Superpixel-guided graph convolution for spatially continuous prediction of deep-sea manganese nodule coverage from multibeam acoustic data

Deep-sea manganese nodules, containing critical metals, have emerged as key resources for marine mineral exploration. Accurate and spatially continuous quantification of nodule coverage is essential for evaluating resource potential and environmental assessment. Multibeam echosounder (MBES) systems provide full-coverage acoustic observations whose backscatter intensity and terrain attributes are linked to nodule size, density, and sediment roughness, providing a basis for inferring nodule distribution. However, most existing acoustic-based approaches rely on discretized coverage classes, limiting their ability to represent the continuous and highly heterogeneous nature of deep-sea nodule fields. To address this limitation, we propose a spatial prediction framework for deep-sea manganese nodule coverage based on superpixel-guided graph convolution, representing a novel application of graph neural networks in deep-sea mineral assessment. First, terrain and backscatter features are extracted from MBES data, and deep representations are learned using a ResNet-18-based module. Second, handcrafted features are compressed via PCA to generate a pseudo color representation, subsequently segmented using superpixel segmentation. Finally, a graph convolutional network is subsequently trained to achieve spatially continuous and high precision coverage prediction. Experiments using MBES data from the western Clarion-Clipperton Zone demonstrate that the proposed method achieves an RMSE of 0.0312, and an R 2 of 0.9545. Compared with RF, KNN, and CNN regression models, the RMSE is reduced by 0.0398, 0.0418, and 0.0304, while R 2 improves by 0.1815, 0.1985, and 0.1099. These results demonstrate that the proposed framework effectively captures the spatial distribution of deep-sea manganese nodules, supporting environmental assessment, resource exploration, and future deep-sea mining operations.