ABSTRACT Nocturnal burrow-nesting seabirds breeding on isolated oceanic islands pose challenges to conventional monitoring techniques, resulting in their frequent exclusion from population studies. These seabirds have been devastated by nonnative predator introductions on islands worldwide. After predators are eradicated, recovery has been poorly quantified, but evidence suggests some nocturnal seabird populations have been slow to return. We evaluated the use of automated acoustic recorders and call-recognition software to investigate nocturnal seabird recovery after removal of introduced Arctic foxes (Alopex lagopus) in the Aleutian Archipelago, Alaska. We compared relative seabird abundance among islands by examining levels of vocal activity. We deployed acoustic recorders on Nizki-Alaid, Amatignak, and Little Sitkin islands that had foxes removed in 1975, 1991, and 2000, respectively, and on Buldir, a predator-free seabird colony. Despite frequent gales, only 2.9% of 2230 recording hours from May to August of 2008 and 2009 were unusable due to wind noise. Recording quality and call recognition model success were highest when recording devices were placed at sites offering some wind shelter. We detected greater vocal activity of Fork-tailed (Oceanodroma furcata) and Leach's (O. leucorhoa) storm-petrels and Ancient Murrelets (Synthliboramphus antiquus) on islands with longer time periods since fox eradication. Also, by detecting chick calls in the automated recordings, we confirmed breeding by Ancient Murrelets on an island thought to be abandoned due to fox predation. Acoustic monitoring allowed us to examine the relative abundance of seabirds at remote sites. If a link between vocalizations and population dynamics can be made, acoustic monitoring could be a powerful census method.
Midiendo el estatus y la actividad nocturna de aves marinas utilizando equipo de grabaciones acústicas: aplicaciones para la restauración en islas
Las aves nocturnas de islas oceánicas aisladas, que anidan en cavidades, presentan un reto para el uso de técnicas convencionales de monitoreo y a su efecto son excluidos, con frecuencia, en estudios poblacionales. Estas aves marinas de islas han sido desbastadas por depredadores exóticos a través de todo el mundo. Una vez los depredadores son erradicados, el recobro ha sido pobremente cuantificado, y la evidencia sugiere que el recobro, de algunas aves marinas nocturnas, ha sido lento. Evaluamos el uso de grabadoras acústicas automáticas y el reconocimiento de voces, para investigar el recobro de aves marinas nocturnas una vez fueron removidas Zorras árticas (Alopex lagopus) en el archipiélago de las Aleutianas, Alaska. Comparamos la abundancia relativa de aves marinas entre islas examinando el nivel de la actividad vocal. Colocamos grabadores Nzki-Alaid, Amatignak y Little Sitkin en islas en donde fueron removidas las zorras en el 1975, 1991 y 2000, respectivamente, y en la colonia de Buldir, que estaba libre de depredadores. Pese a frecuentes vientos de galera, solo el 2.9% de 2230 horas de grabación, (tomadas de mayo a agosto de 2008 y 2009), no fueron de utilidad debido al ruido producido por el viento. El éxito del modelo en la calidad de las grabaciones y el reconocimientos de llamadas fueron mayores cuando el equipo de grabación fue colocados en lugares con protección al viento. A mayor periodo de tiempo, desde la erradicación de las zorras, detectamos mayor actividad vocal en Oceanodroma furcata, O. leucorhoa y Synthliboramphus antiquus. Además, detectando las llamadas de pichones, utilizando grabaciones automatizadas, confirmamos la reproducción de Synthliboramphus antiquus en una isla que se creyó fue abandonada por estos, ante la depredación de parte de zorras. El monitoreo acústico nos permitió examinar la abundancia relativa de aves marinas en islas remotas. Si se puede producir un vínculo entre vocalizaciones y dinámica de poblaciones, el monitoreo acústico puede ser un método de muestreo poderoso.
On remote islands, the success of introduced-predator eradication efforts has accelerated in the past few decades. There is now an urgent need for innovative methods to monitor avian population recovery posteradication. The conservation biology literature is replete with studies that describe the devastating effects of introduced predators on island avifaunas, especially seabirds (Towns et al. 2006, Rauzon 2007, Jones et al. 2008). Limited evidence suggests that targeted predator-eradication programs are now benefiting seabirds, yet population recovery after eradication has been poorly studied (Lavers et al. 2010). Several factors have contributed to this lack of research, including the difficulty of reaching oceanic islands and the expense of maintaining a survey team. In addition, surveying hundreds of islands over areas spanning thousands of kilometers is impractical, and island ecosystems are sensitive to human disturbance.
