Teton Raptor Center
Spring is in the air for high altitude forest raptors! The owls have been advertising their territories and looking for mates while the diurnal raptors call to their partners from time to time at dawn. For the past six weeks, the Teton Raptor Center crew has been deploying Wildlife Acoustics's SM3 units graciously donated to the project to record all these sounds. Coupled with the recordings, the crews have been skiing and snowshoeing all night in the backcountry of the Greater Yellowstone Ecosystem surveying for owls in the same locations as the recorders. While we sift through all of the recordings with the help of SongScope over the next few months, we'll be starting to piece together how recorders and surveyors compare and begin to understand how the raptors distribute themselves across the southern Greater Yellowstone Ecosystem.
Sifting through over 6 terabytes of data isn't easy! All spring and summer, our dedicated team of trained staff and volunteers have been reviewing sounds gathered this spring around Great Gray Owl territories. Because Great Grays call around 250 dB, automatically detecting the faint, low calls has not been possible using automated processing software. So we're manually sifting through the hours. We have already documented several types of owl calls not previously recorded for this elusive sensitive forest owl. We also have confirmed using Song Meters to be an effective method for simultaneous detecting other forest raptors, such as Northern Goshawks, Cooper's Hawks, Northern Pygmy Owls, and Long-eared Owls. One aspect from our study that has become apparent, the Song Meter SM4 recorders are better at detecting owls than traditional callback survey techniques! We'll be continuing to review recordings over the coming months to describe the vocal behaviors of Great Grays and other owls and to create a monitoring strategy using recorders.
All spring and summer our dedicated team of trained staff and volunteers have been reviewing sounds gathered this spring around Great Gray Owl territories.
Because Great Grays call around 250 dB, automatically detecting the faint, low calls has not been possible using automated processing software. So we're manually sifting through the hours. We have already documented several types of owl calls not previously recorded for this elusive sensitive forest owl.
We also have confirmed using recorders to be an effective method for simultaneous detecting other forest raptors, such as Northern Goshawks, Cooper's Hawks, Northern Pygmy Owls, and Long-eared Owls.
One aspect from our study that has become apparent, the recorders are better at detecting owls than traditional callback survey techniques!
We'll be continuing to review recordings over the coming months to describe the vocal behaviors of Great Grays and other owls and to create a monitoring strategy using recorders.
Forest owls are typically surveyed for and censused using nighttime callback survey methodologies, which require traveling to a territory and broadcasting a conspecific call to elicit a response from territorial breeding owls. However, the effectiveness of surveys is not well documented and our previous data on Great Gray Owls suggests they may not be adequate to assess territory occupancy. In 2016, we simultaneously conducted callback surveys and continuously recorded all sounds within 18 known Great Gray Owl territories in western Wyoming to compare the two protocols to determine occupancy. We also investigate several recording and data analysis systems. Using two surveys in a week, we would have estimated that 40% of territories were unoccupied, whereas recorders detected 100% occupancy rates in weeks currently investigated. We found that manual analysis is the most precise way to determine if and when owls are calling, but automated band limiting energy detectors in conjunction with random forest classifications to determine positive calls may be the most effective methods for determining Great Gray Owl occupancy. We are currently continuing data analysis to further refine results and effectively evaluate methodologies, but the data are clear that nighttime surveys is not an appropriate method for determining occupancy of nesting Great Gray Owls.
Boreal Owl, Northern Pygmy-Owl, and Great Gray Owl are all designated Wyoming Species of Greatest Conservation Need (SGCN) that occur in boreal and montane forest habitats. These habitats are at risk due to increasing frequency and intensity of wildfires, large-scale insect infestations, and disease. In addition to these risk factors, which are exacerbated by ongoing climate changes, timber harvest and forest thinning treatments can reduce and fragment stands of high quality habitat. Boreal Owls (NSS3) are classified as vulnerable due to severe limiting factors. Few scientific studies have been conducted on this species. Great Gray Owl and Northern Pygmy-Owl populations are thought to be stable but limited. These owls are classified as NSSU (unknown status) because populations are difficult to assess given current monitoring methods.
