Comparing imaging, acoustics, and radar to monitor Leach’s storm-petrel colonies

Introduction

Globally seabird populations are declining; however this conclusion is largely driven by data on well monitored populations, the majority of which are surface nesting seabird species (Paleczny et al., 2015). Surface nesting seabirds are relatively easy to monitor at their breeding colonies because active nests can be identified and counted (e.g.,  Gaston, 2003). In contrast, much less is known about population trends of burrow nesting seabirds. Although species with this nesting behavior make up ∼50% of threatened seabird species (Spatz et al., 2014), these seabirds are much more cryptic at breeding sites, making surveys difficult (e.g.,  Piatt, Roberts & Hatch, 1990; Renner et al., 2006). Even in large colonies, studies are restricted to locations where nest cavities can be repeatedly accessed or where mark-recapture studies are feasible (e.g.,  Sheffield et al., 2006). In some cases, the mere presence of humans can damage nesting habitat or assist predators in locating nests (Rodway, Montevecchi & Chardine, 1996; Blackmer, Ackerman & Nevitt, 2004; Pollard, 2008). Additionally, there are often safety, logistic, and financial limitations to accessing remote seabird colonies. Therefore, alternative monitoring methods are needed. Most seabirds have k-selected life-history strategies (i.e., long-lived, few offspring per breeding attempt), and will skip breeding when conditions are poor, which means population monitoring programs need to be sustainable over decades to be viable. Thus, monitoring methods should be easily repeatable, statistically rigorous, and low cost. Additionally, methods need to be sufficiently robust to detect short-term changes in order to inform managers of population fluctuations prior to substantial declines. This need is desperate, as limited evidence suggests that burrow nesting seabird populations are declining globally, in part due to threats accrued on land (e.g., introduced predators, habitat degradation) (Croxall et al., 2012; Spatz et al., 2014). A lack of monitoring of this suite of species inhibits conservation prioritization and adaptive management to address global seabird declines.

Traditionally, burrow occupancy during the breeding period is used to measure reproductive effort. Burrow occupancy is determined using a ‘grubbing’ technique: visual or tactile inspection of burrows (sometimes with a burrow scope) in a given area. Though widely considered a reliable method (Parker & Rexer-Huber, 2016), potential for error remains, especially when investigators visit sites infrequently (e.g., once per season or less). Surveys are conducted during targeted periods of the breeding cycle (e.g., chick hatch), however annual variation in the timing of breeding can compound errors; for instance, if breeding failures occur prior to surveys this leads to underestimates of occupancy. Most colonies are large, requiring sub-sampling of burrows within standardized quadrats along a transect; this can lead to error if habitat is heterogeneous and not adequately sampled. Burrows can be convoluted networks with multiple entrances or unreachable nest chambers, leading to a high proportion of burrows with unknown status. Finally, despite efforts to standardize methods between surveys, observer biases have the potential to influence results through subtle differences in interpretation of protocols, and site-specific knowledge (e.g.,  Anker-Nilssen & Rostad, 1993). In spite of these caveats, burrow occupancy remains the simplest and most widely used metric of reproductive effort against which to assess the efficacy of newer monitoring methods.

In-situ environmental sensors have transformed ecological studies at all scales from individuals to populations (Burger & Shaffer, 2008; Marques et al., 2012). For population abundance monitoring, the use of acoustic recorders and remote cameras is rapidly growing, particularly for cryptic species (Blumstein et al., 2011; Trolliet, 2014; Gasc et al., 2016). These instruments can be deployed in the field and left to sample for long durations without costly human assistance. The number of species-specific vocalizations or images detected per unit of time can be used as an index of activity within the reception range of the recorder. Likewise, mobile marine radar can be used to monitor the activity of flying animals (Cooper et al., 1991). Advances in image capture and analytical programs have improved data acquisition from radar units (e.g.,  Assali, Bez & Tremblay, 2017). Improvements in current recording equipment and development of additional sensors will undoubtedly continue. Acoustic recordings, automated cameras, and radar have great potential for monitoring seabirds, but their relative efficacy has not been previously fully evaluated.

