We propose DiffLiG, a graph-based model that addresses station heterogeneity and oversmoothing via spatiotemporal encoding and diffusion-based ensemble refinement for accurate, structure-aware precipitation interpolation.