This research presents an attempt to build an automatic recognition system for individualhumpback whales and bottlenose dolphins based on their audio records. Such recognition systemsare an improvement upon bioacoustics-based tracking methods for population census ofendangered species for two reasons. Firstly, when coupled with tracking, they provide a uniqueautomated population assessment tool which is free of human biases and allows large areacoverage during ecological surveys conducted for conservation of endangered species. Secondly,these methods provide a more robust substitute for inconvenient and error-prone photo-ID toolsused for monitoring migratory patterns of species like humpback whales.

Although both species have a large repertoire of sounds [2,3] which serve a number of uniquefunctions, this work targets the humpback whales songs and the clicks of bottlenose dolphins toaccomplish this task. Although both type of vocalizations, obtained from unrestricted individualsand recorded using passive systems in our case, pose a challenging classification problem in adynamically changing environment yet this selection is made because the frequency ofoccurrence for clicks and songs is higher than any other type of vocalizations in respectivespecies.

Classification of humpback whale

The architecture of humpback classification system is shown in Fig. 2(a). Audio data, recordedusing a single hydrophone (shown in Fig. 1), is pre-processed into smaller file-segments anddown-sampled at 8 KHz for further processing. This is followed by song unit detection using barcrossing rate (BCR) algorithm [4,5] . Cepstrum feature vectors are extracted from each song unitas explained in Fig. 2 (b). The feature vectors coming from individual humpback whales areseparated and used to train the support vector machine (SVM) model, to be used later forclassification. SVM was used because of its accepted optimal performance in literature [1] (Fig.3). During classification stage, feature vectors in test set are grouped together to make utterancevectors. The final classification decision for assigning any utterance vector to a particular whaleclass is taken by following three threshold rules.

〓,here D is the test utterance vector length and F represents the decision taken by multi-classSVM model, based on one-against-one voting approach, for single feature vector from test set .

For humpback whales, we have tested our algorithms on both clean and noisy data [4,6].Moreover, the accuracy of the system is tested for data sets of different individuals, recorded indifferent years [5]. The suggested system is robust enough to classify whales with accuracygreater than 90%. It is observed that a test utterance vector corresponding to two minute durationis sufficient to achieve this classification accuracy for good quality songs [4]. The noisyconditions may impact the accuracy but use of suitable filtering and spectral subtraction, based onsilence parts of recording, shows remarkable improvement in accuracy results. The only problemis the classification of animal records with a gap of more than a few years [6]. In that case, thesystem may not be able to perform best. But still accuracy remains above 50%. Similarly, usingpooling and threshold-based decision methodology, provision is provided to incorporate arrival ofnew animals.

Classification of bottlenose dolphin

The architecture for bottlenose dolphin tracking and classification system is shown in Fig. 5. Theacoustic data was recorded from a 5-hydrophone array as shown in Fig. 4 (a). The data wascollected in a target-touch experiment in which two dolphins were trained to approach and toucha target of 2 MHz hydrophone place at a distance of 2 m from the 5-hydrophone array. Fig. 4(b)shows the scenario of target-touch experiment with a subject dolphin approaching the target (2MHz hydrophone), held by trainer. The direct and reflected clicks, above a user-specified energythreshold, are successfully detected using a running energy sum algorithm. Time and path delaysare approximated with an improved correlation-based method applied to 5-channel data. Targetposition was localized using following formulae for range r, horizontal angle 0 and elevationangle o〓。

Short-duration classification of different dolphins is performed using track-association of eachunlabelled click using three approaches which considered position information, clickcharacteristics (similarity pattern in time domain) and hybrid features (position/clickcharacteristics) respectively. Comparison suggests superiority of hybrid feature technique overtwo former techniques. For track-independent classification, a number of features such as spectralcoefficient, cepstrum coefficients, temporal click patterns, beam patterns of each click andcorresponding equivalent aperture size for dolphin transmission system have been analyzed.

Classification problem of bottlenose dolphin clicks is especially challenging because of highlydirectional beam patterns which results in different click characteristics (energy, temporal clickshape, dominant frequencies) at different orientation [7]. Even with multi-hydrophone system, such as one described above, established conventional algorithms for signal detection and delayestimation cannot be trusted due to complexities resulted by click reflections in shallow depthsand narrow beam pattern of the clicks. Therefore a semi-autonomous analysis was adoptedwherein derived click parameters (time delays, position coordinates and their one and twodimensional projections in space, dominant frequencies on three hydrophones along verticalbaseline etc) could be extracted automatically and later examined using graphical user interface(GUI) as shown in Fig. 6. The GUI has options to change or re-calculate some of these features.Moreover new clicks can be added or automatically identified clicks can be removed if desired.Provision is provided for detection of low-energy clicks (earlier missed by automatic method) inmanual as well as in semi-autonomous mode where a user is allowed to provide a reference clickand a relative energy threshold. For bottlenose dolphin, the system ensures reliable observationand identification in short-duration. However, identification over long periods is still open forfurther exploration and research.

