Acoustic Monitoring Technology

Acoustic monitoring and sampling technology

C. Ryan Allen, Shannon E. Romeling and Lynn W. Robbins

Department of Biology

Missouri State University

Springfield, MO

Abstract

Acoustic detectors have been used for monitoring flight activity of bats since the G.W. Pierce developed sound capture technology in 1938. Recently, significant progress has been made in the areas of portability, weather resistance, and the collection and storage of large data sets over extended periods of time. This progress includes the continued development of new and potentially more accurate means of collecting the information contained within each call sequence, as well as more accurate and repeatable ways to identify the species making these calls. The two main categories of detectors used to collect these data are zero-crossing and full spectrum detectors. This study included three commonly used detectors; the zero-crossing Anabat (Titley Electronics, Inc.), and two full spectrum detectors, AR-125 (Binary Acoustic Technology), and SM2 (Wildlife Acoustics). Side by side comparisons were conducted for 34 nights during 2010 throughout Missouri. These data were used to compare average memory consumption, total files collected, total bat passes, species and species group identifications, quality of the call sequences, and reported call parameters. In addition, two automated call identification software packages were used for comparison; BCID (Bat Call Identification, Inc.) and Sonobat 3 NE (Sonobat). All recorded call files were passed through the automated software packages. Furthermore, full spectrum calls from the SM2 recorder were converted into zero-crossing call files allowing the software packages to analyze the same files. Species composition, calls parameters and processing times were measured for each block of files. A total of 140,968 files were collected resulting in 22,228 identified bat passes and 117,680 noise files from the 3 detectors. Results suggested that these detectors are not interchangeable. There exist clear differences in the amount and type of data they record and therefore projects conducted with one are not necessarily comparable to projects done with another.

Introduction

The use of ultrasonic detectors to record echolocation calls has become an important part of studying bat ecology. With the presence of endangered species of bats, and the increased awareness of bat activity in industries such as wind energy, mines, road construction, power lines, and timber, accurate identification of local bat fauna is imperative. Increasingly, the use of bat detectors to passively monitor these sites has become the preferred manner in which these surveys are conducted.

In the 1938, Donald Griffin and Robert Galambos used sound capture technology developed by physicist G.W. Pierce that resulted in the discovery that bats produce and hear sounds in octaves above audible human hearing. After several years of experimenting with bats and the use of ultrasonic sound, Griffin, in 1944, coined the term echolocation to describe the phenomenon they were observing. Echolocation is a process by which an animal orients itself, or identifies the location, character, and perhaps movement of objects, by emitting high-frequency sounds and interpreting the reflected sound waves (Whitaker and Hamilton 1998).

Modern bat detectors use full spectrum or zero-crossing acoustic sampling techniques to record ultrasonic sound. Beginning in the 1980’s, zero-crossing detectors, specifically the Anabat, were increasingly used because of the low data consumption rates, field adaptability, and relative low cost. While full spectrum detectors did exist during this time, due to the lack of advanced computer technology and limited storage capacity, they were not often used as passive monitoring systems. With the rapid advancement of computer technology, it has recently become feasible to use full spectrum technology under field conditions. This naturally leads to the question, which system or detector leads to the most accurate and repeatable results in a user-friendly manner?

While the use of acoustic technology is currently possible in a long term monitoring situation, the large data sets require automated identification. Several attempts have been made to automate bat species identification using techniques such as discriminant function analysis, neural networks and weighted classification trees. The most notable attempts using these methods have been made by Allen 2010; Betts 1998; Britzke et al. 2010; Corcoran 2007; Fenton and Bell, 1981; Gruver et al. 2010; Krusic and Neefus 1996; Parsons and Jones, 2000; Szewczak, 2010.

The purpose of this study was to compare both the hardware and the software of full spectrum and zero-crossing acoustic bat technology in a manner consistent with the manufactures recommended use. While this introduced many variables to the comparison, it was the only way to satisfy the goal of comparing results when using standard techniques. For this study, we chose three commonly used bat detectors and the two known acoustic software packages that have graphical user interfaces.

Methods

This study included three commonly used detectors; Anabat (Titley Electronics, Inc.), AR-125 with an FR-125 recording unit (Binary Acoustic Technology), and SM2BAT (Wildlife Acoustics). Detectors were aligned next to each other on pelican cases on a table approximately one meter off the ground (Figure 1). Detectors recorded for between 4 and 8 hours each night. Data were collected from a variety of locations throughout Missouri (Figure 2). Detectors were placed in a variety of habitats including: fields, near ponds, forested roads and trails.

