Four species were chosen based on their phylogenies and call characteristics: Diopsittaca nobilis (non-Passeriformes) and Sicalis flaveola (oscines-Passeriformes), whose vocalizations show greater influences of learning components, and Crypturellus soui (non-Passeriformes) and Thamnophilus doliatus (Suboscines-Passeriformes), whose vocalizations show strong genetic components (sensuMarler, 2004). These species have wide distributions with favorable spatial structures for this type of investigation (several suitable habitat patches surrounded by unsuitable environments), and sound recordings are available for them.

C. soui belongs to the family Tinamidae. It inhabits both humid and dry environments, such as the edges of dense forests, dry forests, sandbanks, etc. It occurs from Mexico to Bolivia and Brazil (including the states of Espirito Santo, Rio de Janeiro, Minas Gerais, and Mato Grosso). These birds are adapted to terrestrial life, and their flights tend to be short and heavy. Their song is composed of a mild chirp with a tremulous tone with rising long sequences and descending whistles (Sick, 1997; Fig. 2). The learning component of the Tinamidae call is virtually nonexistent (Bertelli and Tubaro, 2002).

D. nobilis belongs to the family Psittacidae. It has a wide distribution that includes Venezuela, Suriname, Guyana, and Brazil (Mato Grosso State) to Bolivia, inhabiting savanna formations and forest borders (Collar, 1997; Sick, 1997). Most parrots have large vocal repertoires (de Araújo et al., 2011; de Moura et al., 2011), and we selected the flight call of this species (sensu Moura et al., 2011) for the study because it encodes species-specific information (de Araújo, 2011; de Araújo et al., 2011). The flight call of D. nobilis species is composed of two notes with similar structures, with an upward modulation followed by a downward modulation (Fig. 1). Parrots are known to be open-ended learners that can alter their vocalizations throughout their lives (Salinas-Melgoza and Wright, 2012).

Fig. 1.

Sonograms of the target species indicating the vocal measurements used in the analyses. FI, initial fundamental frequency; FF, final fundamental frequency; FP, fundamental frequency of the first peak; FUP, fundamental frequency of the last peak; Fmax, maximum fundamental frequency; Fmin, minimum fundamental frequency; NM, number of modulations; DN1, duration of first part; DN2, duration of the second part; DN3, duration of third part; I.N., inter-note interval; DNM, duration of the unmodulated part; D.M., duration of the modulated part.

(0.65MB).

S. flaveola is songbird belonging to the family Thraupidae. It is commonly found in open and semi-open areas in South America, including northern Colombia, Venezuela, Trinidad and Tobago, western Peru, Ecuador, eastern Brazil, Bolivia, Paraguay, Uruguay, and northern Argentina. The song of this species is composed of quick modulated notes that form phrases. As a songbird, the species is expected to possess a strong learning component in its acoustical communication (Sick, 1997).

T. doliatus belongs to the family Thamonophilidae. It inhabits both forests and open areas from Mexico throughout Central America, and a large part of South America to northern Argentina, Bolivia, and Paraguay (Sick, 1997). The species’ song is composed of phrases in which the notes are quickly modulated and uttered at a rising rate. The vocalizations of Thamonophilidae are considered to be innate and to reflect their genetic structure (Isler et al., 2005).

Ecological Niche Models (ENM) were used here to classify the relative suitability of habitats based on their abiotic factors across the total geographic range of those birds. Ecological Niche Modeling uses localities with known occurrences of a given species and the environmental characteristics of those areas to determine its environmental preferences, that is, the intervals of environmental variables in which species is most likely to occur. The maps generated by Ecological Niche Modeling (ENM) are usually composed of patches of apparent high environmental suitability separated by areas of less suitable environments. These maps represent hypothetical distributions of those species, highlighting isolated and disjunct populations that could diverge genetically, morphologically, and/or acoustically. Hence, if ENM can explain a species’ distribution, and if we can determine to which suitable patch a bird belongs, we could identify (with reasonable accuracy) the best areas to release a recovered specimen.

We used Maxent 3.3.3k software, which has performed well in diverse sets of modeling scenarios and has been used in a great number of ecological, biogeographical, and conservation studies (Elith and Leathwick, 2009). The Maxent algorithm uses the maximum entropy principle to fit a probability distribution for species occurrence to a set of pixels across the region of interest. For a detailed explanation on how the maximum entropy principle applies to ENM, please referred to Elith et al. (2011). The reference localities were obtained from records retrieved from the portals ORNIS (http://ornisnet.org and ornis2.ornisnet.org) and Species-link (http://splink.cria.org.br). During our search, we also used the term Ara nobilis, to account for the previous taxonomic designation of D. nobilis. A total of 90 independent occurrence points were obtained for D. nobilis, 95 for C. soui, 96 for S. flaveola, and 117 for T. doliatus.

