INTRODUCTION

The Northern Bobwhite (Colinus virginianus) is both a declining species of concern and a managed game species. Bobwhite decline has progressed for decades and is linked to habitat loss and population fragmentation due to urbanization and changes in agricultural practices (Brennan 1991, Church et al. 1993, Guthery et al. 2000, McKenzie 2009, Hernández et al. 2013). A range-wide adaptive management approach, including research, monitoring, education, and habitat management, is needed to increase populations (Brennan 1991, Palmer et al. 2011, Hernández et al. 2013). In accordance with recommendations from the National Bobwhite Conservation Initiative (NBCI), the South Carolina Department of Natural Resources is collaborating with other states to conduct breeding season monitoring via standardized point counts (Morgan et al. 2016, Chapman et al. 2020). In 2017, we added 14 autonomous recording units (ARUs) to supplement traditional point counts. We predicted that ARU and traditional surveys would produce similar occupancy results and could be an efficient way to expand survey efforts. Here, we provide a comparison of practicality and results using ARU and traditional point count methods. We use this assessment to make recommendations for how each survey technique could improve our ability to monitor bobwhite and other bird populations.

Population monitoring is especially important as it determines effectiveness of restoration and management strategies (Green et al. 2017). Despite improvement in research and monitoring techniques for bobwhite, we still need large-scale methods that can be implemented within an adaptive management strategy (Brennan 2002). Research suggests that the incorporation of ARUs improves monitoring in situations including rare species, large geographical areas, or shortage of trained surveyors (Borker et al. 2015, Pankratz et al. 2017, Darras et al. 2019, Haselmayer and Quinn 2000).

Traditional monitoring is time consuming and requires technicians to travel to survey sites within a narrow range of dates and times. Variability in observer experience and hearing ability is often ignored (Brewster and Simons 2009, Digby et al. 2013). Recent technological advances have produced ARUs capable of results similar to, or better than, traditional surveys (Digby et al. 2013, Borker et al. 2015, Shonfield and Bayne 2017, Darras et al. 2019). Autonomous recording units can survey multiple points over large areas at the same time. Microphones can be left to record automatically for weeks. Files can be reviewed repeatedly by single or multiple observers to reduce identification errors and observer bias (Celis-Murrilo et al. 2009, Holmes et al. 2014, Darras et al. 2019). Observers can examine recordings and identify calls as quickly and accurately as a real-time survey (Digby et al. 2013, Darras et al. 2019). Furthermore, ongoing research is producing promising programs for automatically detecting bobwhite calls (Nolan et al. 2023).

Autonomous recording unit surveys require specialized equipment and software (Darras et al. 2019, Shonfield and Bayne 2017). They allow for collection of copious audio data, which is challenging to sort and archive (Digby et al. 2013, Darras et al. 2019, Shonfield and Bayne 2017). Individual ARUs cannot determine distance and direction, so it is difficult to estimate bird density or abundance. However, solutions to this problem are being studied (Celis-Murillo et al. 2009, Shonfield and Bayne 2017, van Wilgenburg et al. 2017, Pérez-Granados and Traba 2021). Density of birds per area is the current NBCI recommended metric for bobwhite population estimation (Evans et. al 2011). The utility of occupancy surveys for producing population estimates is under consideration (Evans et al. 2011). However, presence/absence surveys are already helpful in identifying study sites and in tracking changes in habitat use or calling phenology (Evans et al. 2011, Borker et al. 2015, Darras et al. 2019, Furnas and McGrann 2018).

METHODS

Autonomous Recorder Surveys

Fourteen Song Meter SM4 Bioacoustics Recorders (Wildlife Acoustics, Inc., Maynard, Massachusetts, USA) were deployed in areas of interest for bobwhite habitat management. Our study sites were at Carolina Sandhills National Wildlife Refuge (Carolina Sandhills) and within the Indian Creek Woodland Savanna Restoration Initiative area (Indian Creek) both located in South Carolina, USA (Fig. 1). Recording took place between 1 April and 31 August 2017. Microphones were located at existing bobwhite survey points. Three microphones were placed at Carolina Sandhills, which is ranked as a NBCI “Tier I” site. A “Tier I” site is a focal area of interest nested within a focal landscape and region, with all levels having appropriate habitat and potential to support a long-term bobwhite population (Morgan et al. 2016). Eleven microphones were placed at Indian Creek, a cooperatively managed area of public and private lands. Indian Creek is ranked as an NBCI “Tier II” site, a focal area nested within a focal landscape or region (Morgan et al. 2016).

