1 Echolocation in bats and dolphins Chapter 64, pp 484-492
Transcription
1 Echolocation in bats and dolphins Chapter 64, pp 484-492
Echolocation in bats and dolphins Chapter 64, pp 484-492 WHO’S CALLING? - ACOUSTIC BAT SPECIES IDENTIFICATION REVISED WITH SYNERGETICS Obrist, Martin K.; Boesch, Ruedi; 1 Flückiger, Peter, F. and 2 Ulrich Dieckmann Swiss Federal Research Institute WSL, Zürcherstrasse 111, CH-8903 Birmensdorf, Switzerland 1 present address: Fledermausschutz Kt SO. c/o Naturmuseum Olten, CH-4600 Olten, Switzerland 2 Fraunhofer Institute for Integrated Circuits, Am Weichselgarten 3, D-91058 Erlangen, Germany Introduction Nocturnal activity, small size, and secretive roosting habits make species of the Microchiroptera among the most elusive of mammals. As bats constitute about 20% of mammalian species, the assessment of their occurrence and distribution should be part of any efforts to evaluate, conserve and monitor biodiversity, but bats are often overlooked. Some species of microchiropteran bats are however, conspicuous because they echolocate to detect, track, and assess airborne prey, usually insects. Since the early years of echolocation research (Griffin 1958) considerable knowledge and understanding has been collected about ecology, echolocation and physiology of bats (Busnel and Fish 1980; Fenton et al. 1987; Nachtigall and Moore 1986; Neuweiler 1999). Advancing technology allowed characterization of echolocation calls of many species. In addition a moderate financial investment in electro-technical tools allows interested laypeople to identify some species of bats by their distinctive echolocation calls. One complication is the fact that bats vary the structure of their calls (Kalko and Schnitzler 1993; Obrist 1995; Rydell 1993) and can emit signals similar to those of other species when facing comparable orientational tasks, such as flying close to clutter. The variety of echolocation calls that any one species of bat produces can lead to partial or complete overlap of the time and/or frequency structure of calls of 1 other species, making species identification impossible despite the use of very expensive tape recorders and analytical tools. In an extensive study on the variability and flexibility of bat echolocation, Obrist (1995) repeatedly discriminated signals not discernible by standard parameter measurements (highest, lowest, loudest frequency, duration, interval, etc.) only by the "pattern" or "shape" of the frequency contour in a spectrogram. Recent approaches to signal recognition tryed to holistically identify bioacoustic signals by pattern recognition algorithms, artificial neural networks, and methods like decision trees in different organisms like birds (Kogan and Margoliash 1998; Mills 1995), frogs (Taylor et al. 1996), whales (Mellinger and Clark 1997), other marine mammals (Fristrup and Watkins 1995), and in humans (Gish and Schmidt 1994). Few studies have tackled bats (Gannon et al. in press; Herr et al. 1997; O'Farrell et al. 1999). Automatic and robust acoustic species identification would be extremely helpful in a variety of research areas. A method capable of recognizing "patterns" of speciesspecific signals has the potential to identify environmental, behavioural, or individual differences in sonar or communicative sounds from whales to birds to crickets. Therefore, it would be a tremendous tool in scientific research and ecosystem monitoring. With this project, we wanted to identify all Swiss bat species by their sonar calls using signal features like variations in sweep rate or shape of sonograms. Even in such acoustically similar species as the members of the Myotis group, we predicted that a new pattern recognition approach should allow species recognition. The study had three goals: 1) find a suitable and economical method to record bat echolocation calls in the field, 2) evaluate a pattern recognition algorithm with the potential for fast and automated identification of bat species and optimize it for correct identification of Swiss bats, and 3) make the whole "recognizer" operable under field conditions, preferably in an unsupervised, automated mode. This text will focus primarily on the second topic. Material and Methods Recording Capture of animals All recordings were made in northern Switzerland between 28 June and 27 September 1995. From 12 species a total of 172 recorded sequences (equals 2 individuals, assuming extremely unlikely rerecording of individuals), comprising 816 calls, were included in the analysis (Table 64.1). Calls were recorded from either individual bats when emerging from "single-species" roosts or from bats captured at cave roosts. The latter recordings were made shortly before dawn when releasing the bats. We tried to minimize environmental noise (wind, leaves, echo clutter) by choosing an appropriate release location. Only the more species-specific search calls or commuting calls (Griffin 1958), starting a few seconds after release were recorded, avoiding the first few very short and rapid calls, typical of bats reorienting after release. Figure 2 shows waveforms and spectrograms of calls typical of