A field test of Indiana bat acoustic identification

Transcription

1 A field test of Indiana bat acoustic identification Joe Szewczak Leila S. Harris

2 Assessing bat presence and species composition...never easy Joe Szewczake

3 Acoustic detection can work

4 but many things work against it. 1) The bats themselves 2) Recording conditions 3) Recording deployment 4) Differences in recorder hardware Garbage in, garbage out. And all of these factors contribute garbage.

5 Bird calls have undergone selection for distinction from other species- they function as identifiers, and have no constraints on complexity. Southwest willow flycatcher

6 In contrast, the calls of echolocating bats serve the utility function of information acquisition. This works best with simple, short chirps.

7 And, many different bat species with similar information gathering needs often use very similar chirps. Myotis californicus Myotis yumanensis

8 Bats adjust their calls to suit the task at hand (call plasticity). clutter open All of these calls from bats of the same species. (Canyon bat calls from different bats in different habitats.)

9 Substantial overlap of call features.

10 Some acoustically similar species have at least some distinctive call types.

11 Some have frustrating overlap.

12 Bats move.

13 Bats adjust their calls to avoid frequency overlap with conspecifics. Sometimes you record both, sometimes you don t. Bat A Bat B

14 Bats speak at high pitches The operational frequency influences the interaction of the wave with targets. A target must be > λ/2 to generate an echo.

15 Ultrasound behaves differently than the sounds we hear. Shorter wavelengths of higher frequencies more affected by air turbulence, convection, etc. Higher frequencies attenuate more rapidly than lower frequencies.

16 Also, bats vary the amplitude through a call. As distance increases, higher frequencies and lower amplitude parts of calls get lost. high amplitude low amplitude low call fragment

17 Distance obscures details farther closer farther

18 Distance obscures details Obvious with visual observations, but how do you recognize distance with sound data?

19

20

21 The problem: some fragments can mimic fully-formed calls of other species. How do you know when you have a fragment?

22 Yes, it s a bird, but what type? We can often say we have bats, but can t say what kind.

23 Recording conditions can obscure details

24 Noise happens. Thermal convection, wind, clutter

25 Tadarida brasiliensis Echoes happen.

26 The quality and accuracy of recognition depends upon the quality of the recorded signals. Recording from the ground, near flat surfaces, or through tubes will contribute distortion.

27 Detail view of call from the sequence in the previous slide.

28 Really?

29 Acoustic equivalent of DNA contamination, or poor lab technique. Do we need a certification program?

30 Elevated microphones can listen up, and down toward the ground, and avoid distorting effects near the ground and surfaces.

31 Indiana bats vs. little brown bats.

32 Goal: Recognize presence or absence of Indiana bats. How can we test whether any of our acoustic methods accurately recognize Indiana bats? In particular, how can we assess our rate of false positives for Indiana bats? Our methods based on captured and tagged bat recordings. Do these work for real bats out in the wild? Like those we record passively?

33 We can do cross-validation with the data on which we build classification systems, but even better if we can test these systems against the real bats we record in the same way we do our passive field recording. If we could, we d have something of a super cross-validation for Indiana bat recognition.

34 So how can we test for false positives? Where we know we have no Indiana bats, but the others. Mylu Myso

35 Case study: Maine- beyond Myso range. Recordings made near a known little brown bat roost in southern Maine approximately 100 miles beyond the reported range of Indiana bats.

36 Case study: Maine- beyond Myso range. full-spectrum data acquired from three sites using Pettersson D500X detectors. Analyzed using SonoBat 3.1 NE and recordings converted to Anabat format using Myotisoft ZCANT for analysis using EchoClass 1.1.

37 Results Three recording sites yielded 112, 177, and 73 high frequency bat passes.

38 Results EchoClass reported twice as many Indiana than little brown bats at site one, 10 times as many at site two, and 1.7 times as many at site three. SonoBat reported 4% Indiana to 88% little brown bats at site one, and no Indiana bats at sites two and three.

39 Results, recording site near barn roost 99% likelihood 0% likelihood (5% Myso less than ~7% error rate)

40 Results, garden recording site 99% likelihood 0% likelihood (no consensus Myso)

41 Results Note large proportion of red bats in EchoClass results

42 Discussion Manual inspection of files revealed the 8% red bats files reported by SonoBat to be consistent with red bats SonoBat: Labo EchoClass: Labo

43 and the much larger proportion of Myotis sequences reported by SonoBat to be consistent with Myotis. Apparently most red bats files reported by EchoClass were Myotis files. SonoBat: Mylu EchoClass: Labo

44 Z-C transmogrifies little brown bats into red bats. Same call in full-spectrum and Z-C turn down = Myotis turn up = red bat Call trending by Z-C misses Myotis toes.

45 full-spectrum zero-crossing With strong signals and no confounding additional signals or noise, full-spectrum time-frequency trending and zero-crossing produce similar results.

46 Fully rendered time-frequency trend using full-spectrum data and processing. Call fragment rendered by Z-C processing of the same signal. noise Fully rendered time-frequency trends provide more confident and higher quality extraction of call characteristics (e.g., characteristic frequency, Fc). Higher quality data leads to better and more confident species discrimination.

47 (Results, part 2) Case study: NJ- Myso capture releases. (No Mylu in recordings, only Myso recordings.) SonoBat: 100% likelihood of Myso. (no consensus Mylu)

48 Conclusions Recording beyond the range of Indiana bats can provide a test for assessing the rates of false positives for Indiana bats using authentic field recording data. A preliminary test revealed that EchoClass produced sufficient false positives at all sites to conclude 99% likelihood of occurrence, despite no expected Indiana bats in the samples. EchoClass missed the majority of Myotis present, reducing the available sample to determine Myso presence.