Nowhere are the problems of surveying avian responses to pest management better reflected than in the Aleutian Islands, an 1800 km long archipelago stretching between Alaska and Russia (Byrd et al. 2005). Beginning in the 18th century, Arctic foxes (Alopex lagopus) were introduced to over 450 Aleutian Islands for fur farming (Bailey 1993, Croll et al. 2005). Although the effects were not quantified, foxes eliminated or severely reduced the native avifauna on nearly every island where they were present (Murie 1959, Bailey 1993, Byrd et al. 1994, 2005). Due to their small size and ground-nesting habits, nocturnal burrow-nesting seabird species, such as Fork-tailed Storm-Petrels (Oceanodroma furcata), Leach's Storm-Petrels (O. leucorhoa), and Ancient Murrelets (Synthliboramphus antiquus), were among the most affected by fox predation (Bailey and Kaiser 1993). In the late 1940s, a program to eradicate Arctic foxes and restore seabird populations began and, by the 1990s, an intensive program was underway (Ebbert 2000, Ebbert and Byrd 2002). As of April 2009, foxes had been removed from 34 Aleutian Islands (J. C. Williams, pers. comm.). Research suggests that many avian populations are recovering after fox removal (Byrd et al. 1994, Williams et al. 2003), but the recovery of nocturnal burrow-nesting seabirds has been poorly documented.
Nocturnal burrow-nesting seabirds are difficult to count directly (Keitt 2005). Conventional methods, such as day-time vessel and beach surveys, are unreliable because detection rates of nocturnal seabirds are likely to be low. Methods such as searching for burrows and call-playback are unrealistic across the expansive Aleutian archipelago. When burrows are found, occupancy for most species is low (e.g., 68% for Leach's Storm-Petrels; Sklepkovych and Montevecchi 1989) producing inaccurate estimates of abundance. Also, reaching into burrows to determine occupancy is invasive, often reducing hatching success (Ambagis 2004, Blackmer et al. 2004).
An alternative to physical surveys of seabird presence is the use of acoustic monitoring. Digital recording technology has been used to monitor a diverse array of secretive taxa (e.g., anurans, Peterson and Dorcas 1994; cetaceans, Marques et al. 2009; birds, Swiston and Mennill 2009). The advantages of acoustic monitoring are that devices can be easily deployed and retrieved, requiring only two brief visits to sites, and they can simultaneously record at multiple sites over entire seasons, facilitating spatial and temporal comparisons of activity. Moreover, recent advances in call-recognition algorithms are facilitating the automation of species–specific call identification and analysis of long-term trends in vocal activity from the huge volumes of acoustic data recorded (Peterson and Dorcas 1994).
Acoustic recording is especially advantageous for monitoring nocturnal seabirds, due to their conspicuous night-time vocalizations (Robb et al. 2008). Colonies of nocturnal burrowing seabirds are noisy places, where vocalizations replace visual displays (Brooke 1986). Calls must therefore contain information about species, sex, and individual identity (McKown 2008). Vocal repertoires of many nocturnal species have been categorized and linked to behavior, such as mate attraction, burrow territoriality, signaling suitable nesting habitat for prospectors, and chick begging (Danchin et al. 1998, Robb et al. 2008). Moreover, nocturnal vocalizations have been used as an indicator of general activity (e.g., Ancient Murrelets; Gaston et al. 1988, Jones et al. 1990) and have been suggested to be a useful indicator of seabird status and relative abundance (Whittington et al. 1999, Keitt 2005).
We examined the possibility of using automated acoustic recording and vocal activity as a cost-effective method of examining recovery of populations of nocturnal seabirds following fox eradication in the Aleutian Islands. Specific objectives were to: (1) determine if automated recording devices could record seabird calls in the harsh, windy, wet environment characteristic of the Aleutian Islands, (2) determine if call-recognition software could identify calls of three seabird species in recordings with varying levels of background noise and vocal activity, (3) provide an index of nocturnal seabird abundance based on call activity on four islands and test the hypothesis that variation in call activity can be predicted by time since fox eradication, and (4) consider how these methods can be used to monitor recovery of nocturnal seabird populations.
