Detection And Classification Based On Beluga Whale Whistle

In recent years,underwater passive acoustic monitoring(PAM)has been used to capture the acoustic signals emitted by various cetaceans to study their behavior and interactions in their habitat.At the same time,with the increase of the memory of the acquisition sensor,the long-term acoustic observation record can provide researchers with richer information about the species and their habitat environment.PAM contains three key technologies: location,detection,and classification.Detection and localization are usually used to assess the population of whales and dolphins and to determine the location of animals in their habitat on a large spatio-temporal scale,while classification is used to study their habitat and behavior by using the different frequency-modulation characteristics of recorded acoustic signals.In this paper,two key technologies of PAM are studied,namely whistle event detection technology and whistle classification technology.Delphinapterus,known as the canaries of the sea,produce two main types of sound signals:a wideband click signal for echolocation and an FM whistle signal for communication.In this paper,the whistle frequency modulation signal is studied.The complex and highly developed communication system of Delphinapterus reflects the complexity of their social relationships.Different whistle types are related to Delphinapterus behavior but are also affected by juvenile,foraging,and environmental noise.Therefore,detecting and classifying beluga whale whistles are essential in studying its social behavior and protecting it from human activities.This paper studies a complete whistling detection and classification system based on 1000 minutes of Delphinapterus whistle signals collected from Harbin Polemuseum and develops a set of integrated call data storage,detection and classification software.The software provides the basis for future connections between Delphinapterus whistles and behavior.The main research contents of this paper are as follows:1.Research on detection methods of unsupervised whistle events.This part studies whistle-event detection algorithms from the time domain,time-frequency domain,and Mel cepstrum domain,respectively,and compares the f1-Score performance of each algorithm through simulation.Among them,the time domain detection algorithm mainly used the autocorrelation and cross-correlation detection method;In the time-frequency domain detection algorithm,the combination of spectral smoothing and adaptive thresholds are mainly used.Principal component analysis(PCA)and K-means clustering are mainly used in the Mel cepstrum domain.By testing the detection results under different SNR,PCA,and K-means clustering methods can obtain the optimal whistle event detection results when the SNR is-15 d B.Its F1-Score performance at frame level reaches the highest 0.93.2.Study on whistle classification method.Delphinapterus whistles are divided into five categories according to their time-frequency characteristics.This part mainly uses different deep network models to classify five categories of whistles.First,the whistle classification performance of the VGG model,Res Net model,Dense Net model,and the integrated model of the three models based on fine-tuning migration is studied.Secondly,a method based on semantic segmentation and a generalized linear classifier is proposed to classify white whale whistles.The fully convolutional network(FCN)structure is used to complete semantic segmentation of whistles to extract whistle FM features in the time-frequency domain.In contrast,the support vector machine(KSVM)is used to classify extracted FM features further.This is the first time the semantic segmentation model has been applied to the study of Delphinapterus whistles.The method finally achieves 95% classification accuracy in the test set,the same as the ensemble model with three deep networks classification results.The results show that this method can effectively reduce the computational complexity and does not need much training data while ensuring classification accuracy.3.A set of Marine biological call database software was developed.The software can complete the management of a large number of data,and the feature value mark of a large number of data contained in the software can quickly complete the task of keyword retrieval for a single data and complete the task of data set generation.The development software uses Qt(C++ language)to design the database interface while using My SQL to build the database model.