Ethics statement

This project is run under the coordination of KLS and MIM (Instituto Nacional de Ciência e Tecnologia em Áreas Úmidas (INAU); ‘Programa Ciência sem Fronteiras’ (CsF) of the Brazilian government). Research was conducted under permission of the Brazilian Ministério do Meio Ambiente (MMA), Instituto Chico Mendes de Conservação da Biodiversidade (ICMBio), Sistema de Autorização e Informação em Biodiversidade (SISBIO, permit no 39095–4 issued to KLS). Automated acoustic monitoring is a non-invasive survey method that does not negatively affect the studied animal communities.

Study area and data acquisition

Our study was carried out in the northern Pantanal region, municipality of Poconé, Mato Grosso, Brazil. Audio recordings were collected with Song Meter SM2+ recorders (Wildlife Acoustics; www.wildlifeacoustics.com) in the Fazenda Pouso Alegre area (–16.50303 S, –56.74533 W; 115–126 m a.s.l.; c. 110 km²; recording period: July 2012 through October 2013) and at the Estância Ecológica SESC Pantanal: Reserva Particular do Patrimônio Natural, Baía das Pedras (–16.49879 S, –56.41309 W; 119–131 m a.s.l.; 878.7 km²; recording period: November 2013 until June 2016). In total we gathered c. 120 TB of continuous soundscapes.

For the purpose of the present research, we automatically analyzed the acoustic activity of Vanellus chilensis for the period 29 October 2012 to 26 October 2013, i.e. 12 months of recordings. As of 6 October 2013 we gathered 32 242 soundscapes of 14 min, with intervening pauses of 1 min. Moreover, until the end of the recording period at Fazenda Pouso Alegre we gathered another 959 soundscapes of 30 min without pauses. Very few gaps in time coverage occurred due to logistical problems: c. 4 days around New Year’s Eve 2012/2013, 27 h between 11 and 12 January 2013, and 20.5 h between 17 and 18 August 2013. Eventually, we performed a comprehensive data analysis only for April to September 2013 (section Validation datasets and assessment procedures). The latter subsample contained 17 493 soundscapes of 14 min each (c. 2.7 TB = 4373 h).

All data originate from PPA001, one of nine permanent recording stations deployed at the property of Fazenda Pouso Alegre. PPA001 was located in an extensive, seasonally-flooded, and savannah-like cattle pasture (S1 and S2 Figs). In 2013 the flooding persisted between mid-January and mid-June, whereas in other years floodwaters may recede as early as March. Starting with 1 April, we took water level measurements on every seventh day until 9 June, i.e. on 11 occasions until the floodwaters completely receded. We measured the water level at a distance of 300 m from PPA001, after readings taken at both sites demonstrated that they were almost identical.

The Song Meter SM2+ recorder was configured for 48 dB hardware amplification and 6 dB software amplification and we used two-channel recording with sampling rate 48 kHz, resolution 16 bits per sample, and no WAC compression. Here we processed only channel A, corresponding to the left microphone (i.e. omnidirectional SMX-II). We did not explicitly measure the perceptive range of the SM2+ recorder. Likewise, information about the source level and directionality is not available. Since the noise floor level varied over time due to changing environmental conditions (e.g., S1 and S2 Figs), the detection distance for V. chilensis sound events also varied. However, the signal amplitude of the weakest V. chilensis calls detected in the soundscapes showed little variation in the study period (section Recognizer performance assessment). Therefore we regard the effect of the time-varying ambient noise as negligible. We estimate that the weakest V. chilensis signals recorded in this study (c. –50 dB), correspond to a distance of 300 and 500 m from the microphone. This was the minimum distance between station PPA001 and some higher-ground areas, such as dirt roads and cattle gathering sites, which had already fallen dry in early April (cf. S1A Fig, showing the extent of drying open areas in pinkish color as of 4 May).

The target species

Southern Lapwing Vanellus chilensis (Molina, 1782) is a common and widespread Neotropical wader of open habitats, such as savannahs, steppes, floodplains, riverbanks, lakeshores, boggy terrain with short matted vegetation, pastures, city parks, and ploughed fields. Its four recognized subspecies are distributed from Nicaragua south to Tierra del Fuego, Argentina [44,45,56,57]. The subspecies present in Central Brazil is V. c. lampronotus (Wagler, 1827) (S3 Fig).

Vocalizations.

