以自動錄音裝置與人工智慧辨識探討棉花嶼大水薙鳥發聲活動

Abstract

管鼻目鳥類面臨多種威脅面臨族群下降，但其繁殖地位於偏遠島嶼且多在夜間活動並選擇洞穴為巢，因此研究者及管理單位缺乏其資訊與監測。為解決此困境，本研究嘗試以被動式聲學監測中的自動錄音裝置作為工具，並結合以深度學習為基礎的人工智慧辨識模型探討大水薙鳥（Calonectris leucomelas）於棉花嶼的發聲情形。本研究架設 6 台自動錄音裝置自2023 年 2 月至 11 月間收錄日落到日出的連續錄音資料，並採用以卷積循環式神經網路為基礎的自動辨識工具 Go Go Owl Ranger 進行模型訓練。在閾值分數為 0.2362 時精確度、召回率與 F1 分數皆為 0.9058，儘管有誤判成其他物種但仍顯示出極佳的辨識表現。偵測結果顯示島的東北側發聲頻率最高，同時大水薙鳥平均鳴叫與錄音機周圍有無巢穴沒有明顯差異，推測可能有潛在的繁殖巢穴。另外，本研究發現大水薙鳥有兩個鳴叫高峰時段，判斷這可能是跟其繁殖間的關鍵期與月相週期間的變化有關。本研究旨在借助聲學監測工具和人工智慧辨識，進行。若未來能在島上執行長期聲學監測，並持續優化辨識模型，將有助於推進大水薙鳥聲學研究及其保育應用。

Although birds of the Procellariiformes face various threats and population declines, researchers and management organizations lack information and monitoring on their breeding status on remote islands where they are nocturnal and burrow-nesting. This study attempted to investigate the vocalizations of Streaked Shearwater (Calonectris leucomelas) on Mienhua Islet using an automated recording device as an optimal solution and an artificial intelligence recognition model based on deep learning. This study collected continuous audio recordings from sunset to sunrise over eight months, and the Go Go Owl Ranger, a software based on a convolutional loop neural network, was used to train the automatic sound recognition model. The accuracy, recall, and F1 score were all 0.9058 at a threshold score of 0.2362, which showed excellent recognition performance despite the misidentification of other species. Results showed that the highest frequency of vocalizations was on the northeastern side of the island and that there was no significant difference between the average call of the streaked shearwater and the presence or absence of nests in the vicinity of the recorder, suggesting that there may be potential breeding burrows. In addition, two peak calling periods were observed in this study, and it was concluded that this may be related to the critical period between breeding and also may affected by the lunar cycle. This study aims to create a comprehensive seabird survey framework in Taiwan with the help of acoustic monitoring tool and artificial intelligence. Suppose long-term acoustic monitoring can be carried out on the island and the identification model can be continuously optimized. In that case, advancing the acoustic study of the nagaimo and its conservation application will be helpful.