the 12 species recorded. Signal acquisition All ultrasonic recordings were accomplished with a D140 ultrasonic bat detector (Lars Pettersson Electronics AB, Uppsala, Sweden). The detector's built-in electret microphone picks up the frequencies from 10 kHz to 120 kHz, but sensitivity drops progressively above 80 kHz. Using the available digital time expansion mode, signals of 0.87 s where sampled at 350 kHz with 8 bit resolution. Such samples contained between 3 (Nyctalus noctula ) and 15 (Myotis mystacinus) calls and are further referred to as a sequence. Replays performed at one-tenth of the original sampling rate were recorded using a Sony WM-D6C Walkman Professional for subsequent analysis in the laboratory. Direct recording To allow future field use of our setup, we also implemented software on a Macintosh PowerBook computer to make direct recording of ultrasound signals using a high speed analog-to-digital converter PCMCIA-card. Recording and pattern recognition algorithm were integrated in one software package, BATIT (BioAcoustic Taxa Identification Tool), and the provided interface to the Macintosh scripting language (AppleScript) allows one to implement an automatic acoustic "specieslogger". Conventional signal analysis Cassette recordings were digitized using an Apple Macintosh PowerPC 8500 with 22.05 kHz sampling rate at default 16-bit sampling depth. After extraction with the automated algorithm (see Preprocessing, below) signals submitted to the synergetic computer (SC, see below) also were analyzed with Canary software (Charif et al. 1995) to measure parameters from the sonograms (Table 64.1). Time 3 and frequency measurements (duration, DUR; lowest frequency, LFR; highest frequency, HFR; frequency of maximum energy, MFR) were taken from spectrographic representations of the calls. Spectrograms contained temporally overlapping windows, which possibly led to systematic, but minimal error in measuremnts of call durations (time-smearing) when compared to published signals of the species in question. However, analysing a spectrographic display was preferred over analysing an amplitude display and a spectrum display, because lowintensity parts of calls invisible in the latter, led to decreased readings for high frequency, low frequency and duration. Signal analysis with a synergetic computer Pattern recognition For the identification of bat calls we used the classical pattern recognition system. It consists of several modules (FIG. 64.1.). In general, the system is able to learn a representative set of calls from each species. Using these sets, the system assigns newly recorded calls of an unknown species to one of the learned species. The output consists of a position table on which the most likely species is on top followed by other possible species in decreasing order of probability. Preprocessing Call sequences stored in binary format with Canary or BATIT can directly be input to preprocessing and the classification algorithm of the SC. Recorded sequences show three main different parts: the call itself, the echo (not always present) and noise. Although the intercall interval or repetition rate can be speciesspecific in some bat populations (Fenton and Bell 1981), we neglected this parameter and concentrated on analyses of single vocalizations for the following reason: preliminary tests with short time Fourier transformations (STFT) of whole sequences showed bad recognition results because a) too much noise was included, b) the time interval between two consecutive calls often varied too widely, c) signal characteristics of echolocation calls did not resolve enough in the available spectrogram resolution (memory restrictions), d) long silent intervals dominated spectrograms of complete sequences, leading to overall less discernible segments. To analyze and classify, it was therefore necessary to select each call out from a sequence. Call extraction Peaks were identified in a signal previously smoothed with a 256-point moving 4 average filter. To ensure a good signal-to-noise-ratio and the detection of all suitable calls, only values 16.9 dB in amplitude (10*log(50)) above the average signal level were considered as peaks. This level was commonly encountered in quiet recording situations, or achieved in noisy recordings by high-pass filtering above 10 kHz. After detection of a peak, 3072 samples before and 5120 samples after the peak's position were extracted, thereby creating short signals of 8192 samples. Separating the actual call from its echo was attempted in some species (e.g. Nyctalus noctula in FIG. 64.1). All modes of manipulation such as filling in with noise or cutting before echo introduced artifacts. This led us to leave extracted signals untouched. Therefore, the extracted calls often contained a considerable amount of interfering echos from ground- or water reflections, and ambient noise, wind, Doppler shifts and other interferences (Pye 1993). Fourier Transformation After detection, the STFT for each call was calculated (amplitude spectra = sqrt(real*real + imag*imag), 256 samples, Hamming window, 81% overlap). These calculated windows were composed