Brood parasites use the parental care of others to raise their young and sometimes employ mimicry to dupe their hosts. The brood-parasitic finches of the genus Vidua are a textbook example of the role of imprinting in sympatric speciation. Sympatric speciation is thought to occur in Vidua because their mating traits and host preferences are strongly influenced by their early host environment. However, this alone may not be sufficient to isolate parasite lineages, and divergent ecological adaptations may also be required to prevent hybridisation collapsing incipient species. Using pattern recognition software and classification models, we provide quantitative evidence that Vidua exhibit specialist mimicry of their grassfinch hosts, matching the patterns, colours and sounds of their respective host’s nestlings. We also provide qualitative evidence of mimicry in postural components of Vidua begging. Quantitative comparisons reveal small discrepancies between parasite and host phenotypes, with parasites sometimes exaggerating their host’s traits. Our results support the hypothesis that behavioural imprinting on hosts has not only enabled the origin of new Vidua species, but also set the stage for the evolution of host-specific, ecological adaptations.

Materials and methods

Fieldwork

During January–April 2013, 2014, 2015, 2016 and 2017, data were collected on nestling morphology, begging calls and postural movements over an area of about 40 km² on and around Musumanene and Semahwa Farms (centred on 16°47′S, 26°54′E) in the Choma District of southern Zambia. The habitat is a mixture of miombo woodland, grassland and agricultural fields.

Visual mimicry

Photographing Vidua and grassfinch nestling mouths

Eggs were taken from nests in the wild and placed in a Brinsea Octagon 20 Advance EX Incubator at 36.7°C and 60% humidity. Nestling mouths were photographed within a few hours of hatching in the incubator. The chick was held below a prism until the mouth naturally opened, and the mouth then pressed gently over the apex of the prism (PEF2525 equilateral prism, UV fused silica, 25 x 25 mm aperture, Knight Optical, Kent, UK). This allowed the angular interior surfaces of the chick’s mouth to be projected onto the prism face opposite this edge. A wooden block secured the prism and held a 40% Spectralon grey standard (Labsphere, Congleton, UK) in a consistent position. Photos were taken with a Micro-Nikkor 105 mm lens and a Nikon D7000 camera that had undergone a quartz conversion (Advanced Camera Services, Norfolk, UK) to allow sensitivity to both human-visible and UV wavelengths, by replacing the UV and infrared (IR) blocking filter with a quartz sheet. The camera was placed on a tripod and pointed vertically down onto the flat surface of the prism at approximately 50 cm distance. The chick was gently held between thumb and forefinger as it bit on the prism. For each individual nestling, two photos were taken, each with a different filter. UV photographs were taken with a Baader UV pass filter (transmitting 320–380 nm). Human-visible photos were taken with a Baader UV-IR blocking filter (transmitting 420–680 nm). For each photograph the aperture was set to f13, and the shutter speed varied with exposure. A flash (Metz 76 MZ-5 digital) was attached to the camera body via a lateral bracket and had been modified by removal of its UV blocking filter, such that it emitted both visible and UV light. The flash was set to under-expose by 3 stops for the “visible” images, and to over-expose by 3 stops for the “UV” image. ISO was set at 400 and images were taken in RAW (NEF) format. All images were taken indoors in a dark room to minimise ambient light. The setup is shown in Figure S1. Once the photographs had been taken, the chicks were returned to their nests.

Pattern mimicry

Measurements of overall similarity between mouth marking patterns of different species were carried out using NaturePatternMatch (NPM) (Stoddard et al. 2014). NPM is a computer vision program that uses the Scale Invariant Feature Transform (SIFT) algorithm to detect local features in images and gives each pairwise combination of images a similarity score (Lowe 1999, 2004). These features are thought to correspond to those used by birds in real object recognition tasks (Soto and Wasserman 2012) and have been shown to be important in pattern recognition and egg rejection decisions in another host species, the tawny-flanked prinia (Prinia subflava) (Stoddard et al. 2019). Each image was scaled to the same size, using the width of the prism as a reference, such that the edge of the prism was 1500 pixels long. This value was chosen because it approximates the smallest image in the dataset, and thus minimizes any information loss or artefacts caused by scaling up. Only the green channel was taken from each image, as this corresponds most closely with the spectral sensitivity of the double cones in bird vision, thought to be influential in the processing of pattern information (Cronin et al. 2014). The background and the edge of the prism were masked out and the images cropped to size. NPM calculates pairwise pattern differences between images. As a measure of host-parasite similarity, we calculated the mean distance between each Vidua species and each grassfinch species (raw distance). We additionally submitted these pairwise distances to classical multidimensional scaling, which embeds points in an n-dimensional space in which the Euclidean distances between the points are maintained. This allowed a centroid to be calculated for each species (the average of all positions of all samples from that species). We measured the distance between each Vidua species and each grassfinch species in this space (centroid distance). The qualitative results and conclusions were the same for both methods (Table S1). Sample sizes are summarised in Table S6.

