Seafloor Terrain Semantic Segmentation for Autonomous Deep-Sea Mining Operations

Abstract Deep-sea mining vehicles rely on forward-looking sonar systems to collect critical environmental data for mapping, navigation, and operational safety. Real-time semantic segmentation of seabed terrain from sonar images is essential for accurately interpreting the underwater environment. However, these images present several challenges, including low resolution, substantial noise, and indistinct textures and shapes. Additionally, seabed terrain features vary across multiple spatial scales, which complicates the extraction of meaningful information using conventional image processing techniques. Such limitations hinder the precision and reliability of mapping efforts necessary for effective deep-sea operations. To address these challenges, a multi-scale semantic segmentation model has been developed specifically for sonar images. The model incorporates a downsampling module based on Discrete Wavelet Transform (WT-DM), which suppresses noise while preserving critical terrain features, ensuring clearer input for subsequent analysis. Moreover, a TriScale Attention Module (TAM) is introduced to capture features at varying spatial resolutions. This module enhances segmentation performance by allowing the model to focus on relevant patterns across different scales. In the absence of publicly available datasets, experiments were conducted using a custom-built seabed terrain dataset. The proposed model achieved a precision of 82.4% and a mean intersection over union (MIoU) of 79.6%, demonstrating its effectiveness. Comparative evaluations with other mainstream models further confirmed the superior performance and competitiveness of the proposed approach.