Generating a database of song bouts

We downloaded recordings of Spotted and Eastern towhees from the Macaulay Library (Cornell Lab of Ornithology 2009) by requesting access to the relevant song files, as well as from Xeno-canto [33] using the WarbleR package in R to query the database [34]. We filtered the recordings to eliminate those without song by specifying “Sounds/Type = Song”. We also requested recordings marked as “Spotted × Eastern Towhee (hybrid)” and “Spotted/Eastern Towhee (Rufous-sided Towhee)” from the Macaulay Library to analyze separately as a “hybrid/unsure” category—of the 27 hybrid/unsure recordings we obtained, 24 recordings had photos and/or a description of the morphology of the individual (Table A in S1 Text). For further analysis, we constructed a spreadsheet of metadata for each recording, which included species, date, location, latitude, longitude, and recordist. To minimize the chance of including repeated song recordings of the same bird, we discarded any duplicate entries with the same recordist, date, and location. For recordings that named a location but not geographic coordinates, we estimated the latitude and longitude by identifying the location on a map. We opened each recording file in Audacity version 3.1.3 (https://www.audacityteam.org/) to manually extract one song bout from each recording. Files from Xeno-canto are generally stored in mp3 format and files from Macaulay Library in wav format; we resampled recordings at 44,100 Hz if they were recorded at a different sampling rate and exported each bout as a.wav file to standardize the recordings for analysis, and we noted whether the sampling rate and/or file format had been converted. During resampling, Audacity implements a low-pass filter at half of the sampling rate to prevent aliasing. We considered the breeding season to be between April and August based on the “Breeding” section of the Birds of the World database [35,36], and we conducted analyses both on recordings from any month (see Results) and on recordings from only the breeding season months (see S1 Text).

To assign the zone of song overlap between the ranges of the two species, for each degree longitude we plotted the proportions of Eastern and Spotted towhees relative to the total number of both species combined, which allowed us to identify where the two species co-occur most frequently and how their relative abundances shift geographically. In other words, at each degree longitude we calculated the number of recordings of Spotted towhees divided by the total number of breeding-season recordings of both species combined, and we repeated this calculation at each degree longitude for the Eastern towhees. We then plotted these proportions on a graph where the x-axis represents longitude and the y-axis represents the proportion of each species (Fig A in S1 Text). At the most eastern and western longitudes, the recordings were all from Eastern and Spotted towhees, respectively. Based on our findings, we chose to identify the area between 102°W and 91°W as the zone of song overlap, which contained the majority of the area where recordings of both species seemed to co-occur (Fig A in S1 Text). In addition, we downloaded from eBird all sightings of putative hybrids (eBird query: “Spotted × Eastern Towhee (hybrid) - Pipilo maculatus × erythrophthalmus”) during the breeding season and plotted their locations; we observed that reported hybridization events primarily occurred in this region (Fig B in S1 Text). We also determined an approximate frequency of hybridization in the zone of overlap by finding the area with the highest density of hybrid sightings between 2013 and 2023 (Nebraska) and dividing the estimated number of unique sightings of hybrid towhees in Nebraska by the number of unique Spotted towhee or Eastern towhee sightings in Nebraska during the breeding season, filtering out those from the same date, latitude, and longitude.

Song analysis (syllable features and syllable segmentation)

To begin analyzing the features of the syllables within these bouts, we used the Chipper software [37], which was designed to facilitate syllable segmentation and analysis of field recordings with different levels of background noise. This software performs automated syllable segmentation by attempting to parse each song into syllables, defined as periods of sound separated by periods of silence, and then allowing the user to adjust the signal-to-noise ratio, enable high-pass and low-pass filters, and modify the segmentation when syllables have been incorrectly parsed, such as when background noise occurs during the inter-syllable silences (Fig 2). To prevent lower-amplitude syllables from being segmented incorrectly, we normalized the amplitude across each song. Additionally, we adjusted the signal-to-noise threshold, minimum syllable duration, and minimum silence duration to most accurately define the syllables in the song. Since our recordings are from nature and have wide variation in their background noise profiles, we view each spectrogram to assess what is background noise versus the bird singing and manually adjust these parameters so that the syllables are parsed in a way that appears to be most accurately defined with minimal noise; this procedure has been shown to have high within-user repeatability and high between-user reproducibility [37]. If any syllables appear to be parsed incorrectly after this procedure, Chipper allows the user to manually modify the syllable segmentation. If, when visualized in Chipper, we determined that the song overlapped with other birds singing or had excessive background noise, we removed the recording from the analysis. We provide our code for statistical analysis and visualization of the song feature data and metadata about each song recording at https://github.com/CreanzaLab/TowheeAnalysis.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Spectrogram illustrating the definition of specific song elements.