Comparison of upper palate spot size between parasites and hosts was carried out using the R package patternize (Van Belleghem et al. 2017), which quantifies variation in colour patterns from digital images. Analysis was carried out using R v3.4.4 (R Core Team 2018). Homologous regions of the mouth in each photograph were identified by placing five landmarks on reference points around the mouth, and the images were aligned to an arbitrarily chosen reference image. This allowed patterns to be compared among images even if there were slight differences in the distances between camera and chick and in the positioning of the chick within the image. To extract the black upper palate markings, thresholds were manually adjusted for red, green and blue colour channels for each image and their success at extracting black patterns assessed. Some manual adjustment of thresholds was needed between images to account for differences in lighting conditions and ensure that patterns were accurately extracted. Shaded regions that had been erroneously identified as pattern were manually removed from the selection. To compare spot size between hosts and parasites, the number of pixels in the standardised images that each of the upper palate spots contained was calculated for every individual. The spot size was then calculated relative to the overall size of the mouth. Comparisons were performed with Wilcoxon tests in R (R Core Team 2018). The sample sizes for the comparison of spot sizes were the same as for the analysis of pattern mimicry (see Table S6).

Colour mimicry

Raw pixel values from the red, green and blue channels for both the visual and the UV images were extracted from regions of interest (ROIs) in nestling mouth images using the Multispectral Image plugin in Image J (Schneider et al. 2012; Troscianko and Stevens 2015). Chosen ROIs were: 1) gape flanges, 2) outer upper palate (distal to medial palate spot), 3) inner upper palate (proximal to medial palate spot), 4) medial palate spot. ROIs 1, 2 and 3 were selected separately on right and left-hand sides of the chick’s mouth and a mean score of the two values was used. The medial palate spot lies along the bilateral line of symmetry for the chick’s mouth and so only a single ROI was required. Raw pixel values were converted into avian cone capture values based on the cut-throat finch (Amadina fasciata) visual system (Hart et al. 2000a) using Microsoft Excel version 15.30. The cut-throat finch is the most closely-related grassfinch species to the hosts of Vidua finches for which visual sensitivities have been calculated (Olsson and Alstrom 2020).

Cone-capture values for each image were analysed with a discriminant function analysis (DFA) using the MASS package in R (Venables and Ripley 2002). A multinomial logistic regression (MLR) was also carried out on the same dataset. While both DFA and MLR can be used to address questions about categorisation, MLR has fewer restrictive assumptions than DFA. However, DFA is thought to be a better approach when sample sizes are small (Pohar et al. 2004). For DFA and MLR, the models were initially trained on cone capture values of the images from the 10 co-occurring grassfinch species we photographed at our study site. The results from both MLR and DFA were similar (Table S2) and so only the DFA results are reported in the main text. Sample sizes are summarised in Table S6. MLR was implemented using the multinom function from the R package nnet (Venables and Ripley 2002). DFA was implemented using the lda function from the R package MASS ((Venables and Ripley 2002). The observed versus expected percentages were compared using the binom.test function in R base stats package (R Development Core Team 2017).

The DFA/MLR models were initially trained on cone-catch values of the estrildid data.  The training data consisted of 3 locust finch (Paludipasser locustella), 32 common waxbill, 10 blue waxbill (Uraeginthus angolensis), 7 green-winged pytilia (Pytilia melba), 5 orange-winged pytilia, 4 red-billed firefinch (Lagonosticta senegala), 15 Jameson’s firefinch, 5 zebra waxbill (Amandava subflava), 5 African quailfinch (Ortygospiza atricollis) and 9 bronze mannikin (Spermestes cucullatus) individuals (see Table S6). The models were then tested using the cone-capture values from the parasite species data. If the ROI colours of parasites match those of their host more closely than any other sympatric grassfinch, parasite data should be classified by the discriminant function as an instance of its specialist host species more frequently than would be expected if the parasite data were randomly assigned to any of the host species. These testing data were extracted from images from 17 pin-tailed whydah (Vidua macroura), 5 purple indigobird (V. purpurascens) and 1 broad-tailed paradise whydah (V. obtusa). The reason for the small sample size for broad-tailed paradise whydah is that it is an uncommon species whose host’s nest is difficult to find. To our knowledge our photographs and sound recordings are the first ever taken of this species’ nestlings in the wild.