The high temporal and spatial resolution data recorded by electronic monitoring equipment makes manual review to identify species detections impractical. The lack of easily implemented automated methods to process these datasets currently limits the ability of ecologists to capitalize on the amount of data stored in image, video, acoustic, or radar files (Weinstein, 2017). However, data processing methods are rapidly improving. Machine learning and other automated detection algorithms can be applied to image and acoustic data sets to extract meaningful metrics (Deng, Hinton & Kingsbury, 2013; Cires˛an & Meier, 2015; Krizhevsky, Sutskever & Hinton, 2017). For instance, Deep Neural Networks (DNNs) can be used to detect sounds from recordings with spectro-temporal properties that are similar to signals produced by target species. Currently, some level of manual review is necessary to label data patterns corresponding to detections that are then used by classification algorithms to count all such patterns in a dataset. As new tools emerge, there is great promise in reanalyzing archived images and recordings. This, in itself, is an improvement on monitoring programs reliant solely on human collected data, as no visual or audio archive exists and time-series from these efforts are often constrained by evolving data collection methods.

It is important to compare the results of monitoring methods as each method has its own strengths and limitations. To date, there are few examples that compare abundance metrics for seabirds. Using visual counts of birds to infer population abundance is not a new approach (Bednarz et al., 1990; Clarke et al., 2003), but is context dependent. At the colony, counts can be used to document attendance patterns (e.g.,  Harding et al., 2005; Huffeldt & Merkel, 2013), however, doing so for species that arrive and depart from their colonies at night adds an additional challenge. Trials with infra-red video recording equipment at a European storm-petrel (Hydrobates pelagicus) colony found that flying birds were poor predictors of occupancy estimates derived from play-backs, likely because not all flying birds were related to the sample plots (Perkins, Bingham & Bolton, 2017). Acoustic call rates have rarely been compared to visual counts at seabird colonies. Acoustic call rates of surface nesting Forster’s terns (Sterna forsteri) yielded significant correlations with colony counts, however colony sizes were small, ranging from 15-111 breeding pairs (Borker et al., 2014). Finally, radar surveys have been used to measure changes in relative abundance of seabird populations for decades (Day et al., 2003; Cooper, Raphael & Peery, 2006; Raine et al., 2017), yet validation or comparative studies exist for a limited number of species (Bertram, Cowen & Burger, 1999; Cragg, Burger & Piatt, 2016).

The Leach’s storm-petrel, Hydrobates leucorhoa, (LESP) is a small burrow nesting seabird that lives 20 years or more (Morse & Buchheister, 1977). It breeds in dense colonies, typically located on off-shore rocks or islands across the Northern Hemisphere, with small numbers breeding in the Southern Hemisphere (Huntington, Butler & Mauck, 1996). Leach’s storm-petrels are highly vocal and nocturnally active at their breeding colonies, making them an ideal candidate species for comparing acoustic, image, and radar monitoring methods (Watanuki, 1986; Buxton & Jones, 2012). Additionally, conservation needs are pressing. Mark-recapture studies and play-back surveys in the North Atlantic have highlighted declines (Newson & Mitchell, 2008), but little is known about populations in the North Pacific. Infrequent surveys along the Oregon coast (1978, 1988, 2008) indicate that populations on some islands have dramatically declined while others have increased (Kocourek et al., 2009). However, the magnitude of the changes in population estimates, spatial inconsistencies, under-sampling of available habitat, and variability in survey timing have led to uncertainty about population status along the Oregon coast (Kocourek et al., 2009). As a result, researchers and managers are left with insufficient data to assess the conservation status and develop management plans. Therefore, there is a pressing need for monitoring methods that can be applied more frequently at multiple breeding colonies where effective conservation actions can be taken.

Here we compared four data collection methods to monitor abundance of Leach’s storm-petrels at their breeding colonies: monitoring by manual nest inspections (grubbing), infra-red cameras, acoustic recorders, and radar. We compared rates of burrow occupancy, to plot level abundance metrics (detection rates) from images and acoustics. We did not make a similar comparison with radar data because the radar collected information at a different spatial scale (i.e., colony-wide versus plot level). Then we compared the remote methods to each other. These monitoring methods provided data on different aspects of the population and at different spatial and temporal scales. Thus, we expected the strongest correlations among metrics that represent breeding performance (e.g., burrow occupancy and acoustic ground calls) or among metrics that represent population size (image and radar counts of flying birds and acoustic aerial calls), and poor correlation between these two groups. Finally, we discuss recommendations for future monitoring efforts of burrow nesting seabirds.