On this spectrogram of an Eastern towhee song, we indicate a syllable, silence, bout, maximum syllable frequency, and minimum syllable frequency. The color of the syllable indicates its unique syllable type. Spectrogram generated from Macaulay Library recording 52506381, https://macaulaylibrary.org/asset/52506381.

https://doi.org/10.1371/journal.pcbi.1013135.g002

After segmentation, we examined a subset of 20 song bouts per species in Chipper to estimate a noise threshold, or the minimum number of matrix elements in the spectrogram that a note must contain in order to not be discarded as noise (Spotted towhee = 87, Eastern towhee = 70, hybrid/unsure = 65), and syllable similarity threshold, or the percent syllable overlap that determines whether two syllables are considered the same or different (Spotted towhee = 26.6, Eastern towhee = 25.5, hybrid/unsure = 25.0), using Chipper’s “Noise Threshold” and “Syllable Similarity Threshold” widgets, respectively. We calculated these values for both species to allow Chipper to make more accurate measurements for syllable features and song syntax specific to our set of Spotted and Eastern towhee songs. We then ran the song analysis function in Chipper, which uses the spectrogram and the syllable segmentation data to provide automated measurements of numerous song features, which were used for statistical analysis. We used 16 song-feature outputs from Chipper: bout duration (ms), number of syllables, rate of syllable production (calculated as the number of syllables divided by bout duration (1/ms)), smallest syllable duration (ms), largest syllable duration (ms), average syllable duration (ms), number of unique syllables, degree of syllable repetition (calculated as the number of syllables per number of unique syllables), average syllable upper frequency (Hz), average syllable lower frequency (Hz), maximum syllable frequency (Hz), minimum syllable frequency (Hz), overall syllable frequency range (Hz) (calculated as the overall maximum syllable frequency minus the overall minimum syllable frequency), largest syllable frequency range (Hz), smallest syllable frequency range (Hz), and average syllable frequency range (Hz) (calculated as the average of the maximum minus the minimum frequency of each syllable); see [37] and the Chipper manual for more detail on song feature calculations. We then analyzed these song features across 1067 Spotted towhee song bouts, 1718 Eastern towhee song bouts, and 27 “hybrid/unsure” song bouts, using various statistical methods to assess geographic patterns and variation in songs and machine learning approaches to investigate the degree to which songs can be accurately classified using song feature data. We use multiple different machine learning methods to assess whether there is convergence in the results of the models or whether there are potential idiosyncrasies tied to each unique method.

Data Preparation

To investigate the geographic variation in birdsong of the Spotted and Eastern towhees, we first determined that the song features were not normally distributed, so we calculated the natural log of each feature for analysis. Our analysis thus contained a total of 1067 Spotted towhee samples and 1718 Eastern towhee samples and data for 16 song features. To visualize the data, we plotted each log-transformed feature against the longitude and latitude coordinates of its recording (Figs C–E in S1 Text). All the base maps used in the figures were created using Natural Earth data (http://www.naturalearthdata.com/about/terms-of-use/) via the rnaturalearth package in R.

Generalized Linear Models

To assess the relationship between the song feature data and potential interacting variables, we used generalized linear models (GLM). First, we generated a dataframe that includes the song feature data, geographic coordinates, and species classification for all the samples. We also include the overall conversion status of the recording files for each sample, which was denoted as ‘unconverted’ if the file was not converted before analysis and ‘converted’ if the original recording file had to be converted to a.wav file (generally from.mp3 or.m4a) or if the sampling rate had to be converted to 44,100 Hz. We then use the ‘glm’ function (R package: stats) to fit GLMs to our raw song feature data, using a log link function and including scaled longitude, scaled latitude, species, and the overall file conversion status as fixed effects.

