Underwater Acoustic Detection and Classification for Cetaceans’ Vocalizations

被動聲學已被視為觀察海洋生物及進行長時間水下環境監測的一項重要工具，由於聲學的錄音資料量龐大，故需要即時有效的自動化偵測器，快速辨識聲音訊號、篩選重要數據。本研究之聲音偵測器，主要可分為「特徵萃取」和「分類」兩項模組；「特徵萃取」模組，利用熵偵測器搭配時域端點偵測法，截取重要音訊，並在其時域及頻域上萃取訊號的特徵參數；「分類」模組，運用倒傳遞類神經網路作為辨識工具，本研究樣本來源採集康乃爾大學網路資料庫並以臺灣東北海域及龜山島附近海域常出現之海洋哺乳動物聲音作為辨識對象。將所萃取的特徵參數作正規化及設定目標值後，輸入類神經網路作訓練，直到網路穩定收斂，再進行測試。 目前本方法可以獲得良好的正確判斷率，未來如能增進自動化偵測器之辨識準確度，還可應用於辨識其他未知音訊，透過自動化偵測及分類的能力，輔助專業人工辨識員，達到「省時、省工、省力」等多項優點。Passive acoustics has been well recognized as an important tool for observing marine animals and long-term underwater environment monitoring. Since the amount of data is enormous, it is needed to have an effective auto-detector to select critical features and classify their patterns from the recorded acoustic signal. In this study, we had developed an automatic detector with both the feature extraction and classification modules. In the feature extraction module, we select features from the entropy and end-point of the time signal. Then, we normalized the extracted features as inputs for the classification module based on the theory of back propagation neural network (BPNN). The BPNN will be trained and tested using the cetaceans’ acoustic signals from database of Cornell University Macaulay Library Marine Collection until the network becomes stable and convergent. The selected samples are commonly found cetaceans from the northeastern offshore of Taiwan and Guishan Island. Currently, our detector had obtained fairly good recognition rate for classifying cetaceans. In the future, our automatic detector can be applied to classify similar acoustic signals if we can improve the accuracy. We believe the proposed automatic detector will be a robust tool, which supersedes the experienced human operators due to less time consuming and low labor cost