Multi-scale Transformer-based classification of bathymetric LiDAR data in shallow water environments

Bathymetric LiDAR data plays a crucial role in mapping underwater topography, enabling applications in coastal monitoring, environmental assessment, and seabed classification. However, the inherent complexity and noise in 3D bathymetric point clouds pose challenges for accurate classification. To address this, we propose a voxel-based method for efficient classification of bathymetric LiDAR data, moving beyond traditional point-wise processing of unstructured point sets. In our approach, 3D points are aggregated into structured voxel grids, and their features are embedded within each voxel. To capture spatial dependencies between voxels, we employ a window-based attention mechanism that partitions voxel features into local windows where self-attention is applied. To enhance contextual learning across regions, we adopt a shifted window strategy inspired by Swin3D, allowing voxels near window boundaries to interact with adjacent regions and reducing the locality limitation of fixed windows. To improve computational efficiency, we use a voxel selection mechanism. Using HDBSCAN, we cluster voxel features within each window based on density and retain representative voxels with distinct characteristics. This reduces redundant attention operations while preserving critical structural information. Furthermore, to capture both fine-grained and large-scale spatial patterns in bathymetric data, we design transformer heads grouped by scale. Each head group processes voxels from windows of varying sizes, enabling the model to learn multi-scale representations. The fused output captures both detailed local variations and broader contextual cues. Experimental results demonstrate the effectiveness of our method, achieving an overall classification accuracy of 75.4% on bathymetric LiDAR datasets, highlighting its capability in underwater terrain analysis.