Research On Automatic Classification Method Of Bird's Song Oriented To Bird's Sound Sensor Network

Birds have been widely used as biological indicators which are extremely sensitive to the changes of habitats and environment.At the same time,birds are important indices of the monitoring and evaluation of the biological diversity.In addition to the existing radar,optoelectronics and manual detection methods,vocalization-based bird species monitoring is a very important complement.An non-invasive bioacoustics monitoring method is to establish a birdsong sensor network system.In the sensor network system,each sensor node collects bird sound data and sends to the sink node for pre-processing,the sink node uploads the processed data to remote monitoring terminal.Since the computing capacity,storage capacity and endurance of the sink node are limited,it is a reasonable choice for the sensor node to pre-process the collected data before uploading to the sink node,which can effectively reduce the network transmission pressure.Therefore,it is meaningful for automatic bird sound detection and classification based on a sensor node.In view of the above situation,this thesis focuses on the automatic bird sound detection and classification based on a single sensor node.Firstly,several commonly used methods of automatic segmentation of continuous acoustic monitoring data are introduced,including the dual-threshold method based on short-time energy and short-time zero-crossing rate,the method based on morphological filtering and the method based on Gaussian mixture model(GMM).Then the Mel frequency cepstral coefficients(MFCC)feature and two texture features are introduced,in addition,a new Mel subband parameterization(MSP)feature is proposed which characterizes the spectral pattern of the bird sound.This new feature parameterizes the output energy sequence of each Mel subband by an autoregressive(AR)model.Aiming at the problem of species identification,this thesis analyzes and compares several classifiers,including random forest(RF),support vector machine(SVM)and hidden Markov model(HMM).Finally,the performance analysis and comparison based on the field recordings demonstrate that the proposed method is more robust.Considering that the datasets of the existing research results only contain the expected species,while this thesis uses the field recordings which not only include the expected species,but also include the environment noise and the sound of other unknown species(denoted as "unknown" class),the method proposed in this thesis is more suitable for the task of automatic analysis of field continuous acoustic monitoring data.The results of this study will lay a good foundation for building a bird ecological monitoring system in natural environment.