抹香鯨喀嗒聲偵測方法研究

Abstract

本研究利用花蓮外海拖曳式錄音與台東外海底錠式長期監測資料建立抹香鯨（Physeter macrocephalus）喀嗒聲之自動化偵測流程，並比較 Teager–Kaiser Energy Operator（TKEO）、EMD（Empirical Mode Decomposition）與 SE（Sample Entropy）三種方法在不同音檔中的偵測表現。透過 12 筆人工標記音檔之參數掃描結果顯示，EMD 結合脈衝寬度條件（width_percent）在跨音檔表現上具較高穩定性。 方法比較後選定IMF = 1、α = 2 與 width_percent = 0.6 作為最終參數，並結合 TKEO 與 EMD之階層式偵測流程，以提升長時間資料分析之運算效率。該流程應用於台東外海連續三個月（約 2118 小時）之實際錄音資料，可穩定在合理運算時間內完成長時間偵測。於2023/07/23 06:00–07:00之完整人工標記錄音資料中，所建立之偵測流程可達 Recall = 0.916、Precision = 0.865、F1-score = 0.889，再進一步針對三個月長期偵測結果進行日夜分析，抹香鯨喀嗒聲在日夜時段之出現頻率與每日、每小時變化程度存在差異，呈現可觀察之日夜分布趨勢。此外，本研究亦針對偵測結果進行喀嗒聲類型分類與喀嗒聲間隔（Inter-click Interval, ICI）分析並結合聲學傳播模型（ASORPS）進行偵測範圍之估算作為長期被動式聲學監測結果輔助依據。因此透過建立自動化偵測與分析流程，可作為臺灣東部海域抹香鯨長期被動式聲學監測研究之基礎參考。

This study establishes an automated detection framework for sperm whale (Physeter macrocephalus) clicks using towed acoustic recordings from offshore Hualien and long-term bottom-moored acoustic monitoring data from offshore Taitung. Three detection methods—Teager–Kaiser Energy Operator (TKEO), Empirical Mode Decomposition (EMD), and Sample Entropy (SE)—were evaluated across different recordings. Parameter scans based on 12 manually annotated datasets indicate that EMD combined with a pulse-width constraint (width_percent) provides higher stability in cross-file performance. Based on the comparative evaluation, IMF = 1, α = 2, and width_percent = 0.6 were selected as the final parameter set. A hierarchical detection framework integrating TKEO and EMD was then developed to improve computational efficiency for long-duration analyses. The proposed framework was applied to approximately three months (about 2118 hours) of continuous recordings from offshore Taitung, demonstrating stable performance and feasibility within reasonable computational time. Using one hour of fully annotated data from 06:00 to 07:00 on July 23, 2023, the detection framework achieved a Recall of 0.916, a Precision of 0.865, and an F1-score of 0.889. Further statistical analyses of the three-month detection results revealed differences in the occurrence frequency and temporal variability of sperm whale clicks between daytime and nighttime periods, indicating observable diel distribution patterns. In addition, click-type classification and inter-click interval (ICI) analyses were conducted based on the detection results. Detection range estimates were further derived by incorporating an acoustic propagation model (ASORPS), providing supplementary information for interpreting long-term passive acoustic monitoring results. Overall, the proposed automated detection and analysis framework can serve as a reference for long-term passive acoustic monitoring studies of sperm whales in eastern Taiwan waters.