Two automated call identification software packages were used for comparison; BCID 10 (Bat Call Identification, Inc.) and Sonobat 3 Northeast (NE) version (Sonobat). Sonobat 3 is a full spectrum discriminant function analysis identification software recently developed for several regions of the U.S. The northeast version was used in this study because a midwest version was not available. BCID 10 is a zero-crossing weighted classification tree analysis developed in 2007 and updated in 2010. It currently covers most of the northeast and midwest species of the U.S.

The AR-125 and Anabat microphones were set at approximately 45° angles and 6” apart. Due to the unique configuration of the SM2BAT detector, it was set vertically next to the other two detectors. The SM2BAT was used with both microphones with the thought that most users would utilize the two microphones, taking advantage of this unique feature. Additionally, as stated in the introduction, this project was set-up with the purpose of duplicating standard techniques. Throughout the experiment, Anabat II with a ZCAIM, Anabat SD1 and Anabat SD2 units were randomly chosen each night. Due to cost constraints, only one SM2BAT and one AR-125 with an FR-125 recording unit were available for use.

Anabats were set with the sensitivities calibrated equally for all units and all units were synced in time. A division ratio of 16 was used for all test nights with a standard microphone. Anabat data were collected on a CF card and processed through CFCread version 4.2.1 with default settings. The AR-125 was set to a time-expansion (TE) of 10, duration of 15.0 second, idle of 3.0, delay of 0.0, low-frequency of 15.0 kHz and high-frequency of 90.0 kHz. Data were collected on an 8 GB flash drive and then run through the Sonobat Batch Scrubber 3 using default settings. The SM2BAT was set in accordance with the SM2BAT supplemental manual (Wildlife Acoustics, Inc, 2009-2010). An SMX-US microphone was used for both microphones and used in 192 kHz stereo. Data were run through WAC2WAV (Wildlife Acoustics Audio Compression Converter) version 2.9. WAC2WAV was set with default settings and split triggers, skip noise and SMX-US compensation filter were selected.

Data were used to compare average memory consumption, total files collected, total bat passes, and reported call parameters (mean Fmax, mean Fmin, mean duration, Fc and Fk). Recording time varied each night; therefore, all nightly data reported were based on a 10 hour time period. This was done by calculating an average per hour and multiplying by 10. Full spectrum calls from the SM2BAT recorder were converted into zero-crossing call files using WAC2WAV software allowing Sonobat 3 NE and BCID 10 to analyze the same files. Direct comparisons could then be made between the Anabat and SM2BAT as well as the two software suites. To do these direct comparisons, 5 randomly chosen nights of data were used due to the large volume of call files. These same 5 nights of data were also used in the parameter comparisons. Parameter comparisons were done for eastern red bats, tri-colored bats and silver-haired bats because they represent the full range of frequencies and call characteristics. Additionally, there were a large number of them available in the data for analyses. Myotis species could not be analyzed for parameter differences due to the low number of calls available. For these 5 nights of data, SM2BAT files converted to zero-crossing were compared to the Anabat files using BCID 10 (to compare detectors) and the un-converted SM2BAT and converted SM2BAT files were analyzed using the two different software packages in order to directly compare them.

Processing times of software packages were measured for each block of files when possible. Fewer data points exist for the full spectrum files due to extremely large SM2BAT files that would cause stack overflows and crash the software. Sonobat 3 NE was used to identify all full spectrum call files using default settings. Bat passes were calculated by the high/low tally from the output file given by Sonobat 3. The column MeanClassification was used for identification to species. BCID 10 was used to identify all zero-crossing call files using default settings. Bat passes were calculated with the minimum number of calls was set to 1 and species identification was calculated with the minimum number of calls set to 4.

Results

Comparisons were conducted for 34 nights from May 17 to July 17, 2010 throughout Missouri. A total of 140,968 files were collected resulting in 22,228 identified bat passes and 117,680 noise files from the 3 detectors. A total of 4,980 of these files were identified to species by the two acoustical software packages. An average of 0.02 MB/hr of data was collected from the Anabats, 2.06 MB/hr from the AR-125 and 3.55 MB/hr from the SM2BAT (Figure 3).

Analyzing these data with 27 identical computers running BCID 10 and Sonobat 3 NE resulted in processing times of approximately 582 files/minute by BCID 10 and 3 files/minute by Sonobat 3 NE (Figure 4). The parameter comparison using 5 randomly chosen nights of data for eastern red bats resulted in significant differences in the mean Fmax and Fk of all three detectors (Figure 5). For tri-colored bats, significant differences were found among all three units for Fmax and Anabats differed significantly from the full-spectrum detectors for both duration and Fk (Figure 6). Silver-haired bats produced significant differences in Fmin among all three detectors (Figure 7). The AR-125 significantly varied from the Anabat and SM2BAT in Fmax and Fc and the SM2BAT varied significantly from the other two in duration for silver-haired bats as well.