Monitoring of these species and their habitats has typically depends on detection of territorial owls during the early spring courtship period using nighttime callback techniques. Federal land managers are tasked with determining if species of conservation concern occur in project areas and typically use callback surveys to detect Boreal, Northern Pygmy-, and Great Gray Owls when planning habitat treatments. Callback detection methods include playing recorded territorial calls of target owl species at night during the courtship and nesting seasons and documenting responses. Night surveys are often difficult to conduct in Wyoming due to the rugged, remote terrain typically occupied by owls, lack of trails and roads, heavy snow loads, snow melt conditions, and severe temperatures during February, March, and April. Night surveys also create safety hazards for field personnel and are expensive to conduct, which severely limits the number and scope of surveys that can be completed each year. Further, detection rates related to callback techniques remain unknown, making it difficult to assess occupancy and population status accurately.
During our studies of Great Gray Owls in western Wyoming from 2012-2016, we documented 25 nesting territories within the study area using a variety of techniques, including callback surveys, radio-tracking, systematic nest searching, and opportunistic sightings. Our data strongly suggest that not all nesting pairs respond to callback surveys and other owls also respond in areas without nests. Both issues have significant implications for population monitoring and habitat treatment assessments.
Detection rates have not been determined for any of the forest owl species so it is unknown what proportion of the nesting population is being measured and what portion is missed during callback surveys. Recent owl studies suggest that using automated recording devices can be effective in monitoring owls within designated areas (e.g., Grava et al. 2008, Rognan et al. 2008, Goyette et al. 2011). These devices can detect all calling owls present within an area over multiple nights and weeks in contrast to the traditional callback survey method where surveys in a particular area are often restricted to only one or two nights during a season. Furthermore, recorders can be set up and checked during the day and in favorable weather conditions which can reduce risk with field work.
In 2016, we conducted a study to test and evaluate the use of automated recording devices for monitoring Great Gray Owls during the breeding season. We tested two recording systems, simultaneous callback and recording surveys, and different analysis software and methodologies. Our goals were to determine the effectiveness, ease of use, and methodologies of automated recording systems to monitor forest raptors by testing the systems in known Great Gray Owl territories.
We designed our surveys to simultaneously capture data from automated recording systems and traditional callback survey methodologies for Great Gray Owls. Because our goal was to test the effectiveness of automated recorders, we only surveyed in known, previously active territories within our existing study area. In each territory, we conducted two survey periods to correspond with the early and late calling periods for Great Gray Owls in western Wyoming. The early survey period was a three week window from 25 February — 16 March and the second period extended from 16 March — 6 April, 2016. During each period, we deployed three automated recorders in a triangular array, 150m from the 2015 nest site (Figure 1). Because the survey distance of recorders was unknown, we chose a conservative detection distance of 150m to conduct this test.
We tested two different recording units (Figure 2). We used one array of Wildlife Acoustics Song Meter SM-3 recording devices donated by Wildlife Acoustics. We set these recorders to record from sunset to sunrise to ensure the batteries would last one week. The SM- 3s recorded sounds in 16-bit .WAV format and saved separate files each hour. We also custom built our own recording systems after discussions with and advice from other sound ecologists, such as Shan Burson from Grand Teton National Park. To make this system, we used Roland R- 05 hand-held sound recorders placed in waterproof housings with two ultra-low noise floor microphone electrets to record in stereo. We also tested two rechargeable batteries for different lifespans. These systems continuously recorded sound in 128 kb/s .MP3 format during the entire deployment and saved separate files at 64 mb size (ca. 1.1hr). We also tested these systems side- by-side in trials prior to deployment and tested detection distances using our broadcast survey callers.
We deployed three recorders in five territories per week and re-deployed in each territory during the second survey period. This gave us a total sample of 15 territories that were recorded for one week each in the early and late periods. We also deployed arrays at three additional owl territories to continuously record during the entire calling period with no calling period (control territories), for a total of 18 territories sampled in 2016.
While recorders were deployed, we also conducted traditional nocturnal callback surveys. We conducted callback surveys on two nights each week the recorders were deployed in each territory with at least three days between surveys. Call back survey locations were predetermined in a triangular fashion, opposite the recording devices, 200m from the nest (Figure 1).
