CNN Multibeam Seabed Sediment Classification Combined with a Novel Feature Optimization Method

The classification of seabed sediments is an essential aspect of marine spatial planning and management. Multibeam echo sounders (MBESs) have been widely used for efficient and reliable sediment classification. However, feature optimization is a complex and difficult task that requires considerable effort and domain expertise. In addition, selecting the optimal number of features and avoiding redundancy can pose challenges and ultimately lead to suboptimal classification performance. To address these limitations, this paper proposes a new approach to classifying seabed sediments by combining the parametric uniform manifold approximation and projection (PUMAP) feature optimization method with a convolutional neural network (CNN). MBES data are used to generate the main textures and bathymetric features, which are then transmitted via the CNN network. PUMAP is used to construct a representative graph from the extracted features and optimize the graph in a low-dimensional space using the weight of an autoencoder. The proposed approach was tested with MBES data from an Irish Sea survey. The results show that the proposed approach outperforms conventional sediment classification methods, with total accuracy of 97.20%, an F1 score of 97.31%, and a kappa of 0.9638.