Methods

Plot design

On Goat Island, we deployed cameras and acoustic recorders and sampled burrow occupancy in three adjacent rectangular plots (10 m × 25 m) on the north vegetated slope (GIA, GIB, GIC, Fig. 1, Table 1). Each plot had three near-infrared cameras with self-contained covert illuminators (No-Glow) capable of taking single frame time-lapse or near video images (2 frames/sec). All cameras faced upslope to the south, away from the prevailing winds to minimize lens fouling; posts with reflectors marked the edge of the IR camera range (10.48 m) and are visible in the images. The cameras used were: Reconyx PC900 HyperFire (3.1 megapixels) and Moultrie M-990i (4 megapixels). Photos from the Moultrie cameras were of insufficient quality, potentially due to the IR flash, so were not used in the analysis. Cameras were powered by deep cycle 12V marine batteries, but recording was limited by 32GB SD cards. Cameras were attached 1.5 m above the ground to metal fence posts, padded with foam pipe insulation to minimize collision injury risk to birds. A single acoustic recorder (Song Meter SM3 with built-in microphones, Wildlife Acoustics, Concord, MA, U.S.A.) was positioned at the center of each plot at 0.3 m above the ground. Song Meters were powered with internal batteries or external battery packs depending on desired recording duration and contained 4 SD cards (128 GB each). At Saddle Rock, two circular plots (SRN and SRS) were set up with an acoustic recorder and three cameras mounted on a single post in the center (Fig. 1). Circular plots, with a radius of 7 m, were chosen at Saddle Rock because this generally matches the survey shape sampled by the acoustic recorders’ omnidirectional microphones. The area surveyed by an acoustic sensor is dependent on many factors including: ambient noise (wind, rain, surf), presence of target and non-target species, amplitude of target calls, relative humidity, and equipment and settings on the equipment. The sensor records a sphere. In a quiet location, that sphere might have a 100⁺ m radius, in contrast to a small 10 or 15 m radius in a loud location. The proximity of the plots (∼25 m) means they are not independent replicates in representing the entire island for comparison with radar, but they are mostly independent replicates for comparisons of cameras and acoustic recorders. Both islands were visited periodically throughout the summer to exchange SD cards and batteries and ensure equipment was operating correctly.

Results

Acoustic monitoring

From the full dataset, the aerial call model identified 4,691,985 events above the 99% probability threshold. In total, 92,200 ground call events were detected after manual removal of false positives. Because each 2 s clip was determined to contain a detection, or not, a maximum of 30 detections could occur in one minute. Of the minutes that contained aerial calls, 56.1% had >20 detections, which indicated frequent signal saturation for aerial call rates. For the ground calls, only 0.08% of minutes with calls had >20 detections.

The final model explaining aerial call rates included night of year (by year), hour (22-4), moon illumination separated by whether the moon was above the horizon or not, presence of ground calls, and median flux sensitivity ( Table S1 and Fig. S5). This model explained 76.1% of the variation in aerial calls (Fig. 2), but retained a complicated structure of autocorrelation of fluctuating 2–3 day cycles (identified by visual inspection of autocovariance plots). Aerial call rates were significantly higher in 2014 than in 2015 (p < 0.001); similarly, hourly aerial call rates were relatively saturated (≥20 calls/min) more frequently in 2014 (61%), than in 2015 (23%) (X² = 871.43, p < 0.001). The final binomial GAM model including night of year (by year), hour, moon illumination separated by whether the moon was above the horizon or not, aerial call rate, and median flux sensitivity explained 41% of the variation in ground call rates. At Goat Island, ground calls occurred more often in 2015 than in 2014 (p < 0.001). Ground call rates during hours when ground calls occurred were higher in 2015 (0.61 calls/min), than in 2014 (0.43 calls/min).