We began surveys no earlier than one half hour after sunset and typically completed before 02:00 hours. In order to maximize surveys during the study and because of concurrent study objectives for Boreal Owls (Aegolius funereus), we called for both Great Gray Owls and Boreal Owls at all survey locations using a FOXPRO caller (Foxpro NX4). Each calling period consisted of a 2-min listening period, followed by the Boreal Owl territorial call, a 1-min listening period, the Great Gray Owl territorial call, a 1-min listening period, the Great Gray Owl territorial call again, and a final 2-min listening period. When we detected a Great Gray Owl at a territory, we did not continue surveying any of the remaining points at that territory and instead determined it occupied and proceeded to the next territory.
We processed recordings using a variety of techniques. First, we manually analyzed recordings by plotting spectrograms in a free software called AudacityÂ®. We plotted recordings as spectrograms and visually searched for all owl calls during from sunset to sunrise, recording species, gender (when known), call type, frequency of call, time, duration, and number of calls.
Second, we investigated the use of Song ScopeÂ® software produced by Wildlife Acoustics and Raven ProÂ® software produced by Cornell University. The process involved four steps; 1. Creating a band limited energy detector (BLED) in Raven to select individual notes in a pre-set frequency range, 2. Construct training data sets by classifying selections from step one as true Great Gray Owl territorial notes or false positives to train the random forest (RF) classification, 3. Predict true positive territorial notes using a random forest classification in R, and 4. Manually classify predictions as true detections or false positives.
We use BLEDs constructed by Medley (2013) to identify selections — notes in the same frequency Great Gray Owl territorial calls. Medley constructed separate BLEDs for male and female Great Gray Owls in California. In our audio recordings, we found that Medley's male detector selected more true female territorial calls than Medley's female detector selected, so we used his male detector for both sexes.
To identify territorial owl calls in selections provided by the BLEDs, we trained random forest (RF) (Breiman 2001) supervised learning algorithms, in conjunction with Recursive Partitioning (Rpart; Therneau and Atkinson 2015) and Random Forest (Breiman and Cutler 2015) packages. To train the RF classifier, we classified all BLED selections from approximately 30 hours of recordings from 9 sites as Great Gray Owl notes or not Great Gray Owl notes. We classified all Great Gray Owl notes as territorial (t), agitated (a), defensive (d), contact (c). Only territorial calls were abundant enough in our training recordings to train the RF algorithms.
Using a custom R Script (Medley 2013), we used the resulting RF classifiers (territorial, false) to predict GGOW territorial notes from unlabeled BLED selection tables output from Raven ProÂ®. Using R, we combined notes less than two seconds apart into sequences, decreasing the amount of time necessary for technician classification and enabling us to count territorial sequences, rather than individual notes.
Recorders Function and Performance
In our trial tests of the two recording systems, we found the SM-3 had greater functionality to our custom-made units but increased sensitivity and ease of finding calls with our system. The three SM3 units can be programmed with custom recording schedules, automatically assess sunrise and sunset by day, and have an integrated light sensor. They only record in .WAV format that at 329mb/hr, requiring >5 times as much digital storage as our custom recorders. We were able to custom make 25 recorders, which recorded sounds 24 hrs/day with less digital storage requirements.
We were able to document owls simultaneously at multiple recorders, suggesting the detection radius of the units was >200m. In our initial testing, we could detect calls up to 300m, but were unsure of how the volume on the broadcast caller directly relates to wild owls calling. We have been unable to confirm a Great Gray Owl detection on all three recorders in the array simultaneously and most detections are only recorded on one unit, suggesting that a 200-250m detection radius is appropriate for Great Gray Owls.
We conducted a total of 60 play-back surveys surrounding 15 known Great Gray Owl nest sites while simultaneously recording (two surveys/week in each early and late deployments, Table 1). We detected Great Gray owls during 18 of our surveys and most detections were at the first survey point within a territory (13 of 18 detections). Owls were not detected until the second survey point in three instances, and in two instances, owls were not detected until the final survey point. We only recorded a pair of owls in three territories; all others were single individuals.