to one feature vector of 20352 floating point numbers normalized to 1, which is equivalent to a spectrographic display of 159 spectra each containing 128 data points (see FIG. 64.2). Phase was therefore neglected from analysis and only magnitude considered. Synergetic Computer Synergetics is an interdisciplinary field which deals with self-organizational phenomena in nature (Haken 1978; Kohonen 1984). These phenomena have in common that many microscopic parts in an unsorted order (chaos) transform themselves in a sorted order. The importance of each part is minor; only the properties of the whole system are relevant and can be described through synergetic differential equations. The synergetic computer (SC) is not an actual computer per se, but a new set of algorithms emanating from this interdisciplinary field, and has only recently been used for classification tasks (Haken 1988; Haken 1996; Wagner et al. 1993; Wagner et al. 1995). For the classification of bat calls we used an algorithm termed SCMELT (Dieckmann 1997). One significant advantage of this algorithm is its ability to combine several training patterns per class into one feature vector without losing any information about the training patterns. The training patterns are melted into one prototype. The prototype has the same dimension as the training vectors and is normalized to length 1. Because of this ability the SC can handle big dimensions in 5 contrast to artificial neural networks (ANN). ANN also can handle big dimension, but the computational power needed to train an ANN with input vector of 16384 features is prohibitive. Description of the synergetic computer using adjoint prototypes (SCAP) are given in Hogg (1996) and Wagner (1993). The adjoined prototypes of the SC-MELT are achieved through simple addition of the corresponding, adjoined prototypes. One of the most interesting properties of the SC is that it emphasizes pattern contents that are unique among all others and it diminishes of pattern contents common to all others. The learning time of the SC is easily determined and very fast. The generation of the prototypes of the bat calls takes only a couple of minutes on a 200 MHz 603e RISC processor (such as an Apple Macintosh PowerBook 3400). The classification is even faster because it is simply a scalar or dot product. Classification Assume we let the SC training 3 calls of each of 10 species (classes), resulting in 10 prototype feature vectors. We now test 50 echolocation calls of an unknown species contained in the training base. The SC computes the scalar product of each test call with each class prototype, resulting in 10 values per call, varying between 1 (identical call as in training base) and 0 (no resemblance to any of the training calls). The training class with the highest scalar product identifies the species where the calls most likely came from. Some signals invariably get assigned to wrong species, resulting in a frequency distribution over the 10 classes. The peak in the frequency distribution finally identified the best fitting species (FIG. 64.4 A). Dependencies of recognition success To evaluate the effect of training call selection on recognition success, we repeated the training and classification process 115 times with our data set. Calls included in the training base were drawn from 1 to 9 sequences. The number of calls per sequence was varied from 1 to 5. Each combination of sequence-call selection was repeated 5 times with different sequences and calls each time. These selection criteria were restricted by available computer memory, not allowing for more than 12 calls per species (144 training calls). For two species (Myotis blythii and Myotis bechsteini) we had to use identical training bases repeatedly due to a lack of recordings. For each of the resulting 115 repetitions, we counted for every species the number of correctly assigned calls, the number of misclassified calls and whether the correct species was the most common classification or not. These classification results were analyzed by multivariate analysis of variance (MANOVA) for the effect 6 of number of sequences, number of calls, and number of calls per sequence on recognition success. As described above, the higher the maximum scalar product, the clearer the recognition of a tested call. Similarly, a great difference between the highest and second-highest scalar product indicated a clear differentiation between classes. Both parameters, maximum scalar product (MSP) and difference to second-best (DSB) were used to delimit the number of valid calls. We systematically tested all combinations of the MSP (0, 0.4, 0.5, 0.6) and the DSB (0, 0.1, 0.2, 0.3), to reject calls from the classification. Similar to a rejection criterion like signal-to-noise-ratio, this eliminated a number of calls from the original base of calls, thereby changing the number of cases in subsequent analyses. Statistical analysis Using DataDesk software (Data Description Inc., Ithaca, New York, USA) we performed a linear regression analysis to test whether the increase in delimiting parameters (MSP, DSB) improved classification. The influence of