Binomial GAM models were run to test for differences in call rates between Goat Island and Saddle Rock in 2015 from May 5 to August 14 (Table 2). Both aerial call rates and ground call rates were significantly higher at Goat Island than at Saddle Rock (p < 0.001). The final model for presence of ground calls explained 71% of the variation and included island, night of year, hour (20-6), moon illumination separated by if the moon was above the horizon or not, aerial call rate, and median flux sensitivity (Table S1 and Fig. S5). The final model of aerial call rate explained 87% of the variation with same predictor variables except ground call presence replaced aerial call rate as a predictor variable.

Method comparisons

Annual estimates of camera detections and call rates of both aerial and ground calls were significantly related to burrow occupancy for each plot (Fig. 3). Camera detections were the most strongly correlated with burrow occupancy (F_1,3 = 44.78, p = 0.007, R² = 0.916), followed by aerial calls 120–240 min past sunset (F_1,6 = 51.5, p = 0.0004, R² = 0.878), aerial calls 60 to 120 min prior to sunrise (F_1,3 = 18.21, p = 0.024, R² = 0.811), and ground call rates (F_1,6 = 13.56, p = 0.010, R² = 0.642). Acoustic energy was not significantly related to burrow occupancy (F_1,6 = 5.936, p = 0.051, R² = 0.414). Because aerial calls during the post sunset time period neared call saturation, with >20 calls/min in 2014 (Fig. 3C), the aerial call rates prior to sunrise were used for further method comparisons. Island and year sample size was too limited to compare radar abundances with burrow occupancy, however higher radar detection rates mirrored higher burrow occupancy at Goat Island.

Using data from both islands, radar detection, camera detections, acoustics (call rates and energy) were significantly and positively correlated with the exception that radar detections did not correlate with either aerial or ground call rates (Table 3). Camera detections were significant and linearly correlated to all other metrics of abundance (Table 3). Acoustic energy was significantly related to both camera and radar detections; however, these correlations had relatively low R² values. Aerial calls and camera detections were the most strongly related abundance metrics, and this was the only significant relationship between data collection methods using the high-density Goat Island data alone (Table 3). Generally, ground and aerial call rates were not highly correlated (Table 3, Fig. S4). Finally, to assess if acoustic energy was capturing additional information when acoustic call rates were saturated we fit non-linear models to the relationship of nightly aerial call rates with acoustic energy at Goat Island. Non-linear models improved AIC scores and provided a better fit to the data than linear models (Fig. 4).

Table 3:

Summary of instrument comparisons for measuring nightly abundance of Leachs storm-petrels, Hydrobates leucorhoa, on Goat Island and Saddle Rock, Oregon.

Linear models were run with both islands combined and then on Goat Island separately to test method comparisons at a high-density site only. Model estimates are sorted by R² values; significant p-values are in bold. Radar models were restricted to 2015, while other models include data from both years.

Comparison	R2	R2(adj.)	F	p	df
Both Islands
Aerial calls ∼ Camera	0.649	0.639	66.58	<0.001	36
Radar ∼ Camera	0.572	0.529	13.35	0.004	10
Aerial calls ∼ Acoustic Energy	0.512	0.511	991	<0.001	946
Ground calls ∼ Camera	0.312	0.297	20.86	<0.001	46
Camera ∼ Acoustic Energy	0.270	0.255	17.05	0.0006	46
Radar ∼ Acoustic Energy	0.207	0.175	6.526	0.0171	25
Aerial calls ∼ Radar	0.166	0.114	3.18	0.093	16
Ground calls ∼ Acoustic Energy	0.106	0.105	112.6	<0.001	946
Ground calls ∼ Aerial calls	0.102	0.101	107.8	<0.001	946
Ground calls ∼ Radar	0.048	−0.012	0.801	0.384	16
Goat Island only
Aerial calls ∼ Acoustic Energy	0.559	0.558	1048	<0.001	827
Aerial calls ∼ Camera	0.457	0.427	15.14	0.0011	18
Radar ∼ Camera	0.188	−0.015	0.925	0.391	4
Camera ∼ Acoustic Energy	0.125	0.093	3.983	0.0558	28
Aerial calls ∼ Radar	0.119	0.039	1.48	0.249	11
Ground calls ∼ Aerial calls	0.092	0.091	83.94	<0.001	827
Ground calls ∼ Radar	0.094	0.012	1.141	0.308	11
Ground calls ∼ Acoustic Energy	0.084	0.082	75.44	<0.001	827
Ground calls ∼ Camera	0.075	0.042	2.258	0.1441	28
Radar ∼ Acoustic Energy	<0.001	−0.059	0.002	0.968	17