Linear Discriminant Analysis of song feature data

Using the raw song feature data, we conducted a Linear Discriminant Analysis (LDA) that uses the 16 song features to maximize the separability of the two species and minimize the variation within each species class using the ‘lda’ function from the MASS package [38] in R. We obtained a random subsample (75% for training and 25% for testing) of the entire data set and then downsampled the training set to obtain a balanced subsample of Eastern towhees and Spotted towhee song bouts (N_{Spotted_towhee} = 796 and N_{Eastern_towhee} = 796). Thus, the model was trained on 1592 total song bouts and tested on a subset of 697 song bouts that were not in the training set. We calculated the overall model accuracy , balanced accuracy , and Cohen’s kappa (κ) using the ‘confusionMatrix’ function from the caret package [39] in R for this model and all other classifier algorithms in this manuscript. We also used a permutation test to assess whether each machine learning model performed better than random chance. To do this, we randomly shuffled the species labels and compared the permuted accuracy of 1000 permutations to the actual accuracy of the model.

For these analyses and all other classifiers described below, we also assessed the effect of file conversion status on the results (see Supplemental Methods in S1 Text).

Principal components and Procrustes analyses of song feature data

We used a Principal Component Analysis (PCA) to reduce the dimensionality of the dataset and visualize the variation in the song data, and we measured the proportion of variance explained by each principal component (PC). Using the ‘prcomp’ function in the stats package in R, we centered and scaled the log-transformed data and performed the PCA. We assessed whether the two species could be distinguished in PC space by testing a simple linear classifier using the ‘lda’ function in the MASS package in R. This type of Discriminant Function Analysis finds a linear function that best discriminates the ‘Species’ classification of the song recordings based on PC1 and PC2. In other words, this classifier finds a line through the two-dimensional PC-space such that songs on one side of the line are more likely to be from Spotted towhees and on the other side of the line are more likely to be from Eastern towhees, indicating the separability of the species after PCA [40]. Further, to determine whether the PCA showed geographic structuring, we followed up with a Procrustes analysis using PC1 and PC2 from the PCA of our song data. For this analysis, the matrix of PC scores was rotated and transformed onto the target matrix containing the longitude-latitude coordinates of the samples, such that the sum of squared distances between the corresponding points of the transformed and target matrix was minimized and plotted on a map using the ‘procrustes’ function from the vegan package [41] in R.

UMAP visualization

We used a Uniform Manifold Approximation and Projection (UMAP) as another method of reducing dimensionality and visualizing variation in our song-feature data. We rescaled the log-transformed data using the same technique as in the PCA, and we used the ‘umap’ function in the umap package [42] in R to generate a two-dimensional projection of the data. Since most of the points occupy a single cluster in the resulting UMAP projection, we then tested a simple linear classifier as above to determine how well the songs could be separated by species in the UMAP projection. We included the sample of 27 “hybrid/unsure” towhees in the UMAP analysis, but we excluded these points when we tested the species classifier.

Random forest model

We used a random forest model (RFM) to assess how accurately Spotted and Eastern towhees can be distinguished from one another by their song features. First, we log-transformed the song feature data. Next, we separated the recordings into a training set and a test set in multiple ways, as explained below, and trained the random forest classifier on the 16 song features from the training set containing species identifiers, using the ranger package [43] in R. For each model, the algorithm then averaged multiple decision trees to make predictions of the species identity in the test set. We used our random forest algorithm to train two models: 1) a model trained on a geographically unbiased subset of the entire data set and 2) a model trained on a subset of song bouts obtained from the non-overlap zone. For the former model, we obtained a random subsample (75% for training and 25% for testing) of the entire data set, and we then downsampled the training set to obtain a balanced subsample of Eastern towhees equal to the number of Spotted towhee song bouts (N_{Spotted_towhee} = 796 and N_{Eastern_towhee} = 796). We tested the model on 697 song bouts, none of which were included in the training set. For the latter model, we separated the song bouts recorded in the zone of overlap from those in the non-overlap regions, and trained only on a subset of the non-overlap song bouts, retaining a sample size equal to that of the zone of overlap for testing. Once again, our training set was downsampled, such that we obtained a balanced subsample of Eastern towhees equal to that of Spotted towhees in our training set (N_{Spotted_towhee} = 796 and N_{Eastern_towhee} = 796). We tested this model on a sample of towhee songs from the non-overlap zones (N = 216), an equally sized sample of towhee songs from the zone of overlap (N = 216), and the sample of “hybrid/unsure” towhee songs (N = 27). For both models, we created a 10-fold cross-validation control and tuned the hyperparameters using a 10 x 16 grid using the caret package [39] in R, which allows us to evaluate every possible combination of the number of predictors to randomly sample at each split and the minimum number of samples needed to keep splitting nodes using Gini impurity to split nodes. We also repeated the random forest model using 100 different random seeds to assign recordings to the training and test sets, and we used these trials to estimate the overall accuracy of the classifier as well as a 95% confidence interval.