Stereo channel recordings were collected using a 24 kHz sampling rate and were saved as .WAV files. We used the unit’s internal microphones set to a 16-dB gain (which includes 26 dB of pre-amplification). Autonomous recording units were programmed to start recording at sunrise in 15-min on/off intervals for two consecutive hours. Recording points were visited once to deploy ARUs and once to collect them. We only analyzed data from the interval starting 30 min after sunrise. This is consistent with the timing of traditional point counts and with the time frames determined in other studies to have the most bobwhite calling (Elder 1956, Hansen and Guthery 2001, Lituma 2017). A technician used Raven Pro Interactive Sound Analysis software (available from the Cornell Lab of Ornithology; Bioacoustics Research Program 2017) to identify bobwhite calls by visually scanning for calls and listening to audio to confirm suspected vocalizations. Each call was tagged with its start time. The analyst also ranked the quality of each recording: good recordings were clear with little to no background noise, fair had repetitive noise that did not interfere with spectrograms, and poor had noise that impeded observer ability to detect calls. Identifiable rainfall was ranked as light, moderate, or heavy. For detailed weather information, we used measurements from the Columbia South Carolina Metropolitan Airport, the closest weather station with comprehensive hourly data. The station was 65 km from Indian Creek and 108 km from Carolina Sandhills, so we averaged conditions from an hour before to an hour after recording time. We also examined the correlation of weather station wind and precipitation data with recording quality and occurrence of rain observed on audio files.

We used all available ARU data to test if recommended seasonal timing and survey length of traditional point counts are appropriate for our region. We summed the daily ARU calls from all points, then used a 7-d moving average to determine the seasonal peak of calling activity. Using recordings where at least one bobwhite was detected, we examined average time to detection and determined the proportion of calls that would have been detected using theoretical survey times of 5, 10, or 15 min. To assess the time required for audio file analysis, we re-analyzed a randomly selected sample of 100 5-min audio segments. The technician used a stopwatch to record the combined time required to navigate to the file in the Raven program, select and tag calls, and save the selection table.

Traditional Surveys

We conducted unlimited radius, 5-min, single-observer point counts based on NBCI guidelines (Morgan et al. 2016, Chapman et al. 2020). Surveys took place at 14 ARU points and 34 additional points. Two surveys were conducted at each point during the month of June for a total of 96 surveys. Observers recorded the number of bobwhites calling at any distance in 1-min time bands, as well as start times, temperature, percent cloud cover, ranked wind speed, and ranked noise. Surveys occurred within 3 h of sunrise and did not occur during rainfall, if windspeed was above 13 km/h, or with constant background noise.

Statistical Comparison of Survey Methods

To compare the performance of ARUs and traditional surveys as directly as possible, we limited the traditional survey data to points with corresponding ARU data. We limited ARU data to recordings within our observed seasonal calling peak (25 May–10 July). This time span is consistent with the approximately 6-wk range recommended by NBCI for point count surveys (Morgan et al. 2016). We compared presence/absence results for 658 peak season ARU and 28 traditional surveys using a two-tailed t-test as well as calculating average encounter rates for each survey type. We consider the encounter rate to be the probability of a surveyor finding at least one bird per survey (Strimas-Mackey et al. 2020). We also assessed seasonal ARU and traditional survey results using Wilcoxon signed rank tests paired by point to compare overall occupancy status and average encounter rate throughout the season at each of the 14 survey points. To determine if points where birds were detected on traditional surveys were those with higher calling rates, we used a two-tailed t-test to compare the number of ARU calls at points with and without birds detected on traditional surveys. Recording time of ARUs directly overlapped with an in-person survey on four occasions, which we compared to see if observers and ARUs located the same birds.

Random Forest Modeling

We used random forest modeling, a supervised machine learning approach, to build separate model sets exploring the relative importance of point and survey variables to traditional and ARU results. As we were primarily interested in comparing different methods to detect birds rather than modeling occupancy, it was appropriate to use random forest encounter rate models (Guillera-Arroita et al. 2015, Strimas-Mackey et al. 2020). For these analyses, we used the full set of 658 ARU survey occasions occurring at peak season and the full set of 96 traditional surveys from all 48 points. For both model sets, we built classification-type models using the randomForest package for R version 4.0.4 (Liaw and Wiener 2002). Random forest modeling is sensitive to class imbalance (Chen et al. 2004, Strimas-Mackey et al. 2020). Both traditional and ARU surveys resulted in fewer bird detections than non-detections, so we used the weighted random forest method, which penalizes incorrect classification of the minority class more strictly (Chen et al. 2004). We weighed the detection class inversely to the proportion in which it occurred (1:4 in traditional data and 1:3 in ARU data) so the model prioritized correct classification of detections three or four times higher, respectively, than non-detections. This allowed us to keep our full data sets rather than re-sampling the classes evenly, which was especially important for the smaller traditional data set.