We used autonomous digital recording devices called Song Meters (Model SM1, Wildlife Acoustics Inc., Concord, MA; firmware version 1.5.0 in 2008, 1.7.0 in 2009). Gain on both left and right channel microphones was set to the default of +42.0 dB with a sensitivity of −35 dBV/pa. Detection of calls depends on background noise levels and call properties, but, under ideal conditions, Song Meters can detect calls upto 50 m away (RTB, pers. obs.). Because most seabird calls are below 7 to 8 kHz, Song Meters were set to a sample rate of 16 kHz in stereo. The available 32 GB of memory and predicted battery life (high capacity 12,000 mAh batteries) of 100 h allowed for ∼32 nights at 3 h/night of recording. Song Meters were programmed to record in 15 min on/off cycles from dusk (00:30; Hawaii-Aleutian Standard Time) to dawn (06:15). This allowed recording throughout the night (for 3 discontinuous hours), with batteries needing to be changed only once every 32 d. Song Meters recorded from 31 May to 4 August 2008 and 2009 on Amatignak and Buldir, 2 June to 2 August 2008 on Little Sitkin, and 31 May to 31 July 2009 on Nizki-Alaid. We were unable to replace batteries during the 2009 season at Nizki-Alaid Islands so Song Meters were programmed to record in 15 min on/off cycles from only 01:30 to 04:30. This allowed recording during the time of peak activity of nocturnal seabirds (as determined from recordings in 2008), with batteries needing to be changed only once every 64 d.
Study area and history of fox removal We placed four Song Meters on each of three islands (Amatignak [51.27°N, 179.10°W], Little Sitkin [51.95°N, 178.54°E], and Nizki-Alaid [52.75°N, 173.94°E]), and one on Buldir Island (52.35°N, 175.92°E; Fig. 1). Foxes were removed from Nizki-Alaid in 1975, Amatignak in 1991, and Little Sitkin in 2000. Nizki and Alaid are a pair of islands joined by a sandbar and thus considered one site (Byrd et al. 1994). Buldir Island was used as a control site because it has large colonies of nocturnal burrow-nesting seabirds of known size and foxes were never introduced (Byrd and Day 1986). All islands had treeless windswept landscapes dominated by a subarctic grass and scrub ecosystem (Byrd et al. 2005). Characteristic weather included steady winds >30 km/h and frequent fog, rain, and wind storms.
Deployment protocol Song Meters were fastened to 1 m wooden stakes and placed 50 to 150 m from shorelines at elevations under 400 m. Song Meters were either placed near the four cardinal points of each island (Fig. 1) or at sites far enough apart to avoid recording birds in the same areas (∼1 km). We attempted to place Song Meters within suitable burrow-nesting seabird habitat, but also near tall grassy slopes, potentially providing shelter from wind and wave noise (see later).
Recording quality Recording files (.wav) were reviewed using Song Scope 2.3 (Wildlife Acoustics Inc. 2009). Recording quality was evaluated and categorized based on visual scans of spectrograms. We scored each night on a scale from 1 to 5 based on the amount of continuous broad band ‘white noise’ or wind noise that obstructed spectrograms: (1) no background noise, (2) limited noise that obscured low frequencies (0.1–2.0 kHz), (3) 1 to 5% of recording time obscured by bouts of noise (0.1–8.0 kHz), (4) 5 to 50% of recording completely obscured by noise, and (5) >50% of recording obscured by noise and considered unusable. To assess Song Meter performance relative to wind velocity, category assignments were compared with wind direction and wind speed (m/s) data from the National Data Buoy Center (NDBC–NOAA, Stennis Space Center, MS). Recording quality at Amatignak and Little Sitkin was compared with data from weather buoy 46071 (51.16°N, 179.00°E) located 130 km southwest and 88 km south of each island, respectively. Recording quality at Nizki-Alaid was compared to data from weather buoy 46070 (55.00°N, 175.28°E) located 266 km northwest of the island. We compared recording quality among sites and islands.
To determine if nearby hills provided wind shelter and produced clearer recordings, we examined the relationship between recording quality and the number of nearby grassy slopes. Nearby grassy slopes were defined as hills covered in Leymus sp. grass within 5 m of a Song Meter and at least 10 m higher than the Song Meter. We separated the presence of grassy slopes into four categories: (1) no nearby grassy slopes, (2) one adjacent grassy slope away from the shore, (3) one grassy slope separating the Song Meter from the shore, and (4) one adjacent grassy slope away from the shore and one slope separating the Song Meter from the shore. No other combination of grassy slopes around Song Meters was observed.
