 |
|
ISSN 1320-0682 | ||||
| Volume 4 | 1997 | ||||
Foraging poses various challenges to the echolocation abilities of bats, including orientation, collision avoidance, location of potential prey and the filtering of interference from other bat calls - for example, Fenton (1990), Kalko and Schnitzler (1993). Variability in bat echolocation is partly influenced by the environment, since bats are able to adjust their sonar signal according to the nature of the environment in which they are foraging - for example, Aldridge and Rautenbach (1987), Schumm et al. (1991), Simmons et al. (1978). Bat echolocation varies individually, intraspecifically and geographically (Brigham et al. 1989; Kutt 1994; Neuweiler 1984; Obrist 1995; Thomas et al. 1987). For example, the call frequency of the Hoary Bat Lasiurus cinereus occurring in Arizona differs from those of same species occurring in Manitoba (Barclay 1986; Fenton & Bell 1981).
However, the flexibility and adaptability of bat sonar signals is limited, and is species- or genotype-dependent (Simmons & Stein 1980). This does not mean that every bat call is identifiable, but rather that each taxonomic unit probably has a typical pulse parameter combination which could be detected. Given sufficient field experience and knowledge of the local bat fauna, identification of bats by their calls is possible at least for some species (Ahlen 1981; Fenton & Bell 1981; Woodside & Taylor 1985).
This study is the first published attempt to identify bat species from their ultrasonic calls using a decision tree based classification. The methodology applied in this investigation eliminates the subjectivity of call analyses, instead relying on the mathematical properties of the ultrasonic calls. Reference calls for parts of the south-eastern Australian region are already available (Herr & Klomp 1995) and provide the background for this attempt to automate bat call identification.
The intention of this study is to determine the potential of machine learning based systems to identify species (or higher taxonomic level) of bats based on echolocation calls. Decision tree classification techniques have already delivered results in sound pattern analysis, such as in the call recognition of frogs and birds (Taylor 1996). Machine learning based systems are also successfully used in speech recognition, as the amount of available software illustrates (http://www.itl.atr.co.jp/comp.speech/Section6/Q6.5.html). This type of software is ideal for bat call analyses because each echolocation call has repetitive patterns that are probably taxon specific - for example, Fenton and Bell (1981), Jones and Corben (1993), Neuweiler (1984), Woodside and Taylor (1985).
All echolocation calls were recorded from the eight species of bats released by hand and free-flying in natural habitat following capture at various sites in the western slopes of the Australian alps. All samples included males and females, except Falsistrellus tasmaniensis, which were all males. Only those calls displaying typical 'search phase' sequences of pulses were used - for example, Fenton and Bell (1981). The use of regional reference calls avoided the variability caused by geographic differences in species - for example, Thomas et al. (1987). Reference calls of released bats were recorded with the ANABAT V system (Titley Electronics, Ballina, NSW, Australia). The frequencies of ultrasonic calls were divided by a factor of 16 and stored on a magnetic tape together with a calibration tune. The data produced were then zero-crossed and stored digitally on computer.
The output signal of the ANABAT system is a portion of the original sound, thereby eliminating the technical constraints for recording (and hearing) the high frequency sound. All amplitude information and harmonics are lost. This is a drawback when slight differences in harmonics and power (amplitude) are important for the research. However, since analyses of bat calls mainly focus on frequency change over time, this technique is suitable for the identification of bats - for example, Conole and Baverstock (1995), Kutt (1995), Richards (1996), Spencer and Coles (1996). Further, the ANABAT system is the most widely used bat detection and recording equipment in Australia, so methods automatically analysing these stored calls are particularly required.
Reference calls were prepared by determining the mathematical parameters of each pulse in the calls. Any noise distorting the clear definition of a pulse was excluded manually. The ANALOOK (version 4) call analysis program was used to determine all the mathematical parameters of the pulses that could be precisely defined and measured (Figure 1). Time between each pulse (Tbp), duration (Dur), maximum frequency (Fmax), minimum frequency (Fmin), average frequency (Fmean), and the characteristic frequency (Fc, the frequency at the point of the lowest slope) of each pulse within all the calls were determined.
The C4.5 classification method (fully described in Quinlan (1993) and references cited therein) was employed to create and train the decision tree using 160 reference calls from eight bat species.
The program SIPINA was used to establish a decision tree based classification system. In addition to the automated model creation and array of user controls, the advantage of this program for the user is the provision of the sets of decision rules used to discriminate each group, thus elucidating the process and final acceptance of the final classification - for example, Diederich and Tickle (1995).
The SIPINA program used the training data to create a decision tree with 21 nodes, eleven leaves and a maximum depth of ten layers. The classification was then evaluated using an additional three reference calls of each species that had not been included in the training data set.
