Hierarchical active learning for efficient semi-supervised seagrass image classification

Author Identifier

Md Kislu Noman: https://orcid.org/0000-0001-6281-1464

Syed Mohammed Shamsul Islam: https://orcid.org/0000-0002-3200-2903

Riaz Ul Haque Mian: https://orcid.org/0000-0001-6550-5753

Document Type

Conference Proceeding

Publication Title

Proceedings - 2024 25th International Conference on Digital Image Computing: Techniques and Applications, DICTA 2024

First Page

553

Last Page

560

Publisher

IEEE

School

School of Science

Funders

Shimane University / Edith Cowan University

Comments

Nawayai, F. A. B. M., Noman, M. K., Islam, S. M. S., & Mian, R. U. H. (2024, November). Hierarchical active learning for efficient semi-supervised seagrass image classification. In 2024 International Conference on Digital Image Computing: Techniques and Applications (DICTA) (pp. 553-560). IEEE. https://doi.org/10.1109/DICTA63115.2024.00086

Abstract

Seagrass meadows play a crucial role in marine ecosystems by providing habitat and food for numerous marine species, stabilizing sediments, improving water quality through filtration and sequestering carbon dioxide from the atmosphere, thus contributing to climate regulation. Image classification for seagrass is essential for monitoring and understanding its distribution, health, and ecological significance in coastal and marine environments. Mislabeled and limited labeled data impairs image classification performance, especially in applications like seagrass detection. This paper proposes a semi-supervised hierarchical active learning-based technique (HALT) to address this challenge. The proposed HALT combines strategic data selection from active learning and knowledge transfer from a pre-trained model to overcome performance degradation in semi-supervised models. The experiment presented in this paper on the publicly available 'DeepSeagrass' dataset demonstrates that HALT can improve the overall accuracy of the base models InceptionV3 and ResNet50 by 1.4% when they are trained on only 30% labeled data. This approach reduces the effort required for human labeling, effectively addresses mislabeled data, and helps achieve high classification accuracy in semi-supervised learning.

DOI

10.1109/DICTA63115.2024.00086

Access Rights

subscription content

Link to Full Text

Share

 
COinS