Southern Lapwing is vocally active all months of the year and at day and night [44,54,58]. Their sounds are particularly conspicuous during flapping courtship and territorial flight displays as well as in threat situations [36,49]. At least some of the variation in alarm calls is functional [48,49]. Choruses are frequently heard during the nonbreeding season, when flocks up to a few hundred birds congregate [44,46,55,59]. The spectral structure of typical call notes emitted by adult and immature birds is quite complex and consists of multiple harmonics with most of the energy concentrated in the frequency range 1.1–10 kHz (cf. [36]: Fig 1, Table 1). We point out that only the aforementioned vocalization types were available for the compilation of the V. chilensis training library ([36]: section 2.1.3.2) used for detector development (section The automated acoustic detector). Other less well-known lapwing call variants undoubtedly exist, including distinct short series of dry and mechanical-sounding warning rattles, low-volume/low-frequency contact calls lacking the characteristic upper harmonics of other vocalizations of adults, and calls emitted by chicks. Because the latter call types were not represented in the V. chilensis training library, they were neither detected by the automated sound recognizer nor considered in the analyses of the manually annotated validation library (section Validation datasets and assessment procedures). Explicitly, in the present study we focus only on the characteristic multi-harmonic lapwing vocalizations (Fig 1, S1–S9 Audios; and [36]: Fig 1).

Download:

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

Table 1. Cumulated monthly number and precision of Vanellus chilensis raw detections.

https://doi.org/10.1371/journal.pone.0169041.t001

Download:

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

Fig 1. Spectrograms of exemplary Vanellus chilensis vocalizations.

All spectrograms represent unfiltered 4-s segments of WAV soundscapes passively recorded at station PPA001 in the Fazenda Pouso Alegre area (windowing function: Blackman-Harris; resolution: 512 bands; decibel range: 85 dB; the strongest target signals were equalized to –3 dB). A: flight calls of a lapwing patrolling its inundated territory; 28 Apr. 2013, 17:38 h (S1 Audio). B: display and agitated cackling of a pair flying over their inundated territory; 29 Apr. 2013, 16:35 h (S2 Audio). C: short series of cackling alarm calls of perched V. chilensis close to its nest; 31 May 2013, 14:31 h (S4 Audio). D: short series of agitated alarm calls of a lapwing flushed from its nest or close to it; 30 May 2013, 06:13 h (S6 Audio). E: excerpt of a long alarm call series of a V. chilensis flock mobbing a predator to defend their chicks; 16 Jun. 2013, 15:35 h (S8 Audio). F: segment of a long alarm call series of a pair or flock mobbing a potential predator or food competitor; 21 Jun. 2013, 12:28 h (S9 Audio).

https://doi.org/10.1371/journal.pone.0169041.g001

The automated acoustic detector

For the automated recognition of Vanellus chilensis vocalizations we used the species-specific acoustic detector reported in Ganchev et al. [36]. The automated species-specific detector consisted of two main processing stages, audio parameterization and pattern recognition (Fig 2).

Download:

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

Fig 2. Overall block diagram of the Vanellus chilensis recognizer, a Gaussian Mixture Model—Universal Background Model (GMM-UBM)-based detector.

https://doi.org/10.1371/journal.pone.0169041.g002

The audio parameterization consisted of audio preprocessing, feature extraction, and feature standardization steps. The preprocessing comprised: (i) rescaling the amplitude of each soundscape recording to peak levels [–1, +1] through adaptive gain control (AGC) amplified weak audio events that otherwise would be disregarded due to their low energy; (ii) down-sampling the signal to sampling rate 24 kHz reduced the computational demands in the subsequent processing steps and removed high-frequency insect sounds ≥ 12 kHz; (iii) filtering with a Butterworth high-pass filter (order 10, cut-off frequency 1000 Hz) suppressed low-frequency interference, such as wind noise and some anthropogenic noise of distant cars and airplanes; all remaining sound signals were treated uniformly regardless of their origin; and (iv) segmenting the audio to short 20-ms segments with 75% overlap.

For each segment we computed a set of 90 audio descriptors, including 35 Linear-Frequency Cepstral Coefficients (LFCC) [67] and their first and second temporal derivatives, referred to as ΔLFCC and Δ²LFCC respectively. Here we kept only the first 20 Δ²LFCCs, as the last 15 were found to be less informative. All feature vectors were post-processed to uniform dynamic range in order to facilitate the pattern recognition process. During the signal parameterization, we treated all target sounds and noises uniformly, so that the abovementioned set of 90 audio descriptors was computed for each segment of audio. Based on the post-processed feature vectors, an acoustic activity detector selects for further processing only those parts of the audio with considerable acoustic activity. Here we set the threshold to –50 dB with respect to the full range input of the SM2+ recorder. The selected segments are used in the pattern recognition process. We only considered detections of acoustic events with amplitudes of at least –50 dB because below that level V. chilensis signals were to faint for being discerned and annotated with confidence, even for a bird sound expert (section Recognizer Performance Assessment).