Traditional survey variables included in model sets were point identity, survey start time, day of year, observer, and site type, as well as ranked windspeed, cloud cover, and noise. For ARU models, we considered point identity, day of year, recording quality, site type, and weather station variables (ranked cloud condition, wind speed, and precipitation). In both cases, temperature was positively correlated with day of year and was not included. We used mean decrease in accuracy (MDA) to determine variables with the greatest impact on encounter rate. During model testing, we began with the full model and dropped variables with MDA less than zero, or the lowest ranked variable, in each step until only two variables remained. If variables to be dropped had similar MDA, we tested all alternatives. For traditional surveys, site and observer were highly correlated, so we began with two full model sets. For both traditional and ARU models, point identity served as a surrogate for habitat differences. We assessed survey models using AICc rank (Akaike Information Criterion adjusted for small sample size), mean square error, percent accuracy, specificity (ability to correctly classify non-detections) and sensitivity (ability to correctly classify detections).

RESULTS

Autonomous Recorder Data

We analyzed 30,015 min of automated recordings from the full April–August data set. Of the 2,001 surveys, the analyst rated 74.26% as good recording quality, 22.14% as fair, and 4.10% as poor. The most common noise interference was from rainfall and accompanying wind. Of the poor-quality recordings, 74.39% had moderate or heavy rainfall, whereas only 0.03% of the good-quality recordings had even moderate rainfall. Other interference included insects, frogs, traffic, calling from other species, and microphone static.

In 15-min recordings where bobwhites were detected, we found that the mean time to first detection was 5.76 min (median 5.34, range 0.002–14.98). We divided the recordings into 5-min time bands and found 48.45% of first detections occurred in the first 5 min, and 77.02% within 10 min. This implies that had we used a 5-min recording time, comparable to in-person surveys, we would have detected birds on less than half of the occasions found using a 15-min recording.

We determined the length of the calling season at each point by calculating the number of days between the first and last bobwhite calls detected. Mean calling season was 74 d (median 70, range 32–105). Seasonal variation in calling, smoothed by a 7-d moving average, shows that peak calling occurs in June and early July (Fig. 2). There is a small secondary peak in late July (Fig. 2). The highest average call rate was 141 calls per day centered around 14 June. For statistical comparisons using ARU data, we used data from 25 May–10 July to encompass the peak observed calling season. The June survey window for our traditional point counts fits within the peak calling period and should appropriately sample bobwhite populations in our area (Fig. 2). However, surveys limited to June may, over time, miss any seasonal shift in calling phenology and do not include the secondary peak in July.

Our timed trials indicated that, on average, it took an analyst 1 min, 34 sec (median 1:20, range 0:31–5:04) to select bobwhite calls within a 5-min recording. The average difference in number of calls detected between the timed re-analysis and the original analysis was 0.33 calls per file, with 95% showing no difference in occupancy status.

Comparison of Autonomous and Traditional Surveys

When we compared raw occupancy results from individual surveys (658 ARU and 28 traditional) at the same 14 points, we found no significant difference between the two methods (t = -1.26, df = 29.92, p = 0.22). However, when point occupancy after multiple surveys is considered, the ARU method produced a higher number of occupied points. Birds were detected by ARU at 13 of the 14 points (Table 1). At the five points where birds were detected by traditional surveys, they were also detected by ARUs. The two methods agreed on raw occupancy status at six points (Table 1). At the remaining eight, the ARU method located calling, whereas no birds were detected using the traditional method. The average probability of encountering at least one bird per point over the course of a season using our traditional method was 0.36, whereas the probability using our more survey-intensive ARU method was 0.93 (Table 1). A Wilcoxon signed rank test, with observations paired by point, indicates that birds are significantly more likely to be detected using our ARU method than our traditional method (V = 0, p = 0.006). When we performed a similar paired Wilcoxon signed rank test using average point-based encounter rates, to account for differences in survey effort, we again found no significant difference between the methods (V = 26, p = 0.18). The mean point-based encounter rate was 0.32 (range 0.00–0.51) for ARUs, and 0.21 (range 0.00–1.00, Table 1) for traditional surveys.

Of the 28 sets of surveys that occurred at the same place and time, both methods returned six detections for equal encounter rates of 0.21. However, survey sets did not agree on which points were occupied on which days. Only four survey occasions had a 5-min traditional survey that occurred fully within a 15-min ARU recording. Of these occasions, three agreed, with neither in-person observers nor ARU analysis detecting any birds. On one occasion, a bobwhite call was found by the ARU method but was not simultaneously detected by the observer in the field.