Imperfect colour mimicry of hosts by parasites was investigated by comparing the hues of corresponding mouth structures in parasites and hosts. As in the colour mimicry analysis, gape flange, upper palate (inner and outer) and medial palate spot colours were compared in hosts and parasites. To test for differences in hue in each host-parasite pair, multivariate analysis of variance (MANOVA) was carried out, using the manova function in R (R Core Team 2018), with the four cone catch values as the response and species identity as the explanatory variable. To compare luminance of these structures in each host-parasite pair, a t-test was carried out on the double cone channel values. The double cone channel (the sum of the medium and long wave cone catch values) is thought to be a good proxy for luminance vision in vertebrates (Pignatelli et al. 2010; Cronin et al. 2014).

Vocal mimicry

Recording nestling begging calls

Chicks were removed from their nest and placed in an artificial nest inside a box. The artificial nest consisted of a plastic bowl, used as a nest platform in aviculture, tightly lined with nesting material from abandoned grassfinch nests. Chicks were left in the artificial nest for a few minutes to allow acclimation. To stimulate begging, the chick was tapped gently with forceps on the bill. Recordings were made using an Audio-Technica ATR35s tie-clip microphone (or a Sennehiser ME-66 shotgun microphone for part of the 2014 field season) held by hand approximately 3 cm away from the focal bird’s mouth. Vocalisations were recorded in WAV format on a Tascam DR-05 portable recorder. Recordings were made for around 2 minutes or until sufficient begging calls had been obtained (at least 10 seconds of continuous begging where possible). After recordings, the chicks were returned to their nests. Sonograms were produced and analysed using the default settings in Raven Pro 1.5 (Bioacoustic Research Program 2014).

Testing for mimicry in begging calls

Classification models were used to test the hypothesis that nestling Vidua mimic the begging calls of their hosts. To do this, 13 parameters were extracted from each call: frequency bandwidth, bandwidth 90% (the frequency range containing 90% of the total call energy), call duration, duration 90% (the period of time containing 90% of total call energy), peak frequency, centre frequency, minimum frequency, frequency 5% (the frequency above which 95% of the total call energy is contained), maximum frequency, frequency 95% (the frequency below which 95% of the total call energy is contained), total energy, aggregate entropy and average entropy. We used all these parameters to maximise the amount of information given to the model, and so allow it to characterise the host calls as well as possible. Many of these parameters have been used previously to characterise the vocalisations of birds, particularly to compare the begging calls of avian brood parasites and their hosts (Langmore et al. 2008; Anderson et al. 2009; De Mársico et al. 2012). Calls were defined as the basic repeated unit within a bout of begging. For most species, this represented a single uninterrupted trace on the sonogram, except for common waxbill and pin-tailed whydah which give a two-note call (transcribed as “we-chee”) that is repeated rapidly. This call was described as these two units combined.

Both a discriminant function analysis (DFA) and a multinomial logistic regression (MLR) model were then trained on begging call parameters from locally-occurring grassfinch nestlings (for explanation of the relative merits of DFA and MLR see “Colour mimicry” above). This created a function, built from the 13 parameters, which best separated the begging calls of each host species. The training data included calls from five common waxbill, one African quailfinch, four blue waxbill, two bronze mannikin, two Jameson’s firefinch, three green-winged pytilia, three orange-winged pytilia and two zebra waxbill individuals (see Table S6). To maximise the discriminatory ability of the DFA/MLR, individual call notes, rather than means for individuals, were used as input data points. This allowed the maximum quantity of data to be used in the creation of the classification function. It also means that the model was exposed to parameter values from actual calls rather than to abstract “mean calls”.

