Deep learning-driven automatic detection of mucilage event in the Sea of Marmara, Turkey

Hacıefendioğlu K., Başağa H. B., Baki O. T., Bayram A.

NEURAL COMPUTING & APPLICATIONS, vol.35, no.9, pp.7063-7079, 2023 (SCI-Expanded) identifier identifier

  • Publication Type: Article / Article
  • Volume: 35 Issue: 9
  • Publication Date: 2023
  • Doi Number: 10.1007/s00521-022-08097-1
  • Journal Indexes: Science Citation Index Expanded (SCI-EXPANDED), Scopus, Academic Search Premier, PASCAL, Applied Science & Technology Source, Biotechnology Research Abstracts, Compendex, Computer & Applied Sciences, Index Islamicus, INSPEC, zbMATH
  • Page Numbers: pp.7063-7079
  • Keywords: Convolutional neural networks, Grad-CAM, Deep learning, Mucilage, MARINE MUCILAGE, PHYTOPLANKTON COMPOSITION, ADRIATIC SEA, DYNAMICS, WATERS, BAY
  • Karadeniz Technical University Affiliated: Yes


A slimy and sticky structure is formed in sea surface due to the excessive proliferation of plantlike organisms called phytoplankton, which is formed by the combination of many biological and chemical conditions, the increase in sea temperature and bacterial activities accordingly. The rapid detection of this structure called mucilage is very important in terms of early intervention and cost determination. Remote sensing methods have been used quite frequently in recent years for the automatic classification and localization of such events with the help of satellite images. Deep convolutional neural networks (DCNNs) trained on mucilage images are applied as a very successful method thanks to their ability to automatically extract superior features. The studies carried out for the target point detection obtained as a result of extracting the visual features from natural images with these networks have reached the goal. In this study, transfer learning methods are proposed to improve the detection of mucilage areas from the satellite images. The Sea of Marmara, which has been difficult times due to the mucilage events, was selected as the study area. The dataset was trained to classify mucilage images with the convolutional neural network (CNN) models and then reused to localize mucilage areas. Residual networks (ResNet)-50, visual geometry group (VGG)-16, VGG-19, and Inception-V3 were used for individual CNN models. Gradient-weighted class activation mapping (Grad-CAM) technique was used to visualize the learned behavior. A custom CNN model was created, and comparisons were made with the real mucilage areas with the intersection over union considering the most efficient convolutional layer to better localize the mucilage areas. It was concluded that the custom CNN model has showed superior localization performance compared to other models.