Peter F. FLÜCKIGER Bat Protection Kt SO. c/o Museum of Natural History Olten, Kirchgasse 10, CH-4600 Olten (Switzerland) - PDF

Variability in echolocation call design of 26 Swiss bat species: consequences, limits and options for automated field identification with a synergetic pattern recognition approach Martin K. OBRIST Ruedi

Please download to get full document.

View again

of 16
All materials on our website are shared by users. If you have any questions about copyright issues, please report us to resolve them. We are always happy to assist you.

Social Media

Publish on:

Views: 28 | Pages: 16

Extension: PDF | Download: 0

Variability in echolocation call design of 26 Swiss bat species: consequences, limits and options for automated field identification with a synergetic pattern recognition approach Martin K. OBRIST Ruedi BOESCH Swiss Federal Research Institute WSL, Research Department Landscape CH-8903 Birmensdorf (Switzerland) Peter F. FLÜCKIGER Bat Protection Kt SO. c/o Museum of Natural History Olten, Kirchgasse 10, CH-4600 Olten (Switzerland) Obrist M. K., Boesch R. & Flückiger P. F Variability in echolocation call design of 26 Swiss bat species: consequences, limits and options for automated field identification with a synergetic pattern recognition approach. Mammalia 68 (4): KEY WORDS Echolocation, pattern recognition, Chiroptera, monitoring, Switzerland. ABSTRACT Pattern recognition algorithms offer a promising approach to recognizing bat species by their echolocation calls. Automated systems like synergetic classifiers may contribute significantly to operator-independent species identification in the field. However, it necessitates the assembling of an appropriate database of reference calls, a task far from trivial. We present data on species specific flexibility in call parameters of all Swiss bat species (except Nyctalus lasiopterus and Plecotus alpinus). The selection of training-calls for the classifier is crucial for species identification success. We discuss this in the context of echolocation call variability differing between species and its consequences for the implementation of an automated, species specific bat activity monitoring system. Publications Scientifiques du Muséum national d Histoire naturelle, Paris. 307 Obrist M. K. et al. MOTS-CLÉS Echolocation, signaux, Chiroptera, reconnaissance automatique, Suisse. RÉSUMÉ Variabilité des signaux d écholocation de 26 espèces de chauves-souris suisses : conséquences, limites et options pour un système automatisé d identification avec une approche de reconnaissance synergique. Les algorithmes de reconnaissance de patron représentent une approche prometteuse pour identifier les espèces de chauves-souris à partir de leurs signaux d écholocation. Des systèmes automatisés tels que les classifications synergiques peuvent contribuer à une identification sur le terrain indépendamment de l expérience de l opérateur. Cependant il est nécessaire de constituer une solide base de données de signaux de référence, une tâche délicate. Cet article présente des données sur la variabilité des paramètres des signaux de toutes les espèces de chauves-souris suisses (à l exception de Nyctalus lasiopterus et Plecotus alpinus). La sélection des «signaux d apprentissage» est déterminante pour une bonne identification des espèces. Ceci est discuté dans le contexte d une variabilité des signaux d écholocation, différente selon les espèces. Les conséquences de ces résultats sont évaluées dans la perspective de réalisation d un système automatisé de suivi d activité spécifique des chauves-souris. INTRODUCTION Due to their active orientation system, echolocating bats are conspicuous. Therefore, by acoustically monitoring bats we can eavesdrop on their behaviour. In the course of evolution, bats have occupied the nocturnal niche of almost every possible habitat. In parallel with their evolution, wing morphology, hearing abilities and echolocation call characteristics adapted to the specific habitats and foraging method (Neuweiler 1999). Today signal variability covers constant frequency and frequency modulated calls of various composition and temporal structure. Additionally, bats adapt their calls to the particular situation, e.g., using short, broad-band signals at high repetition rate when hunting close to clutter, or long narrow-band signals when foraging in open air space (Kalko 1995). Despite this flexibility, recognition of bat species by their calls can be tackled. This is an old theme with new variations. In 1958 already, Griffin described echolocation calls differing between species (Griffin 1958). But it was not until 23 years later, when the first publications dedicated to acoustic bat species identification where published by Ahlén (1981), and Fenton & Bell (1981). The topic gained new momentum in recent years in the light of affordable biodiversity monitoring and conservation. Advances in technology further facilitate registering and analysing of ultrasound signals (Parsons and Obrist 2004). Different approaches can be taken to quantitatively or qualitatively analyse echolocation signals (Parsons et al. 2000). Heterodyning detectors look at a narrow frequency band. Calls