林思民林子皓Lin, Si-MinLIN, Tzu-Hao徐子妍Hsu, Tzu-Yen2023-12-082028-07-312023-12-082023https://etds.lib.ntnu.edu.tw/thesis/detail/68d71ad177ebd2fdd61c905ca24630b2/http://rportal.lib.ntnu.edu.tw/handle/20.500.12235/121342近年來，聆聽水下聲音逐漸被視為一種監測魚類群聚生態的新興技術，但受限於資料庫的匱乏與缺乏能夠有效降低噪音干擾的聲音分析工具，魚類聲音的研究仍有許多挑戰。本研究透過兩個方面探討被動式聲學監測魚類群聚生態的可行性。首先，透過國內外的聲音資料庫蒐集魚類聲音資料，並針對可能會出現在台灣東北海域的礁岩性物種進行聲音特徵分析，並以流形學習探討聲音多樣性與物種多樣性之間的關係。結果顯示，礁岩性魚類的聲音在時序和頻率特徵上有很大的變化，可以透過聲音特徵的種間差異將各科魚類區分成不同的聲音功能群，顯示種間差異是貢獻礁岩性魚類聲音多樣性的主要來源之一。為了實際觀察魚類聲音多樣性在礁岩生態系的長期變化趨勢，本研究於基隆市望海巷潮境海灣資源保育區內設置水下錄音測站，透過聲源分離模型自動化偵測魚類聲音，分析魚類聲音豐度的季節性變化模式，並透過頻率和時序特徵在降維後的分布密度變化量測聲音多樣性。結果顯示魚類發聲活動與聲音多樣性的高峰集中在春夏兩季，但同時也受到極端海溫、颱風以及人為活動等因子影響，因而導致在2020至2022三年間觀察到的季節性變化趨勢略有不同。本研究結果證明運用機器學習分析長時間水下錄音可以有效評估魚類聲音多樣性，是了解礁岩性魚類群聚組成變化和探討其受到環境、人為變動影響的關鍵指標。此技術的後續應用將有助於保育主管機關與研究人員長期監測礁岩性魚類群聚生態，並為魚類多樣性的保育管理提供重要參考依據。未來建議擴充聲音資料庫並增加長期錄音樣點，探討魚類聲音多樣性在不同棲地、時間的變化，更進一步了解魚類群聚面對環境變遷和人為干擾的反應。In recent years, passive acoustic monitoring (PAM) has emerged as a promising technique for monitoring soniferous fish. However, the application of PAM in fish community assessment still faces challenges due to limited databases and a lack of effective tools to avoid noise interference. This research sought to address these challenges by integrating machine learning techniques into the PAM of fish community. First, audio data of reef-associated species were collected from online sound databases. Subsequently, manifold learning was applied to measure the spectral and temporal rhythmic features of fish sounds and to analyze the relationship between sound diversity and species diversity. The results showed that three functional groups of families can be classified depending on the inter-specific variations observed in the temporal and spectral domains. This suggests that inter-specific variation is a major source of fish sound diversity. To investigate the long-term trends of fish sound diversity in reef ecosystems, an underwater recording station was established in Wanghaixiang Chaojing Bay Resource Conservation Area, Keelung, Taiwan. Fish sounds were automatically detected using a source separation model, and the seasonal patterns of sound richness and diversity were analyzed. The results showed that the peak of sound richness and diversity occurred in spring and summer, but were still influenced by factors such as extreme seawater temperature, typhoons, and human activities. These findings suggest that the use of machine learning in the analysis of long-duration underwater recordings facilitates the assessment of fish sound diversity, which serve as an important ecological indicator for understanding community structure and associated changes in response to natural and anthropogenic stressors. Future recommendations include expanding fish sound databases and increasing the number of long-term recording sites will provide opportunities to explore the spatial-temporal dynamics of fish sound diversity, thereby improving conservation management of fish community.被動式聲學監測礁岩性魚類聲音多樣性特徵擷取群聚生態學生態擾動passive acoustic monitoringreef-associated fishsound diversityfeature extractioncommunity ecologyecological disturbanceetd