The SM2BAT recorded the highest number of bat call files, noise files and bat passes, however it had the fewest sequences identified to species by Sonobat 3 (Figure 8). Anabat files identified by BCID 10 were dominated by mid-frequency species (Lasiurus borealis, Nycticeius humeralis, Perimyotis subflavus) and the SM2BAT and AR-125 were dominated by high-frequency species, which includes all of the mid-frequency species plus the Myotis spp. There is no mid-species category when using Sonobat 3 (Figures 9-11). While the species distribution among all three methods was relatively consistent when looking at the entire data set, there were obvious differences when compared on a nightly basis (Figure 12).

There was a large amount of variability in the species level identification of call files. Sonobat 3 in conjunction with the full spectrum detectors identified many more low-frequency calls to species (Figure 13), but relatively few Myotis spp. Only two M. sodalis and no M. septentrionalis were identified by Sonobat 3 NE, while BCID 10 identified 27 files belonging to these two species (Figure 15). Identification of P. subflavus was nearly equal among all three detectors, but BCID 10 identified many more N. humeralis and L. borealis (Figure 14). On average, the Anabat in conjunction with BCID 10 and the SM2BAT in conjunction with Sonobat 3 NE, found eastern red bats and big brown bats to be the most common species (Figures 16 and 18). The AR-125 in conjunction with Sonobat 3 NE found hoary bats and big brown bats to be the most common species (Figure 17).

After analyzing the same randomly chosen 5 nights of SM2BAT files converted to zero-crossing files and Anabat files using BCID 10, there were clear differences in the species groups and species detected by the two detectors. The SM2BAT detected more high-frequency and low-frequency species than the Anabats; whereas, the Anabats detected more mid frequency species (Figure 19). The most apparent difference in the species comparison was the much larger number of tri-colored bats detected by the Anabats (Figure 20). The use these same 5 nights of data with the un-converted SM2BAT files and the converted SM2BAT files, allowed for a direct comparison between the software packages. For the species group composition comparison, BCID 10 identified more of both high and low frequency species (Figure 21). The largest difference in the species comparison was the higher number of tri-colored and eastern red bats identified by BCID 10 (Figure 22).

Discussion and Conclusions

The overall results of this study suggest that these detectors are not interchangeable. There exist clear differences in the amount and type of data they record and therefore projects conducted with one are not necessarily comparable to projects done with another.

The full spectrum detectors clearly collect more data (Figure 3). This may make them more useful when attempting to collect calls from a rare, quiet, or difficult species. However, data processing times can be quite extensive (Figure 4). More bat passes appear to be identified using full spectrum equipment as well, but it is unknown at this time if this is an artifact of noise being attributed to bats or actual bat calls. There is some qualitative evidence that this is the case, but a full statistical analysis has yet to be done. It does appear that additional noise may play a role in the ability for software to identify a call to species. More bats appear to be identified to species using the BCID 10 software which is likely attributable to more extraneous noise present in full spectrum calls, as well as the conservative nature of Sonobat 3 NE (Figures 8). Additional filtering techniques are in the process of being developed which should eliminate some of this discrepancy (Joe Szewczak, pers. comm.).

While it is no longer a problem to store extremely large amounts of data, processing times are still an issue. The processing time for the zero-crossing call files for this project was approximately two hours, while the full spectrum call files took well over 200 hours. It is recommended that future software developers of full spectrum identification suites look into parallel processing as an alternative programming design. While this type of programming (e.g. CUDA) typically requires specific hardware for the user, the time saved could be well worth it.

Automated call identification is still being developed, but it is likely the future of acoustic sampling. The software developers recognize the current limitations and are continually expanding and improving upon their software. This study indicates that surveys analyzed with different software packages should not be considered comparable data for abundance and species composition type analysis, however richness appears nearly equal across all variables over time with the exception of some difficult to distinguish Myotis spp. not being identified by Sonobat 3 NE. This issue is currently being addressed in new versions of the Sonobat software (Joe Szewczak, pers. comm.).