with differing energy content or different temporal sweep structure in this frequency band will sound differently in such a detector (Fig. 1). Set to 45 khz, a Pipistrellus pipistrellus will be easily detected as strong plop, while Vespertilio murinus will probably not be heard at all and Eptesicus serotinus might just be audible as weak tick. Species with short, broadband signals will mostly tick very similar, possibly with different intensity depending on the energy 308 Variability in echolocation call design of 26 Swiss bat species Eptesicus serotinus Vespertilio murinus Pipistrellus pipistrellus bechsteinii myotis mystacinus 115 khz 35 ms FIG. 1. Symbolic representation of the acoustic filter which a heterodyning bat detector implies. Echolocation calls of six species are given as spectrograms. The central black rectangle enframes the few khz around 45 khz, at which the detector is tuned. Eptesicus serotinus will be heard as weak tick, Vespertilio murinus will not be heard at all and Pipistrellus pipistrellus will sound as full tock. The three species will all sound very similarly as loud short tick. myotis nattereri 156 khz 25 ms FIG. 2. Spectrographic (dark grey) and zero-crossing (white dots) representation of echolocation calls of myotis and nattereri. Period-plots (zero-crossing displays) do not show harmonics and dots may jump between harmonics of similar intensity (see middle part of call of nattereri). The display becomes random for low intensity signals (beginning of calls). The black frame indicates parameters measured in Canary for statistical analysis: duration, highest and lowest frequency. The intersection of the black cross indicates the measurement of the frequency of peak energy. maximum of their calls and their repetiton rates. Interpretation and valuation of all these parameters heavily depend on the experience of the observing person. A visual approach to echolocation call identification can be taken by either calculating spectrograms or period-plots (Fig. 2). The latter is achievable in real time with a limited amount of hardware (e.g., ANABAT system, Titley Electronics 1998; O Farrell et al. 1999a), while spectrogram calculation usually necessitates a fast computer. However, period- or zero-crossing plots do perform badly with weak or noisy signals, blurring the display considerably (Fig. 2). This complicates the recognition in some situations and species, and their use is discussed very controversially (Barclay 1999; O Farrell et al. 1999b; Fenton 2000; Fenton et al. 2001). 309 Obrist M. K. et al. If real-time performance is not required, offline analysis offers more accurate methods to species recognition. It involves in most cases interaction with an operator, who measures in the amplitude or the spectrogram display critical parameters like highest and lowest frequency, frequency of main energy and duration of the signal (Fig. 2, black frames with cross). With these discrete measurements of spectral and temporal parameters, further analyses with statistics, artificial neural networks or decision trees can be performed, giving good performance (Herr et al. 1997; Parsons & Jones 2000; Russo & Jones 2002; Obrist et al. 2004). Still, in many cases, measurements will be identical for calls from different species. In some of these cases, the spectrogram may show characteristic signatures in the shape of the frequency sweep. Here, a pattern recognition approach using shape analysis may help to discriminate signals. In a pilot study, Obrist et al. (2004) had analyzed echolocation signals of 12 species. With a less advanced recording setup and a limited data set, they had reached average recognition rates of 80%. Here we report on data of 26 species, which were all recorded directly to digital media. We tested the hypothesis, that a pattern recognition approach based on a synergetic algorithm will outperform classical statistical analysis of parametric measurements, even for larger species assemblages. MATERIAL AND METHODS RECORDING All recordings were done with a Pettersson D980 bat detector connected through a voltage amplifier stage to a PCMCIA data acquisition card (ComputerBoards PC-CARD-DAS16/330) in an Apple Macintosh PowerBook laptop computer (PowerBook 3400 or PowerBook G3) and driven by custom made software. Thus, digital recordings of 26 bat species echolocation calls where acquired, mostly when releasing identified bats, rarely in front of previously inspected roosts of known species occupancy. Tadarida teniotis was recorded in free hunting flight. 