Points varied in calling activity from no calls to 1,125 calls per point, with an overall total of 5,887 calls detected by ARUs. The mean number of birds detected per point on in-person surveys was 0.50 (range 0–2) birds (Table 1). There was no significant difference in the number of calls detected by ARUs at points where no birds were detected on in-person surveys and at points where one or more birds were detected (t = -1.92, df = 4.63, p = 0.12). However, at points where traditional surveys detected one or more birds, the mean count of calls from ARUs was higher, (mean 510.00) than the mean at points without birds (mean 210.22).

Random Forest Model Results (Traditional Surveys)

For traditional surveys, random forest models were relatively poor at predicting actual survey outcomes, however, our main interest was in relative importance of survey variables. Patterns emerged indicating that point-based variables, survey timing, and observer were most important. The top ranked model includes point identity, survey start time, and day of year. The top model has the best mean square error (MSE = 0.46), the highest percentage correct (54.17%), and the highest specificity (0.60) in the model set but has low sensitivity (0.25) (Table 2). The second and third models (id+obs and id+site) are interchangeable and are competitive with the top model within 1.49 AICc. Site and observer were highly correlated, as each observer surveyed at only one site. The second ranked models have slightly worse MSE (0.54), lower percentage correct (45.83%), and lower specificity (0.40), but much higher sensitivity (0.75) (Table 2). The top model maximizes the ability to classify negative detections, whereas the secondary models maximize the ability to classify positive detections. Point identity is in all three models and contributes the most to within-model mean decrease in error (MDE). This is to be expected as it accounts for the unmodeled differences in habitat.

In the top model, point identity contributes 10.45 toward MDE, survey day contributes 6.71, and survey start time contributes 3.54. A partial dependence plot for survey day of year shows that surveys taking place earlier in the June survey window have a higher relative probability of detecting birds, with the probability declining most steeply around 28 June (Fig. 3a). A partial dependence plot for survey time indicates that surveys conducted at least 50 min after sunrise have a higher relative probability of detecting birds, with a slight decrease in probability of detection around 135 min (ca. 2.5 h) after sunrise (Fig. 3b).

In the competitive id+obs and id+site models, point identity contributes 20.48 and 19.38, respectively, toward MDE. Observer contributes 8.96, and site contributes a similar 7.93. The Carolina Sandhills site had a higher relative probability of bird detection than the Indian Creek site (Fig. 4a). Observers A and B, who surveyed at Carolina Sandhills had a higher relative chance of detecting birds than observer C, who surveyed at Indian Creek site, although there is overlap in range between observers B and C, who conducted most surveys (Fig. 4b).

Random Forest Model Results (ARU Surveys)

Random forest models for ARU surveys had better fit and better predictive ability than models for traditional surveys. Fit, accuracy, and specificity were moderate, and sensitivity was high (Table 3). Models also provided a clear pattern of variable importance. There was only one top ARU model, including point identity and survey day of year. The top model has the best MSE (0.30), the highest percentage correct (69.51%), the highest specificity (0.63), and good sensitivity (0.85) (Table 3). Point identity again contributes the most toward MDE within individual models.

Within the top model, point identity contributed 44.97 toward MDE, whereas day of year contributed 23.95. As expected, different points have different relative probabilities of detection. The partial plot for survey day of year shows that surveys taking place between approximately 26 May–20 June (days 145–170) have a higher and more consistent relative probability of detecting birds than surveys outside this window (Fig. 5).

DISCUSSION

Practicality of Autonomous Recorder Surveys

Overall, ARUs were reliable, with no malfunctions, and produced a high proportion of good-quality recordings. For bobwhite detection at our sites, a moderate amount of background noise was not problematic, and only rain consistently caused poor recording quality. Much like the restrictions placed on traditional surveys, poor-quality recordings, or those from days with adverse weather conditions, could be removed from an ARU analysis. The time required to process a 5-min section of recorded data averaged only 1 min, 34 sec, considerably less than the 5 min required to do the survey in person plus the additional time needed to move between points. As we could conduct at least three ARU surveys for every in-person survey, ARUs would be helpful for increasing survey effort.

Pros and Cons of Autonomous and Traditional Surveys

Our data suggest that a point count or recording session should probably be longer than the 5-min counts commonly used in roadside transects and currently recommended by the NBCI (Elder 1956, Rollins et al. 2005, Morgan et al. 2016). On average, recording for only 5 min would result in detecting occupancy at less than half of points found to be occupied using a 15-min ARU survey. A 5-min in-person survey might also be expected to detect fewer than half of occupied points on a given day. As birds react to human presence and take time to return to normal calling behavior, two or three 5-min surveys over multiple days may not be an equivalent effort to one 10- or 15-min ARU survey (Shonfield and Bayne 2017). Use of ARUs can increase survey effort by increasing the number of days surveyed and the length of each survey as well as removing the constraints of covering many points within a short time.