There were some differences in parameters recorded by the three detectors and reported by the software packages; however, the majority of them are not significantly different (Figures 5-7). It was expected that duration and Fmax would differ significantly from zero-crossing to full spectrum due to the sensitivity of full spectrum detectors and the differing sampling rates. The call files we chose for comparison were all identified using the software packages and visually to ensure that we were comparing the same species. However, both the BCID 10 and Sonobat 3 NE software rely heavily on the call parameters falling within a narrow band in order to make an identification. This effectively reduced the standard deviation of these data sets making the error rates appear extremely low. This subsequently showed some statistically significant differences between the hardware systems that may or may not actually exist. It has been shown that the natural variation of call parameters within a species greatly exceeds these error calculations and therefore the three detectors are likely comparable for reporting call parameters (Murray et al. 2001). It is suggested that more research be conducted in this area using unknown call files recorded simultaneously or artificial sound, eliminating the bias of the software systems. Another major problem that may lead to the differences in software identifications is the differences in the call libraries, which include species in the library, sample sizes of these species, and methods used to collect the data. We recommend that all data included in call libraries that are used for species identification be available for peer review, and all identifications using these libraries include identification probabilities and confidence limits of these species or species group identifications.

While this study has produced valuable insights into the behavior of these hardware and software systems, it has opened the door to many more questions. Future work still needs to be conducted to determine how these systems vary when most or all of the confounding variables have been removed. The overall impression is the hardware all performs adequately in general but fails to standardize echolocation research as a group. This is somewhat expected given the complicated nature of recording high frequency sounds and the different designs (i.e. microphones, sampling rates, etc.) among the detectors. This in turn has a profound effect on the performance of any software packages. At the same time, standard levels of acceptable confidence have yet to be developed for automated software and there will always exist a trade-off between quantity over quality in the identification of bat echolocation. Currently the software is being tailored to specific hardware, which is likely why Sonobat 3 NE is much more conservative than BCID 10. The hardware it is used with records a lot of extraneous noise, thus making the filtering process much more difficult. On the other hand a zero-crossing recorder can only record one sound at any given instance and likely misses some important information such as harmonics. The future direction of echolocation research will likely be more influenced by normal market conditions (cost, availability, time, ease of use, etc.) rather than specific technological advancements.

Literature Cited

Allen, C.R., L.W., Robbins. 2010. Efficient repeatable approach to quantitative call identification. Oral presentation, Ozark Summit 2010 Living on Karst: Sustainable Management of Ozark Ecosystems.

Betts, B.J. 1998 Effects of interindividual variation in echolocation calls on identification of big

brown and silver-haired bats. Journal of Wildlife Management 62:1003-1010.

Britzke, E.R., J.E. Duchamp, K.L. Murray, R.K., Swihart, L.W. Robbins. 2011. Acoustic identification of bats in the eastern United States: a comparison of parametric and nonparametric methods. In Press.

Corcoran, A.J. 2007. Automated acoustic identification of nine bat species of the eastern United States. M.A. Thesis. Humboldt State University.

Fenton, M. B., and G. P. Bell. 1981. Recognition of species of insectivorous bats by their echolocation calls. Journal of Mammalogy 62:233-243.

Gruver, J., S. Howlin, C. Nations, T. McDonald. 2010. Using discriminant function analysis and other quantitative techniques to classify bat echolocation calls. West Inc. Oral presentation, North American Symposium on Bat Research.

Krusic, R. A., and C. D. Neefus 1996. Habitat associations of bat species in the White Mountain

National Forest. Pages 185-198 in Bats and Forests Symposium. R. M. R. Barclay and R.

M. Brigham, editors. Victoria, British Columbia: British Columbia Ministry of Forests.

Murray, K.L., E.R. Britzke and L.W. Robbins. 2001. Variation in search-phase calls of bats. Journal of Mammalogy. 82(3): 728-737.

Parsons, S. and G. Jones. 2000. Acoustic identification of twelve species of echolocating bats by discriminant function analysis and artificial neural networks. J. Exp Biol. Sep (Pt 17): 2641-56.

Szewczak, J. Sonobat: Software for Bat Call Analysis. Obtained from: www.sonobat.com. Accessed November 26, 2010.

Whitaker, J.O. and W.J. Hamilton. 1998. Mammals of the eastern United States. Cornell University Press, Ithaca, NY.

Wildlife Acoustics Inc. 2009-2010. Song meter SM2BAT. Revised 8/13/10.

Result Figures:

Figure 1. Set up of detector comparison. Anabat SD1 on left in this example, AR125 with FR125 in the middle and SM2BAT on the right.

Figure 2. Locations of detector comparison test locations.

Figure 3. Average hourly data consumption for each detector 0ver 2 weeks based on 10 hours of recording. This equates to: Anabat – 2.8 MB, AR125 – 288.6 MB, SM2 – 496.9 MB. Note: Will vary drastically by site.