We used those points to produce the Ecological Niche Models (ENM) employing MaxEnt software; the models were evaluated based on the area under a receiver operating characteristic (ROC) curve (AUC; Costa and Schlupp, 2010; Lozier et al., 2009; Marini et al., 2010; Peterson et al., 2007). We used environmental layers from Project Worldclim (Hijmans et al., 2005) at a spatial resolution of 30″ (1km2). The layers extended between longitudes 78°31′W to 35°W, and between latitudes 12°47′N to 34°46′S, corresponding to most of South America. To ensure data independence, we removed self-correlated environmental layers from the analyses based on correlations greater than 0.9 (a similar procedure can be found in de Araújo et al., 2014). We repeated the process until no correlations higher than 0.9 were found, leaving the following environmental layers: Bio 3 (isothermy), Bio 4 (temperature seasonality), Bio 7 (annual temperature range), Bio 10 (mean temperature of the warmest quarter), Bio 11 (mean temperature of the coldest quarter), Bio 14 (precipitation in the driest month), Bio 15 (precipitation seasonality), Bio 16 (precipitation in the wettest quarter), Bio 17 (precipitation in the driest quarter), and altitude. Once the models were built, the maps were analyzed and classified according to the presence of isolated patches of high habitat suitability (which were then numbered).

Call recordings were obtained from reference collections at Fonoteca Neotropical Jacques Vielliard (Unicamp), Macaulay Library (Cornell), Wikiaves, and Xenocanto. After examining the available recordings, we selected 540 files based on their overall quality (signal to noise ratio). These recordings were filtered (high pass filter at 160Hz) and the amplitudes were normalized to 0dB before the analyses. This procedure standardizes the recordings and reduces problems arising from comparing parameters between recordings (Zollinger et al., 2012). Editing and sound measurements were performed using Cool Edit Pro 2.0 software.

Selecting specific acoustical parameters to be measured is a complex process and depends upon the physical structure of each signal. Also, before we could choose acoustical parameters, we had to select a specific component from the vocal repertoire of the species. Only a portion of the vocalizations within a repertoire encodes the species-specific information, and among the repertoire components we have chosen the songs, for it contains the species information that could guarantee call homology, while it is commonly heard in the field (Vielliard, 1987). As for the macaw species, we decided to use contact calls here defined as the call uttered while in flight (de Moura et al., 2011), for it encodes species information and it is also very common (de Araújo et al., 2011). In order to choose the acoustical parameters, we examined a small sample of the dataset to ensure that the parameters would be measurable throughout the sample. Even though some bird species have noisy and pulsed calls (e.g., Marques et al., 2010), bird calls are typically composed of a melodic structure in which call frequencies follow a harmonic series and are modulated over time. There are at least six parameters that can be measured in virtually every melodic call, and we used all of them to create the measurement map. The initial (FI), final (FF), maximum (FMax), and minimal (FMin) fundamental frequencies and call durations (D) were measured directly from a spectrogram, while the dominant (FD) frequency was measured from a power spectrum (Fig. 1). Additionally, we selected landmarks on the spectrogram to refine our measurements and improve the data. The flight call of D. nobilis, for example, has two notes, and each note can be divided in three portions, an upward frequency modulation portion (D1), a quick modulated portion (D2), and a final downward modulated portion (D3; Fig. 1, upper right sonogram). By using those landmarks, we were able to measure the frequencies at the end of the first (FP1) and second portions (FP2; Fig. 1, upper left sonogram). Additionally we counted the number of modulations found within the middle portion and, as there are two notes, likewise measured the Inter Note Interval (IN) of the flight call.

We used a similar approach to create a measurement map for C. soui. We were able to divide its call into a modulated and a non-modulated portion, and measured their durations (Dm and Dnm respectively; Fig. 1, upper right sonogram). Additionally we measured the initial (FI), final (FF), maximum (Fmax), and minimal (Fmin) fundamental frequencies and dominant (FD) frequency from the power spectrum (Fig. 1, upper left sonogram).

