Geometry-Aware Graph Fusion Improves Rainfall Reconstruction

Fine-scale rainfall reconstruction is crucial for accurate urban flood modeling, but existing sensing systems provide data with incompatible spatial supports: point measurements from gauges, path measurements from microwave links, and gridded areas from radar or satellites. Current heterogeneous graph approaches often reconcile these sources in feature space, overlooking the critical geometric differences of their supports. This research proposes a novel geometry-aware multi-support heterogeneous graph neural network. It represents each observation type—0D points, 1D lines, or 2D grids—as distinct node layers. These layers are then fused through cross-support message passing into a point-support prediction layer, enabling precise field reconstruction. An inductive masked-node formulation allows the model to reconstruct rainfall at any user-defined target location or grid, independent of the sensing resolution. Tested on Singapore data, the proposed method achieved a 23.2% reduction in RMSE compared to classical interpolation baselines and consistently outperformed other neural architectures. A generalization study using Sydney data revealed that multi-support fusion is most beneficial when the rainfall field is undersampled relative to its spatial correlation length, providing significant gains where data sparsity is an issue.