data, first, we randomly sampled a subset of a minimum of 45 min within the dawn hours between 05:30 and 09:30 a.m. from each recorder location across all days sampled per year.In the event of malfunctions or bad weather encountered in the audio segment, we analyzed additional segments (Table S2 provides the total number of files and minutes annotated).To simplify the annotation process, we split the original minute-long recordings from the 45-min subsample into 10-s segments (N recordings analyzed: 20,700).Two authors, who are regional eBird reviewers with expertise in avian species and their wide vocalization repertoire found in the study region, used Raven Pro 1.5 (Cornell Lab of Ornithology, 2014) to manually inspect the 10-s segments for annotation.We used a presence/absence matrix to note whether a particular avian species was heard in the segment or not.Avian vocalizations that could not be identified were sent to other eBird experts, and the correct identification was then finalized based on the majority opinion (Table S3 provides a list of all species acoustically detected in the sites).Each observation in the models presented in this study (N = 20,700) represents one 10-s audio segment.

| Computation of acoustic indices
We analyzed the acoustic data using the R programming language (version R 4.3.1)(R Core Team, 2021).
Six indices, ACI, BI, ADI, AEI, H, and NDSI, were computed using the multiple_sounds function in the soundecology package (Villanueva-Rivera et al., 2011).For indices where we could set limits for minimum and maximum frequency within the function, such as ACI and BI, we computed the indices for the frequency range  For each 10-s audio segment, we also computed the NP, an acoustic index that counts the number of frequency peaks within a spectrum (Gasc et al., 2013).For this, first, we created a mean spectrum of each 10-s file from all sites and across all days by computing a shorttime Fourier transform using the meanspec function from the seewave package (Sueur et al., 2008).We set the sampling frequency of each WAVE file as 48,000 Hz and the length of the window for analysis as 512 (Hanning window).We set the argument "norm" as FALSE to obtain the absolute amplitude values.Thus, we had 256 time/ frequency bins with absolute amplitude values.We then used the fpeaks function in the same package to detect the peak frequencies in each mean spectrum, where we set the frequency distance threshold to 0. This resulted in a two-column matrix with spectral frequencies in the first column and amplitude values in the second.To calculate the total NP for each 10-s segment, we counted all peaks present in each segment across the frequency range of 2-11 kHz.We then aggregated the exact time of recording for each 10-s segment into forty-eight 30-min bins.We defined NP in each soundscape recording as the total NP present in each 10-s segment.Table 1 includes all indexspecific frequency ranges that we considered in this study (Table S4 provides the range and mean ± SD for all indices computed).

| Statistical analysis
For all statistical models, as a response variable, we used avian species richness.The acoustic indices were the predictor variables in the models.All measures were calculated for each observation of 10-s soundscape segments.
To test the direction and magnitude of the correlation between each acoustic index and the response variable, we used Pearson's correlation coefficient.In addition, we also used a non-parametric regression analysis using Spearman's rank correlation test to examine the nonlinear monotonic relationship between each acoustic index and the response variable, if any.
To test if a combination of different indices affects the response variable, we used a generalized linear mixedeffects model (GLMM) in R (R Core Team, 2021) using the lme4 package (Bates et al., 2015) with Poisson distribution (Figure S4).We assigned space (sampling sites) and time and day of recording as random effects in the model to account for spatial and temporal variation in the data.A combination of scaled non-collinear acoustic indices was fixed variables in the GLMM (Figure S1).To fit the GLMM, we used the MuMIn package (Barton, 2020).We set the threshold as 0.5 at a 0.05 significance value to determine collinearity among predictor variables (acoustic indices) and then dredged the models using the dredge function in the same package.We chose the model with the lowest Akaike Information Criterion (AIC) and highest R 2 values as our final model.We used the avian species richness as the response variable with all acoustic indices computed for all default and alternate frequency ranges as predictor variables.Additionally, while the sites are statistically comparable according to the overstory, to accommodate for small differences in the understory between the sites, we added a variable "Understory density" with three levels (high, low, and none) as a fixed effect in alternate models.
As an additional precautionary step, we used the Moran's I test estimates to ensure there was no spatial autocorrelation in our data (Figure S5).We also ensured there was minimal or no impact of the audio length selected, 10-s, on our results.For this, we redid all analyses using the original 1-min-long recordings and provided supporting evidence in Figures S6, S7 and Table S9.
Additionally, we used Spearman's rank correlation method to test the relationship between the indices and the avian species richness and found that three out of the seven indices showed a weak negative correlation with coefficients -0.18 (NP) and -0.36 (ADI).Four of the seven indices (ACI, AEI, BI, and NDSI) showed a weak positive correlation with coefficients ranging from 0.067 (NDSI) to 0.41 (BI) (Figure S2 and Table S6).