DOI: 10.7717/peerj.6721/table-3

Discussion

The metrics derived from acoustics, radar, and images each provided unique information on the abundance patterns of Leach’s storm-petrels, as none of the metric comparisons resulted in an R² greater than 0.65. Counts of birds within the small field of view of the IR trail cameras were significantly correlated with all other metrics. Acoustic recorders were easily implemented throughout the season, and yielded three metrics of abundance: ground call rates, aerial call rates, and acoustic energy. Acoustic energy and aerial call rates were strongly related, but neither related strongly to ground calls. While nightly attendance patterns were similar between call rates and radar, overall correlation was low, potentially due to the scale of inference from each method or bird behavior (e.g., frequency of calls vs. flight). Long-term deployments of all methods have the potential to be the foundation of a robust monitoring program, if they can be implemented at the spatial and temporal scale necessary to document the population of interest. Further investigation is warranted to better understand the scale of change, or the sensitivity of each method to changes in abundance.

Trail cameras provided a promising new method for quantifying colony abundance of nocturnal seabirds. Counts from photos were strongly correlated with both the colony scale radar abundance index and the plot level burrow occupancy. Adding cameras to a monitoring program is low cost and requires minimal additional effort. With maturing automated image analysis, this tool will only improve. Likewise, implementing IR cameras as a stand-alone option seems viable as a low budget option, however there is significant species variation in flight patterns around colonies and which would likely influence the efficacy of this metric for other species (e.g., Perkins, Bingham & Bolton, 2017). Leach’s storm-petrels are small, agile fliers and appear to repeatedly take flight during the night when attending the colony. In contrast, heavier wing-loaded species such as sooty shearwaters (Ardenna grisea), or tufted puffins (Fratercula cirrhata) circle a colony, potentially relative to wind directions, and then land, exhibiting little localized flight behavior. These species-level differences in flight behavior are presumably also relevant considerations for both radar and acoustic metrics. An added benefit of IR cameras is that this method allows for detection of unexpected species and behaviors. In our case, we identified additional species including potential avian predators. This could also be done with acoustic recordings by running a suite of detectors for other vocally active species.

Overall, our findings suggest that among the current technology tested, acoustic monitoring has the greatest potential to monitor abundance of LESP at fine temporal resolution with multiple indices over long survey periods. While time-lapse photography and radar both could be scaled to deliver seasonal data, they provide a metric of aerial activity, albeit on different spatial scales per sensor (with one caveat, occasionally birds in images were counted on the ground). The information-dense acoustic recordings could yield a greater variety of indices beyond what we calculated. For instance, Leach’s storm-petrels exhibit sexual dimorphism in the frequency of their chatter calls that could provide insights into sex attendance patterns (Taoka et al., 1989), and detection of chick calls, heard during spot checks of the recordings, could provide an index of breeding success and a reference for breeding phenology. Additionally, unaccounted for short-term daily correlation remained in the binomial GAM analysis of acoustic call rates (e.g., 2–3 day cycles). This pattern may be related to foraging trip durations or social facilitation of colony attendance and further investigation is warranted (e.g., Granadeiro et al., 2009).