Gradient boosting machine

We used a gradient boosting machine as another machine learning technique for classification of the song bouts. This technique is similar to random forest in that a large number of decision trees are trained on a subset of data and then tested on a withheld set of test data, but whereas these trees are independently generated with RFMs, a gradient boosting machine sequentially produces trees that are informed by the errors in previous trees. This technique can produce more accurate models, but has an increased risk of overfitting. As with the previous model, we obtained a random balanced subsample of 1592 song bouts (N_{Spotted_towhee} = 796 and N_{Eastern_towhee} = 796) to train the model and tested the model on 697 song bouts. The song feature data was centered and scaled. We used the ‘xgb.train’ function from the xgboost package [44] in R to train the model on the 16 song features using a multi-class softmax objective with early stopping (early_stopping_rounds = 10) to prevent overfitting. We set our model parameters to a maximum of 100 boosting iterations and used a multiclass log loss evaluation metric to calculate the classification error, such that the model attempts to improve each subsequent iteration based on how well the previous iteration is able to classify the observations.

Deep learning model

Additionally, we used a deep learning model via the Torch package [45] in R that uses a multilayer perceptron (i.e., fully connected feedforward neural network) for classification of the song bouts. Again, we used a training set consisting of a random balanced subsample of 1592 song bouts (N_{Spotted_towhee} = 796 and N_{Eastern_towhee} = 796) and tested the model on 697 bouts. The data was centered and scaled prior to integration with the model. The network contains an input layer, two hidden layers, and an output layer. The input layer has a size of 16, defined by the number of song features. We used 64 neurons in the hidden layers, aiming to balance model performance and computational efficiency without overfitting. Each fully connected layer applies a linear transformation to the data, mapping it to a new space using the ‘nn_linear’ function. The first hidden layer maps the input features to the 64 hidden neurons, followed by a second hidden layer of the same size, both of which use the ReLU activation function to introduce non-linearity. The final fully connected layer outputs the model’s raw predictions, which are then passed through a log-softmax activation function to compute class probabilities. The entire training dataset was passed through the model 100 times (num_epochs = 100) during the training process to allow the model to progressively learn from the training data to improve its accuracy and performance. Within each epoch, we used a small batch size of 32 for more stable convergence, such that 32 training samples were used per iteration so that the model frequently updated its parameters during the training process for more efficient training. We used the ‘nnf_cross_entropy’ loss function for the model to compute the unweighted cross entropy loss for the classification tasks and used an Adam optimization algorithm for Adaptive Moment Estimation with a 0.001 learning rate using the ‘optim_adam’ function to adjust the model’s feature weights during training.

Convolutional neural network using spectrogram images

In addition to the other machine learning models using song feature data, we also applied a convolutional neural network (CNN) on spectrogram images of the towhee recordings to test species classification using a different type of input data and computational approach. We modified an Audio Classification CNN model from [46]. To apply the model to our data, we first created spectrograms from.wav files using the librosa and matplot packages in python and applied low and high pass filters (high pass filter = 1000 Hz; low pass filter = 10000 Hz). Our model accepts images of 224x224 pixels with 3 color channels (RGB) in the input layer. The model then has a series of convolutional layers with 32, 128, 128, and 128 filters, each of size 3x3 pixels, which help to extract increasingly complex features from the input images. Each of the convolutional layers use ReLU activation functions to introduce non-linearity after each convolution. To reduce the dimensions of the feature maps, we use MaxPooling2D layers with a pool size of 2x2 after each convolutional layer to reduce the spatial dimensions of the feature maps, improving efficiency of the model and helping prevent overfitting. The features are then flattened into a one-dimensional vector using the Flatten layer. This vector is then passed to a Dense layer with 1024 neurons and a ReLU activation for further abstraction and learning. The final output layer has 2 neurons (i.e., the Spotted and Eastern towhee classes), using the Softmax activation function to output class probabilities.

To ensure that the results would be comparable to our other models, we used the same training set, such that we trained the CNN on the same balanced set of towhee recordings (N_{Spotted_towhee} = 796 and N_{Eastern_towhee} = 796). We normalized our data by dividing the input data by 255 pixels and applied one-hot encoding on the labels. We trained the model using the ‘model.fit’ function in Python using a batch size of 32 to update the model’s weights after each set of 32 samples, and we passed the entire training dataset through the network 10 times (epochs = 10). We also repeated this model in three different ways: 1) repeating this exact model 10 times, 2) using a filter size of 5x5 pixels and running that model 10 times, and 3) using a filter size of 7x7 pixels and running that model 10 times to capture a larger section of the images (Fig H in S1 Text).

