Semi-Supervised Learning for Sediment Classification Using Convolutional Neural Networks with Digital Elevation Model and Backscatter Data

Abstract

Accurate sediment classification is crucial for advancing marine research, environmental monitoring, and sustainable seabed use. However, acquiring large amounts of labeled data in such settings is often challenging, expensive, and time-consuming. To address this limitation, a semi-supervised learning framework has been proposed that leverages convolutional neural networks for sediment classification using both labeled and unlabeled data. The approach utilizes pseudo-labeling, where confident predictions on unlabeled samples are iteratively incorporated into training to enhance model generalization. The method is applied and evaluated on a dataset that includes multi-modal inputs such as a digital elevation model and multibeam sonar backscatter data. Experimental results indicate that semi-supervised learning with convolutional neural networks can achieve high classification accuracy in scenarios characterized by limited labeled data and a large volume of unlabeled data. This approach highlights the potential of deep learning combined with semi-supervised strategies for efficient underwater environment classification.