A Semisupervised Classification Method for Riverbed Benthic Sediments Using Integrated Superpixel Segmentation and Confident Learning Sample Enhancement

Riverbed sediments are crucial in river ecosystems, significantly impacting water quality, environmental protection, and water resource management. With the increasing ease of obtaining high-resolution river data, integrating multibeam technology and field sampling has become one of the popular methods for identifying riverbed sediments. However, limitations in underwater field sampling techniques result in insufficient sampling data, leaving many unlabeled high-resolution acoustic data underutilized, which hinders practical training and deployment of classifiers. To address this issue, this article proposes a sediment sample enhancement method based on superpixel segmentation, confident learning (CL), and active learning (AL) to improve model performance. First, superpixel segmentation is employed to expand field sampling data and increase the number of samples. Next, CL removes noisy labels from the expanded samples, improving data quality. Finally, the centers of superpixel blocks are introduced as unlabeled data and AL techniques are used to select high-quality unlabeled data for labeling. Experiments combining multibeam data from the FuchunRiver region with actual sediment sampling data classify five riverbed sediments. The proposed method expands the original 18 sampling points to 8333 effective sampling points. Comparison results show that the proposed method achieves a classification accuracy of 85.3%, an improvement of 2.9% to 9.9% compared to traditional classification methods, offering a new approach for large-scale benthic environment detection in river basins.