In this study, marine radar provided data at the colony scale, yet was most correlated to counts from cameras with a small spatial area of inference, potentially because both are metrics of flight activity. This, and the lack of correlation between radar and ground calls suggests that flight behavior measured at the colony scale may not be tightly linked with reproductive effort. This is perhaps not surprising as individual flight-paths may be convoluted and at the colony-scale includes non-breeding birds. However, further investigation is needed to disentangle this. Additionally, at the colony scale, breeding seabirds continue to attend even when reproductive success is low and non-breeding seabirds may visit colonies throughout the breeding cycle, often with a seasonal peak in the return of failed breeders and prospecting immatures (Boulinier et al., 1996). Little specific information is available for colony attendance of non-breeding birds for Leach’s storm-petrels and this complicates interpretation of any abundance metric tested in our study. As implemented, radar created the least disturbance, since colonies were not physically visited and remains the sole option for inaccessible sites. With the advances in technology that have occurred since we conducted our study the maintenance visits for cameras and acoustic recorders would not be necessary; regardless, colonies need to be accessible for deployment and retrieval of recording devices. Given the varying results of the method comparisons, coupled with the costs of purchasing and maintaining radars and the greater complexity of radar operations relative to camera and acoustic units, we do not recommend additional radar studies for LESP along the Oregon coast at this time.

A reliable long-term monitoring method must be robust to shifts in breeding phenology. For burrow nesting seabirds, surveys are traditionally timed to coincide with assumed periods of high colony attendance (Ratcliffe et al., 2010), yet seabird breeding phenology is known to shift due to changing environmental conditions (Gjerdrum et al., 2003; Frederiksen et al., 2004). This consideration is certainly pertinent for the Oregon coast, as recent oceanographic conditions—a marine heatwave (Bond et al., 2015), are unprecedented (Frölicher & Laufkötter, 2018). Very little is known about how these changes in marine conditions could impact storm-petrel breeding phenology in Oregon, but seabird breeding phenology is often variable (Byrd et al., 2008; Youngflesh et al., 2018). Using a monitoring method that captures the full extent of the breeding season, prevents missed information and misinterpretation. As implemented here, both camera and acoustic monitoring provide the temporal data stream necessary to address this, but further analysis is needed. Stationary radar systems could be set-up to record a similar seasonal record, but this effort is much more logistically complex than deploying acoustic recorders and trail-cameras.

Understanding annual differences in abundance is highly dependent on an understanding of within season variation in attendance patterns. Throughout the season, individual Leach’s storm-petrels visit the colony for different reasons and at different frequencies. For example, during incubation, when mate switches might occur infrequently, aerial calls, photo counts, and radar surveys might underestimate population size, but provide a better understanding of the number of breeders. Conversely, during a colony re-occupation period a higher number of non-breeders might visit for nest prospecting and pair bonding and result in higher population estimates and greater variability among monitoring metrics. For instance, the cross-method comparison revealed low R² values between ground calls and all other metrics, yet correlation with burrow occupancy, suggesting this metric is related to breeding effort rather than colony attendance. Additionally, annual or colony level differences in foraging trip durations (Pollet et al., 2014; Hedd et al., 2018), could lead to differences in abundances of birds attending a colony on a given night. We found a relationship with moonlight for both ground and aerial call rates, this was not completely surprising as avoidance of colonies on full moon nights has been previously documented (Watanuki, 1986; Buxton et al., 2013). Our hourly analysis of acoustic and radar data highlight the night pattern in colony attendance by Leach’s storm-petrels: an increase, plateau, and drop-off by dawn. When testing for annual differences it is important to obtain a season-long record of attendance to avoid mismatches in phenology between years. Other explanatory factors such as weather conditions influence colony attendance patterns of seabirds and detection rates of acoustic recorders and cameras, and inclusion would likely increase the explanatory power of our models.

Analytical methods are rapidly evolving, and could be adapted to overcome two of the analytical challenges we found in this study: overlapping aerial calls and manual counting of photos. We found an annual differences in aerial call rates, however this difference was largely driven by the lower percentage of saturated hours in 2015. Saturation of aerial call rates was the result of how the acoustic data were processed, and in future studies this could be overcome by reducing the 2-second call detection window, applying a sliding window, and/or building detectors that are able to count overlapping calls. Acoustic energy appeared to overcome some of the challenges in aerial call saturation, but provides a less precise metric as ambient sounds, especially wave noise, also contributed to and complicated the record at Saddle Rock. Similar to advances in acoustic signal recognition, methods to automate the recognition and subsequent counting of objects in photos are rapidly evolving. During our study, camera counts were limited in scope due to the extensive amount of time that manual counting required. Ideally, detection algorithms for automated counting of target images that are currently under development will substantially reduce this processing time (Weinstein, 2017).