Song Scope recognition models Using Song Scope, we built call-recognition models to automatically search recordings for the most numerous vocalizations, including Leach's Storm-Petrel chuckle calls (Taoka et al. 1988), Fork-tailed Storm-Petrel flight calls (Simons 1981), and Ancient Murrelet chirrup calls (Jones et al. 1989). Call-recognition models for each vocalization were generated using a three-step process. First, to build a call-recognition model, Song Scope requires “training data” or an assortment of calls to “train” a model to identify specific calls of interest within recordings (Wildlife Acoustics Inc. 2009). Ideally, enough reference calls are used as training data to capture natural variation in call structure. We built a “basic recognition model” for each call type, using prerecorded calls, to search field recordings for more calls to use as training data in more complex recognition models (see step 2 later). Basic models were built using 2–5 high-quality prerecorded reference calls as training data. High-quality storm-petrel reference calls were obtained from Buldir and Egg islands, Alaska, in 2006 (Seneviratne et al. 2009) and Ancient Murrelet reference calls were obtained at McPherson Point, Langara Island, Canada, in 2006 (Major and Jones 2011). High-quality reference calls included either at least three different individuals over 10 min (Ancient Murrelet chirrup) or many individuals over a maximum of 30 min of recording (storm-petrel calls).
Next, using the basic recognition models (from step 1 above), recordings from Amatignak, Nizki-Alaid, and Little Sitkin were scanned for loud, clear calls that were selected and saved. After ∼12 calls (from different individuals and with minimal background noise) were saved, recognition models were revised by incorporating these new calls along with the original reference high-quality training calls (used in step one). Basic models created in step one were then discarded.
Finally, optimal model parameters were determined for each comprehensive recognition model generated in step 2. Model parameters in Song Scope included minimum frequency, frequency range, and sample rate (Hz) based on the properties of each species' call, Fast Fourier Transform window size (frequency bins or samples), dynamic range (i.e., how much signal energy is used to compare call components; db relative to peak call signal db), maximum Hidden Markov Model (HMM) states, and HMM feature vector size (Supplementary Table S1; Agranat 2009). Maximum syllable duration, maximum syllable interval, and maximum song duration (s) specific to each species' call were also included in each model. To set parameters appropriately, a random group of field recordings was selected and reviewed using Song Scope's relative sound pressure scale (dB) with a signal detection algorithm (visual representation of sound pressure; Wildlife Acoustics Inc. 2009). Calls of varying quality were identified and parameters adjusted so the strongest signal (highest dBs) was visible. In this way, we attempted to minimize the level of background noise and interference by other species, and maximize average call energy.
Once generated, Song Scope recognition models scanned all 845 nights of recordings (with 15 min periods batched together) to identify the calls of each species. Identified calls were checked visually by a human observer to remove false positives (background noise or nontarget calls incorrectly identified as a call). After scanning recordings, Song Scope generated a spreadsheet with the time each sound was identified by the model and linked each time to a spectrogram displaying the identified sound. In this way, we could quickly review calls identified by the recognition model and remove incorrectly identified calls. Therefore, false positives were not considered a relevant measure of model success. Recordings were also checked for false negatives, or targeted calls not identified by the Song Scope recognizers. False-negative calls were identified by visually scanning spectrograms. Calls missed by recognizers with all frequency components fully visible were noted as false negatives, whereas low volume and distant calls (barely visible on the spectrograms and syllables or harmonics not visible) were not noted. A false-negative rate was then calculated by dividing the number of target calls missed by Song Scope recognition models (false negatives) by the sum of false negatives and correctly identified target calls (total calls). To determine the circumstances under which recognition models were most successful, false-negative rates were compared with recording quality, species, sites, wind shelter (number of grassy slopes), and mean nightly call activity (number of calls/night).
