MST-Net: A Seabed Sediment Classification Framework Based on a Multiscale Fusion and Transformer Model Using Airborne LiDAR Bathymetry Data and WorldView-2 Images

Seabed sediment data can reflect the sedimentary and ecological features of a seabed and are highly important for scientific research and applications, such as marine engineering construction, ecological protection, and resource exploration. Airborne LiDAR bathymetry (ALB) systems can obtain high-precision shallow underwater full-waveform data and terrain point clouds, providing new opportunities for sediment classification. The particularity and complexity of seabed environments make it difficult for ALB data to fully reflect sediment features. Satellite remote sensing images provide multispectral information, which can be combined with ALB data to accurately classify seabed sediment. However, multisource data fusion usually causes feature redundancy, increases model computing overhead, and affects the accuracy of sediment classification. To address the above issues, a multiscale fusion and transformer neural network (MST-Net) is introduced for seabed sediment classification in shallow sea areas. First, ALB data are fused with WorldView-2 multispectral images. From the fused dataset, the terrain, waveform, and spectral features are extracted to enhance the diversity and representativeness of the input features. Second, a transformer-based self-attention module is designed to assign adaptive weights to the 25-D input features, thereby emphasizing the most informative features. Finally, a multiscale fusion module is constructed, where feature vectors extracted at different scales are upsampled using a multilayer perceptron and subsequently fused, enabling the model to effectively capture geometric structures at multiple spatial scales. The proposed MST-Net model was evaluated using fused data from the shallow coastal areas surrounding Wuzhizhou island, Hainan Province. The proposed MST-Net yielded excellent performance metrics, with an overall accuracy of 0.985, a mean intersection over union of 0.957, and a kappa coefficient of 0.934. With the precision, recall, and F1 score all exceeding 0.97, the model successfully categorized seabed sediments into coral reef, gravel, and sand. Following training, the optimal model was applied to map the sediment distribution across the shallow areas of Wuzhizhou island. The robustness and efficacy of this approach were further substantiated through rigorous comparative and ablation experiments.