Recommendations

The success of a long-term monitoring hinges on a robust sampling design and the availability of reliable resources to sustain the monitoring effort. We did not include a comprehensive analysis of the cost of each of the methods since they are continually changing and survey effort impacts costs, but equipment costs (cameras < acoustic recorders < radar), field costs (cameras ≈ acoustic recorders < radar), data processing (radar < automated acoustic call rates < manual camera counts), and data interpretation (radar ≈ camera counts < acoustic data) all contribute to method feasibility and how easily a given monitoring method could be implemented on a scale representative of seabird populations.

Whereas acoustic recorders are currently more commonly used in remote monitoring of seabird colonies, we recommend a new emphasis on deployment of cameras for a number of reasons. First, camera derived counts were correlated with acoustic metrics, radar surveys and burrow occupancy estimates (more data is needed to better understand these relationships). Thus, camera derived counts could be particularly helpful for monitoring less vocal species such as tufted puffins (Fratercula cirrhata). Second, photos have the ability to detect rare events or species (e.g., predators) that could be informative for management in unanticipated ways (Anderson et al., 2017). The same is true for acoustic datasets since the sampling sphere is larger; however, particular sounds need to be filtered through detection algorithms or spectrograms need to be visually reviewed. In contrast, photos may be more easily scanned for the unexpected. We found camera counts were highly correlated over the duration of a burst. It seems likely that photographs taken at 1–10 min intervals throughout the night would yield as much or more information on abundances than the targeted counts of this study. A more evenly spaced recording scheme throughout nighttime hours would allow trail cameras to record for a longer duration. Finally, since we purchased trail cameras, improvements on image capacity and battery life make this application more feasible without maintenance visits, and increased demand may foster purpose-built camera systems.

However, we do not imply validation and interpretation of detection rates from acoustics and images are complete. Additional validation studies should be done concurrently with monitoring over the longer-term because datasets can be re-analyzed as methods progress. Our method comparisons appeared to be related at large effect sizes (i.e., when data from both colonies was included). Likewise, our sample sizes were small and additional calibration at intermediate densities is needed, along with a plot design that better reflects instrument sampling range. Our analysis identified significant differences between years. These differences may be reflective of variation in attendance patterns related to colony success. Our analysis was limited to 2–3 paired camera and acoustic recorders per-colony, per-year. This appears adequate to capture large-scale differences in abundance, but spatial placement and the number of recorders relative to the size of the colony are key considerations. The number of recorders that are needed per unit area still needs validation and is beyond the scope of this study, but likely is dependent on species characteristics and colony topography (e.g., Oppel et al., 2014). Placement of recorders at colony edges versus colony centers could lead to different conclusions depending on how population changes occur (colony expansion/contraction versus more uniform changes in density). Additionally, placement of acoustic recorders must consider other sources of ambient noise that could interfere with recording quality such as wind and waves (Buxton & Jones, 2012). Likewise, placement of cameras needs to consider prevailing winds and weather to maximize collection of usable photos. In our study, we coupled the two to make direct comparisons, but deploying instruments in an alternating grid could capture more spatial variation.

Our results indicate that IR camera and acoustic monitoring have potential for low cost monitoring of dense colonies of burrow or crevice nesting species. Additionally, results of our acoustic monitoring suggest that with additional analysis and advances in technology acoustics could provide an efficient monitoring tool in these scenarios. Considering the lack of information on population trends and the need to identify declines before they are catastrophic, we urge managers to employ both methods while working towards refining detection and data analysis methods. Indeed, the USFWS Inventory and Monitoring Programs of the US West Coast regions are currently developing seabird monitoring protocols under the new Pacific Seabird Program and are considering these methods as they reach the full implementation stage for each species. Remote methods analyzed in this study have the potential to provide low impact, repeatable, and low cost monitoring techniques for annual monitoring and, thereby, provide an early warning indicator and the opportunity to discover and mitigate the root causes of population changes.

Supplemental Information