The V. chilensis detector used in this study explicitly modelled the target species and the acoustic environment through purposely developed Gaussian Mixture Models (GMMs). In particular, based on 75 h of soundscape recordings, we built a Universal Background Model (UBM) via an Expectation–Maximization (EM)-based training of a 512-component GMM. The UBM represents the model of everything that is not Southern Lapwing. Next, we created the V. chilensis model via means-only adaptation of the UBM and a purposely-designed training dataset, containing characteristic lapwing vocalization types (section The target species). The training dataset consisted of manually tagged and selectively filtered recordings of the target species ([36]: section 2.1.3.2). Subsequently, each vector of audio descriptors is scored against both models and the degree of similarity is computed. The log-likelihood estimator scores were compared with a predefined threshold in order to make decision on the presence or absence of V. chilensis vocalizations in the particular audio segment. Based on this decision, we assigned each audio segment represented by a set of audio descriptors to the target species or to the background noise. In this sense, we treated audio segments containing sounds of wind, rain, and other interferences no differently than any other portion of audio. The only difference was that they showed lower scores when tested against the target species model and, on some occasions, higher scores when tested against the UBM. Therefore segments dominated by wind, rain, and other non-target noises were typically classified as not representing V. chilensis acoustic activity.

In a last step, we processed the sequence of scores and decisions with a temporal smoothing algorithm that incorporates domain knowledge about the structure and duration of the target V. chilensis sound events. Similar to Ganchev et al. ([36]: Table 2), we implemented a set of simple rules: (a) the minimum duration of a call was 50 ms and (b) the maximum duration of pauses between calls in a call series was arbitrarily set at 0.6 s. The first rule translated to convolution of the sequence of decisions with a 50-ms window function that smoothed fragmentation due to interferences in the detection of individual calls, facilitating reliable reconstruction of the start and end of V. chilensis call events. The second rule led to convolution of single calls with a window function of duration 600 ms, which ideally resulted in detection of cohesive sequences of calls as a single call series.

Download:

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

Table 2. Validation library: cumulated number of confirmed and unconfirmed Vanellus chilensis call events according to the loudest call amplitude in the vocalizations.

https://doi.org/10.1371/journal.pone.0169041.t002

The recognizer output for each soundscape recording is a list of start/end timestamps of V. chilensis sound events, representing isolated single calls and call series. We point out that at the current stage of development we made no effort to build a recognizer that can distinguish between call series emitted by single birds, pairs, and groups of V. chilensis. Moreover we considered any split of call series into fragments undesirable. Thus, overlapping choruses emitted by different birds or groups of birds usually were merged within a single detection. Thereby we avoided that long call series were given more weight than isolated single calls and short call series. Otherwise the calling activity of floating nonbreeding lapwing flocks could have masked the expected increase in vocal activity during the main breeding period. Consequently, in the present study the term vocal activity refers to the cumulated number of detected sound events per time unit and does neither make a statement about the number of vocalizing lapwings nor about the number of call notes contained in each call series. We analyze the cumulated counts of V. chilensis acoustic events per hour, day, and month, as well as per daytime and nighttime.

Validation datasets and assessment procedures

A bird sound expert prepared all training and validation datasets and assessed the detector results. Inspecting the raw detection results of the complete annual cycle of recordings (Table 1), we decided to focus our analysis on April to September 2013. The latter period includes the supposed main breeding season with the expected higher than average number of V. chilensis detections (cf. assumptions A1-A4).

We evaluated the V. chilensis recognizer performance through a two-pronged approach: (I) For the determination of the false positive rate we drew representative random samples from the V. chilensis detections computed for each month. (II) For all other performance statistics, we selected 26 soundscapes containing at least one correct positive detection. Detailed descriptions of the two validation datasets follow:

1. Random sample.—The minimum number of assessed detections was 150 for months with low activity levels and 2% of all detections for months with high acoustic activity of the target species (S1 Table). We implemented a very strict definition of false positives, counting as insertions even those sound events where the start/end timestamps referred to non-target species rather than to the faint V. chilensis signals discernible in the background of the same segment. In the case of doubts with respect to the identification of non-target bird species, we consulted xeno-canto (www.xeno-canto.org) and several comprehensive audio publications on Neotropical birds [68–70]. This way we could identify all but one non-target bird species that provoked detections. In a few cases, it remained unclear whether the detection was triggered by V. chilensis or by a non-target species, and as a result we excluded eight detections from the analysis. Subsequently we assessed the recognition results in terms of precision measured in percentages: (1) where H (hits = true positives) indicates how many times the V. chilensis recognizer correctly detected a target vocalization and I stands for insertions (false positives). Here H+I is the valid sample of target and non-target sound events, i.e. the number of randomly selected detections minus the positives for which it remained unclear whether they were triggered by V. chilensis or by another species (S2 Table).
2. Validation library.—Determination of the total number of target events in soundscapes is a laborious task. For practical reasons, we took the following approach for selecting a proportionate subsample of soundscapes: (a) We considered only recordings for which the presence of V. chilensis sound events was confirmed on the basis of the aforementioned random sample. (b) For each month, we selected a minimum of four recordings, about one per week. (c) The target value of recognizer detections was 10 to 12 per recording, however we chose variant numbers of detections when no soundscapes were available in that range. (d) The target sample size of expert-confirmed V. chilensis sound events per month was ≥ 100 for the dB-level range 0 to –50 dB. (e) For September the sample size was only reached after annotating six soundscapes, resulting in a total of 26 tagged 14-min soundscapes for the entire study period (S3 Table).