Great Gray Owls were detected at 11 of the territories across the entire study period and we did not detect any owls at four of the surveyed territories using this methodology. Typically surveys are conducted only once or twice within territories. When looking at the one-week survey windows (early and late) for each territory (this assumes we only did two surveys within a territory, not four), callback surveys would have estimated that 40% of territories were unoccupied using the early season window. Using the late season window, we did not detect owls in 53% of territories.
Of the four territories in which we did not detect an owl, all had an active nest and successfully fledged young in 2016. Across the four survey nights within each territory, we detected owls in 25% of the territories only night, 25% two nights, and 19% three nights. We did not detect owls in all four survey nights within any territory. Owls were detected in nine territories in the early round of surveys and seven in the late round. At five territories, owls were detected in both rounds of surveying. (Table 1)
Of the observations of Great Gray Owls during play-back surveys within known territories, three detections were visual with no accompanying vocalizations by the owls. Of all vocalizations detected, 16 were territorial calls, three were agitated calls, and one was a defensive hoot.
Manual Sound Analysis
We gathered a total of 32,391 hours of recordings for this project; 28,112 hours from our recorders and 4,278 hours from SM3 units. Of the total, roughly 14,800 hours were recorded from the three control territories at which recorders were placed for the entire courtship period. Excluding daytime recordings (roughly 12 hours/day) and the control nests, we have gathered ca. 10,934 hours of recordings from dusk to dawn at known Great Gray Owl territories. Currently, we have manually analyzed 2,607 hours of those recordings. We have recorded all known call types and several previous undescribed calls or variations of calls from Great Gray Owls. All territories that have been analyzed had positive detections for Great Gray Owls within each week surveyed. We detected >7,899 calls (not notes) across the hours analyzed ranging in frequency from 150-450 Hz. Males and female territorial calls can be relatively easy to distinguish within a territory but difficult to distinguish among territories. As many as three individual Great Gray Owls have been detected within one night at a single territory and determining individuals by frequency and call pattern exhibits great promise.
We also detected Boreal Owls, Northern Pygmy Owls, Saw-whet Owls, Great Horned Owls, Long-eared Owls, and Northern Goshawks during our analysis. For Great Gray Owls, we detected 4,658 territorial calls, 2543 territorial calls as part of a duet, 276 agitated calls, 167 begging calls, 50 defensive calls, 27 begging calls, and 1 chitter call. From the data analyzed to date, the longest stretch in which we did not detect any Great Gray Owls calling was six days.
However, this stretch of inactivity was recorded after March 25th, 2016 and that particular pair initiated incubation on March 25th in 2015 (the nest was not located in 2016 so incubation date was unknown). Most owls reduce calling after the onset of incubation, which may explain the reduced calling after March 25th in 2016. Excluding that territory, the longest stretch of calling inactivity was two days.
Owls were detected within all territories we have reviewed, meaning that we have 100% territory occupancy rates. At each site, we also detected at least two owls at all sites. Using only one recorder per territory to eliminate the potential of sampling a call twice, we found that Great Gray Owls call throughout the night, but have a greater frequency of calling from 11pm — 4am (Figure 3).
Automated Recorder Analysis
We determined that Song Scope Â® was not going to be adequate for the analysis of Great Gray Owl calls because it would not allow for user configurations in key analysis components. Songs are individually identified by the user and unknown controls within the software create the algorithm to search for calls. We wanted greater control of variable control and decided to first use Raven ProÂ® software.
To date, we have analyzed 942 hours of recordings using the Raven detector and RF classification for Great Gray Owl territorial calls on one recorder within each array from eight different territories. Across the 94 sampling nights, the detector found owl calls in 46 nights (49%; Table 3). The automated detector classified a total of 2,456 territorial calls (after notes were aggregated into calls). Similar to our manual detections, the highest frequency of calls was in the 3:00 and 4:00 hours (Figure 4).
Raven only analyzes .WAV files and conversion from .MP3 was one of the most time- consuming steps that requires a vast amount of temporary storage, followed by running the BLED. However, both the conversion process and running BLEDs are passive processing times running on a computer that can be run on large file batches overnight to maximize processing time. Person-time is needed to annotate calls in the BLED selection tables and verifying results of the RF classification. We recorded time needed to annotate and verify calls for 626.77 hours of recordings and it required a total of 7.8 person-hours, or ca. 45 seconds/hour of recording.