species, number of learned calls, and number of sequences (individuals) on the classification was tested using ANOVA. To compare the identification abilities of the SC, we performed multivariate analyses on the four parameters measured with Canary. We used DataDesk for a MANOVA with subsequent post-hoc tests (Sheffé) to separate means of species’ call characteristics. With the same software we performed a single linkage cluster analysis. SAS (SAS Institute Inc., Cary, NC, USA) statistical software was used for a discriminant function analysis with reclassification for a cross-validation. A significance level of p ≤ 0.05 was applied in all statistical tests. Results Cluster analysis The cluster analysis clearly separated Nyctalus noctula from the other species due to the bandwidth and the duration of its calls (compare spectrograms in FIG. 64.1). Pipistrellus pipistrellus also stands apart from the other species, but less clearly. Vespertilio murinus, Eptesicus serotinus, and Plecotus auritus are separated from the Myotis (FIG. 64.3). Multivariate analysis of variance MANOVA showed significant variance in all four call parameters. The Scheffé post-hoc test separated most species by one or more of their mean call characteristics, 7 most often by MFR and LFR followed by HFR and DUR, reflecting increasing variability (CV) in these parameters (Table 64.1). However, Myotis mystacinus, M. bechsteini, M. blythii, and M. brandti were difficult to separate just by the measurements taken with Canary. The reason remains unclear for the first species, but it might be the relative small sample size available from the latter three species (5, 16 and 19 calls respectively; see Table 64.1). Discriminant function analysis Using only parameters in Table 64.1, discriminant component analysis properly identified 46% of all Myotis-calls and 60% of all species’ calls, when splitting the data pool in half and testing the second half. Considering only non-Myotis species, a recognition success of 82% was achieved. Splitting the data pool in four parts and testing the remaining three quarters of the signals recognition success deteriorated to 45% (Myotis group), 58% (all species) and 78% (non-Myotis species). Synergetic computer Using the synergetic computer to classify patterns of spectrograms of echolocation calls was surprisingly effective. In some cases a single call was sufficient to characterize the species, as seen in FIG. 64.4. However, most species showed considerable variations in call contour (see CV in Table 64.1). Accordingly, at least five, optimally chosen vocalizations had to be learned from every species to cover a reasonable part of the flexibility and make classification reliable and results consistent. We performed a variety of exploratory statistical analyses to evaluate the influences of different variables on classification. In the remaining statistics Myotis bechsteini, Myotis blythii and Myotis brandti were not considered, because too few recordings from these species were available for statistical analysis. As expected, ANOVA indicated a highly significant influence of species on the classification success. Averaged over 115 trials with 1 to 9 training calls per species, Nyctalus noctula scored best with 75% correct classifications while calls of Myotis daubentoni had a correct classification rate of only 32%. Example results of tests with 5 and 9 training calls are given in Table 64.2. Classification success (percentage of correctly classified calls) increased with the number of calls in the training base, except in Nyctalus noctula, where classification success reached a plateau at 100% with only 1 training call. The general impression, that the increase in success rate started to level out towards 9 training calls per species could not be further investigated due to lack of recordings and computer memory 8 restrictions (FIG. 64.5, FIG. 64.6). In all species, the recognition success was dependent on the number of sequences included in the training base, probably due to the inclusion of more variability occurring among recordings of different individuals. Increasing the number of training calls drawn per sequence leads to a marginal increase in recognition success, underlining high inter-individual and lower intraindividual variability in echolocation calls (Obrist 1995). Recording and hardware restrictions again prohibited further analysis; it was not known wether increasing the number of calls per sequence would saturate or even decrease recognition success at higher training call numbers. We used the maximum scalar product (MSP) and the difference of the secondbest scalar product (DSB) to evaluate a combination of both which achieves highest classification results with lowest rejection of calls. For all species, any increase in MSP (0, 0.4, 0.5, 0.6) and any increase in DSB (0, 0.1, 0.2, 0.3) resulted in a significantly higher recognition success (p ≥ 0.001; calls of Myotis blythii and Myotis bechsteini not considered), at the cost of rejecting a number of calls (FIG. 64.5, FIG. 64.6). While 57% of all calls were correctly classified (0% call rejection) when MSP and DSB