|Abstract: ||烏魚(Mugil cephalus)為臺灣重要經濟性魚種，然而近年來烏魚漁獲量有明顯下降之趨勢，勢必需要漁業基礎資料用於漁業管理。本研究採集臺灣冬季洄游群烏魚，進行種類判定、年齡結構分析及年齡成長分析，期望能提供漁業管理參考。為解決傳統耳石定齡作業耗時的問題，本研究收集烏魚生物學特徵及耳石外部形態，並利用隨機森林演算法 (Random Forest) 建立推估年齡之模型，及評估其應用之可行性。本研究自2014至2019年每年冬季烏魚漁汛期間 (12月中旬至翌年1月中旬)，於臺灣東北部海域採集共380尾烏魚樣本，定齡樣本為218尾。為了解該漁場隱蔽種類組成，本研究進行基因快篩鑑定，結果顯示樣本種類皆為西北太平洋三個隱蔽種中的中國北方種 (NWP1)。 分析樣本基礎漁業生物資料結果顯示，雌、雄魚體長組成範圍分別為44-67公分、40-57.5公分，體重組成範圍分別為610-3200公克、697-2034公克。利用魚體尾叉長及魚體去內臟重計算體長體重關係式，結果顯示性別間不具顯著差異，因此將雌、雄體長體重關係式合併表示：WE = 2.27 × 10-2 × LF 2.7923。定齡樣本是以耳石切片判讀年齡，樣本中判讀到最大年齡為10歲 (n=2，佔0.9%)，最低年齡為2歲 (n=28，佔12.8%)，年齡峰值為3歲 (n=70，佔32.1%)，整體年齡頻度集中於2-5歲 (n=189，佔86.7%)。利用耳石輪寬回推體長之方式增加樣本數至821尾，再利用其年齡及體長資料計算VBGE成長方程式，並將成長參數計算生長表現指數 (φ’) ，雌魚：Lt = 54.7 × [1-e -0.532 (t+0.2) ] φ’ = 3.20。雄魚：Lt = 50.1 × [1- e -0.566 (t+0.2) ] φ’=3.15，經檢定結果顯示，性別間成長方程式具有顯著差異。 建置隨機森林模型使用之特徵變量為耳石重量、魚體全長、尾叉長、魚體重、性別，因變量為年齡。以皮爾森相關係數觀測各特徵變量與年齡之相關係數顯示，以耳石重量與年齡最為相關 (0.74)，其次為魚體重(0.56)、全長(0.55)及尾叉長(0.52)。隨機森林模型建立結果顯示，參數設置為總數量Ntree = 3000，及每棵決策樹使用變數之數量mtry=1，訓練集袋外錯誤率(OOB, Out of bag error)為58.17%，最佳分類的特徵變量為耳石重。由檢定比較隨機森林模型定齡與耳石定齡兩方法間無顯著差異，再以隨機森林定齡之優缺點及應用結果，進一步評估應用之可行性及發展潛力。|
This study collected spawning Mugil cephalus (grey mullet) from surrounding waters in the northeastern Taiwan, conducted species identification, age structure analysis and age growth analysis, and provided a reference for fishery management. In order to solve the problem of time-consuming preparation of otoliths, this study collected the biological characteristics of fish and the external morphology of otoliths, and then used the Random Forest algorithm to establish a model for estimating age and evaluating the feasibility of its application. In this study, a total of 380 mullet samples were collected in the northeastern Taiwan from 2014 to 2019 at every winter fishing season (from mid-December to mid-January), with 218 otolith aged. The rapid screening results showed that all specimens belonged to NWP1 among the 3 cryptic species in the Northwest Pacific. The general fishery biological data showed that the body length of female and male fish was 44-67 cm and 40-57.5 cm, and the body weight range was 610-3200 g and 697-2034 g, respectively. Using the fork length and the body weight to calculate their relationship, the results showed no significant difference between sexes. Therefore, length-weight relationship of both sex was pooled and is WE = 2.27 × 10-2 × LF 2.7923. The fish age was estimated by otolith slices. The oldest fish in all sample was 10 years old (n=2, 0.9%), and the youngest on age was 2 years old (n=28, 12.8%). The age composition peaked at age 3 (n=70, 32%), and the age frequency is concentrated in 2-5 years old (n=189, 86.7%). In order to enlarge age dataset, method of back-calculation of fish length at each year from otolith radius were applied, and all 821 data points were used to calculate another VBGE growth equation. Also, the index (φ') for the growth parameter by sex was further calculated. The results are female：Lt = 54.7 × [1-e -0.532 (t+0.2) ], φ’= 3.20, male：Lt = 50.1 × [1- e -0.566 (t+0.2) ], φ’=3.15, and a significant difference was found between sex. The characteristic variables used in the establishment of the Random Forest model are otolith weight, body length, fork length, fish weight, and sex from the fish, and the dependent variable is age. The results of the random forest model establishment show that the Out of bag error (OOB, out of bag error) is 58.17%. The important predictor variable is otolith weight. There is no significant difference in the VBGE between otolith ageing and random forest model ageing. A comprehensive comparison between the two methods and their feasibility were discussed for future potential development of the application.