Recognition models could not be used for recordings from Buldir Island or for Fork-tailed Storm-Petrel flight calls at the western site on Amatignak due to abundant overlapping calls of all species. To determine an average nightly call rate at these sites, we counted calls on 16 random nights manually by visually scanning spectrograms. Fork-tailed Storm-Petrel flight calls and Leach's Storm-Petrel chuckle calls were recorded constantly during peak nocturnal hours on Buldir, with few or no periods of silence. These numerous calls were counted by subtracting periods of silence from a ‘constant call rate.’ Both calls last for ∼1 s, resulting in a constant call rate of 900 (60 s × 15 min for each 15 min recording period). Intermittent seconds with no calling were subtracted from this total of 900 to give the number of calls per 15 min period. Also, recognition models were not used to count less common vocalizations. Instead, we counted Leach's Storm-Petrel purr and screech calls (Taoka et al. 1988), Fork-tailed Storm-Petrel three-syllable male calls (Simons 1981), and Ancient Murrelet chick calls and songs (Jones et al. 1989) by visually scanning spectrograms.
Minimum recording nights To determine the number of nights of recording needed to capture highly variable nightly call activity (i.e., minimum recording nights in a season), we graphed cumulative means for the most common call types (Leach's and Fork-tailed storm-petrel chuckle and flight calls and Ancient Murrelet chirrup calls) at each site on Amatignak Island, the only island with 2 yr of data. We looked for an asymptote, when cumulative mean calls per night approached a stabilizing point. We used a bootstrapping method, randomly selecting, with replacement, two nights (from a total of 92, 116, 95, and 106 nights, respectively, at each site on Amatignak), and taking an average number of calls on these nights. We repeated this process 200 times, because 200 nights or ∼ 7 months, is the approximate number of nights it would take to fill two 32GB flash cards recording at 15 min on/off intervals during peak nocturnal activity (01:30–04:30). We summed the cumulative means then observed the number of repeats or nights it took for random means to stabilize around an asymptote. This would provide crucial information for future study design because it indicates the deployment duration required to capture a more accurate mean call activity level for nocturnal seabirds.
Identifying and indexing species To characterize seabird activity on each island, we used a hierarchical classification scheme based on the frequency of each call type identified by the recognizers and noted during visual scans of recordings. We first determined if each call was present or absent. We then calculated the percentage of nights each call was noted during the total recording period (detection rate), the total number of calls in all recordings, and the mean number of calls per night. We used mean calls per night as an index of activity and relative abundance, and used the three most common calls (chuckle, flight, and chirrup) to compare seabird activity with number of years since fox eradication. Call rate does not give an accurate count of individuals or population (Gaston et al. 1988), but acts as an index of activity (Jones et al. 1990) and relative abundance (Keitt 2005). We assumed large numbers of calls were associated with large numbers of individuals and vice versa (Gaston et al. 1988).
We also considered context associated with each call type recorded as an indicator of breeding status. Leach's Storm-Petrel chuckle, Fork-tailed Storm-Petrel flight, and Ancient Murrelet chirrup calls are given in a wide variety of contexts by adult or prospecting birds (Simons 1981, Taoka et al. 1988, Jones et al. 1989). Leach's Storm-Petrel purr and screech calls, Fork-tailed Storm-Petrel flight calls, and Ancient Murrelet songs are given from in or near burrows and are associated with mate advertizing, burrow defense, and courtship (Simons 1981, Jones et al. 1989, Huntington et al. 1996). Finally, calls of Ancient Murrelet chicks are associated with family departures from nest sites and confirm breeding (Jones et al. 1987).
Statistical analysis The effect of wind speed, wind direction, and wind shelter on recording quality score was compared using an ordinal logistic regression with wind speed, wind direction, their interaction, grassy slope category, and recording sites as explanatory variables in R 2.11.1 (Design library; R Development Core Team 2010). False-negative rates were compared among recording qualities, species, recording sites, grassy slope category, and with call activity using a binomial GzLM with a log link. False negative calls, scored as a 1, and correctly identified calls, scored as a 0, formed the response variable. We compared the mean number of chuckle, flight, and chirrup calls per site relative to years since predator eradication using a poisson generalized linear mixed model (GzLMM), using site as a random variable, in R 2.11.1 (lme4 library). Values are presented as means ± SE.
We obtained 876 h of recordings on 292 nights (459 h from Amatignak Island, 240 h from Little Sitkin Island, and 177 h from Buldir) in 2008 and 1354 h on 553 nights in 2009 (768 h from Amatignak Island, 427 h from Nizki-Alaid, and 159 h from Buldir).