For the manual tagging of the soundscapes we followed the approach of Ganchev et al. ([36]: section 2.1.3.3). Methodological details specific to the present work are outlined here: (A) For the purpose of manual tagging we made use of an FFT filter to reduce competing noises such as wind and insects by –100 dB in the frequency ranges 0–640 Hz and 12.6–24 kHz. (B) We annotated directly in Adobe Audition version 3.0 (www.adobe.com) by setting start and end time marker cues for each target event. (C) The recognizer-generated timestamps were not changed. Where various detections referred to different segments of the same call event, we regarded only the first detection as valid and subsequent detections as double hits (S3 Table). Likewise, we counted false positive detections separately. (D) For strong acoustic signals, we noted the compound dB-values of the loudest call (i.e. signal plus background) [cf. 36]. By contrast, for weaker target events below about –35 dB, we selectively filtered the background noise of the frequency bands between the harmonics of the loudest calls to avoid overestimation of the dB-values. (E) We repeated the manual screening for potential V. chilensis sound events and subsequent identification several times. Depending of the complexity of the soundscape, the time effort per 14-min sound file was between about 16 and 40 h.

For the performance assessment, we considered the following annotations for analysis (Table 2): (i) Confirmed target events, corresponding to the characteristic multi-harmonic type of lapwing vocalizations emitted by adults and immature birds (section The target species). (ii) Unconfirmed events, comprising audio signals that we could not identify with certainty as target or non-target, either because they were too faint or because background noise levels were too high.

We evaluated the recognition results in terms of four performance metrics: accuracy (Eq 2), correct (Eq 3), missed (Eq 4), and precision (Eq 1) measured in percentages (S1 Appendix): (2) (3) (4) where H stands for hits, I for insertions, and N is the total number of confirmed target events according to the validation dataset. We calculated the precision (Eq 1) and accuracy (Eq 2) only for the sake of completeness, although it is evident that the number of annotated soundscapes was too small and not representative with respect to false positive detections when compared with the results of the random sample of detections.

Acoustic activity patterns

During the six months covering the presumed main breeding period, the false positive rates fluctuated widely between months (Table 1). In order to account for these pronounced changes, we split the study period into phases with similar error rates, calculated averages per period, and used the resulting percentage values to correct the number of daytime and nighttime detections (Table 3, Fig 3). Daytime detections included phases of increased activity around dawn and dusk, i.e. the entire period between one hour before sunrise and 45 min after sunset (Fig 4).

Download:

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

Table 3. Cumulated daytime and nighttime detections per acoustic activity period of Vanellus chilensis and calculation of precision.

https://doi.org/10.1371/journal.pone.0169041.t003

Download:

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

Fig 3. Corrected cumulated daily acoustic activity and variation of nocturnal activity of Vanellus chilensis in relation to lunar phases and water levels.

Upper part: blue bars = diurnal activity; red bars = nocturnal activity; dashed orange lines = full moon nights. Black dotted lines mark key dates and periods in the V. chilensis life cycle, which in the text are referred to as vcA, vcB, etc.: A = occupancy of territory; B = egg-laying; C = day 12 of incubation; D = earliest date of hatching; E = change from downy to juvenile plumage; F = acquisition of flight (fledging), and G = attainment of independence by juveniles. Lower part: blue line = water level; dashed lines and circled numbers = periods for the calculation of average false positive rates, hereafter and in the main text referred to as ap for ‘acoustic activity period’; gray circled numbers = ap with false positive rates ≥ 10%; black circled numbers = ap with false positive rates < 10%. Original free image for the design of moon phase icons by courtesy of PhotoVaco.com (www.photovaco.com/photos/animals/full-moon-2-628).

https://doi.org/10.1371/journal.pone.0169041.g003

Download:

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

Fig 4. Cumulated diel acoustic activity of Vanellus chilensis in relation to sunrise and sunset in June (n = 20 498), July (9367), and August 2013 (4697).

A: detections per hour for all three months; B-D: monthly proportion of detections per quarter hour. Hours of sunrise and sunset correspond to average values (dashed orange lines). Original public domain pictogram for the design of sunrise and sunset icons by courtesy of Icône.com (http://en.icône.com/tag-sun-0).

https://doi.org/10.1371/journal.pone.0169041.g004

Uncorrected diurnal acoustic activity was analyzed only for those three months for which average false positive rates were below 10% (i.e. Jun.–Aug. 2013; Table 1), because for the remaining months unevenly distributed false positives may have resulted in distortions of the activity patterns. We plotted per hour and per quarter-hour views to demonstrate the effects of time resolution on the interpretation of diurnal activity patterns (Fig 4).