643 sequences (3.6 hours), containing calls of 362 hand-identified specimens were recorded. Prior to analysis, high-pass filtering (7.5 khz) was applied and single echolocation calls (26 ms = 8192 data points) were cut from the sequences. Most species calls fit into this window size (except Rhinolophidae) and only rarely there were more than one signal enclosed in a single cutout. SIGNAL ANALYSIS AND STATISTICS signals were cut from the recordings by help of a simple integrating detector algorithm of these were visually qualified as noise or otherwise inadequate for the analysis. Of the remaining 14354, 5153 where qualified suitable for identification purposes and 9201 additionally as suitable for training. Of these in a randomly selected subset of 2398 calls we calculated high resolution spectrograms (0.6 khz resolution, 87.5% window overlap) with Canary (Cornell University, Ithaca). Duration (DUR), highest frequency (HFR), lowest frequency (LFR) and frequency of main energy (MFR) were then manually extracted for later statistical analysis with a discriminant component analysis with resubstitution (PROC DISCRIM) in SAS (SAS Institute Inc.). The discriminant function was ten times repeated with 12.5%, 25% or 50% of all 2398 considered calls and tested with the remaining ones. Percentage values of coefficients of variation (CV) were transformed arcsin ( CV) for statistical comparison (Zar 1984). SYNERGETICS We apply a synergetic algorithm, which is also used e.g., in product control (Haken, 1988; Wagner et al. 1993; Wagner et al. 1995; Haken 1996). The classification of bat calls is achieved with an algorithm termed SC-MELT (Wagner et al. 1995; Hogg & Talhami 1996; Dieckmann 1997). This algorithm combines several training patterns per class into one feature vector, which has the same dimension as the training vectors. This ability enables the synergetic algo- 310 Variability in echolocation call design of 26 Swiss bat species rithm to handle big dimensions in contrast to artificial neural networks (ANN). We calculated spectrograms each consisting of point-spectra for the pattern recognition approach. The computational power needed to train an ANN with an input vector of size (spectrogram of dimension 159x128) is prohibitive. The synergetic algorithm emphasizes unique pattern content of training signals and diminishes of pattern contents common to all others. Learning times are very fast: the generation of the prototypes containing hundreds of bat calls takes only a couple of minutes on a current G4 processor (Apple Macintosh Power Mac G4/867). The classification of a signal is even faster because it is simply a scalar product. Each signals classification result consists of an array of scalar product values, which reflects the signals match to each stored class (i.e. the 26 acquired bat species). First, we performed learning tests on random subsets of 520 echolocation calls, 20 calls per each of the 26 species. The random sets were issued from the full database of 9201 signals which where visually approved as good for learning purposes, i.e. which did not contain e.g., calls immediately after take-off or strong echoes. We tested against a second set of 520 different signals from the same database. We performed three selections of learning calls and tested each training set against three other selections of classification calls. Apart from the raw classification rates we also calculated recognition rates with a filter criteria, which rejected classifications with maximum scalar product values smaller than 0.6 or with a difference to the next best scalar product of less than 0.2. We then tried to optimize the training base by picking from every species the calls from those random sets, which had achieved the highest classification rates. Thus, we generated a fourth training base consisting of a simplistically optimized sub-selection. Again we tested the three classification sets against this training base. RESULTS STATISTICS Echolocation call parameters Figures 3 and 4 illustrate the range of echolocation calls considered in this study for every investigated species. The results of the parametric spectrogram measurements are given in Table 1. For further analysis we split the species in two groups: the genus and all other species. have significantly shorter signals (DUR) with significantly higher starting frequencies (HFR; Table 1, t-test, the columns below the mean values). To test the within-species variability against the between-species variability of echolocation call parameters, we calculated the coefficients of variation (CV) for every species and parameter. We then compared CVs averaged over all species against CVs averaged over all other species with a t-test. Both groups show comparable within-species variation in the duration and highest frequency of their calls. However, species show significantly higher variation in the lowest frequency and the frequency of peak energy (Table 1, t-test, the columns below the CV values). To compare between-species variation we performed a variance ratio test, again between and non- species. Here, the latter show significantly higher variance in all four parameters compared to the genus (Table 1, Variance ratio test). Thus, variation within species is larger than in other species, but species differ less among each other relative to the non- species. Discriminant function analysis With discriminant function analyses (DFA) we explored the classification power of parametric measurements to be able to compare against the power of the pattern recognition approach. Increasing the number of calls to calculate the function did result in a general increase of the percentage of correctly reclassified calls from 68% to 75%. However, for bechsteinii and brandtii a decreased classification success 311 Obrist M. K. et al. Barbastella barbastellus Plecotus