Multi-Granularity Graph Neural Network for Satellite-Assisted Marine Environmental Field Reconstruction over Sparse Observation Grids

Marine remote sensing combines broad-coverage satellite observations of the ocean surface with sparse in-situ and underwater observations. However, reconstructing continuous three-dimensional environmental fields remains challenging because of irregular sampling, vertical non-stationarity, and bathymetric barriers. In this paper, a multi-granularity graph collaborative neural network (MG-GCNN) is proposed for high-fidelity field reconstruction and situational awareness in sparse monitoring scenarios. The network abstracts discrete observation points as heterogeneous nodes in the topological graph structure. By incorporating high-resolution bathymetry as a geometric prior, terrain-aware graph construction, vertical feature integration, multi-granularity aggregation, and self-supervised masked node reconstruction are jointly used to capture spatial and vertical dependencies under limited observation availability. Experimental results show that MG-GCNN significantly outperforms baseline interpolation and convolution models in terms of reconstruction accuracy, especially in regions with complex underwater terrain and extreme sampling sparsity. The reconstructed environmental fields can potentially provide three-dimensional environmental inputs for subsequent ocean-acoustic propagation modeling, underwater sensing, and related marine applications.