Statistical tests

For testing the anticipated negative correlation between water level and V. chilensis vocal activity, we performed a one-tailed Spearman’s rho test, using the last ten water level measurements and the corresponding five-day averages of corrected lapwing detections. The water level reading from 1 April (= 42 cm) was omitted because with respect to the averages of acoustic activity this measurement lay outside the study period. Here we opted for a non-parametric correlation test because histogram plots revealed irregular distribution of the data.

Comparing the anticipated seasonal patterns (cf. assumptions A1-A4 and section Duration of the breeding cycle) with Fig 3 we found seven life-cycle phases of V. chilensis acoustic activity for which we present means and standard deviations for small samples ().

Finally, to test for differences in variance of V. chilensis vocal activity between nights with and without moonlight we formed three samples, each containing 13 nights around (a) the new moon preceding the full moon phase to be tested, (b) the full moon phase itself, and (c) the next new moon. We excluded one or two nights around first quarter and last quarter moons, depending on the total length of the lunar phases. The data were log₁₀-transformed before performing a one-way ANOVA (analysis of variance) with significance level at p = 0.05. Subsequently we used post hoc Tukey HSD (honestly significant difference) tests with significance level at p = 0.01 for pairwise multiple comparisons.

Recognizer performance assessment

For the six-month period covering the supposed main breeding season, the automated sound recognizer detected 48 879 supposed target events (Table 1, S1 Table). Of these we validated the identification of 1250 detections, embedded in 1063 original soundscapes. The validation of the random sample of detections showed that the recognizer precision varied considerably during the breeding cycle (Table 3). Fourteen avian species triggered false positives, besides some other sound sources (S2 Table). The highest false positive rate was reached in April (52%), when V. chilensis vocal activity was particularly low. The lowest rates were found for the period June to August (3–8%) with high acoustic activity of the target species. In the period April to June the majority of misidentified audio signals were emitted by common non-passerine waterbirds, viz. Limpkin Aramus guarauna (n = 64), Wattled Jacana Jacana jacana (32), Black-bellied Whistling Duck Dendrocygna autumnalis (27), and White-backed Stilt Himantopus melanurus (10). By contrast, in the late dry season months July to September common landbirds triggered most false positives, namely Great Kiskadee Pitangus sulphuratus (20), Monk Parakeet Myiopsitta monachus (15), and Guira Cuckoo Guira guira (10).

The ratio between detected, missed, and unconfirmed target events demonstrates the limitations of the V. chilensis recognizer in comparison with human experts (Fig 5A and 5B). The percentage of correct hits rapidly decreased with attenuating signal strength, resulting in a rather low overall recognizer performance (S1 Appendix). However, monthly false negative rates in the amplitude range of 0 to –50 dB showed little variation at 72.7 ± 6.9%, with more misses in the period May to July (74.4–80.2%) and less in the other months (62.7–71.3%). Likewise, the amplitude of the weakest event detected per soundscape was rather invariable (–39.3 ± 5.4 dB, n = 26 out of 281 correct detections), with May (–35.3 dB) and June (–37.3 dB) being the months when signal amplitude had to be higher than in the other months to trigger detection (S3 Table). Unexpectedly, in a few recordings extremely weak target vocalizations were also detected, e.g. in July and August. Visual inspection of the corresponding audio files revealed that the recognizer had assigned positive hits to compound signals, consisting of loud cicada songs and faint target signals. The monthly rate of repeated detections of single V. chilensis call events was also rather steady (22.7–28.8%), with the exception of August when the rate was lower (13.8%; S3 Table).

Download:

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

Fig 5. Recognizer performance, according to the loudest call amplitude: number of detected, missed, and unconfirmed Vanellus chilensis events in the validation library.

A: per 5-dB category; B: cumulated number. The dashed lines indicate a change in scale. The ranges –51 to–∞ and 0 to–∞ dB show the effect of target vocalizations with an amplitude < –50 dB, which are barely identifiable even for a bird-sound expert.

https://doi.org/10.1371/journal.pone.0169041.g005

Seasonality of vocal activity

For the complete annual cycle of recordings the V. chilensis sound recognizer time-stamped a total of 62 292 call events (Table 1). As predicted by assumptions A1 and A3, the number of lapwing detections fluctuated around an average value most of the year, i.e. 3000 ± 900 call events (Nov. 2012-Apr. 2013 and Aug.-Oct. 2013), and markedly increased above average levels only for a three-month period (May-Jul. 2013). The latter period was sufficiently long to cover the main V. chilensis breeding season [46,47]. The predicted decrease in acoustic activity during the height of the flood cycle (Feb.-Apr. 2013) was not discernible in the raw detection data (Table 1), likely due to the high activity of waterbirds triggering false positive detections. The 52% false positive rate found in the random sample of detections for April supports this interpretation (Table 1).