The peak calling season observed in our ARU data is similar to the mid-June to July peaks observed in research conducted in Missouri, Kansas, Oklahoma, Georgia, and Florida (Bennitt 1951, Robel et al. 1969, Wilson 2000, Hansen and Guthery 2001, Terhune et al. 2009). Our study, like several others, shows a later secondary calling peak (Robel et al. 1969, Hansen and Guthery 2001, Terhune et al. 2009). If calling is associated primarily with breeding and nesting, the highest calling rates should correspond to peak breeding season (Speak and Haugun 1960, Terhune et al. 2009). The secondary calling observed in July could be a result of birds re-nesting after a failed attempt or producing a second brood (Sermons and Speake 1987, Terhune et al. 2009). The peak calling season, according to ARU data, occurs during the June survey window used for traditional surveys. Calling season, however, may vary from year to year. Researchers in several regions have concluded that variation in site, year, and weather had an impact on breeding, call timing, and behavior (Speake and Haugen 1960, Wilson 2000, Hansen and Guthery 2001, Terhune et at. 2009). There is also evidence for changes in bird breeding phenology based on long-term climate change (Forchhammer et al. 1998, Parmesan and Yohe 2003, Charmantier and Gienapp 2014). Using ARU surveys that extend to the edges of the calling season would ensure that peak season data are captured every year and could be used to track shifts in breeding season timing.

On a point-by-point basis throughout the year, the ARU method outperforms the traditional method by finding birds at 57% more survey points. On a survey-by-survey basis, however, when the same number of surveys on the same days of the year are compared, both methods provide similar results. Both methods had the same number of detection occasions and had equal encounter rates of 0.21, but daily comparisons did not agree on which points were occupied on which days. On any given day, the probability of detecting a bird at a point is low, so when survey times do not match exactly, we would not necessarily expect equal results. The combination of similar success rate of individual surveys, higher success rate of ARUs throughout the season, and low encounter rate suggests that the difference in seasonal results between methods is primarily due to the larger number of repeated surveys possible with the ARU method.

Traditional surveys allow for more direct estimation of bobwhite abundance as it is possible for observers to identify several birds calling simultaneously. The number of birds estimated on traditional point counts was low, ranging from zero to two. This means the occupancy rate may not differ drastically from abundance. Additionally, for easily detectable species, encounter rate closely approximates actual occupancy (Strimas-Mackey et al. 2020). Based on their loud and identifiable call, it is possible that encounter rates for bobwhite do closely reflect occupancy and abundance. In our results, sites where at least one bird was detected by a traditional survey did not have significantly more ARU calls than sites only found to be occupied by ARU surveys. However, each ARU season consisted of 47 survey replicates, whereas each in-person season consisted of only two. If birds are not calling reliably within the 5-min in-person surveys, this could easily confound results. Further study into the relationship between occupancy and abundance for bobwhites is needed.

Random Forest Model Interpretation

Random forest models indicated that for both traditional and ARU surveys, the point identity was most important in modeling relative probability of encountering bobwhites. This suggests that, regardless of survey type, habitat variables are the main drivers of encounter rate. Variables having to do with weather conditions and other survey-based variables were less important than point-based variables in final models. Although it would be better to use weather data from a closer station for ARU models, Pearson tests showed slight correlation between greater precipitation at the weather station and severity of rain heard on recordings (r = 0.305, df = 2009, p < 0.00) and between higher windspeed at the weather station and lower recording quality (r = -0.17, df = 2009, p < 0.00). Weather variables from in-person surveys were taken directly at the time of the survey and were also unimportant in top models. It is likely that point occupancy rather than survey-based detectability is the primary driver of encounter rate. In traditional surveys, site type was also an important modeling variable. However, site and observer were related, so it is difficult to determine if site type, observer ability, observer bias, or a combination affected encounter probability. In ARU models, we eliminated observer bias by using the same recording units and having a single observer analyze all data. Site type was not important in ARU models.

Survey timing was important to traditional surveys. Results showed that relative probability of encounter rate increased about 50 min after sunrise and started to decline about 2.5 h after sunrise. This suggests that it may be necessary to adjust the current protocol in which surveys begin at sunrise and to narrow the 3-h time window. The ARU method allows for all surveys to occur at the same time so it can maximize effort at the optimal time. We examined ARU data from 30–45 min post sunrise, which is slightly before peak detection, so timing of future ARU studies could also be adjusted. In both traditional and ARU models, survey day of year was an important factor. Traditional surveys were limited to June, but relative probability of encountering birds decreased throughout the month, and sharply toward the end. Automated recording unit surveys analyzed were limited to 25 May–10 July based on the peak season from the entire data set. Surveys conducted in late May to mid-June had higher and more consistent relative probabilities of encounter than those earlier in May or in late June or July. Both methods agree that the best survey dates for our region of South Carolina are from early to mid-June. However, the actual study period should likely extend beyond this time period.