| Combination of indices using a generalized linear mixed-effects model
The combination of acoustic indices also had an overall weak effect on the response variable avian species richness.The top model that predicted the avian species richness (R 2 = 0.54) included the indices ACI, BI, H, and NDSI (refer to Figure S3 and Table S7 for more details).We found BI was the strongest predictor occurring in all top 10 models.ACI was in 8 out of the top 10 models predicting avian species richness.
The alternate models (refer to Table S8) with the understory density as a fixed effect showed similar results, suggesting no effect of the differences in the understory density on the performance of the indices.

| DISCUSSION
Technological advances, such as PAM, can help improve biodiversity conservation and management practices.However, inconsistent results from different biogeographic regions lead to a need for data-driven insights from different biomes, especially tropical forests (Buxton et al., 2018), before the indices are widely used at large spatio-temporal scales (Gasc et al., 2015;Maass et al., 2005;Mammides et al., 2017).
BI is a function of both sound level variation relative to the quietest frequency band and the number of frequency bands utilized (Boelman et al., 2007).We expected BI, which measures the inherent irregularity of avian vocalizations and allows computational flexibility for a desired frequency range, to have a strong relationship with the avian species richness.In line with our expectation, we found that BI showed a stronger correlation than all other indices (Figure 2) and contributed efficiently to the combination of indices in the top GLMM model predicting avian species richness (Table 2).We found that, individually, BI performed better than all other indices across all statistical models.BI was also the strongest predictor found in all the top 10 GLMMs.Thus, in comparison with other indices, BI showed relatively stronger evidence of predicting avian species richness in our sites and may be a viable option, although with limitations, to use for a rapid biodiversity assessment using avifauna as indicators of biodiversity and ecosystem health.

| Predicting avian species richness
Acoustic indices such as AEI, ADI, and H measure the evenness of the distribution of acoustic energy and/or activity in the soundscape.Higher species richness is expected to be associated with an abundance of sound signals across the frequency bands within a soundscape, and thus, a wider distribution of acoustic energy.Therefore, we expected the indices to correlate with avian species richness in our sites.However, contrary to our expectations, we found weak evidence of a linear correlation between these indices and the avian species richness (Figure 2).
On the other hand, lower species richness is expected to be associated with fewer acoustic signals, which leads to an increase in the complexity of soundscapes with sharp contrasts between louder signals against a quieter soundscape.Thus, we speculate that one may find higher ACI and NP values in regions where the soundscape consists of fewer instantaneous species vocalizations.With higher species richness, ACI and NP values might be lower as instantaneous acoustic signals lie closer in frequency and time, thereby decreasing the variation in the intensity of sounds (Bradfer-Lawrence et al., 2019).Even though this inverse relationship of ACI and NP with avian species richness is evident from the negative correlation between these two indices and the avian species richness (Figure 2), ACI and NP do not exhibit a strong potential for representing avian species richness in our study sites.Additionally, given that our sites being socioecological systems have a human presence in them, we included NDSI in this study with the expectation that it would efficiently predict avian species richness between 2 and 8 kHz while accounting for anthropogenic presence in the soundscapes in the 1-2 kHz.Fuller et al. (2015) also found the highest correlation between species richness and NDSI compared to other indices in a forest in Australia.However, we found no correlation between NDSI and species richness in our sites, which we believe could be because of the higher richness of avifauna species, such as the Spotted dove (Spilopelia chinensis) and Coppersmith barbet (Psilopogon haemacephalus), vocalizing in the lower frequency range in our sites.

| A combination of indices
Every index mathematically summarizes different components of the acoustic soundscape (Gasc et al., 2015).Previous studies demonstrate the higher efficacy of an additive combination of indices for efficiently predicting avian species richness (Buxton et al., 2018;Towsey, Parsons, et al., 2014;Towsey, Wimmer, et al., 2014).For instance, Eldridge et al. (2018) found that compound indices predicted avian species richness better in a temperate forest in the United Kingdom and a tropical humid forest in Ecuador.Contrarily, our results show that a combination of indices only weakly predicted avian species richness in a tropical dry forest (Table 2).Buxton et al. (2018) found that a combination of indices effectively predicted species richness in a North American forest (R 2 = 0.86).Similarly, Towsey, Parsons, et al. (2014) and Towsey, Wimmer, et al. (2014) also reported a higher predictive ability of a combination of indices in an Australian forest (R 2 = 0.81).Compared to these studies, a combination of indices does not effectively represent species richness in our sites as well.We find the hypothesis presented by Buxton et al. (2018), who found that acoustic models capture avian species richness more accurately in regions with either low or high species diversity per unit area, may apply to our findings.We speculate that in regions such as tropical dry forests, with moderate diversity per unit area in comparison to species-rich tropical humid forests and species-poor temperate forests (Morales et al., 2015), the acoustic indices may prove to be less reliable.