Overall, both systems are expected to be useful for recording valuable information such ascetacean movement, behaviour and their population density estimation in a particular region andprovide a reliable alternative for currently prevalent techniques in the field for cetaceanconservation.

Fig. 1: System used for humpback whale song recording

Fig. 2 (a): Algorithmic diagram for humpback whaleclassification

Fig.2 (b): Extraction of cepstrumcoefficients

Fig. 3: Support vector machine used for optimal classification ensures selection of largemargin classifier

Fig. 4 (a): 5-Hydrophone systemfor bottlenose classification

Fig. 4 (b): 5-A scenario of targettouchexperiment

Fig. 5: Algorithmic block diagram forbottlenose tracking and classification

Fig. 6: A snapshot of graphical user interface used in acoustic feature analysis of data of bottlenose dolphins

クジラ類の生態学的な研究は、生活環境が水中であるということから、陸上生物に関する研究とは違って、観測手法が限られ、困難であり、我々が知っている知識は極めて少ないといってよい。生態学的な研究の第一歩は、水中行動の観測であり、また、それがどの個体がおこなっているかという個体識別をおこなうことである。クジラ類は、イルカと呼ばれる小型歯クジラ類を含めて、水中生活のために、数種類の音を出し、食料の捕獲、障害物の発見、会話などをおこなっている。本論文では、クジラ類が出す音響(鳴音)に注目して、それを利用して個体識別をおこなう手法の研究を、特徴的な2種類の鳴音、すなわちザトウクジラのソングおよびハンドウイルカのクリックスを例にとっておこない、それぞれの個体識別手法を提案し、その有効性を検証している。

本論文は、前文、ソングを利用したザトウクジラの個体識別、クリックスを利用ハンドウイルカの個体識別、および結言の、4つの章に分かれている。

第一章では、現在広くおこなわれているクジラ類研究を個体識別という観点から議論し、本論文の音響的な研究の位置づけをおこなっている。

第二章では、雄のザトウクジラが発する可聴域のソングを、Support Vector Machine (SVM)を導入して、Supra-Classifier Voting Methodologyを用いて解析する手法を提案している。小笠原海域において録音され、かつ目視観測より個体が識別されている複数のザトウクジラのソングの音響データをベンチマークとして解析し、提案するシステムの有効性を示した。先行する研究では、個体識別に、Gaussian Mixture Model (GMM)が多く用いられているが、その問題点を指摘し、GMMでは分離できなかったデータをSVMで分別して、識別することに成功している。また、録音日、および年の違う、同一個体からのデータのSVM分析結果を比較し、環境ノイズや録音システムなどの録音環境の違い、個体の音声特性の年変化などが、SVM解析に及ぼす影響を検討している。

第三章では、歯クジラ類がおこなうエコーローケーション(音響定位)のためのクリックスを利用して個体識別をおこなう手法を提案し、伊豆・三津シーパラダイスにて飼育されている2頭のハンドウイルカについて、その有効性の検証をおこなっている。クリックスは前出のソングとは異なり、広帯域の信号であり、前章で提案したSVMを適用して個体識別することは困難である。そこで、ハイドロフォンアレーを用いてクリックスを検出し、到達時間差から高い精度で位置検出をすると同時に、音響ビームのエネルギ、中心周波数、指向性等の計測を同時におこない、個体を時系列的に識別する方法を開発した。この手法により、遊泳中の小型クジラ類から発せられるクリックスから、海面や海底からの反射による信号をとりのぞき、半自動的にクジラ類の位置計測および個体識別が可能となった。

第四章では、これらをまとめて、提案した手法をさらに発展させる方向を示している。

以上のように、本論文は、クジラ類の二種類の鳴音を利用して、個体識別をおこなう手法を開発することにより、これまで観測できなかったクジラ類の生態あるいは行動の一面を捉えることが可能とするものであり、水中観測工学に新たな知見をもたらすものである。ザトウクジラの鳴音はその種に特徴的なものであるが、ハンドウイルカのクリックスについては、他の小型歯クジラ類、とくに絶滅が危惧されているカワイルカやカワゴンドウにも適応可能なものであり、これまでほとんどできなかったそれらの生態観測と保護に役立つものであると期待される。

よって本論文は博士(工学)の学位請求論文として合格と認められる。