Based on seasonal shifts in the occurrence of sound emissions of V. chilensis and the non-target species we distinguished between eight different acoustic activity periods (S2 Table). Subsequently we determined period-specific false positive rates (Table 3) to correct the cumulated number of daily V. chilensis detections (Fig 3). This procedure barely affected the observed general pattern of seasonal acoustic activity (S4 Fig). In later sections below, prefix symbols for dates and periods given in parenthesis refer to those shown in Fig 3, termed hereafter ap for ‘acoustic activity period’, and vc for certain crucial events in the V. chilensis life cycle. All comments on the behavioral contexts of V. chilensis vocalizations (Fig 1A–1F) and the corresponding audio files provided in the Supporting Information (S1–S9 Audios) refer to sounds found in the random sample of 1250 detections.

Daily activity patterns

Once we had accounted for erroneous detections, a one-tailed Spearman’s rho test confirmed a significant negative correlation between the height of the water level and the intensity of V. chilensis acoustic activity (r_s = –0.95, p < 0.004). As predicted by assumption A2, lapwing vocal activity was lowest in the first two weeks of April 2013, with 19 ± 6 call events detected per day (Fig 3: ap1). Correct detections of lapwings referred to birds vocalizing at a great distance to the microphone, resulting in faint signals. The behavior of V. chilensis began to change on 18 April when the water level fell to c. 30 cm and the species started to fly around recording station PPA001 (vcA). Thereafter lapwings patrolled the site regularly (Fig 1A and 1B, S1 and S2 Audios). Between 26 April and 20 May the water level dropped to 9 cm and acoustic activity stabilized around 110 ± 35 detections per day (ap2), with a slight decrease perceptible around mid-May (vcB). Subsequently, V. chilensis call frequency rose quickly from 21 May onwards (begin ap3) before it stabilized a few days later at a level of 467 ± 104 sound events per 24 h (end ap3 and ap4). Between 18 May and 12 June, we repeatedly noticed distinct short series of agitated alarm and cackling calls, indicative of adults at the nest or close to it (Fig 1C and 1D, S3–S7 Audios). During these weeks the floodwaters retreated completely (cf. S1A–S1C Fig). A sudden and sustained increase in V. chilensis activity was noted on 13 June and, based on assumption A3, interpreted as the earliest day of hatching (vcD). Thereafter daily detection rates fluctuated at 766 ± 133 hits for about two weeks (ap5) with a maximum value of 1117 call events reached on 21 June. Most detections in that period referred to extensive alarm call series of adults mobbing potential predators and food competitors in defense of their chicks (Fig 1E and 1F, S8 and S9 Audios). On 29 June vocal activity fell again to about 489 ± 84 detections per day (end ap5 and begin ap6; vcE). A pronounced drop in daily activity occurred on 9 July (end vcE). During the subsequent seven weeks, vocal activity slowly decreased further, with averages for the entire period varying at 220 ± 50 events per day (ap6 and ap7). Finally, in September V. chilensis acoustic activity was almost as low as in early April, with daily detections of 43 ± 18 (ap8 and vcG).

During the three lunar phases from June to August, V. chilensis nocturnal acoustic activity was higher in moonlight nights when compared with nights around the new moon (Fig 3; ANOVA, F_2,36 = 22.9 to F_2,36 = 116.9, p < 0.001). However, a post hoc Tukey HSD showed that at the end of the breeding period differences in nocturnal activity means were not significant for the new moon and full moon phases in August, possibly because of the recording gap in the night from 17 to 18 August.

Diel activity patterns

Diurnal vocal activity was analyzed in two time resolutions, showing that per-hour sums of detections (Fig 4A) mask many interesting details discernible in per-quarter-hour views (Fig 4B–4D). Generally speaking, V. chilensis acoustic activity peaked around dawn and dusk. However in the per-quarter-hour projections, but not in the per-hour view, it is perceptible that vocal activity varied in time even between the activity peaks at dawn and dusk; that is, activity tended to peak periodically every 45 to 120 min.

The most striking difference in diurnal activity patterns between months is the high level of daytime activity in June, when detection rates fluctuated around 45% of maximum activity even during midday heat. Furthermore, in the afternoon a sustained increase of activity occurred as early as 2.5 h before sunset. According to assumption A4, we regard an all-day-active pattern as indicative of lapwings tending chicks. The situation was similar but less pronounced in July, when the proportion of call events decreased to 30% of maximum activity in the late morning hours and to 25% in the early afternoon. Likewise, the strong increase of afternoon activity began about 45 min later than in June. By contrast, in August V. chilensis daytime activity dropped to less than 5% between 10:00 and 15:30 h and a strong increase in afternoon activity set in only about 1.5 h before sunset.

How well do the recognizer results reflect the real vocal activity patterns of Vanellus chilensis?