auritus Plecotus austriacus Hypsugo savii Pipistrellus kuhlii Pipistrellus nathusii Pipistrellus pipistrellus Pipistrellus pygmaeus Miniopterus schreibersii 26 ms Eptesicus serotinus Eptesicus nilssonii Vespertilio murinus Nyctalus leisleri Nyctalus noctula Tadarida teniotis Rhinolophus hipposideros Rhinolophus ferrumequinum 156 khz FIG. 3. Spectrograms of three exemplary echolocation calls of each investigated bat species not belonging to the genus. In many species signals show high plasticity, ranging from broad-band, multiharmonic calls to very narrow-band, quasi-constantfrequency vocalizations (e.g., Pipistrellus sp., Vespertilio murinus). Most signals in the Rhinolophidae where truncated by the chosen window length (see methods). was observed (33% to 31% and 33% to 28% respectively). The averaged results of 10 DFA, calculated with equal numbers of data calls and test calls are given in Table 2. Reclassification results varied from 27% to 100%. Species of the genus always scored considerably worse than did the other species. SYNERGETICS Table 3 summarizes the classification results for different settings. The fusion of a sub-selection of well-performing training calls to a new training base increased the average classification rate from 70% to 73% for raw classifications and from 86% to 88% in filtered classifications. Filtering lead to a rejection of 37% of the calls of the random set. The sub-selection slightly decreased this number to 31%. Even though a reduced number of signals qualify for classification, the filtering approach improves classification rates by 12-13%. Figure 5 illustrates the percentage of correct classifications with and without filter. The effect differs strongly between species but is never negative. species consistently show lower classification rates than other species. DISCUSSION ECHOLOCATION CALLS Digital recording at khz with 12 bit data depth ensured a high fidelity of the recordings (Parsons and Obrist 2004). The description and illustration of echolocation call characteristics of 312 Variability in echolocation call design of 26 Swiss bat species bechsteinii blythii brandtii capaccinii daubentonii 26 ms emarginatus myotis mystacinus nattereri 156 khz FIG. 4. Spectrograms of three exemplary echolocation calls of each investigated bat species in the genus. Variability between species is considerably lower compared to Fig Swiss bats species given here compare well with other published reports (e.g., Ahlén, 1981). Recordings made directly after releasing handidentified bats differ from those recorded later in search flight. By recording sequences of 20 seconds in a post-trigger mode and by monitoring the acoustic behaviour of the animals, we could verify and later select the type of calls contained 313 Obrist M. K. et al. TABLE 1. Echolocation call parameters of 26 swiss bat species. Mean, standard deviation (StdDev) and coefficient of variation (CV) are given for call duration (DUR), lowest call frequency (LFR), frequency of peak energy (PFR) and highest frequency (HFR). Values given in italics do not properly indicate species parameter, as the measurement window of 26 ms was truncating total call lengths. Range, mean and variances averaged over all species of the genus and over all other species are indicated below. Results of t-tests and variance ratio tests are given at the bottom (see text for further details). DUR [ms] LFR [khz] PFR [khz] HFR [khz] Species N Mean StdDev CV Mean StdDev CV Mean StdDev CV Mean StdDev CV bechsteinii 65 4,3 0,8 18% 26,5 2,9 11% 48,4 6,6 14% 103,5 12,7 12% blythii 100 3,3 0,7 20% 24,5 4,4 18% 53,2 10,9 20% 106,2 14,1 13% brandtii 100 4,6 1,1 24% 27,5 3,1 11% 45,7 4,9 11% 103,6 12,8 12% capaccinii 100 5,2 1,1 21% 32,0 2,6 8% 45,1 4,3 10% 86,8 8,4 10% daubentonii 100 3,9 0,9 22% 27,3 3,0 11% 42,7 3,5 8% 81,2 8,0 10% emarginatus 100 3,6 0,7 19% 36,3 2,8 8% 54,5 7,4 14% 113,1 12,5 11% myotis 100 6,0 1,7 29% 22,2 2,6 12% 37,1 4,0 11% 86,0 11,3 13% mystacinus 100 3,6 0,5 15% 27,9 3,5 12% 46,8 5,6 12% 99,7 12,6 13% nattereri 100 4,1 1,1 27% 14,0 4,0 29% 40,4 8,8 22% 108,6 18,6 17% Barbastella barbastellus 100 4,3 1,0 24% 25,7 2,2 8% 36,0 4,8 13% 48,3 3,9 8% Plecotus auritus 100 2,9 0,6 19% 22,7 1,7 7% 37,7 5,1 13% 55,7 5,6 10% Plecotus austriacus 100 5,8 1,4 25% 18,0 2,3 13% 27,6 2,5 9% 45,3 3,3 7% Hypsugo savii 72 7,3 1,0 13% 28,8 0,8 3% 34,9 2,0 6% 48,3 7,2 15% Pipistrellus kuhlii 100 6,3 0,9 13% 33,6 1,3 4% 39,5 1,8 4% 63,6 12,8 20% Pipistrellus nathusii 100 6,9 1,4 20% 36,1 1,1 3% 41,3 2,2 5% 61,5 13,9 23% Pipistrellus pipistrellus 100 6,3 0,9 15% 42,6 1,4 3% 47,4 2,0 4% 73,8 15,9 21% Pipistrellus pygmaeus 100 6,0 0,9 14% 51,5 1,8 4% 56,2 2,4 4% 84,1 16,5 20% Miniopterus schreibersii 100 6,2 0,8 13% 47,4 1,2 3% 53,9 3,8 7% 87,3 11,0 13% Eptesicus serotinus ,9 2,4 22% 22,4 1,2 5% 26,8 1,8 7% 47,2 7,4 16% Eptesicus nilssonii ,7 1,6 15% 24,6 1,1
Related Search
Similar documents
View more...
We Need Your Support
Thank you for visiting our website and your interest in our free products and services. We are nonprofit website to share and download documents. To the running of this website, we need your help to support us.

Thanks to everyone for your continued support.

No, Thanks