To explore an efficient strategy for intelligent bedrock mapping that can be applied in the areas with coexisting Quaternary coverages and bedrock outcrops, a graph convolutional network (GCN) was implemented for bedrock classification using stream sediment geochemical samplings in the Chahanwusu River area, Qinghai Province, China. The sampling points were organized into a terrain weighted directed graph (TWDG) using Delaunay triangulation to capture the upstream-downstream relationships among the geochemical sampling points. The experimental results indicate that the semi-supervised GCN models, only using 20% of the labeled sampling points, achieved accuracies of 68.20% and 78.31% in ten-type and five-type bedrock discrimination, respectively. In conclusion, it is feasible to map the bedrock type through the concentrations of elements on the stream sediment geochemical sampling points. The proposed data-driven GCN bedrock classification method not only improves the efficiency of bedrock mapping but also may be applied in a large area.