外文摘要：This study explores the application of convolutional neural networks (CNN) for age estimation in Pacific Bluefin Tuna (Thunnus orientalis) using otolith images. The objective is to assess the feasibility of CNNs as a cost-effective tool for fish age determination, while evaluating the potential improvements with imputing missing values in the auxiliary dataset and image augmentation techniques. Additionally, a user-friendly web tool is developed to enable public access to the CNN model. Three trained models, Baseline, Otolith Mass Imputation (OMI), and Otolith Mass Imputation and Image Augmentation (OMIA), are compared and evaluated based on performance metrics. The results highlight the superiority of the OMIA model, achieving the highest accuracy (+/- 1-acc = 72.81%) and lowest coefficient of variation (CV=7.38%). The model's predicted age distribution closely resembles the ground truth, as well as the parameters of the von Bertalanffy growth function. Heat maps reveal that the attributes used by the model, particularly the opaque zones on the ventral arm of the otolith, mimic the age identification strategies employed by human experts. However, the study identifies challenges, including poor performance and negative impacts on predictions due to data imputation for age groups (ages 4-5 and 25-27) with limited samples. Despite these limitations, this study represents a significant step towards machine learningbased age estimation, serving as a valuable aid in traditional fish aging studies. The implications for management and future research directions are also discussed.
外文关键词：deep learning;CNN;Bluefin tuna;Fish otoliths;Age estimation;Web tool
作者：Kuo, Yan-Fu;Ma, Tsung-Hsiang;Chang, Yi-Jay;Shiao, Jen-Chieh;Jin, Chien-Bang
作者单位：Natl Taiwan Univ;Minist Agr
期刊名称：FISHERIES RESEARCH
期刊影响因子：0.0
出版年份：2024
出版刊次：274
原文传递申请：江苏省科技资源（工程技术文献）统筹服务平台