Seabed Sediment Mapping Based on Physical Property-Guided Graph Fusion Network and Multibeam Data

Seabed sediment mapping is extremely useful in the exploration of underwater mineral resources etc. Although traditional surveys produce high-precision results, they cannot meet the needs of large-scale mapping. Acoustic technology is preferred owing to the large detection range and characterization of sediment’s backscattering features. However, inherent spatial heterogeneity makes acoustic data-based sediment feature extraction challenging, limiting the mapping accuracy. To address the issue, a physical property-guided graph fusion network was proposed and tested on a self-built dataset, which was based on multi-beam acoustic and driven topographic data. The network has three modules: (1) Dense dilated pyramid and attention-based dual-branch encoder. This is designed to gradually capture and fuse sediment features at various scales and neighborhoods, within those two data-based branches. (2) Graph fusion-based multimodal spatial context feature extraction. The stacked multimodal encoder features were used as graph node features, and to adaptively generate the adjacency matrix, better characterizing sediments’ spatial relationships. Through graph convolutions, spatial context features were further learned and multimodal features were thoroughly fused from the spatial and channel dimensions. (3) Physical property-guided optimization. Considering the correlation of sediment types and physical properties, these two label data-based decoders were performed and corresponding loss summation were used to optimize the network. The proposed network outperforms eight mainstream ones, at least achieving the OA and mIOU improvements of 3% and 8%, and obtains satisfactory seabed sediment mapping result. Overall, the physical property-guided graph fusion network and multimodal dataset were positive for seabed sediment mapping. Our codes are available at https://github.com/WhyBlue/PGG-Net.