Genetic analysis

We downloaded fasta files of sequences of the cytochrome oxidase subunit I regions of the Spotted and Eastern towhee mitochondrial genomes available in the Barcode of Life Data Systems (BOLD) database and the NCBI database. We obtained a total of 18 Spotted towhee and 5 Eastern towhee sequences. We aligned these sequences using MAFFT [47], and we used a pairwise F_ST (fixation index) in Arlequin [48] to quantify the level of genetic differentiation between the two species by measuring the proportion of genetic variation that is due to differences between populations. It also measures the extent to which individuals within populations are similar to one another, where a larger F_ST describes a greater difference in allele frequencies within a population. We also did an Analysis of Molecular Variance (AMOVA) to quantify the proportion of genetic variation that is due to differences between versus within populations.

PCA and Procrustes analysis of genetic data

Using the ‘glPCA’ function from the adegenet package [49] in R, we ran a PCA on the aligned fasta files from our genetic analysis. We followed with a Procrustes analysis to transform the two-dimensional plot of the first two principal components onto the target geographic plot of the longitude-latitude coordinates of the samples using the ‘procrustes’ function from the vegan package [41] in R. We then performed a resampling test using the ‘protest’ function in the vegan package to assess whether the real song data were more closely associated with geography than expected by chance.

Assessing isolation by distance

We used a Mantel test to assess the correlation between pairwise distance of a geographic distance matrix and both a song feature distance matrix and genetic distance matrix. We computed a pairwise geographic distance matrix, with the distance between each pair of recording locations estimated by the ‘distm’ function from the geosphere package in R (geodesic distance ‘distGeo’ with WGS84 ellipsoid parameters). To obtain the song feature distance matrix of the raw song data, we used the ‘dist’ function from the stats package and calculated the Euclidean distance for each pairwise comparison of the 2785 song bout samples. We then performed a Mantel test of the geographic distance matrix versus each song feature distance matrix using the ‘mantel’ function (with parameters method = “spearman” and the default permutations = 999) from the vegan package.

We use the same distance estimation as above to obtain a geographic distance matrix for the sampling locations of the genetic sequences. Using the fasta file of the aligned genetic sequences from the Spotted and Eastern towhee mitochondrial genomes (N_Spotted = 18 and N_Eastern = 5), we compute a pairwise distance matrix using the ‘dist.alignment’ function from the seqinr package [50]. We then performed a Mantel test of the geographic distance matrix versus the genetic distance matrix using the same method as above.

Data and code availability

Our data and code for all analyses described in this manuscript are available at https://github.com/CreanzaLab/TowheeAnalysis.

Generalized Linear Models

We used generalized linear models (GLMs) to assess the degree to which longitude, latitude, species identification, and file conversion status contribute to the song feature trends we see in the Spotted and Eastern towhees (Table 1). Our analysis suggests that latitude is a significant predictor (p < 0.05) for nine song features, while longitude is a significant predictor of all song features except for average syllable frequency range. Species classification was significant in all song features except bout duration and rate of syllable production. Finally, we assessed whether converting the sampling rate or converting the file type to wav format for consistency had a meaningful effect on the song feature outcomes by including conversion status in the generalized linear models. The models provide evidence that the overall file conversion status was statistically significant only in five song features. For these five features, file conversion had similar t-values to other significant factors, suggesting that the effect of conversion on the song feature has a similar order of magnitude as other factors relative to their standard errors. However, the degree of significance was much less than that of the other variables (i.e., latitude, longitude, and species), indicating that it likely has less predictive power (Table 1). We see similar results when using only samples from the breeding season (Table B in S1 Text). We visualized songs before and after conversion to ensure that converting the sampling rate or file format of the recordings was not altering the timing or frequency of the songs. However, we did observe a geographic bias in file format, with mp3 files overrepresented in the west and thus in Spotted towhee recordings (Fig L in S1 Text), suggesting that conversion status is associated with the geographic variation in song features and is not itself a source of this variation.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Results of generalized linear models for relationship between song feature data and longitude, latitude, species classification, and file conversion status of Spotted and Eastern towhee.