CONCLUSIONS

Automated recording units and human observers are both able to detect presence of bobwhites on individual surveys with similar results. The increased number of replicates possible with ARUs allows for significantly better estimates of occupancy, whereas traditional surveys allow for estimation of abundance. Both methods could be combined into a powerful bobwhite monitoring program.

There are many advantages to ARU use, including the ability to survey more points, for longer surveys, and over a longer season. As analysis of ARU data takes less time than a comparable in-person survey, monitoring efforts could be expanded without adding personnel or hours. Automated recording unit surveys are also able to eliminate most survey-based variation in detectability. Whereas observer, survey day of year and survey start time are important for traditional surveys, ARUs eliminate observer differences and ensure that all surveys occur at the same times. We also found ARU recordings to be a practical way to track seasonal calling patterns, which would be ideal for tracking long-term changes in calling phenology.

As it is difficult to estimate counts of birds with ARUs, they are best deployed for presence/absence surveys. They could easily be used to determine occupancy of marginal sites where bobwhites are not known to occur, or where populations are known to be lower. Current efforts focus on relatively small areas within high priority focal sites, generally with higher occupancy. Automated recording units could be used to expand the survey zone and to target new areas for study or habitat management. Traditional methods could be incorporated to conduct abundance and habitat surveys at sites of interest. New analytical methods under development are improving the ability to automate the ARU call identification process and are on the verge of making it possible to use call counts to model population size. Such advances would allow the ARU technique to be viable as a lone-use method. However, for now, we recommend ARU use as a supplemental method for expanding the spatial and temporal range of traditional bobwhite monitoring.

LITERATURE CITED

Bennitt, R. 1951. Some aspects of Missouri quail and quail hunting 1938-1948. Missouri Conservation Commission Technical Bulletin 2. Missouri Conservation Commission, Jefferson City, Missouri, USA.

Bioacoustics Research Program. 2017. Raven Pro: interactive sound analysis software (version 1.5). Cornell Lab of Ornithology, Ithaca, New York, USA. http://www.birds.cornell.edu/raven

Borker, A. L., P. Halbert, M. W. McKown, B. R. Tershy, and D. A. Croll. 2015. A comparison of automated and traditional monitoring techniques for marbled murrelets using passive acoustic sensors. Wildlife Society Bulletin 39:813-818. https://doi.org/10.1002/wsb.608

Brennan, L. A. 1991. How can we reverse the Northern Bobwhite population decline? Wildlife Society Bulletin 19:544-555. https://doi.org/10.1002/wsb.223

Brennan, L. A. 2002. A decade of progress, a decade of frustration. Pages 230-232 in S. J. DeMaso, W. P. Kuvlesky, Jr., F. Hernández, and M. E. Berger, editors. Proceedings of the national quail symposium: volume 5. Texas Parks and Wildlife Department, Austin, Texas, USA. https://doi.org/10.7290/nqsp05g84k

Brewster, J. P., and T. R. Simons. 2009. Testing the importance of auditory detections in avian point counts. Journal of Field Ornithology 80:178-182. https://doi.org/10.1111/j.1557-9263.2009.00220.x

Celis-Murillo, A., J. L. Deppe, and M. F. Allen. 2009. Using soundscape recordings to estimate bird species abundance, richness, and composition. Journal of Field Ornithology 80:64-78. https://doi.org/10.1111/j.1557-9263.2009.00206.x

Chapman, S. A., M. K. Foley, J. L. Hodges, and J. J. Morgan. 2020. National bobwhite conservation initiative’s bobwhite almanac, state of the bobwhite 2020. National Bobwhite Technical Committee, Knoxville, Tennessee, USA. https://nbgi.org/download/nbcis-state-of-the-bobwhite-2021/

Charmantier, A., and P. Gienapp. 2014. Climate change and timing of avian breeding and migration: evolutionary versus plastic changes. Evolutionary Applications 7:15-28. https://doi.org/10.1111/eva.12126

Chen, C., A. Liaw, and L. Breiman. 2004. Using tandom forest to learn imbalanced data. Statistics Department Research and Industry Technical Report 666. University of California, Berkeley, California, USA. https://statistics.berkeley.edu/tech-reports/666