Timing of verification depends largely on the number of Great Horned Owl or other false positives within the recordings or sites that need verification.
MANUAL RECORDER ANALYSIS VS. CALLBACK SURVEYS
To date, we have manually analyzed a total of 19 nights in which simultaneous callback surveys were conducted. Of the nights analyzed, the manual analysis of the recordings was vastly superior at detecting owls than callback surveys (Figure 5). We detected calling owls on the recorders in 11 nights where surveyors did not detect an owl (Table 3). Only once was an owl detected on a survey that was not detected by a recorder and that was a visual observation of the owl.
It took a total of 48 person-nights to complete all surveys in 2016 (two teams of two people each, 12 nights). Time spent traveling to survey location and surveying by teams was a total of 92.7 hours, or 185.4 person-hours. We have only been able to manually analyze 93 nights of recordings to-date due to the significant amount of time needed to review each hour of recordings. However, due to concurrent objectives of quantifying and describing details of each call from all owl species (not just Great Gray Owls) in addition to reviewing three recorders per territory, describing time taken to analyze the recordings is not possible. If an experienced observer was reviewing recordings simply for presence/absence of only Great Gray Owls, it takes approximately five minutes to review each hour. Time needed deploying and retrieving recorders is highly variable depending on study area, remoteness of territories, terrain, snow cover, etc. Generally, within our study area, we could retrieve and re-deploy all 15 recorders in different territories with 3-4 people in one day.
MANUAL RECORDER ANALYSIS VS. AUTOMATED DETECTOR IN RAVEN
We were able to directly compare a total of 93 nights in which we analyzed entire nights both manually and with the automated detector in Raven (Table 3). This was comprised of 10 weeks (51 individual nights) from different territories where we conducted callback surveys and 36 nights from two of our control territories. Raven correctly classified nights as owls present or absent in 81% of nights (Figure 6). Raven did not find any calls during nights in which our manual analysis did not find calls, but it did miss owls in 23 nights (19%). In determining territory occupancy, Raven would have classified eight of 11 territories as occupied with one week of recorder deployment.
Raven was consistent with our callback survey results on nine of 15 survey nights, but detected owls on five nights in which surveyors did not (Table 3). Only in one case did surveyors detect an owl that the detector missed, which was a visual detection of a non-calling owl. Time needed to analyze files is more than drastically reduced using BLEDs.
We found that traditional callback surveys drastically underestimated occupancy of Great Gray Owl territories in 2016. Even with four surveys conducted per territory, we did not detect owls in 26% of territories surveyed, whereas recorders detected owls in 100% of territories.
Based on our results to-date, the use of traditional callback surveys may not be appropriate to determine occupancy or territory locations for this Wyoming species of concern.
There are several possible reasons why owls did not respond to callback surveys. First, our results to-date indicate that peak calling periods for Great Grays occurs from 11pm-4am. Our callback surveys were conducted from dusk until ca. 1am, but most often completed by 11pm in 2016. It is possible that males may be foraging outside of the immediate nest areas during that time and are not available to respond to playbacks. Alternatively, owls may respond after the survey period is complete. In at least one territory, we detected owls on the recorders ca. 10min after the surveyors had left the area. Finally, it is possible that owls had become habituated to the survey caller after repeated visits to a nesting territory, considering we have surveyed within many of the territories in previous years.
The use of recorders to detect owls can be extremely affective but requires a vast amount of time if reviewing all sound files and recording details of all owl species, as we have been. It appears that three days of recorder deployment may be adequate to detect if owls are present within a territory during the courtship period, prior to incubation. However, deployment for one week is likely a better protocol to ensure detection of owls if incubation has been initiated.
Deployment of recorders requires fewer people than nighttime callback surveys, can be done during the daytime hours for safer backcountry travel, and provides continuous data collection over multiple days. More territories can also be simultaneously surveyed by recorders than an individual. In areas of known nest sites, deployment of one recorder to determine occupancy is likely adequate. This technique can be used to locate new nesting territories or to determine owl presence in unknown areas by setting up arrays of recorders during the courtship period 500m apart (250m detection radius).