The remaining species (S. flaveola and T. doliatus) showed higher levels of organization, as their calls are composed of phrases (groups of notes). Although some studies have opted to measure each note and describe the calls based on their note compositions (e.g., Silva et al., 2000), we chose to restrict our measurements to the macro-structures of the phrases for S. flaveola. We therefore measured the total durations of the phrases, the initial (FI), final (FF), maximum (FMax), and minimal (FMin) fundamental frequencies of each phrase, the dominant (FD) frequency, as well as the mean note rate (calculated from the mean period [1/note period] measured) as shown in Fig. 1. As the last note of the song of T. doliatus presents a distinct structure, we also measured last note's temporal and espectral parameters.

We examined the existence of collinearity among call parameters before performing the analyses, by calculating the variance inflation factor (VIF). Collinearity could make it more difficult to detect any effects, and one strategy consists in sequentially calculating the Variance Inflation Factor (VIF) and dropping the covariate with the highest VIF, recalculating the VIFs and repeating the process until all VIFs are smaller than a pre-selected threshold (Zuur et al., 2010). Here, we removed all of the vocal parameters that presented values higher than 3 (Zuur et al., 2010), with only the remaining variables being used in the analysis. The full list of the parameters kept for each species is presented in Table 1 (the full model).

We used a multinomial logistic regression to classify the calls of different ENM patches based on vocal parameters. We used the year of the recording as a covariate to control for possible changes in call structures over time. We selected candidate models through a stepwise approach in which we compared a full model (all variables included) against partial models built by suppressing a single variable for each partial model. We used second-order Akaike's information criteria (AICc) to select the model with the lowest AICc, and permanently removed the suppressed variable from the model. The selected model was then used as base model in the next round, which further suppressed variables one at a time. We repeated this process until the best model was the one built without the removal of any additional variable. To select the final models, we compared the AICc value of the candidate models (the full model and base models of each round) to a null model. The model with the minimum AICc value is the one which minimizes information loss (Burnham and Anderson, 2002), and thus retains the highest empirical evidence. The final model was used to predict the original patch of the calls with the use of a multinomial regression. In order to examine the model classification efficiency we built a confusion matrix that presents the percentage of correct classification of the calls among patches, and compared these values to that expected by chance through a chi-squared test. Hence, despite we used AUCc metric to select the best model, we chose to test the model's classification efficiency through a chi-squared.

As low differences in AICc values (ΔAICc) indicate support for multiple models (Burnham et al., 2011), we kept the models with a ΔAICc of 4 to test if they could improve call classifications. We also estimated the strength of the evidence for each model by calculating Akaike weights (Wi), which measure the probability that the ith model is the best model; we also estimated the evidence ratios (ER) that measure the empirical evidence for the best model relative to the target model (Burnham et al., 2011. The multinomial models were built using R software (R Core Team, 2015) with the nnet package (Ripley and Venables, 2011), while the models were selected by using the AICcmodavg package (Mazerolle, 2011).

The Ecological Niche Models (ENM) demonstrated satisfactory performances, with the ENM of C. soui having an AUC value of 0.936, D. nobilis an AUC value of 0.907, S. flaveola an AUC of 0.926, and T. doliatus an AUC value of 0.877. The annual temperature range had a 38.1% relative contribution to the C. soui model, and the habitat suitability of C. soui was highest in areas with an annual temperature range lower than 150°F (Fig. 3, upper right). Precipitation during the wettest quarter was the variable that relatively most contributed to the D. nobilis model (29.2%), and habitat suitability was highest in areas with precipitation in the wettest quarter ranging from 500 to 1000mm (Fig. 3 upper right). Isothermality was the variable that relatively most contributed to the model of S. flaveola (21.3%), and habitat suitability was highest around 60 (Fig. 3 lower left). Temperature seasonality was the variable that relatively most contributed to the model of T. doliatus, and low habitat suitability was found for Temperature seasonality higher than 3000 (standard temperature deviation*100). Despite the differences in the response curves, all of the models showed high habitat suitability regions as well as low suitability regions (Fig. 2).

Fig. 2.

Habitat suitability of the target species in terms of environmental variables, showing: a fast drop in habitat suitability for temperature range higher than 150 in Crypturellus soui (upper left graph); the optimum values of the precipitation in the wettest quarter between 500 and 1000mm for Diopsittaca nobilis (upper right graph); a reduced suitability for values of isothermally higher than 60 for Sicalis flaveola; and reduced habitat suitability of values of temperature seasonality for Thamnophilus doliatus. Red region indicate the most suitable habitats while the blue region would refer to unsuitable areas, while green and yellow indicate intermediate habitat suitability.

(0.33MB).