were both set to 0, average recognition rate increased to 62% (with 23% rejection) for MSP > 0.4 and DSB > 0.1. Even more (71%) were correctly classified with MSP > 0.6 and DSB > 0.3, but with these settings 56% of all calls where rejected. Judging from these results, we considered a combination of MSP > 0.5 and DSB > 0.2 as adequately delimiting rejection criteria for a call, offering the best compromise between recognition success and call rejection. Using 9 training calls per species and applying the above rejection criteria, we reached an average recognition rate of 80% at the cost of rejecting 33% of all detected signals. For reasons of call similarity (FIG. 64.1), this result was slightly different between Myotis and nonMyotis species. Increasing scalar product criteria further lead to a loss of more than half the data recorded. Discussion Using a synergetic computer for pattern recognition, we identified calls of 12 bat species from spectrograms of their sonar calls with variable, mostly high, degrees of success. Recognition success varied with species and with the size and variety of the training base. Applying various restrictions helped to improve classification results. Classical recognition tasks Since Griffin first reviewed bat echolocation (1958), species have been known to 9 differ in the structure of their vocalizations. Many publications have since dealt with the identification of echolocating bats by their calls (Ahlén 1981; Fenton and Bell 1981; Vaughan et al. 1997b; Zingg 1990). Despite the fact that Griffin (1958) observed plasticity in call structures of individual bats, this topic only recently has received extensive attention (Betts 1998; Obrist 1995; Rydell 1993; Zbinden 1989). Variability is one problematic aspect of acoustic monitoring, the training of qualified personnel with detectors yet another. The intensity of vocalizations can vary within 30-50 dB between bat species. Therefore, some species will invariably be more conspicuous than others and are over-represented in any acoustic monitoring study. Comparing mist-net catches and acoustically recorded bat passes can show corresponding activity patterns in temperate regions (Kunz and Brock 1975) but this seems not to be the case in southern Africa (Rautenbach et al. 1996). Nevertheless, acoustic monitoring coupled with the ability to distinguish species, makes monitoring studies feasible (Fenton 1970; Hickey and Neilson 1995), at least in temperate regions and when counting all Myotis as a single species. Several thorough studies proved the potential of acoustic monitoring of habitat use by bats, as well as its limitations (e.g. Furlonger et al. 1987; Vaughan et al. 1997a). Limitations include the necessity to have trained personnel spend long nights in the field, and the inability to separate species within the genus Myotis or similarly calling groups (e.g. Phyllostomatidae) in the field. Myotis contribute significantly to biodiversity: nine out of 25 bat species (36%) occurring in Switzerland (Hausser 1995) belong to Myotis. Only recently was an acoustic identification of Myotis-species by statistical analysis achieved (Vaughan et al. 1997b). Pattern recognition approaches In the last decade there has been an increased interest in automatic recognition of acoustic patterns. Human speaker identification is most prominent and voluminous in literature (Gish and Schmidt 1994), and has mainly commercial aspects, in contrast to the identification of species (Mills 1995; Taylor et al. 1996) or behaviours (Mellinger and Clark 1997; Potter et al. 1994). Species recognition approaches generally use decision trees (Quinlan 1993) or artificial neural networks. Synergetic recognition algorithms have so far been applied to speaker recognition (acoustic and visual, Dieckmann 1997) or industrial part identification tasks (Wagner et al. 1993) with impressive results. The ability to learn test patterns unsupervised and to classify full spectrograms at high rates makes synergetics a promising technique for animal vocalization recognition. 10 Statistical analyses can reach high recognition rates for non Myotis speciesVespertilionids. Zingg (1990) achieved 86% using 5 signal parameters, Vaughan et al. reached 89% (1997b, 67% for Myotis) using 6 parameters, including the call interval. Using only 4 call measures, we achieved comparable recognition (82%) with the discriminant function analysis. For the genus Myotis, performance droped below 50% in our tests. However, the a–priori probability of identifying the proper species by chance is only 8%. The success rate of the SC compares favorably to these numbers. When carefully choosing the learning base and rejecting ambiguous calls, the SC reached average call recognition rates of 77% for the genus Myotis and 84% for other Vespertilionids (average 80%). It is possible to improve these results further. First, digital recording of the ultrasound as 12-bit data will substantially improve signal quality. Second, the preparation of the spectrogram (window size and shift) could be further optimized for the SC, and, with more available RAM we will be able to train the SC with