https://doi.org/10.1371/journal.pcbi.1013135.t001

When we plot these song features by longitude and species, the strongest predictors in our GLMs, we did not observe sharp discontinuities in song features in the zone of overlap between the two species (Fig E in S1 Text). The two species showed very similar distributions for some song features in the zone of overlap (e.g., bout duration and rate of syllable production), whereas for some frequency characteristics the two species had partially overlapping distributions and we observed a small cline in the zone of overlap (e.g., average syllable upper frequency). The number of unique syllables showed a clinal pattern in Spotted towhees, increasing in number of unique syllables from west to east, whereas the Eastern towhees had a consistent average number of unique syllables across their range (Fig E in S1 Text). This geographic variation in Spotted towhees could reflect either different selection pressures in the western portion of the range or cultural drift in which a single-syllable song rose to prevalence in the west.

Linear Discriminant Analysis (LDA)

Our linear discriminant analysis of the raw song feature data was able to classify the species of the samples with 86.8% accuracy (balanced accuracy = 86.9%; Table 2). A permutation test suggests that the model performed significantly better than chance (p < 10^-3). We see similar results with song feature data from only the breeding season (Table C in S1 Text). Taking into account the agreement expected by random chance, Cohen’s kappa suggests that there is substantial agreement between the predicted species classification and the actual classification of the samples (κ = 0.73; Table 2). The rate of syllable production had the highest contribution to the discriminant function, followed by the number of unique syllables. While there is some visible separation of species classification in one-dimensional space, there was still considerable overlap (Fig 3).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. Results for machine learning models for analyses of song feature data from song recordings of Spotted and Eastern towhees.

https://doi.org/10.1371/journal.pcbi.1013135.t002

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. Results of a Linear Discriminant Analysis trained on towhee song data.

Plot of LD1 results of a subset of towhee song bouts (N_test = 697) using a Linear Discriminant Analysis trained on raw song feature data from a balanced training set of Spotted and Eastern towhee bouts (N_train = 1592). The model revealed 86.8% prediction accuracy (balanced accuracy = 86.9%; Cohen’s κ = 0.73). Points are jittered vertically for visualization.

https://doi.org/10.1371/journal.pcbi.1013135.g003

Principal Components Analysis (PCA) and Procrustes analysis of song data

A PCA of the song data revealed considerable overlap in the song variation captured by PC1 and PC2. Individual songs in both the range of overlap and non-overlap were scattered throughout PC space; if the individuals in the zone of overlap had songs that showed greater species-level differentiation, we would expect that these points would have greater separation in PC space than those in the zone of non-overlap. PC1 explained 35.6% of the variance, with average syllable duration having the highest loading. PC2 explained 24.8% of the variance, with average syllable upper frequency having the highest loading. With a simple linear partition of this two-dimensional PC space (LDA), we could predict the species of a song with 76.2% accuracy. There was moderate agreement between the actual species classification and the model predictions beyond what is expected by chance (κ = 0.49, permutation test p < 10^-3; Table 2). A Procrustes analysis comparing the first two principal components of the song data to the latitude-longitude coordinates of the recording location indicated that, despite the extensive overlap between the songs of the two species in PC space, there is significantly more geographic signal in the song data than expected by random chance (Procrustes rotation 0.30, p < 0.001, Fig 4A–4C).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 4. Spatial distribution of Spotted towhee and Eastern towhee song data and genetic data.

Song data is shown in panels A–C (N_{total_recordings} = 2785; N_{Spotted_towhee} = 1067; N_{Eastern_towhee} = 1718) and genetic data is shown in panels D–F (N_total = 23; N_{Spotted_towhee} = 18; N_{Eastern_towhee} = 5). (A) Distribution of song recordings in North America. (B) Principal component analysis (PCA) of song bouts using 16 song features. (C) Procrustes analysis of song data using PC1 and PC2 from the PCA in panel B. (D) Distribution of genetic sequences obtained from the Barcode of Life Data Systems database and NCBI. (E) PCA using single nucleotide polymorphisms of aligned sequences of the cytochrome oxidase subunit I regions of the mitochondrial genome. (F) Procrustes analysis of genetic data using PC1 and PC2 from the PCA analysis in panel E. Ellipses indicate 95% confidence intervals. Base maps (panels A, C, D, and F) were made with Natural Earth (http://www.naturalearthdata.com/).