The maps in Fig. 3 show patches of high suitability (warm colors) separated by areas of low suitability (blue) for all of the target species. Based on these maps, we classified the potential distributions into five distinct areas for C. soui: Area 1 (extreme northern South America), Area 2 (northwestern South America), Area 3 (mid-western Brazil), Area 4 (the Brazilian coast), and Area 5 (Trinidad and Tobago) (Fig. 4, right); six suitable isolated patches were identified for D. nobilis: Area 1 (Venezuela), Area 2 (Roraima State, Guyana, and Suriname), Area 3 (Amazonas, Pará, and Amapá states), Area 4 (central Brazil), Area 5 (inland northeastern Brazil), and area 6 (coastal northeastern Brazil) (Fig. 4, left); five suitable isolated patches were identified for S. flaveola: Area 1a (southeastern Brazilian coast), Area 1b (central-eastern Brazilian coast), Area 1c (northeastern Brazilian coast), Area 2 (comprising Venezuela and Colombia), Area 3 (Bolivia); and eight suitable isolated patches were identified for T. doliatus: Area 1 (northern Colombia and Venezuela); Area 2 (northeastern Venezuela), Area 3a (southern Colombia and Equator), Area 3b (Bolivia and southern Peru), Area 3c (southeastern Brazil), Area 4 (northern Brazil and Guiana), Area 5 (central Brazil), Area 6 (northeastern Brazil).

Fig. 3.

Ecological Niche Models of Crypturellus soui (upper left), Diopsittaca nobilis (upper right), Sicalis flaveola (Lower left and Thamnophilus doliatus (lower left). Warm colors indicate areas with higher habitat suitability and thus higher probabilities of species occurrence, whereas blue represents areas with low habitat suitability and thus low probability of species occurrence. The numbers indicate the selected isolated patches.

(0.66MB).

Fig. 4.

Biplot of the first two components of the Principal Component Analysis (PCA) made with the call parameters of Crypturellus soui (upper left), Diopsittaca nobilis (upper right), Sicalis flaveola (Lower left) and Thamnophilus doliatus (lower left). The colors represent the selected geographical patches.

(0.49MB).

We selected and analyzed a total of 540 sound files, including: 139 sound recordings of C. soui, representing all five areas on the map (Fig. 3); a total of 85 sound recordings of D. nobilis, from five of the six areas identified (Fig. 4; areas 1, 2, 3, 4 and 6); a total of 136 sound recordings of S. flaveola, representing all five areas on the map (Fig. 4); and 180 sound recordings of T. doliatus, representing five of the six areas on the map. As the recordings of D. nobilis were biased toward patch 4 (representing 70% of the recordings) we randomly selected 20 (from 59) calls from patch 4 to reduce differences in sampling effort, reducing the data from the original 85 calls to 46.

Principal Component Analysis showed identifiable groupings within C. soui, with vocalizations from the same area (especially from area 4) grouping together (Fig. 4, upper left). A similar pattern was found for T. doliatus, in which areas 1 and 2 seemed to present some overlap within the two first PCs, but were somewhat distinct from the remaining areas (Fig. 4, lower left). The results for D. nobilis and S. flaveola showed less distinctive patterns, with much overlap between the areas (Fig. 4 upper right and lower right, respectively).

A single model predicted areas with substantial empirical evidence (ΔAICc<4) for C. soui, while the second best model showed low probability (Wi of 0.01; Table 1), the evidence ratio (ER) was 148 times stronger for the best model relative to the second. For D. nobilis, S. flaveola, and T. doliatus, however, we found support for multiple models (ΔAICc<4). Although the Wi values for the best models reached 0.74, 0.53 and 0.73 respectively (Table 1), the evidence ratios found for the best models were 3–7 times stronger when compared to the second model. Even so, the inclusion of additional predictors into the best models did not substantially alter their classification matrices, so we used only the best models in our analysis.

The classification analyses showed differences in efficiency among the studied species. While call parameters allowed classifications that were better than expected by chance in T. doliatus for most of the areas, only a few patches had their calls classified as better than expected by chance for C. soui (patch 4), D. nobilis (patches 3 and 6), and S. flaveola (patch 1a), indicating that the calls from some patches are structurally singular when compared to the others (Table 2, Fig. 3). A closer look at the confusion table shows that C. soui calls from area 3 seem to be closer to those of area 2. This result can be observed in the PCA, but is also supported by the fact that the multinomial analysis classified all seven calls from area 3 as belonging to area 2. A similar pattern emerges from the calls of T. doliatus, in which 75% of the misclassified calls from patch 3b is classified by the model as being from patch 3c.