Geometry-Aware Graph Fusion: Rainfall Field Reconstruction

Geometry-aware multi-support graph fusion reconstructs fine-scale rainfall fields, and the authors submitted the paper on 2 Jul 2026 to WACV 2027 (applications track). The model represents observations by their spatial support type and fuses them through cross-support message passing into a point-support prediction layer used to reconstruct the field.

What did the authors build?

They built a geometry-aware multi-support heterogeneous graph neural network that treats each observation type as a distinct node layer: 0D points for gauges, 1D lines for microwave links, and 2D grids for radar and satellite products. The network fuses these layers via cross-support message passing into a point-support prediction layer, then reconstructs the rainfall field from that layer. The authors also introduce an inductive masked-node formulation, which decouples prediction resolution from sensing resolution and allows the trained model to reconstruct the field at user-defined target locations or display grids.

How does it perform on real data?

On Singapore data the proposed method reduces root-mean-square error by 23.2% compared with the classical inverse-distance weighting interpolation baseline. The paper states the model also consistently outperforms other neural architectures, including convolutional fusion and support-agnostic heterogeneous graph baselines. A generalization study using data from Sydney, Australia examines when multi-support fusion delivers gains: the authors find fusion helps most when gauge spacing is large relative to the spatial correlation length of the rainfall field, and delivers little additional skill when the field is already resolved by available sensors.

How does the geometry-aware approach differ from prior methods?

Prior heterogeneous graph approaches reconciled multiple sensor sources in feature space, giving each its own embedding while discarding the geometry of its support. The authors argue that treating measurement support geometry explicitly — 0D, 1D, 2D — keeps the geometric constraints that different sensors impose on the rainfall field. Their cross-support message passing explicitly routes information between support-specific layers into a unified point-support prediction, rather than merging heterogeneous data into a single, support-agnostic embedding.

Why it matters

Fine-scale rainfall reconstruction is a core input for urban flood modeling, and sensor networks commonly combine point gauges, path-integral microwave links, and area-aggregating radar or satellite products. Representing the geometry of those supports directly responds to a structural mismatch between sensors and the field they observe. The 23.2% RMSE reduction on Singapore data demonstrates measurable improvement over classical interpolation, while the Sydney study clarifies where fusion yields value: namely, under-sampled fields relative to their correlation length. For practitioners, that suggests investment in multi-support fusion will pay off most in areas with sparse gauges or rapidly varying rainfall patterns.

What to watch

The authors say code and models will be open-sourced upon paper acceptance; watch for that release and for the paper’s review outcome at WACV 2027. Also check for follow-up evaluations that report absolute RMSE numbers, runtime and compute costs, or experiments across additional climates — those will determine how widely applicable the approach is beyond the Singapore and Sydney datasets.