| Factors affecting indices' performance
In different biomes, sound propagates differently.For example, in a dense tropical humid forest, there is a higher chance of sound attenuation (Bradbury, 2011).A higher density of vegetation causes a scattering effect, leading to higher attenuation of sound signals of biotic and abiotic origins.Therefore, in a vegetatively less dense biome such as tropical dry forests, sounds may travel farther, thereby increasing the number of acoustically detected species in the soundscape.At our sites, we were able to acoustically detect at most 10 avian species per 10-s audio segment.Whereas, although tropical humid forests may have a higher density and diversity of species per unit area, the number of species acoustically detected in the soundscapes may be lower.For example, Eldridge et al. (2018) acoustically detected $12 species per minute in a North Western Ecuadorian forest.We thus speculate that the soundscapes recorded using passive acoustic recorders in tropical dry forests may display a more complex, noisier soundscape with more instantaneous, overlapping sound signals compared to a tropical humid forest, thus hindering the performance of the indices in this particular biome.While in addition to the density of vegetation and climatic variations, the recorder distance may have an impact on the detection of sounds in a biome, our sites displayed homogeneity in terms of vegetation and species richness (Choksi et al., 2023).We thus speculate that the differences seen across biomes may largely be driven by the differences in the vocalizing communities that constitute the soundscape within biomes.Apart from across-biome differences, we also see differences in the performance of indices within similar biomes, which may occur due to variation in the degrees of anthropogenic use within the forests.For example, in the case of the tropical dry forest biome, a study carried out in a protected area in Costa Rica (Izaguirre & Ramírez-Al an, 2018), found that ACI reflects temporal changes in avian richness in the region.In contrast, ACI did not correlate with our response variable at our sites, which are unprotected and used by local communities for subsistence.We, therefore, speculate that the inconsistent performance of acoustic indices may also possibly be driven by, to an extent, the degree of anthropogenic presence.
The indices combined measure all sounds present in the soundscape across all frequency bands (unless specified).We speculate that in sites with human presence such as ours, the acoustic signals from anthrophonic sources, for example, loudspeakers, may also influence the acoustic signals from low-and mid-frequency biophonic sources, including avifauna acoustic signals.An overlap in frequency ranges of anthrophony and biophony (e.g., birds calling in the 1-2 kHz range) makes it difficult for the indices, especially NDSI, to differentiate between the two.Our results suggest that acoustic indices could predict avian species richness in more complex tropical landscapes dominated by humans when combined.Potentially, indices developed in the future may also take the presence of sounds of anthropogenic origin into consideration.

| CONCLUSION
With the rapid human modification of ecosystems, there is a pressing need for rapid assessment tools to monitor biodiversity.This study provides crucial empirical evidence on the application of acoustic indices, which are emerging as popular rapid biodiversity measurements, in a largely underrepresented and undervalued biome-the tropical dry forests-which is threatened globally.We focused on seven acoustic indices that summarize different features of the soundscape and tested their efficacy in predicting an acoustically derived biodiversity metric-avian species richness.We found that the indices are limited in their capacity to quantify biodiversity in a Central Indian tropical dry forest.However, we recommend that BI may be a relevant index and could potentially be used for rapid avifauna-based acoustic monitoring for biodiversity assessment in this biome.Given the inconsistencies in the performance of indices, we recommend adding to the evidence that acoustic indices are not homogenous in their representation of biodiversity across different biogeographic locations, including protected and non-protected areas.We also recommend testing the indices before using them as rapid biodiversity monitoring tools.
classify acoustic indices as α-indices (within-group indices) and β-indices (betweengroup indices), where α-indices are used for examining the acoustic richness of a soundscape and β-indices are used for measuring the differences between two or more soundscapes.For the scope of this study, we refer to Bradfer-Lawrence et al. (2019) for further classification of the α-indices based on the soundscape features used for computation of index values-soundscape evenness, soundscape complexity in frequency and/or time, and soundscape components (biophony, anthrophony, and geophony) (refer to Table

F
I G U R E 1 (a) Map showing the recorder locations (red circles) within each of the 15 sampling sites in the buffer area of Kanha National Park (KNP) in Bicchiya subdistrict, Mandla district, India.Each sampling location within sampling sites is approximately 400 meters from any other sampling location to avoid any overlap in acoustic data collection.(b) Photo from sampling location MH_2, (c) Photo from sampling location PT_3, and (d) Photo from sampling location SB_2.T A B L E 1 Acoustic indices with their technical descriptions, ecological applications, parameters, references, and functions for implementation in R. All indices are calculated on a 10-s clip basis.

F
I G U R E 2 Scatterplots showing the relationship between avian species richness (all variables computed on 10-s segments) and (a) acoustic complexity index (ACI), (b) bioacoustic index (BI), (c) acoustic diversity index (ADI), (d) acoustic entropy index (AEI), (e) total entropy (H), (f) normalized difference soundscape index (NDSI), and (g) number of peaks (NP), tested using the Pearson correlation method.Lines represent the linear regression slope and shaded areas indicate the 95% confidence intervals.The elements in red represent the data from the Year 2020 while the elements in blue represent data from the Year 2021.Elements in black represent the mean response of the data irrespective of the year.
Top 10 models that predict species richness.
T A B L E 2