Sparse Weather Data Boosts Rain Nowcasting in Graph Neural Network Study

A new preprint from arXiv titled 'Pointwise is Pointless? A Multimodal Ablation Study for Precipitation Nowcasting with Graph Neural Networks' investigates how sparse point observations and other data sources affect dense radar-field forecasts. The team (Miralles et al.) built a graph neural network nowcasting system over the Nordic radar domain, predicting rain rate every five minutes up to two hours ahead. They trained the model with various combinations: radar history alone, MEPS numerical weather prediction, Netatmo surface observations, MSG satellite channels, stochastic noise, and CRPS-based ensemble losses. Each configuration was evaluated using complementary diagnostics on the radar grid, at station locations, for rain onset, and through oracle/displacement/amplitude scores.

The ablation reveals that each data modality improves a different aspect of the forecast. MEPS stabilizes radar-only extrapolation, Netatmo observations enhance local station and onset metrics, and satellite predictors reduce station-level biases but can trigger rain too early when used deterministically. CRPS-based training yields the most consistent gains on the radar grid, while combining satellite and CRPS achieves the best overall oracle/DAS score. The authors conclude that sparse point observations are not uninformative—they provide valuable local constraints—but their benefit for radar-like fields depends critically on the training loss, uncertainty representation, and how observation support is encoded. This work highlights the nuanced trade-offs in multimodal weather forecasting.