With the help of two validation datasets, we assessed the effects of five potential sources of bias, which may have had some modulation effect on the determination of time-varying V. chilensis call rates. Here we explain why we think that the identified biases had only a moderate effect on the final results. (1) Time-varying false positive rate (Table 3).—The general vocal activity pattern was not affected by the post-hoc correction of recognizer results, despite the pronounced fluctuation in monthly false positive rates (cf. Fig 3 and S4 Fig). (2) High false negative rate (S1 Appendix, Fig 5A and 5B).—For the amplitude range of 0 to –50 dB, the monthly false negative rates showed only little variation, particularly for the months of highest activity (May-Jul.). Even in the range of 0 to –40 dB, where time variation was evident and most target signals were detected, miss rates were much lower for months with low V. chilensis activity, April (22.7%) and September (27.9%), than for the high-activity months May-July (50.8–63.6%). August (40.8%) was intermediate. (3) Varying recognizer sensitivity (S3 Table).—The signal amplitude of the weakest target event detected in soundscapes was rather stable. Moreover, in May and June call events had to be stronger than in other months in order to trigger detections. (4) Unexpected detection of very weak target signals.—In two soundscapes in the validation library, the recognizer detected very weak V. chilensis signals in the presence of loud and monotonous cicada songs (S3 Table). Similarly, in the random sample we found eight ambiguous cases where low-amplitude target events were detected in combination with stronger signals from competing bird and frog species (S2 Table). Less ambiguous cases of combined detections were either counted as correct hits or false positives (section Validation dataset and assessment procedures). The most probable reason for such combined detections is the training library, which despite of manual filtering still contained some faint audio signals of competing species. However, considering the small number of combined detections found in the random sample, it is unlikely that they had a strong effect on the final detection results. (5) Repeated detection of different segments of a single call event (S3 Table).—The double-hit rate was almost constant, making it improbable that repeated detection of single call events had a strong effect on detection results. We interpret the lower rate in August as an outlier due to small sample size.

In summary, all factors analyzed here indicate that in the high-activity period (May-Jul.) the number of detected target events was rather underestimated than overestimated when compared with the low-activity months (Apr., Aug., and Sep.). Thus the recognizer results provide a good approximation of the general V. chilensis vocal activity patterns, although we did not investigate whether this also holds true for short-term changes like day-by-day variability (cf. Fig 3). Most importantly for the present work, the corrected detections fit the acoustic activity patterns expected to be found during the main breeding season rather well (cf. assumptions A1-A4 and section Duration of the breeding cycle). The other data discussed in the subsequent sections provide additional support for our interpretations, namely water level (Fig 3), periodic occurrence of distinct V. chilensis vocalization types (S1–S9 Audios), seasonal abundance shifts in (sound-producing) insect species (S10–S12 Audios), and published references on the reproductive strategies of other bird species.

Vanellus chilensis life-cycle phases

In the Venezuelan Llanos, a seasonally-flooded savannah very similar to the northern Pantanal, the main breeding period falls in the early rainy season, April-July ([46]: n = 17 nests). Likewise in Central Brazil’s Planalto region V. chilensis breeds in the late dry and early rainy season, July-December ([51]: n = 98 breeding attempts; [50]: n = 74 nests). By contrast, Amazonian populations of charadriids migrate in response to the seasonal inundation of their floodplain habitats and initiate breeding as soon as the waters recede ([54]: p. 222). Our data suggest that in the northern Pantanal V. chilensis shows a seasonality pattern similar to Amazonian populations of charadriids, probably because flooding is extensive in our study area (S1 and S2 Figs). As predicted by assumption A2, we demonstrate that for the six-month period analyzed here, calling frequency was lowest in the first half of April 2013 when the water level still was >30 cm (Fig 3). However, at the end of the flood cycle the monthly number of V. chilensis detections at PPA001 sharply increased (cf. assumption A3). The maximum of 18 151 raw detections in June was about six times higher than the nine-months average value of 3000 ± 900 of the low activity period (cf. assumptions A1 and A3; Table 1). Based on the manually inspected random sample of detections we exclude the possibility that the increase was due to the presence of nonbreeding flocks (cf. S1–S9 Audios; details given in the next sections). In June and July we also noted high levels of parental care behavior and dial vocal activity patterns showed the characteristic all-day active mode (Fig 4B and 4C), predicted by assumption A4 for lapwings tending young [47,48]. In August behavioral patterns already had reverted to nonbreeding mode (Fig 4D).

Post-breeding behavior.

After the breeding season the formation of flocks allows adults and juveniles alike to optimize their time budgets and to save energy [47,48]. During morning hours lapwings forage singly or in small groups, whereas during the midday heat group size increases and less demanding activities, like preening, loafing, and roosting prevail [55]. Finally, in the afternoon the decrease in group size is accompanied by an increase in general activity. Overall, this behavioral pattern explains the nonbreeding mode of diurnal V. chilensis vocal activity in our study area very well (Fig 4D).