more species, and better evaluate the effect of increasing call numbers per species. Finally, the inclusion of the intercall interval in the species’ characterization could further improve performance. Based on results from studies using similar algorithms in industry (99% recognition, Wagner et al. 1993) and "human" applications (93% recognition, Dieckmann 1997), and taking into consideration the substantial variability occurring in echolocation (Obrist 1995), we hope to reach an overall bat species recognition success of approximately 90%. Only then, do we intend to make the SC system available to fellow scientists. For the characterization of, and the quantitative comparison among species or behaviorally specific call characteristics, standard statistical analysis of measured call parameters remains the only choice. But, for monitoring purposes, the pattern recognition approach rivals this method and certainly is faster and more economical. We estimate a total throughput (cut calls, calculate spectrograms, classify, write output) of 5-6 calls per second on a current Macintosh laptop. As this is roughly the rate at which an average bat emits vocalizations the system approximates real time performance, but lags behind real time by the duration of the recording plus the processing time. If none or only some of the signals are kept for later reference, the system could autonomously log species-specific bat activity for a number of nights, powered by a portable 12 V battery. With falling prices of increasingly powerful laptop computers and the availability of high-speed PCMCIA data acquisition cards, direct field recording of ultrasound to computer hard disc will become economic and widespread in the very near future. 11 Our proposed system constitutes a so-called "black box". The user inputs signals and the machine identifies the emitter of the sounds. Many field workers will object to this approach because no control can be attained over species classification. We do not encourage inexperienced people to feed data into our system without prior knowledge of bat echolocation calls and inherent variability. Special care has to be taken when compiling a valid training base for the synergetic computer. Considering geographic variations of sonar characteristics, compilation of a single training base for a certain species will likely never be feasible. Rather, a variable approach will be more realistic, where qualified observers compile their own database for their regional coworkers. However, our system is attractive because of its ease of operation, cost effectiveness, and the potential for automated monitoring of any acoustically conspicuous species assemblage. Acknowledgments We are very grateful to Lars Pettersson for his generous help with recording hardware. Literature cited AHLÉN, I. 1981. Identification of Scandinavian bats by their sounds. The Swedish University of Agricultural Sciences. Department of Wildlife Ecology, Uppsala, Report 6, 56 pp. 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. BUSNEL, R. G., AND J. F. FISH. 1980. Animal Sonar Systems. Plenum Press, New York, London, 1135 pp. CHARIF, R. A., S. G. MITCHELL, AND C. W. CLARK. 1995. Canary 1.2 users manual. Cornell Laboratory of Ornithology, , Report . DIECKMANN, U. 1997. SESAM: A biometric person identification system using sensor fusion. Pattern Recognition Letters, 18:827-833. FENTON, M. B. 1970. A technique for monitoring bat activity with results obtained from different environments in southern Ontario. Canadian Journal of Zoology, 48:847-851. FENTON, M. B., AND G. P. BELL. 1981. Recognition of species of insectivorous bats by their echolocation calls. Journal of Mammalogy, 62:233-243. FENTON, M. B., P. RACEY, AND J. M. V. RAYNER. 1987. Recent advances in the study of bats. University Press, Cambridge, 470 pp. 12 FRISTRUP, K. M., AND W. A. WATKINS. 1995. Marine mammal sound classification. The Journal of the Acoustical Society of America, 97:3369. FURLONGER, C. L., H. J. DEWAR, AND M. B. FENTON. 1987. Habitat use by foraging insectivorous bats. Canadian Journal of Zoology, 65:284-288. GANNON, W. L., M. J. O'FARRELL, C. CORBEN, AND E. J. BEDRICK. (2004). Call Character Lexicon and Analysis of Field Recordings of Bat Echolocation Calls. Pp. 478-484 in Advances in the Study of Echolocation in Bats and Dolphins, Proceedings of the Biosonar Conference 1998 (Thomas, J., C. Moss, and M. Vater eds.). University of Chicago Press, Chicago. GISH, H., AND M. SCHMIDT. 1994. Text-independent speaker identification. IEEE Signal Processing Magazine, 10:18-32. GRIFFIN, D. R. 1958. Listening in the Dark. The Acoustic Orientation of Bats and Men. Yale University Press, New Haven. (1986 reprint by Cornell University Press, Ithaca, New York), 415 pp. HAKEN, H. 1978. Synergetics. An introduction. Nonequilibrum phase transitions and self-organization in physics, chemistry and biology. Springer, Berlin, Heidelberg, New York, 355 pp. HAKEN, H. 1988. Learning in synergetic systems for pattern recognition and associative action. Zeitschrift für Physik B, 71:521-526. HAKEN, H. 1996. Future trends in synergetics. Nonlinear physics of complex systems. Current status and future trends. Springer, Berlin, 388 pp. HAUSSER, J. 1995. Säugetiere der Schweiz: Verbreitung, Biologie, Ökologie. Birkhäuser, Basel, 501 pp. HERR, A., N. I. KLOMP, AND J. S. ATKINSON. 