CUG-STCN: A seabed topography classification framework based on knowledge graph-guided vision mamba network

Abstract

Multibeam sounding is a high-precision remote sensing method for seabed detection. Seabed topography classification is crucial for marine science research, resource exploration and engineering. When using multibeam data for seabed topography automatic classification, the fuzzy boundaries of different topographic entities, redundancy of multimodal data, and the lack of geological knowledge guidance have led to low classification accuracy. Thus, a knowledge graph-guided vision mamba seabed topography classification network (CUG-STCN) was constructed, consisting of three modules: (1) The long sequence modeling mamba-based encoder addresses the fuzzy seabed topography boundary. It uses 2D-selective-scan to create image blocks in different scanning directions. By combining with the selective state space model to capture long-range dependencies and ensure transmission of spatial context information while maintaining linear computational complexity. (2) The cross-modal information interaction and fusion module addresses the redundancy of multimodal information. By employing a bidirectional information interaction mechanism, it captures the correlations of seabed topography between different modalities and achieving feature fusion. (3) The seabed topography knowledge graph-guided semantic perception module guides the geological knowledge. It constructs seabed topography knowledge vectors through entity query and word embedding, using the similarity between vectors to create a similarity measurement matrix. It provides geological knowledge, enhancing the modeling capability of complex seabed topography relationship. CUG-STCN achieves OA of 90.11% and mIOU of 48.50%, outperforming six mainstream networks, which at most, achieve the OA and mIOU improvements of 5.37% and 14.18%. Notably, the application of CUG-STCN in other regions demonstrates its strong generalization performance.