Symmetry-aware graph optimization for scalable multisource remote-sensing data-fusion networks

Abstract

Abstract The multisource remote-sensing fusion can be used to complement LiDAR, Sentinel-1 synthetic aperture radar (SAR) and UAV imaging to enhance the spatial analysis. But it is a problem that for large multimodal datasets it is not cost effective to represent this data as a spatial-feature graph for fusion and distributed processing. In this work, a symmetry-aware graph optimization framework is proposed to eliminate redundant computation in the multisource remote-sensing data-fusion workflow. Registered remote-sensing units are depicted as graph nodes and spatial neighbouring and multimodal feature similarity are represented as graph edges. The ability to recognize structurally equivalent node groups by graph automorphism combined with a partitioning of the orbits enables optimization to be carried out on a reduced orbit graph rather than the node-level graph. The optimized orbit-level solution is then projected in the original graph. The framework was tested with the graph sizes ranging from 1,000 to 50,000 nodes measured. With 1,000 nodes, the execution time dropped from 12.5 s, which was obtained using conventional optimization, to 8.2 s, obtained using distributed optimization, to 5.1 s. For 50,000 nodes, execution time decreased from 890.4 s to 520.8 s to 245.3 s, respectively. It is observed from these results that the reduction of graph based on the symmetry property enhances the computational scalability for large spatial-feature graph employed in the multisource remote-sensing fusion pipelines.