1997. Identification of bat echolocation calls using a decision tree classification system. Complexity International, 4, URL http://www.csu.edu.au/ci/vol4/herr/batcall.html, last update Jan. 1997, last verfied 15.05.2000. HICKEY, M. B. C., AND A. L. NEILSON. 1995. Relative activity and occurrence of bats in southwestern Ontario as determined by monitoring with bat detectors. Canadian Field - Naturalist, 109:413-417. HOGG, T., AND I. TALHAMI. 1996. A competitive non-linear approach to object recognition: the generalised synergetic algorithm. Pp. 47-50, in Proccedings of the Australian New Zealand Conference on Intelligent Information Systems. IEEE, New York, NY, USA, Adelaide, South Australia. KALKO, E. K. V., AND H.-U. SCHNITZLER. 1993. Plasticity in echolocation signals of European pipistrelle bats in search flight: implications for habitat use and prey 13 detection. Behavioral Ecology and Sociobiology, 33:415-428. KOGAN, J. A., AND D. MARGOLIASH. 1998. Automated recognition of bird song elements from continuous recordings using dynamic time warping and hidden Markov models: A comparative study. The Journal of the Acoustical Society of America, 103:2185-2196. KOHONEN, T. 1984. Self-organization and assoziative memory. Springer-Verlag, Berlin, 255 pp. KUNZ, T. H., AND C. E. BROCK. 1975. A comparison of mist nets and ultrasonic detectors for monitoring flight activity of bats. Journal of Mammalogy, 56:907911. MELLINGER, D. K., AND C. W. CLARK. 1997. Methods for automatic detection of mysticete sounds. Marine and Freshwater Behaviour and Physiology, 29:163181. MILLS, H. 1995. Automatic detection and classification of nocturnal migrant bird calls. The Journal of the Acoustical Society of America, 97:3370. NACHTIGALL, P. E., AND P. W. B. MOORE. 1986. Animal Sonar. Processes and performance. Plenum Press, New York, 862 pp. NEUWEILER, G. 1999. Biology of Bats (translated by Covey, E.). Oxford University Press, New York, Oxford, 304 pp. O'FARRELL, M. J., B. W. MILLER, AND W. L. GANNON. 1999. Qualitative identification of free-flying bats using the Anabat detector. Journal of Mammalogy, 80:11-23. OBRIST, M. K. 1995. Flexible bat echolocation: the influence of individual, habitat and conspecifics on sonar signal design. Behavioral Ecology and Sociobiology, 36:207-219. POTTER, J. R., D. K. MELLINGER, AND C. W. CLARK. 1994. Marine mammal call discrimination using artificial neural networks. The Journal of the Acoustical Society of America, 96:1255-1262. PYE, D. 1993. Is fidelity futile? The true signal is illusory, especially with ultrasound. Bioacoustics, 4:271-286. QUINLAN, J. R. 1993. C4.5: programs for machine learning. Morgan Kauffman, San Mateo, California, USA, 302 pp. RAUTENBACH, I. L., M. B. FENTON, AND M. J. WHITING. 1996. Bats in riverine forests and woodlands: A latitudinal transect in southern Africa. Canadian Journal of Zoology, 74:312-322. RYDELL, J. 1993. Variation in the sonar of an aerial-hawking bat (Eptesicus 14 nilssonii). Ethology, 93:275-284. TAYLOR, A., G. GRIGG, G. WATSON, AND H. MCCALLUM. 1996. Monitoring frog communities: An application of machine learning. Pp. 1564-1569, in Proccedings of the Thirteenth National Conference on Artificial Intelligence and the Eighth Innovative Applications of Artificial Intelligence Conference. American Association for Artificial Intelligence, Menlo Park, Portland, Oregon. VAUGHAN, N., G. JONES, AND S. HARRIS. 1997a. Habitat use by bats (Chiroptera) assessed by means of a broad-band acoustic method. Journal of Applied Ecology, 34:716-730. VAUGHAN, N., G. JONES, AND S. HARRIS. 1997b. Identification of British bat species by multivariate analysis of echolocation call parameters. Bioacoustics, 7:189-207. WAGNER, T., F. G. BOEBEL, U. HASSLER, H. HAKEN, AND D. SEITZER. 1993. Using a synergetic computer in an industrial classification problem. Pp. 206-212, in Proccedings of the International Conference on Artificial Neural Nets and Genetic Algorithms. Springer-Verlag, Berlin, Innsbruck, Austria. WAGNER, T., U. SCHRAMM, AND F. G. BOEBEL. 1995. Synergetic learning for unsupervised texture classification tasks. Physica D, 80:140-150. ZBINDEN, K. 1989. Field observation on the flexibility of the acoustic behaviour of the European bat, Nyctalus noctula (Schreber, 1774). Revue Suisse de Zoologie, 96:335-343. ZINGG, P. E. 1990. Akustische Artidentifikation von Fledermäusen (Mammalia: Chiroptera) in der Schweiz. Revue Suisse de Zoologie, 97:263-294. 