https://doi.org/10.1371/journal.pcbi.1013135.g004

UMAP visualization

Projecting the song-feature data with a UMAP analysis showed that the songs of the Eastern and Spotted towhees did not form two species-specific clusters, but instead made one large cluster with Eastern towhee songs overrepresented on one side and Spotted towhee songs on the other side (Fig 5); one much smaller cluster contained predominantly Spotted towhee songs. As in the PCA (Fig 4B), songs sampled from the zone of overlap were scattered throughout the UMAP projection. The species classification of each song could be better discriminated in the UMAP projection than the PCA plot; with a simple linear partitioning of the two-dimensional UMAP projection with n_neighbors = 15 and min_dist = 0.1, we could predict the species of a song with 86.4% accuracy (Table 2). We tested values of n_neighbors up to 50 and values of min_dist up to 0.9 and found similar prediction accuracies (75.3%–86.4%). This model performance showed substantial agreement with the actual species classification, significantly better than what is expected from random chance (permutation test p < 10^-3; κ = 0.71). The “hybrid/unsure” recordings were located throughout the large cluster of the UMAP projection (Fig 5). The same analysis using data from samples recorded only during the breeding season showed similar results (Fig F in S1 Text). In general, the location of points in a UMAP projection are not stable between runs and are influenced by parameters set by the user, so we do not interpret the distance between points or clusters in this projection to be meaningfully linked to song variation [51,52]; however, we did observe that the relative locations of the song-feature data on the UMAP projection showed an association with the species of the recording (Fig 5).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 5. UMAP projection of Eastern and Spotted towhee song-feature data.

Each point represents an analyzed song bout (N_{total_bouts} = 2785; N_{Spotted_towhee} = 1067; N_{Eastern_towhee} = 1718), with Eastern towhee songs shown in shades of red and Spotted towhee songs in shades of blue. The lighter colors represent recordings from the zone of species overlap. Black dots indicate the 27 recordings from individuals that were classified as potential hybrids (“hybrid/unsure”).

https://doi.org/10.1371/journal.pcbi.1013135.g005

Random forest model

Our random forest model trained on samples from the entire geographic range had an accuracy of 89.5% (κ = 0.78; permutation test p < 10^-3) when tested on a subset of song samples that had been withheld from training (Fig 6A and Table 2). The most important feature in the decision trees was the largest syllable duration. For each recording in the test set, we estimated the confidence of the classifier by tabulating the fraction of trees in the random forest that supported the correct classification (Fig 6B); these confidence values tended to decrease toward the zone of overlap, suggesting that the songs in the zone of overlap were more challenging to classify. We repeated this analysis using 100 different subsets as training data, and the average accuracy was 89.1% ± 0.2.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 6. Geographic distribution of random forest model predictions of Spotted and Eastern towhee species identity.

Predictions were based on a model trained on 16 song features from samples of Spotted towhees and Eastern towhees. (A) We trained a model on song data from the entire geographic range of both species (N_{Spotted_towhee} = 796; N_{Eastern_towhee} = 796) and tested how well it predicted the species identification of a subset of all song samples (N_test = 697; accuracy = 89.5%). (B) The same results from (A) are plotted based on the percent of trees in the random forest classifier that supported the correct species identification; when less than 50% of trees supported the correct identification, the classifier made an incorrect species prediction. The average confidence in species classifications tended to decrease toward the zone of overlap for both species. (C) We then trained a second model on a subset of samples obtained only from the non-overlap zone (N_{Spotted_towhee} = 796; N_{Eastern_towhee} = 796) and tested it on a random subsample of song bouts from both the zone of non-overlap (N_{test_nonoverlap} = 216; accuracy = 93.5%) and the zone of overlap (i.e., 102°W - 91°W; N_{test_overlap} = 216; accuracy = 84.3%). (D) The same results from (C) are plotted based on the percent of trees in the random forest classifier that supported the correct species identification; again, the average confidence of species identifications tended to decrease toward the zone of overlap for both species. (E) We used the same model from panels C and D to predict species identity of song bouts from recordings of “hybrid/unsure” towhees (N_predict = 27). The model predicted that 16 of these “hybrid/unsure” recordings were Spotted towhees and 11 were Eastern towhees, with no discernable longitudinal gradient in the predictions. The dotted line represents the zone of overlap determined by the co-occurence of Eastern towhee and Spotted towhee song recordings (102°W - 91°W). Base maps (panels A, C, and E) were made with Natural Earth (http://www.naturalearthdata.com/).