Recording quality Recording quality differed among recording sites on different islands (all P < 0.043), with a mean overall recording quality of 2.2 ± 0.1 (range = 1.7 ± 0.2–2.5 ± 0.2; Supplementary Table S2). Recordings were considered unusable on only one of 292 nights in 2008 due to wind noise. In 2009, 24 of 553 nights were unusable due to wind noise. Across both years, 2.9% of nights were unusable.
Higher wind speeds had a negative impact on recording quality on Amatignak and Nizki-Alaid islands (G1= 20.0, P < 0.001; Fig. 2). Wind direction, the interaction between wind speed and direction, and year had no effect on recording quality (all P≥ 0.25). Recording quality on Little Sitkin was affected by an interaction between wind speed and wind direction (P < 0.001). Recordings were clearer at sites with more wind shelter (Z3=−2.4, P= 0.02); sites with one grassy slope away from the shore and one slope separating the Song Meter from the shore had better recording quality (2.3 ± 0.1) than sites with no grassy slopes (1.6 ± 0.1; Supplementary Table S2).
Recognition model quality Recognition models were generated for Leach's Storm-Petrel chuckle calls, Fork-tailed Storm-Petrel flight calls, and Ancient Murrelet chirrup calls (Supplementary Table S1). Overall, recognition model success, or the false negative rate of a recognizer, was affected by site, wind shelter category, recording quality, species (all P < 0.008; Supplementary Table S2), and call activity (Z1=−16.2, P < 0.001; Fig. 4). There were no significant interactions between variables (P > 0.099), except between call activity and site (Z1=−3.6, P < 0.001).
False negative rates averaged 0.29 ± 0.01 in clear recordings and 0.54 ± 0.02 in “unusable” recordings (Supplementary Table S2). In other words, calls went from being identified an average of 71% of the time to 46% of the time as recording quality declined (Fig. 3). Leach's Storm-Petrel chuckle call, Fork-tailed Storm-Petrel flight call, and Ancient Murrelet chirrup call recognition models identified an average of 67 ± 0.3%, 69 ± 0.5%, and 56 ± 0.6% of the calls in the recordings, respectively (Supplementary Table S2). False negative rates had a negative relationship with call activity (Z1=−14.7, P < 0.001). On nights with fewer calls, fewer calls were identified by recognizers (Fig. 4). However, recognition models were unusable at locations with numerous overlapping calls of different species (e.g., Buldir). At these sites, models identified <10% of calls. Thus, false negative rates were lowest on nights with a moderate level of call activity.
Minimum recording period The minimum number of nights of recording needed to account for nightly variation in calls per night on Amatignak Island was both species and site specific (Fig. 5). The number of nights needed to reach an asymptote in number of calls recorded per night was dependent on level of call activity. For Leach's Storm-Petrel chuckle calls, sites with moderate levels of call activity (north site: 80.4 ± 14.6; east site: 97.1 ± 9.2 calls/night) did not reach an asymptote after 200 nights, whereas sites with low or high levels of call activity (south site: 6.7 ± 1.3; west site: 187.2 ± 23.9) reached asymptotes after 95 and 75 nights, respectively (Fig. 5). Thus, capturing nightly variation in Leach's Storm-petrel chuckle call activity would require recording for ∼3 months at sites with low calling rates, 2.5 months at sites with high calling rates, and more than 6 months at sites with moderate calling rates. For Fork-tailed Storm-Petrel flight calls, only the north site, with moderate calling rates (4.6 ± 0.7), reached an asymptote before 200 nights (100 nights, or ∼3.3 months). We were unable to measure call rates at the western site due to a high number of overlapping calls. Finally, for Ancient Murrelet chirrup calls, sites with moderate to high levels of call activity (north site: 7.4 ± 1.7; east site: 69.3 ± 7.6) did not reach an asymptote after 200 nights, whereas sites with very low levels of call activity (south site: 0.3 ± 0.1; west site: 4.9 ± 0.8) reached an asymptote after 95 and 75 nights, respectively.
Species indexed and identified Leach's Storm-Petrel chuckle calls were recorded at all 13 sites, and Fork-tailed Storm-Petrel flight calls were recorded at just one site on Nizki-Alaid (Fig. 6). Ancient Murrelet chick calls were the least frequently recorded calls (Fig. 6; Supplementary Table S3).