15 TABLE 64.1._ _ _ Recording statistics. Abbreviation of species names (ab), sample size n (seq = number of sequences; call = number of calls included in analysis) and call characteristics of the 12 recorded species. Mean ± standard deviation, as well as coefficients of variation (cv), are given for the following parameters: DUR = call duration, LFR = lowest frequency in call, MFR = frequency of main (highest) energy in call, and HFR = highest frequency in call. Species indicated with an asterisk have low sample sizes. Call parameter n DUR (ms) SPECIES ab seq call mn±sd Eptesicus serotinus ES 21 80 7.2 ±1.7 Myotis bechsteini * MB 1 5 Myotis blythii * ML 3 Myotis brandti * MR Myotis daubentoni cv LFR (kHz) MFR (kHz) cv HFR (kHz) mn±sd cv mn±sd mn±sd cv 24% 25.5 ±1.3 5% 30.3 ±2.1 7% 57.0 ±4.4 8% 4.1 ±0.1 3% 23.2 ±0.9 4% 45.2 ±1.4 3% 96.5 ±4.5 5% 16 2.8 ±0.9 32% 27.7 ±6.0 22% 41.4 ±5.1 12% 72.0 ±14.5 20% 3 19 3.4 ±0.7 22% 27.3 ±1.5 6% 43.4 ±2.8 6% 83.1 ±11.3 14% MD 32 146 3.8 ±0.8 20% 27.2 ±2.0 7% 42.2 ±2.8 7% 76.1 ±10.8 14% Myotis myotis MM 32 155 4.0 ±1.2 30% 23.7 ±2.6 11% 37.0 ±3.6 10% 77.1 ±13.3 17% Myotis mystacinus MY 10 60 3.1 ±0.9 30% 27.8 ±2.7 10% 45.0 ±4.0 9% 85.2 ±9.0 11% Myotis nattereri MN 27 149 3.0 ±0.8 28% 17.4 ±4.6 26% 33.7 ±7.0 21% 78.0 ±14.4 18% Nyctalus noctula NN 12 24 13.0 ±3.0 24% 21.5 ±1.4 6% 24.2 ±1.7 7% 32.5 ±8.6 26% Pipistrellus pipistrellus PP 10 43 5.4 ±0.9 16% 42.5 ±1.0 2% 45.7 ±1.4 3% 84.6 ±8.6 10% Plecotus auritus PA 15 90 3.3 ±0.9 27% 23.0 ±1.8 8% 33.3 ±4.4 13% 53.2 ±2.8 5% Vespertilio murinus VM 6 29 5.5 ±1.8 33% 22.5 ±1.0 4% 28.7 ±2.5 9% 49.0 ±5.3 11% TABLE 64.2._ _ _ Average percentage of correctly classified calls relative to the total number of calls passing the delimiter restrictions. Results are shown for 5 and 9 training calls with delimiter restrictions MSP > 0.5 and DSB > 0.2. The last column indicates the percentage of all calls rejected by the delimiter restrictions. Species indicated with an asterisk have too low sample sizes to be conclusive (see also Table 64.1). % correctly classified Nr of training calls MSP / DSB % rejected 5 9 9 0/0 0.5 / 0.2 0.5 / 0.2 Species Eptesicus serotinus (ES) 42 60 34 Myotis bechsteini (ME) * 74 94 77 Myotis blythii (ML) * 64 100 3 Myotis brandti (MR) * 31 90 18 Myotis daubentoni (MD) 18 42 43 Myotis myotis (MM) 16 53 51 Myotis mystacinus (MY) 34 64 42 Myotis nattereri (MN) 39 93 30 Nyctalus noctula (NN) 71 98 17 Pipistrellus pipistrellus (PP) 60 93 16 Plecotus auritus (PA) 36 80 42 Vespertilio murinus (VM) 58 90 19 Avg. Myotis 39 77 38 Avg. non-Myotis 53 84 26 Average % 45 80 33 Figure captions FIG. 64.1._ _ _ Sonograms and waveform displays of representative calls of the 12 bat species under consideration. Calls are arranged according to the apparent similarity of their spectrograms. FIG. 64.2._ _ _ Schematics of the data processing, using synergetic computer explained in text. FIG. 64.3._ _ _ Single linkage cluster analysis of species according to mean call parameters DUR, LFR, MFR and HFR. FIG. 64.4._ _ _ Histogram of classification results: the percentage of calls assigned to the various species are indicated. In this run, 82 of 148 test calls recorded from Myotis nattereri (55%) were assigned to the correct species. Only 1 call from each species was used as training base. Species abbreviations as in Table 64.1. FIG. 64.5._ _ _ Three example results of the classification of 2 species‘ calls (Eptesicus fuscus and Myotis myotis) with the synergetic computer. 1, 5 and 9 calls were used to train the algorithm. When including 5 or more calls in the training base, the maximum identification score is always (also in the other 10 species not shown here) at the correct species (horizontally striped bars). Misclassified calls (vertically striped bars) most often fall into classes of species, which proved very similar in the cluster analysis (FIG. 64.3). Species abbreviations as in Table 64.1. FIG. 64.6._ _ _ The effect of varying numbers of training calls (all taken from different sequences), and delimiter values (MSP, DSB) on classification performance and rejection of calls. Due to the limited number of recorded sequences and computer memory restrictions, N varies: 25, 10 and 5 runs are averaged for 1, 5 and 9 training calls respectively. Frequency [kHz] 100 Nyctalus noctula Ves pertilio murinus E ptes icus s erotinus P ipis trellus pipis trellus P lecotus auritus Myotis brandti Myotis myotis Myotis nattereri 50 0 Myotis daubentoni Myotis bechs teini Myotis mys tacinus Myotis blythii Amplitude [V] +0.5 0 -0.5 0 5 10 Time [ms] 15 Fig. 64.1 pattern recognition system preproces s ing smooth signal s ynergetic computer locate peaks Calling bat data recording classifier extract single calls short-time Fourier Transformation combine windows feature extraction learning species identity prototypes Fig. 64.2 Calls from Myotis nattereri(MN) 80 60 40 20 VM PP PA NN MY MR MN MM ML ME MD 0 ES Percent classification 100 Species calls were assigned to Fig. 64.4 Calls from Eptesicus serotinus (ES) Percent classificatoins 100 80 60 40 20 PP VM PP VM PA NN MY MR MN MM ML ME MD ES 0 Species, to which calls were assigned to Calls from Myotis myotis (MM) 80 60 40 20 PA NN MY MR MN MM ML ME MD 0 ES Percent classificatoins 100 Species, to which calls were assigned to Percent correct classification Percent wrong classification 1 learn-call per species 5 learn-calls per species 9 learn-calls per species, with scalar product restrictions Fig. 64.5 no delimiters delimiters MSP > 0.5, DSB > 0.2 reject Percent correct identified / rejected 100 80 60 40 20 0 1 5 number of training calls 9 Fig. 64.6