Table 1 Summary of mean pulse parameters (+/- standard deviation) from 160 reference calls of eight bat species from the western slopes of the Australian alps.
| Species | Animals N | Pulses n | Dur (ms) | Tbp (ms) | Fmax (kHz) | Fmin (kHz) | Fmean (kHz) | Fc (kHz) |
| C. gouldii |
|
|
5.5 (+/- 2.8) | 150 (+/- 92) | 43.2 (+/- 10.7) | 30.3 (+/- 1.5) | 33.1 (+/- 2.8) | 30.7 (+/- 1.6) |
| C. morio |
|
|
3.5 (+/- 1.2) | 99 (+/- 53.7) | 65.8 (+/- 11.4) | 50.3 (+/- 1.6) | 53.9 (+/- 2.8) | 50.9 (+/- 1.7) |
| F. tasmaniensis |
|
|
4.6 (+/- 1.9) | 116.1 (+/- 61.7) | 52.2 (+/- 12.9) | 37.5 (+/- 2.4) | 41.3 (+/- 4.7) | 38.3 (+/- 4.1) |
| N. geoffroyi |
|
|
3.2 (+/- 1.1) | 138 (+/- 61.4) | 63.7 (+/- 10.2) | 39.5 (+/- 3.9) | 48.6 (+/- 5.2) | 46.5 (+/- 7.4) |
| N. gouldi |
|
|
2.8 (+/- 0.9) | 142.2 (+/- 162.2) | 62.3 (+/- 9.4) | 39.8 (+/- 4) | 47.9 (+/- 4.8) | 46 (+/- 5.9) |
| V. darlingtoni |
|
|
5.1 (+/- 1.8) | 88.2 (+/- 76.5) | 56.8 (+/- 8.8) | 42.3 (+/- 1.6) | 45.2 (+/- 2.2) | 43.1 (+/- 1.7) |
| V. regulus |
|
|
3.6 (+/- 1.3) | 103.2 (+/- 53) | 62 (+/- 11.8) | 47.4 (+/- 3) | 51 (+/- 4.7) | 48.4 (+/- 3.6) |
| V. vulturnus |
|
|
3.3 (+/- 1) | 87.6 (+/- 64.9) | 58.8 (+/- 9.8) | 45.6 (+/- 1.6) | 48.4 (+/- 2.2) | 46 (+/- 1.5) |
Table 2 shows the identification ability of the decision tree classification system. The decision tree classified four (C. gouldii, F. tasmaniensis, V. darlingtoni and V. vulturnus) of the eight bat species correctly in 95% of attempts, leaving 11% of their pulses unclassified. The highest rate of pulses being left unclassified occurred for C. morio. Five pulses (100% of the classified pulses) were misclassified as V. regulus or V. vulturnus. Most pulses of N. geoffroyi, N. gouldi were misclassified as V. darlingtoni. Although N. geoffroyi had very similar sonagrams to those of N. gouldi, the other species displayed different shapes in their call sonagrams (Figures 2 and 3). Forty-nine percent of the pulses of V. regulus were misclassified as V. darlingtoni. It can be seen from Table 1 that these species have a considerable overlap in their call parameters.
Table 2 The number of pulses identified by the decision tree based classification from three calls of each of the eight bat species (n - number of pulses, N - number of individual bats). The results of species wrongly identified (and number of pulses) are also shown.
| Tested species | N | n | Correct | Incorrect | Unclassified | Correctly classified | Misclassified as |
| C. gouldii | 3 | 58 | 58 | 0 | 0 | 100% | |
| C. morio | 3 | 48 | 0 | 5 | 43 | 0% | V. regulus (2), V. vulturnus (3) |
| F. tasmaniensis. | 3 | 73 | 64 | 6 | 3 | 91% | V. darlingtoni (6) |
| N. geoffroyi | 3 | 69 | 0 | 49 | 20 | 0% | V. darlingtoni (35), F. tasman. (14) |
| N. gouldi | 3 | 54 | 1 | 26 | 27 | 4% | F. tasman. (12), V. darl. (13), V. reg. (1) |
| V. darlingtoni | 3 | 119 | 108 | 3 | 8 | 97% | V. vulturnus (3) |
| V. regulus | 3 | 72 | 34 | 35 | 3 | 49% | V. darlingtoni (34), V. vulturnus (1) |
| V. vulturnus | 3 | 72 | 42 | 5 | 25 | 89% | V. darlingtoni (4), V. regulus (1) |
| Total | 24 | 565 | 307 | 129 | 129 | 70% |
Rather than classifying call parameters as averages of pulses from individual bats, the intention in this study was to characterise (in the decision tree based classification training) and identify (using test calls of species) each pulse as belonging to a bat species. Such an approach is very similar to the way bat researchers analyse the calls visually. Typically, sequences of pulses with species-specific shapes and parameters are selected from a call for identification, and the other pulses are ignored.
The automated identification of bat calls using a decision tree based classification led to good results for species with distinct mathematical differences in call parameters, although three of the eight species were consistently misidentified by the decision tree classification system, and one species could not be identified reliably.