Automated recognizers in programs like RavenÂ® and KaleidoscopeÂ® will be key in reducing time needed to review recordings. Recognizers have to be created for each species and requires hundreds of known calls to train the software. Particularly for Great Gray Owls that have a territorial call 400Hz, creating detectors can be extremely difficult due to normal background noise at low frequencies. The software can have more difficulty determining the decibel difference between a call note and background noise at these low levels. It is likely that detectors for species that call at higher frequencies will perform better than Great Gray Owl detectors.
Currently, the automated detectors would have misclassified two of eight territories as inactive. We have only run detectors for Great Gray Owl territorial calls, which may have lead, in part, to the reduced detections compared to our manual analysis. For example, Raven did not detect any owl calls in territory C (Table 3), when we found 24 territorial calls during the deployment week. However, we also detected 15 agitated, 9 contact, and 2 begging calls at that territory, which the Raven detector did not search for. Further, we only ran detectors on one of the three recorders within a given territory. Presumably, the number of detections using automated software will increase after the creation of detectors for multiple call types and using all recorders within each territory.
Misclassifying or missing calls using detectors would generally be acceptable in instances when the goal was to determine occupancy only if the proportion missed calls still resulted in at least one detection during the deployment period. In our instance, the detector under-performed for one week of deployment using one recorder. The quality of automated detectors relies on the balance of false positives and false negatives. By relaxing the parameters within the detector to find more calls, the detector also will flag more false positives. Utilizing the RF classification helps eliminate some false positives but many flagged detections still are incorrect, requiring verification of all calls flagged as possible detections. More work needs to be conducted to refine the detector for Wyoming Great Gray Owls.
The analysis for the project is not yet complete. We will continue to determine Great Gray Owl presence/absence for all recordings to compare to both callback surveys and machine detections. In a subset of recordings, we will also continue to record details on all owl calls to help assess calling patterns. Using these data, we will investigate the role of weather, moon phase, and cloud cover on calling patterns. We also will investigate the use of multiple recorders in each territory to determine if one recorder is adequate to assess territory occupancy by using our three control nests to help determine this because they had no influence of callback surveys. Similarly, we will look at the effect of callback surveys on call rates by comparing control to surveyed territories.
We will continue the analysis to determine calling patterns in the early and late periods to determine if there is an ideal time to deploy recorders. Further detailed analysis of the sensitivity and ability of Raven to detect of the two recording systems needs to be conducted. We are also continue to work towards discriminating male and female calls and individual call frequencies to document the number of individuals calling each night.
One major aspect of this analysis is creating automated detectors. We have been focused on creating detectors for Great Gray Owl territorial calls using Raven ProÂ® software. We will soon investigate the use of KaleidoscopeÂ® software for ease of use and accuracy. Kaleidoscope differs from Raven by searching for entire calls, instead of the individual notes within the call as Raven does. This may lead to fewer false positive detections. Because we are manually analyzing all recordings, this affords us a unique opportunity to accurately estimate effectiveness of each software. Key aspects to assess software are how often the detector misses calls, the rate at which it inaccurately identifies a call, time needed to verify results, deployment period needed to accurately assess occupancy, and machine learning ability.
This project could not have been completed without the dedication and enthusiasm of team leader, Katherine Gura, and field technician, Nick Ciaravella. Joe Medley provided the inital Raven detector, random forest programming, and critical advice. Carrie Ann Adams conducted the automated recording process. Superstar volunteers Tim Griffith, Steve Poole, and Bev Boyton manually reviewed recordings and helped deploy recorders. Ellen Yeatman, Emily Smith, Neil White, Daniel Gura, Carrie-Ann Adams, Meghan Warren, and Susan Patla all helped conduct owl surveys. Sarah Ramirez, Arron Couch, Josh Seibel, and Josh Metten and also helped deploy recorders. Susan Patla was integral in the logistics of this project. Wildlife Acoustics provided the SM3 recording units and SongScope software. Field work was funded by Wyoming Game and Fish Department.