Abstract
Three-dimensional geological modeling technology has been widely and deeply applied in the fields of mineral, oil, and gas exploration, underground space utilization, and geological disaster forecasting. The intelligent three-dimensional quantitative characterization of complex subsurface structures has become a hotspot in the field of geoinformatics in recent years. In this study, we propose a framework to establish a bond between geological objects and graph structures by using an improved GraphSAGE neural network (namely GGBondSAGE). The improved GraphSAGE network is designed by graphically characterizing three-dimensional geometric and topological relationships of stratigraphic structures. The Bi-mean aggregation mechanism and the dual-channel hybrid splicing strategy proposed in GGBondSAGE not only capture the spatial morphological features of the geological interfaces of strata, but also improve the robustness of the network model. Several experiments were performed on three typical datasets. The results and comparison analysis show that the proposed GGBondSAGE is an efficient framework for characterizing heterogeneous stratigraphic structures.