Scholarlink

<p>The accurate classification of underwater fish species is important for biodiver-sity monitoring, sustainable fisheries management, and aquaculture. However, underwater fish classification remains challenging due to class imbalance, varying imaging conditions, and high visual similarity among species. This study presents a comprehensive evaluation of deep learning architectures, loss functions, and ensemble strategies for underwater fish species classification. Using the Mark Daniel Lampa Kaggle Fish Dataset comprising 13,304 images across 31 species, duplicate images were removed using a combination of MD5 hashing and perceptual hashing, resulting in a cleaned dataset for model training and evaluation. EfficientNet-B0, EfficientNetV2-S, Swin-Tiny Transformer, and ConvNeXt-Tiny were trained using both class-weighted cross-entropy loss and focal loss. The experimental results indicate that class-weighted cross-entropy loss consistently provided stronger performance for individual models, whereas focal loss produced a marginal improvement when used with the soft-voting ensemble. Among the individual models, EfficientNet-B0 achieved the highest accuracy of 94.81%. Ensemble learning further improved performance, with the four-model soft-voting ensemble achieving a test accuracy of 97.21% and a macro-F1 score of 0.9678 on the cleaned Kaggle dataset. An external dataset of 310 internet-sourced fish images was additionally used to assess model generalization under domain-shift conditions, where the ensemble achieved an accuracy of 76.39% and a macro-F1 score of 0.7335. While performance decreased substantially under domain shift, the results highlight the challenges of generalizing fish-species classification models beyond the training distribution. Fish species classification represents an important component of future end-to-end underwater monitoring systems incorporating fish detection and localization modules. The study provides a benchmark for future research on underwater fish-species classification, ensemble learning, and real-world deployment. The source code is openly available on GitHub and permanently archived on Zenodo (DOI: <a href="https://doi.org/10.5281/zenodo.19315919">https://doi.org/10.5281/zenodo.19315919</a>).</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

A comparative evaluation of deep learning architectures, loss functions, and ensemble strategies for underwater fish species classification

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年8月5日

查看全文
  • Prithvi Shenoy,
  • Ramyashree,
  • S. Raghavendra,
  • B. N. Anoop

摘要

The accurate classification of underwater fish species is important for biodiver-sity monitoring, sustainable fisheries management, and aquaculture. However, underwater fish classification remains challenging due to class imbalance, varying imaging conditions, and high visual similarity among species. This study presents a comprehensive evaluation of deep learning architectures, loss functions, and ensemble strategies for underwater fish species classification. Using the Mark Daniel Lampa Kaggle Fish Dataset comprising 13,304 images across 31 species, duplicate images were removed using a combination of MD5 hashing and perceptual hashing, resulting in a cleaned dataset for model training and evaluation. EfficientNet-B0, EfficientNetV2-S, Swin-Tiny Transformer, and ConvNeXt-Tiny were trained using both class-weighted cross-entropy loss and focal loss. The experimental results indicate that class-weighted cross-entropy loss consistently provided stronger performance for individual models, whereas focal loss produced a marginal improvement when used with the soft-voting ensemble. Among the individual models, EfficientNet-B0 achieved the highest accuracy of 94.81%. Ensemble learning further improved performance, with the four-model soft-voting ensemble achieving a test accuracy of 97.21% and a macro-F1 score of 0.9678 on the cleaned Kaggle dataset. An external dataset of 310 internet-sourced fish images was additionally used to assess model generalization under domain-shift conditions, where the ensemble achieved an accuracy of 76.39% and a macro-F1 score of 0.7335. While performance decreased substantially under domain shift, the results highlight the challenges of generalizing fish-species classification models beyond the training distribution. Fish species classification represents an important component of future end-to-end underwater monitoring systems incorporating fish detection and localization modules. The study provides a benchmark for future research on underwater fish-species classification, ensemble learning, and real-world deployment. The source code is openly available on GitHub and permanently archived on Zenodo (DOI: https://doi.org/10.5281/zenodo.19315919).