Church, K. E., J. R. Sauer, and S. Droege. 1993. Population trends of quails in North America. Pages 44-54 in K. E. Church and T. V. Dailey, editors. Proceedings of the National Quail Symposium: volume 3. Kansas Department of Wildlife and Parks, Pratt, Kansas, USA. https://doi.org/10.7290/nqsp03f8nt

Darras, K., P. Batráry, B. J. Furnas, I. Grass, Y. A. Mulyani, and T. Tscharntke. 2019. Autonomous sound recording outperforms human observation for sampling birds. Ecological Applications 29:1247-1265. https://doi.org/10.1002/eap.1954

Digby, A., M. Towsey, B. D. Bell, and P. D. Teal. 2013. A practical comparison of manual and autonomous methods for acoustic monitoring. Methods in Ecology and Evolution 4:675-683. https://doi.org/10.1111/2041-210X.12060

Elder, J. B. 1956. Analysis of whistling patterns in the eastern bob-white Colinus v. virginianus L. Journal of the Iowa Academy of Science 63:639-651. https://scholarworks.uni.edu/pias/vol63/iss1/72

Evans, K. O., J. A. Martin, and T. M. Terhune, 2011. Incorporating effective monitoring into the national bobwhite conservation initiative. Pages 191-204 in W. E. Palmer, T. M. Terhune, and D. F. McKenzie, editors. The national bobwhite conservation initiative: a range-wide plan for recovering bobwhites. Technical Publication, version 2.0. National Bobwhite Technical Committee, Knoxville, Tennessee, USA. https://quailcount.org/documents/nbci2_0/NBCIver2.0.pdf

Forchhammer, M., E. Post, and N. Stenseth. 1998. Breeding phenology and climate. Nature 391:29-30. https://doi.org/10.1038/34070

Furnas B. J., and M. C. McGrann. 2018. Using occupancy modeling to monitor dates of peak vocal activity for passerines in California. The Condor 120:188-200. https://doi.org/10.1650/CONDOR-17-165.1

Green, A. W., D. P. Grimes, G. Hagan, R. Hamrick, C. Harper, P. Keyser, J. Morgan, I. B. Parnell, R. Thackston, T. M. Terhune, II, and J. A. Martin. 2017. Monitoring northern bobwhite populations reduces uncertainty about management effectiveness: a paradigm of empiricism and hope. Pages 33-34 in T. V. Dailey and R. Applegate, editors. Proceedings of the national quail symposium, volume 8. University of Tennessee, Knoxville, Tennessee, USA. https://doi.org/10.7290/nqsp08law4

Guillera-Arroita, G., J. J. Lahoz-Monfort, J. Elith, A. Gordon, H. Kujala, P. E. Lentini, M. A. McCarthy, R. Tingley and B. A. Wintle. 2015. “Is my species distribution model fit for purpose?” Matching data and models to applications. Global Ecology and Biogeography 24:267-92. https://doi.org/10.1111/geb.12268

Guthery, F. S., M. J. Peterson, and R. R. George. 2000. Viability of northern bobwhite populations. The Journal of Wildlife Management 64:646-662. https://doi.org/10.2307/3802735

Hansen, H. M., and F. S. Guthery. 2001. Calling behavior of bobwhite males and the call-count index. Journal of Wildlife Management 29:145-152.

Haselmayer, J., and J. S. Quinn. 2000. A comparison of point counts and sound recording as bird survey methods in Amazonian southeast Peru. The Condor 102:887-893. http://dx.doi.org/10.2307/1370317

Hernández, F., L. A. Brennan, S. J. DeMaso, J. P. Sands, and D. B. Webster. 2013. On reversing the northern bobwhite population decline: 20 years later. Wildlife Society Bulletin 37:177-188. https://doi.org/10.1002/wsb.223

Holmes, S. B., K. A. McIlwrick, and L. A. Venier. 2014. Using automated sound recording and analysis to detect bird species-at-risk in southwestern Ontario woodlands. Wildlife Society Bulletin 38:591-598. https://doi.org/10.1002/wsb.421

Liaw, A., and M. Wiener. 2002. Classification and regression by randomForest. R News 2:18-22. https://CRAN.R-project.org/doc/Rnews/

Lituma, C. M., D. A. Buehler, E. P. Tanner, A. M. Tanner, P. D. Keyser, and C. A. Harper. 2017. Factors affecting availability for detection: an example using radio-collared Northern Bobwhite (Colinus virginianus). PLoS One 12(12): e0190376. https://doi.org/10.1371/journal.pone.0190376

McKenzie, D. F. 2009. Taking the northern bobwhite conservation initiative to the next level. Gamebird 2006: Proceedings of the sixth national quail symposium 6:16-23. University of Tennessee, Knoxville, Tennessee, USA. https://trace.tennessee.edu/cgi/viewcontent.cgi?article=1247&context=nqsp