We found significant differences between Leach's Storm-Petrel chuckle (Z1=−3.1, P= 0.002), Fork-tailed Storm-Petrel flight (Z1=−3.5, P < 0.001), and Ancient Murrelet chirrup (Z1=−2.1, P= 0.035) mean call activity at islands with differing time periods since fox eradication. Buldir Island, a predator-free colony, had the highest call activity (Fig. 6), where all three species most common call types were recorded frequently (range = 86.4–2727.8 mean calls/night) almost every night (92–100% of nights; Supplementary Table S3). For islands formerly inhabited by foxes, call activity was highest on Amatignak (19 years since fox eradication) and lowest on Little Sitkin (9 years since fox eradication; Fig. 6). Forty-nine times more Leach's Storm-Petrel flight calls and 73 times more Fork-tailed Storm-Petrel flight calls (not including call activity at the very busy site to the west of Amatignak) were recorded on Amatignak than Little Sitkin. In addition, no Ancient Murrelet calls or Leach's Storm-petrel mate advertising calls (purr) or territorial calls (screech) were recorded on Little Sitkin (Fig. 6).
Ancient Murrelet chick departures, with adults and chicks calling sequentially, were recorded on 25 nights in 2008 and 51 nights in 2009 at a site on Amatignak Island. These calls were not recorded at any other island formerly inhabited by foxes (Fig. 6).
We found that recording devices (Song Meters) successfully recorded and call-recognition software (Song Scope) identified seabird vocalizations in the windy environment characteristic of the Aleutian Islands. Moreover, our results demonstrate the utility of using acoustic recorders and recognition software for providing data on the presence, vocal activity levels, and relative abundance of nocturnal seabirds on islands after predator eradication.
Wind noise is one of the most significant challenges to making field recordings (Agranat 2009) and our sites experienced wind speeds >12 m/s during 16% of recording nights. Nevertheless, only 2.9% of the total recording time was unusable due to wind-generated noise, indicating that devices function usefully on windy islands used by burrow-nesting seabirds.
We found that sites with the best recording quality were separated from the shore by a grassy slope and adjacent to one or more abrupt grassy slopes, presumably providing shelter from wave and wind noise. In the Aleutians, burrow-nesting storm-petrels and murrelets generally breed close to shorelines on exposed headlands (Byrd and Day 1986, Byrd et al. 2005). To maximize recording quality while still capturing all potential seabird calls, device placement must balance suitable wind and wave noise shelter and proximity to suitable nesting habitat.
We found that call recognition software identified over 50% of the calls of target species. Recognition models were most useful at sites with moderately levels of call activity. We were unable to use models at locations with numerous overlapping calls of different species (i.e., Buldir) because the software was unable to discern individual calls. Software and algorithms that can effectively identify calls from recordings with high call overlap are needed before recognition models can be used at active colonies (Brandes 2008). However, for island restoration, biologists would likely be most interested in locations with moderate or low call rates of species of interest. Call recognition software also performed poorly at sites with few calls per night in our study. We did not use recognition models to count rare calls (<300 calls/60 device nights) because identifying false positives and negatives was more time consuming than identifying calls by manually scanning spectrograms. Similarly, Swiston and Mennill (2009) concluded that, although automated scanning provides a fast alternative, scanning spectrograms of recordings manually is better for studies of birds with low vocalization rates.
Call-recognition models were negatively affected by recording quality. For recognition models to have the greatest success, the number of unobscured recordings should be maximized, further stressing the importance of appropriate Song Meter placement with adequate wind shelter.
The most successful recognizer was the Fork-tailed Storm-Petrel flight call model that identified over 69% of flight calls in recordings. The flight call is a distinctive loud broadband screech, with frequencies ranging from low-frequency fundamentals (1 kHz) to high-frequency harmonics (8 kHz) and often overpowering background noise or calls of other species (Simons 1981). Song Scope software operates by recognizing both the shape and energy of sounds (Agranat 2009). If structural aspects of a vocalization are lost in background noise, calls become more difficult to recognize. Therefore, the high energy and broad frequency range of the flight call make it an excellent candidate for recognition models.
Nocturnal seabird activity is extremely variable on a nightly basis (from zero Leach's Storm-Petrel chuckle calls to over 500 within the same week) due to moon phase (Watanuki 1986) and weather variables (Gaston and Nettleship 1982). We concluded that, to adequately capture variation and arrive at a stable mean in nocturnal seabird activity, recording devices would be deployed ideally for a minimum of 75 nights, or one summer season. However, each site and species took a different number of nights to reach an asymptote at the sample mean, suggesting that the minimum number of recording nights depends on call type and activity level.