https://doi.org/10.1371/journal.pcbi.1013135.g006

Our model trained on the samples from the zone of only non-overlap was tested on all samples from the zone of overlap (216 song bouts) and on a withheld subset of the same number of samples from the zone of non-overlap (216 song bouts) which had an accuracy of 84.3% (κ = 0.68; permutation test p < 10^-3; accuracy_100seeds=83.7% ± 0.2) and 93.5% (κ = 0.87; permutation test p < 10^-3; accuracy_100seeds=89.8% ± 0.4), respectively (Fig 6C and Table 2). The most important feature was the largest syllable duration. As in Fig 6B, the classifier’s confidence in its species predictions tended to decrease toward the zone of overlap (Fig 6D). When we used the model trained on only the samples from the zone of non-overlap to predict the species classification of the samples from “hybrid/unsure” towhee recordings, it predicted 16 Spotted towhees and 11 Eastern towhees (Fig 6E). Additionally, we tested the models using both 100 and 1000 trees and on only the recordings from the breeding season and obtained similar results (Tables C–D and Fig G in S1 Text).

Gradient boosting machine

Our gradient boosting machine model was able to predict species classification with 88.2% accuracy (κ = 0.76; permutation test p < 10^-3; Table 2). The most important feature in the gradient boosting machine was the number of unique syllables, contributing about 25% to the model’s accuracy. Similar to the random forest models, the second most important feature was largest syllable duration, contributing about 18%.

Deep learning model

Out of all our machine learning algorithms, our deep learning algorithm was able to predict the species classification with the highest accuracy of 90.0% and showed substantial agreement between the predicted classifications and the actual species classification, performing better than expected by chance (κ = 0.79; permutation test p < 10^-3; Table 2). Nonetheless, all our machine learning classifiers had similar prediction accuracies (Tables 2 and C in S1 Text) despite using different methods. Additionally, 43.4% of the incorrect predictions from the random forest model 1, gradient boosting machine, and deep learning algorithm were misclassified across all three models.

Convolutional neural network

Using a convolutional neural network on spectrogram images of the towhee recordings as an alternative approach to assessing species classification, we obtained an accuracy of 87.52% (balanced accuracy = 88.51%; κ = 0.7457; permutation test p < 10^-3), which is similar to the other machine learning algorithms we used in this manuscript (Table 2). We also ran this CNN model 1) 10 times using the same 3x3 filter, 2) 10 times using a 5x5 filter size, and 3) 10 times using a 7x7 filter size, obtaining an average accuracy of 88.32% ± 1.02%, 90.06% ± 1.52%, and 89.74% ± 1.07%, respectively (Fig H in S1 Text).

Mantel test

A Mantel test of song feature distance versus geographic distance showed that all song features had a weak positive correlation with geographic distance (p < 0.001; Table 3). Even song features that had an extremely weak correlation with geographic distance (e.g., maximum syllable frequency) were found to be more closely associated with geography than the randomly permuted matrices (r=0.018, p < 0.001).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 3. Results for Mantel tests of each song feature distance matrix versus geographic distance matrix.

https://doi.org/10.1371/journal.pcbi.1013135.t003

PCA and Procrustes analysis of genetic data

We analyzed the limited genetic data available for the most commonly sequenced gene in the Spotted and Eastern towhees, obtaining 23 cytochrome oxidase I sequences (N_{Spotted_towhee} = 18; N_{Eastern_towhee} = 5). A PCA using these mtDNA sequences revealed 3 different clusters. Eastern towhees formed their own cluster, while the Spotted towhees were split into two clusters: one group that contained two individuals from Mexico and a second group that included all individuals from the United States and Canada, along with one individual from Mexico. Eastern and Spotted towhees cluster separately, suggesting that they are genetically different, at least based on SNPs from COI mtDNA sequences. PC1 explained 43.1% of the variance in the genetic data, and PC2 explained 23.2%. A Procrustes analysis comparing the first two principal components of the genetic data to the latitude-longitude coordinates of the sampling location indicated that there is significant geographic signal in the genetic data (Procrustes rotation 0.85, p < 0.001, Fig 4D–4F).

Genetic analysis

Our genetic analysis revealed a pairwise F_ST value of 0.64, with an AMOVA showing that 63% of the variance was between species and 37% of the variance was within species. Although this analysis was performed only on the 23 available sequences, our AMOVA results broadly agree with the genetic PCA analysis, where we see a separate cluster for the Eastern towhees (Fig 4E).