Who's calling? - Acoustic Bat Species Identification Revised: Laptop Computer Surpasses Human Ear in Real Time Signal Identification in the Field.

• Martin Obrist
• Ruedi Boesch
• Peter Flückiger, all Swiss Federal Institute for Forest, Snow and Landscape Research WSL, CH-8903 Birmensdorf, Switzerland
• Ulrich Dieckmann, Fraunhofer Institute for Integrated Circuits, Erlangen, Germany. present address: DCS Dialog Communication Systems AG

Publication in

• Poster presented at the Biosonar Conference 1998 in Portugal

Index

• Introduction
• Methods
• Recording
• Classification
• Figure 1(a-m)
• Results
• Conclusions
• Acknowledgements

Introduction

Aiming mainly at the acoustic side of the identification task, the purpose of this study was threefold:

• Evaluate an affordable method to achieve high quality recordings of ultrasound signals in the field.
• Test a state of the art pattern recognition algorithm for its ability to classify spectrograms of echolocation calls of bats.
• Implement a user adaptable software for the identification of bioacoustic signals.

Methods

Recording

Our data (Tab. 1) were recorded from identified bats (capture - release - record) with a digital time expansion bat detector, and stored at 1/10 speed on cassette tape. With Canary (Cornell Bioacoustics Lab) we measured duration, highest-, lowest- and frequency of main energy from good quality calls. These values were used for statistical analysis. The same vocalizations were used for testing pattern recognition.

In parallel, we implemented an application on a Macintosh Powerbook which allows to digitally record with a PCMCIA A/D-card at 312 kHz to 80 MB installed RAM in pre- or post-trigger mode (Fig. 1 centre). Userdefined detectors (high frequency, amplitude etc.) may be installed to trigger recording. Data are stored in binary form, enabling Canary to directly read them, already time-, frequency- and amplitude-corrected .

Present recording setup for the acoustic field identification of bats.

Classification

For signal classification we use a synergetic algorithm ( SA). Synergetics apply principles of natural self-organization on learning and recognizing of complex information. The SA suppresses information common to all and emphasizes differences between single training objects.

Synergetics offer very short training and recognition times, handle high-dimensional feature vectors, as commonly occurring in signal processing, and allow for unsupervised training and classification. Neural Networks too can handle big dimensions but then need prohibitive computational power.

Single calls found by a smoothing and integrating preprocessing algorithm are cut out and 159x128-point spectrograms of these signals are calculated automatically, leading to a 20352 point feature vector (Fig. 1 top). The SA learns N calls of each of the 12 species, resulting in 12 prototype feature vectors (classes). When classifying, the SA computes the scalar product of each tested call with each class prototype, producing 12 values between 1 and 0 per signal (training call or no ressemblance to training calls).

The higher the 'maximum scalar product' (MSP), the clearer is the attribution of a tested call to one species. The difference to a second-highest scalar product indicates the degree of 'differentiation from a second-best' class (DSB). MSP and DSB thresholds were varied experimentally to delimit the number of calls subjected to the recognition process.

Signals assigned correctly or falsely produce a frequency distribution over the 12 classes. For a given sample of calls, this histogram identifies the species fitting best (Fig. 3). We evaluated the learning-classification process, with 1 to 5 learn-calls per sequence and 1 to 9 sequences the calls were drawn from (115 runs, each with 16 delimiter combinations).

Figures 1(a-m):

Portraits of the 12 species investigated, and representative amplitude displays and spectrograms of their calls as tested with the synergetic pattern recognition algorithm. Timebase is 35 ms, Frequencybase is 0-110 kHz.

Results

• The number of training calls positively correlates with recognition success (p << 0.01), because additional sequences include more of the variablity occuring between recordings of different individuals.
• Increasing the number of training calls taken per sequence marginally affects average classification success, indicating low intra- and high inter-individual variance.
• Species strongly influence classification results (Tab. 1).

Table 1:

Nyctalus noctula scores best and Myotis daubentoni worst (without the 3 'low n species'). Call spectrograms (Figures 1 a-m) and cluster analysis underline (Fig. 2) the echolocation call similarity in the Myotis group.

Figure 2:

• A discriminant component analysis identifies 46% of Myotis-calls properly (58% of all species' calls), when testing the one halve of the signals not used for the discriminant function description.
• The SA identifies 77% of the Myotis-calls correctly, when 9 learn-calls are used and MSP>0.5, DSB>0.2 are applied as scalar product delimiting criteria. 38% of the signals are thereby rejected from classification. Despite the cost of rejecting calls from the analysis, these values proved optimal in eliminating noisy or ambiguous calls, while minimising the loss of calls and maximising the recognition success.
• With 9 training calls, we reach an average overall recognition rate of 80% while rejecting 33% of the calls (Tab. 1).
• 5 training calls (different sequences) assertain a correct maximum score, and therefore identify the proper species (Fig. 3).

Figure 3:

Conclusions

Decreasing price/power ratio of portable hardware help to spread digital ultrasound field-recording in a wider group of users. Quality, speed and economy speak for new approaches like synergetics to acoustic signal recognition when compared to classical statistical analysis of discrete measurements taken from recordings.

Our versatile program demonstrates a field alternative to crude heterodyne or count-down methods for species identification. Unsupervised, automated monitoring of species specific bat activity becomes possible. Furthermore, the software is fully user configurable for the classification of other signals (other taxonomic groups, behaviours).

However, we wish to express our concerns with many field biologists objecting to the system presented here. It constitues a so called 'black box', as only limited control can be attained over the synergetic classifier. We do not encourage lay people to feed data in such a system without prior understanding of the bioacoustics (e.g. bat echolocation) and the content and potential variability of learning and classification signals.

Acknowledgements

We are very grateful for Lars Petterssons generous help with recording hardware.