Prior to our study, little was known about the status of populations of nocturnal seabirds on Amatignak, Little Sitkin, and Nizki-Alaid islands (Byrd et al. 2005). Using acoustic recordings, we were able to detect species of birds present and use the known context of each call type to infer the behavior of birds at each site (Brooke 1986, Taoka et al. 1988). Although our recordings could not be used to determine the number of birds present, our results indicate that seabird populations on the islands in our study are recovering and, in some cases, breeding.
Buldir Island, where foxes had never been introduced, had the highest calling rates in our study. This was expected because Buldir has an active nocturnal seabird colony with large known breeding populations (Byrd and Day 1986, Byrd et al. 2005). Amatignak Island, where foxes were removed 9 years earlier than at Little Sitkin Island, had a higher rate of all species different call types, suggesting a more advanced stage of nocturnal seabird recovery. However, when Nizki-Alaid was included, we found no clear relationship between calling rates and years since fox eradication possibly due to factors such as differences in island size, proximity to source populations, and variation in marine habitat (Stephenson and Irons 2003, Buxton 2010). Factors affecting nocturnal seabird recovery after predator eradication are currently unknown (Lavers et al. 2010), and additional research is needed.
Byrd et al. (2005) examined seabird populations across the Aleutian Islands by boat and beach surveys and no breeding populations of storm-petrels or Ancient Murrelets were detected on Amatignak Island. However, we detected chick departure calls of Ancient Murrelets, confirming the presence of a small breeding population on Amatignak Island. We also detected calls of Leach's and Fork-tailed storm-petrels that are associated with attracting mates, courtship, and defense of nest burrows (Simons 1981, Taoka et al. 1988, Huntington et al. 1996). Although not confirming breeding, these calls indicate that birds were advertising for potential mates and defending burrows. At the site with the highest rate of all types of storm-petrel vocalizations on Amatignak Island, we also found birds in burrows with brood patches (RTB, pers. obs.). Thus, recording devices identified small breeding populations of nocturnal seabirds overlooked by conventional monitoring techniques.
Although automated acoustic recording can serve as a powerful tool to examine relative abundances of nocturnal seabirds among islands, further work is needed to determine if call activity can be used to infer the size of seabird populations. Previous studies have confirmed the relationship between vocalization rates and numbers of cetaceans, anurans, and migrating songbirds (Larkin et al. 2002, Farnsworth et al. 2004, Nelson and Graves 2004, Barlow and Taylor 2005). However, more research is needed to relate call rates to independent measures of seabird numbers (e.g., Haselmayer and Quinn 2000) and to determine what proportion of birds are detectable (e.g., Dawson and Efford 2009). For rare species, numbers of seabirds could potentially be calculated using individual variation in vocalizations (e.g., Gilbert et al. 1994).
Methods of monitoring nocturnal seabirds include beach surveys, burrow grubbing, and playback. However, island archipelagos may consist of hundreds of islands with suitable seabird habitat, and surveying nocturnal seabirds using conventional techniques may be impractical, if not impossible. Eradication of foxes has provided an opportunity to examine patterns of recovery of nocturnal seabird populations, knowledge essential to future island conservation strategies. Yet recovery remains unstudied due to a lack of affordable and consistent surveying techniques for nocturnal burrow-nesting seabirds. Our results indicate that automated recording devices are a robust tool and, if call activity can be linked to seabird numbers, could serve be used to census and monitor nocturnal seabird recovery on remote islands.
We thank I. Agranat, CEO of Wildlife Acoustics, for hours of assistance with his software, S. Seneviratne and H. Major for use of their high-quality field recordings, T. Danelesko, H. Major, C. Curry, J. Williams, L. Kenney, R. Kaler, and A. Chateauxvert for help placing Song Meters in the field, E. H. Miller, M. Rayner, M. McKown, G. Ritchison, and many anonymous reviewers for useful suggestions, D. Schneider and A. Bond for help with statistics, and J. Williams and the crew of the M/V Tila for getting us to the remote Aleutian field sites in one piece. Financial support was kindly provided through a challenge cost share grant with the U.S. Fish and Wildlife Service, travel grants through the Northern Scientific Training Program, and student research awards through the National Science and Engineering Research Council.