Morgan, J. J., K. Duren, and T. V. Dailey. 2016. National bobwhite conservation initiative coordinated implementation program v1.1. Addendum, the national bobwhite conservation initiative: a range-wide plan for recovering bobwhites. Technical publication. National Bobwhite Technical Committee, Knoxville, Tennessee, USA. https://quailcount.org/monitoring/documents/NBCI%20Implementation%20Program%20v1.1.pdf

Nolan, V., C. Scott, J. M. Yeiser, N. Wilhite, P. E. Howell, D. Ingram, and J. A. Martin. 2023. The development of a convolutional neural network for the automatic detection of northern bobwhite Colinus virginianus covey calls. Remote Sensing in Ecology and Conservation 9(1):46-61. https://doi.org/10.1002/rse2.294

Palmer, W. E., T. M. Terhune, and D. F. McKenzie, editors. 2011. The national bobwhite conservation initiative: a range-wide plan for recovering bobwhites. Version 2.0. Technical Publication, National Bobwhite Technical Committee, Knoxville, Tennessee, USA. https://www.quailcount.org/documents/nbci2_0/NBCIver2.0.pdf

Pankratz, R. F., S. Hache, P. Sólymos, and E. M. Bayne. 2017. Potential benefits of augmenting road-based breeding bird surveys with autonomous recordings. Avian Conservation and Ecology 12(2): 18. https://doi.org/10.5751/ACE-01087-120218

Parmesan, C., and G. Yohe. 2003. A globally coherent fingerprint of climate change impacts across natural systems. Nature 421:37-42. https://doi.org/10.1038/nature01286

Pérez-Granados, C., and J. Traba. 2021. Estimating bird density using passive acoustic monitoring: a review of methods and suggestions for further research. Ibis 163:765-783. https://doi.org/10.1111/ibi.12944

Robel, R. J., D. J. Dick, and G. F. Krause. 1969. Regression coefficients used to adjust bobwhite quail whistle count data. Journal of Wildlife Management 33:662-668. https://doi.org/10.2307/3799391

Rollins, D., J. Brooks, N. Wilkins, and D. Ransom, Jr. 2005. Counting quail. AgriLife Communications - Texas A&M System Extension Publication B-6173. Texas A&M University, College Station, Texas, USA. https://wildlife.tamu.edu/files/2012/05/CountingQuail.pdf

Sermons, W. O. and D. W. Speake. 1987. Production of second broods by Northern Bobwhites. Wilson Bulletin 99:285-286.

Shonfield, J., and E. M. Bayne. 2017. Autonomous recording units in avian ecological research: current use and future applications. Avian Conservation and Ecology 12(1): 14. https://doi.org/10.5751/ACE-00974-120114

Speake, D. W., and A. Haugen. 1960. Quail reproduction and weather in Alabama. Proceedings of the Annual Conference of the Southeastern Association of Fish and Wildlife Agencies 14:85-97. https://seafwa.org/sites/default/files/journal-articles/SPEAKE-85.pdf

Strimas-Mackey, M., W. M. Hochachka, V. Ruiz-Gutierrez, O. J. Robinson, E. T. Miller, T. Auer, S. Kelling, D. Fink, and A. Johnston. 2020. Best practices for using eBird data: version 1.0. Cornell Lab of Ornithology, Ithaca, New York, USA. https://doi.org/10.5281/zenodo.3620739

Terhune, T. M., R. G. Hamrick, D. C. Sisson, and H. L. Stribling. 2009. Summer male call index relative to nesting chronology and autumn density of the northern bobwhite. Pages 54-64 in S. B. Cederbaum, B. C. Faircloth, T. M. Terhune, J. J. Thompson, and J. P. Carroll, editors. Gamebird 2006: a joint conference - Quail VI and Perdix XII, managing gamebirds in the 21st century. 31 May–4 June 2006. D. B. Warnell School of Forestry and Natural Resources, University of Georgia, Athens, Georgia, USA. http://talltimbers.org/wp-content/uploads/2016/03/Terhune_2009_MCindex_BOBWHT.op_.pdf

van Wilgenburg, S. L., P. Sólymos, K. J. Kardynal, and M. D. Frey. 2017. Paired sampling standardizes point count data from humans and acoustic recorders. Avian Conservation and Ecology 12(1):13. https://doi.org/10.5751/ACE-00975-120113

Wilson, H. M. 2000. The calling behavior of northern bobwhite males. Thesis, Oklahoma State University, Stillwater, Oklahoma, USA. https://hdl.handle.net/11244/11645