One striking finding was that V. chilensis had already abandoned the area of PPA001 less than two months after the floodwaters receded (S1A–S1F Fig; Fig 3: vcG), most likely because of a food shortage (cf. [46]: p. 508]). Lapwings feed on small invertebrates, such as earthworms, mollusks, spiders, woodlice, hemipterans, ground beetles, weevils, grasshoppers, and hymenopterans, as well as larvae of lepidopterans, coleopterans, and dipterans, using a run-and-pause technique that includes quick pecks and shallow probing on land [44,47,66,76–78] and sometimes foot-trembling in shallow water [45,54]. During the dry season, grassland-inhabiting invertebrates become less accessible to the birds. Declines in invertebrate populations have been described for several tropical ecosystems with a pronounced seasonal climate, such as savannahs [79] and Cerrado, e.g. earthworms migrate to deeper soil layers, enter a diapause, or are completely absent [80,81] and snails go into estivation [82]. Certain arthropods are sensitive to drought as well, including some of the groups known to be crucial food resources for lapwings [83,84]. Sound-producing insects, like orthopterans and cicadas, are conspicuously evident in our wet season soundscapes (Fig 6A and 6B, S10 and S11 Audios). By contrast, dry season recordings are usually very clean in the entire frequency range between 0–24 kHz, indicating a possible absence of most sound-emitting insect species or at least of their imago stages (Fig 6C, S12 Audio).

Download:

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

Fig 6. Spectrograms of exemplary midnight soundscapes.

All spectrograms represent unfiltered 60-s segments of WAV soundscapes passively recorded at station PPA001 in the Fazenda Pouso Alegre area (minutes 5:30–6:30 of the original 14-min soundscapes; windowing function: Blackman-Harris; resolution: 512 bands; decibel range: 85 dB; amplification = +10 dB; cf. S10–S12 Audios). A: at the height of the rainy season several species of sound-producing insects, anurans, birds, and mammals (mostly cattle) fill the entire frequency range 0–22 kHz with audio signals; 25 Jan. 2013, 23:49 h (S10 Audio). B: at the end of the rainy season when V. chilensis reestablished its territories, sound-producing insects and anurans are still present in the frequency range 0–10.5 kHz; in addition, some of the signals below 4 kHz were emitted by birds and Zebus Bos indicus; 25 Apr. 2013, 23:54 h (S11 Audio). C: at the height of the dry season sound-producing insects and anurans are virtually absent in most soundscapes. The white arrows mark the faint signals of a Feral Horse Equus ferus caballus (left) and a Zebu (right); 25 Jul. 2013, 23:51 h (S12 Audio).

https://doi.org/10.1371/journal.pone.0169041.g006

Comparison with published research

Previous studies based on automated recognition have presented diurnal and seasonal acoustic activity patterns of insects [18], anurans [17,18,43], birds [18,19], and mammals [18,42]. However, with exception of Frommolt and Tauchert [19], who presented quarter-hour sums for the nocturnal booming activity of Eurasian Bittern Botaurus stellaris, the time resolution of comparable studies has been low, i.e. diel activity was presented as cumulated hourly detection rates, which may mask interesting details of the calling activity (Fig 4A–4D). Similarly, most studies on seasonal changes in acoustic activity have only presented monthly detection frequencies [17,18,42,43], which conceal short-term activity changes. Therefore none of the latter studies has linked calling activity levels to multiple life-cycle phases of the studied animals. Here we explore this untapped potential of automated acoustic monitoring. The data obtained by means of automated sound event recognition permit the characterization of V. chilensis diurnal and seasonal activity patterns in unprecedented detail. Most strikingly, on the basis of pronounced shifts in the intensity of adult acoustic activity, it seems feasible to determine the start and end of each V. chilensis life cycle phase with presumed error margins of only a few days. We anticipate that passive sound recording combined with field observations will make interpretations of automatically detected acoustic activity even more precise.

Follow-up research

Promising research directions for improving the overall recognition accuracy on field recordings are audio parameterization methods that make use of image processing techniques applied on the spectrogram seen as an image [18,34,89–95]. The consideration of domain knowledge about the structure of the species-specific sound events and the acoustic model of the operational environment in the classification step are other options to improve recognizer performance [36]. Finally, rule- and grammar-based methods for post-processing the output of the log-likelihood ratio estimator could allow calibrating the detector specifically for distinct V. chilensis vocalization types represented in the training library. However at present the availability of training, background, and validation libraries is the principal bottleneck in the development of species-specific and vocalization-specific acoustic recognizers. The compilation of comprehensive high-quality sound libraries is demanding. Only experienced bioacousticians, familiar with the acoustic vocabulary of the target and other sound-producing species, are qualified to build adequate datasets.