Having constructed classification functions, we then used parasite calls as test data. We tested five pin-tailed whydah, two broad-tailed paradise whydah, and two purple indigobird individuals. Ten call notes from each parasite individual were entered into the MLR and DFA classification functions. To assess mimicry, we calculated the proportion of the ten input calls that were classified as belonging to the host species on which each parasitic species is specialised. Each parasite individual was given this “proportion correct” score. If the mean of these scores across individuals of a parasite species was significantly greater than that expected if parasites were randomly allocated to grassfinch species, it would suggest that parasites match the calls of their hosts better than the other sympatric grassfinch species. We quantified a “proportion correct” score for each individual parasitic chick. Sample sizes are summarised in Table S6.

Begging call recordings were taken from chicks in mid to late development, the stage at which their begging calls become most crystallised and stereotyped. Chicks from several grassfinch species in our study gave various call types earlier in development but settled to consistent calls in mid to late development. Mid-development stage was characterised as being the point at which the primaries had erupted from their pins. This has been used as an indicator of developmental stage in other studies of brood parasite begging (Briskie et al. 1999; Ranjard et al. 2010). The nest composition at the time the chick was recorded varied from one to five host chicks.

One species, the pin-tailed whydah, showed four call types throughout development (Jamie et al. submitted). However, one call is made only by nestlings in mid to late development: a distinctive, two note “we-chee” call, whereas the other three are made earlier in the nestling period. Common waxbill nestlings also make a two-note call in mid to late development (Jamie 2017a). To simplify the analysis, only two-note call types of pin-tailed whydahs and common waxbills were included in the analysis. Three of the five pin-tailed whydah chicks used in the analysis of begging call mimicry (individuals 3, 4 and 5 in Table S3) had been raised in the nest of a blue waxbill and not the natural common waxbill nest. These chicks had been transferred to blue waxbill nests as part of transfer experiments for another study (Jamie et al. submitted). If the calls of pin-tailed whydahs raised in a blue waxbill nest are still assigned as most similar to common waxbill calls by the model, this would suggest that the pin-tailed whydah begging call mimicry is largely innate and not dependent on interactions with its specific host.

Testing for imperfections in vocal mimicry

Differences in the structures of parasite and host begging calls were analysed using linear mixed models. We constructed models using the “lmer” function in the R package lme4 (Bates et al. 2015). As explanatory variables, species identity was a fixed factor and individual identity a random factor, thus avoiding pseudoreplication. To assess whether species identity had a significant effect on call structure, we compared the fit of a model which included species identity and individual identity as explanatory variables with that of a model which included only individual identity. To test for differences in the rate of calling between parasites and hosts, we counted the number of begging calls made over a 6s period of consistent begging. This was done at three points of consistent begging across each recording and the mean call rate taken for that individual. Call rates between pin-tailed whydah and common waxbill were compared using a Wilcoxon test.

Postural mimicry

Chicks were filmed on a Canon Powershot SX50 HS Digital Camera while audio recordings were being made of their begging calls, to record the chicks’ head movements during begging. Examples of begging displays of each species are included in the supplementary materials.

Mimicry was quantified by showing human participants (n = 12) a series of silent, unlabelled videos of nestling grassfinch and Vidua chicks begging. There are currently no avian models of movement perception, and it is difficult to accurately extract quantitative data on chick movement given the inconsistent angle and distance between camera and bird. Therefore, we instead made use of humans, naïve to the hypothesis being tested, as natural movement and pattern recognisers.

Participants were asked to categorise three aspects of movement during the begging display: 1) head rotation, 2) tongue movement, 3) wing movement. Head rotation could be classified as being in the pitch, roll or yaw axes, or absent. Tongue movement could be classified as extended, rapid buzzing, or absent. Wing movement could be classified as waving or absent. For each video, participants described the postural aspects of the begging display according to these characters. The videos were unlabelled so participants did not know what species they were being shown. The order of presentation of videos was randomised. Sample sizes of videos presented to participants are summarised in Table S6. Videos of the begging movements of each species are uploaded with the online supplementary material.

We presented videos in a random sequence and asked participants to characterise the head, tongue and wing movements. This approach, rather than asking participants to match a video to a range of possible reference videos, was chosen to prevent participants from using morphological similarity between chicks (which would be apparent in the videos in addition to the movement) to help make the decision rather than focussing only on movement. By presenting them with videos in sequence and asking them to describe the footage, the descriptions of host and parasite movements could be compared without the confounding effect of morphological similarity. The modal description of each movement for each species by the participants is reported in Table S4.