A significant number (35% to 90%) of unclassified calls occurred for the species C. morio, N. geoffroyi, N. gouldi and V. vulturnus. This may be because the calls of these species are so different in the combination of their parameters that they could not be included successfully in the decision tree algorithm and consequently were marked as atypical for any of the species represented in the training data set. Such pulses could be attributed to the individual flexibility of bats in their sonar signal design that enables them to forage in different habitats.
None of the Nyctophilus species was misidentified with each other. This is surprising given that researchers are currently unable to distinguish between the species (Table 3) using ANABAT recorded calls. Although this is a very encouraging result for researchers requiring consistent and reliable call identifications, the decision tree classification system failed to distinguish between what are ostensively very different calls of some species. For example, distinct differences between the sonagrams of N. gouldi and F. tasmaniensis, and N. geoffroyi and V. darlingtoni can be identified visually (see Figures 2 and 3). However, such features could not be represented mathematically in the parameters used in this study. N. geoffroyi is confused in a large proportion of pulses, which may be due to the presence of short pulses (typical of this species), causing a parameter description similar to F. tasmaniensis and V. darlingtoni. Most of the parameters used to describe the features of N. gouldi pulses are easily confused with those of the pulses of F. tasmaniensis and V. darlingtoni.
The pulse frequency of V. regulus overlaps in the frequency range with both other Vespadelus species. The decision tree classification system misidentified 49% of the pulses of V. regulus with V. darlingtoni. Such misidentification is understandable given the range of parameters found to overlap for both species. Only one pulse (2%) was misidentified as V. vulturnus, which is quite surprising given the similarity in pulse shape (Figure 2) and call parameters (Table 1).
Table 3 summarises the misidentification errors of the decision tree based classification and compares it with the likely ability of an experienced bat researcher to distinguish between the calls of a given bat species.
It can be seen that the decision tree classification system described in this study does not offer any significant advantage in call recognition to an experienced researcher, although careful use of it for a limited range of bat species might offer advantage when large numbers of calls require identification, or might offer advantage in species identification by bat researchers inexperienced in call recognition. In the latter cases, the reliability of an automated classification system must be determined by using calls of known species, calls not used in training the system, prior to any identification of unknown calls.
Table 3 Comparison of bat call discrimination difficulties by an experienced bat researcher using sonagrams and the decision tree classification system used in this study.
| Discrimination between | Discrimination difficulties human | Discrimination difficulties the decision tree based classification |
| V. regulus and V. vulturnus | Difficult.
Similar pulse shape, parameters overlapping. |
Good,
despite overlap in averaged parameter values. |
| V. regulus and V. darlingtoni | Good.
Different pulse shape, averaged parameters overlapping, minimum frequency mostly different. |
Difficult.
Averaged parameter values overlapping. |
| N. spp.
and F. tasmaniensis
N. spp. and V. darlingtoni |
Good.
Different pulse shape. |
Difficult.
Averaged parameter values overlapping. |
| V. darlingtoni and V. vulturnus | Good.
Different pulse shape, minimum frequency different. |
Good.
Parameter values enable differentiation. |
| N. geoffroyi and N. gouldi | Difficult.
Pulse shape similar and highly variable. |
Not misidentified,
but high rate of pulses unclassified for each species. |
It is expected that the methods used in this study to identify bat calls using a decision tree will fail to reliably identify short calls with high variability because the method depends on species-specific pulse parameter combinations. In practice, the identification of ultrasonic calls uses short sequences of pulses, as the shape of each pulse (frequency change over time) becomes more important for identification (Fenton & Bell 1981). C. gouldii, for example, has two distinct types of calls that are usually alternated (Jones & Corben 1993). The species can be easily identified by its call frequency and shape change (shown in Figure 3), without further examination of other parameter values.
Several other Australian studies have published information on bat calls (Crome & Richards 1988; Fullard et al. 1991; Jones & Corben 1993), although only two studies (Jolly 1996; Woodside & Taylor 1985) have attempted to use mathematically derived criteria to distinguish between the calls of different bat species. Woodside and Taylor (1985) used equipment that enabled the inclusion of harmonics and amplitude information into their analysis, which may have conferred a discrimination advantage over this study. Their work involved the use of captive bats which is likely to have led to the recording and analysis of atypical echolocation calls. Bats are known to have the ability to memorise spatial information of familiar environments (Grant 1991; Neuweiler & Moehres 1967). However, these previous studies used statistical methods of discrimination. The advantage of using a decision tree classification system over such statistical approaches, is that each single pulse is analysed and outliers and variation of each pulse are included in the analysis system of a decision tree. This enhances the discrimination between pulses as it does not rely on a single analysis step, but includes testing of already defined groups against the newly formed groups, in addition to optimising the separation. The decision tree classification procedure creates a set of rules that optimises the discrimination between the species using all available information from the extracted variables.