Multi-Label Seabed Image Classification: A Comparative Study of Deep Learning and Machine Learning Models

Anis Bel Hadj Hassin, ESPRIT School of Engineering, 1-2 rue André Ampère, 2083 Pôle Technologique, Tunisia AND University of Tunis El Manar, National Engineering School of Tunis, LAMSIN 1002, Tunis, TunisiaFollow
Elyes Omri, ESPRIT School of Engineering, 1-2 rue André Ampère, 2083 Pôle Technologique, TunisiaFollow

Abstract

This study evaluates the effectiveness of deep learning and machine learning models for classifying seabed images into three ecologically significant categories: algae, rocks, and sand. We investigated five models: ResNet50, VGG16, a custom-designed convolutional neural network (CNN), support vector machines (SVM), and random forests (RF). Deep learning models were fine-tuned on 1259 seabed images, with the custom CNN architecture optimized using depthwise separable convolutions and global average pooling. For SVM and RF, features were extracted using a pre-trained VGG16 network. Class weighting was applied to address potential imbalances in the dataset. The custom CNN achieved the highest validation accuracy (82.01%), followed closely by SVM (82.14%). ResNet50 and VGG16, despite their success in general image classification, achieved lower accuracies of 32.27% and 60.71% respectively. The random forest model performed moderately well with an accuracy of 78.17%. These findings highlight the potential of customized deep learning architectures and traditional machine learning approaches in the domain-specific task of seabed image classification.

First Page

Last Page

Recommended Citation

Hassin, Anis Bel Hadj and Omri, Elyes (2026) "Multi-Label Seabed Image Classification: A Comparative Study of Deep Learning and Machine Learning Models," Journal of King Abdulaziz University: Marine Science: Vol. 36: No. 1, Article 3.
DOI: https://doi.org/10.64064/1658-4325.1025