level: research
researchers trained a multimodal graph neural network to predict rain rate every five minutes up to two hours ahead over the nordic radar domain. they tested different input combinations: radar history, meps numerical weather prediction, netatmo surface observations, msg satellite channels, stochastic noise, and crps-based ensemble losses. the study was an ablation to see which operationally relevant sources improve dense radar-field forecasts.
results show that adding sparse point observations from stations gives only small improvements over radar-only models. numerical weather prediction and satellite data provided larger benefits. noise augmentation and crps-based losses also helped, but the main finding is that pointwise data is not very useful for this task. diagnostics included scores on the radar grid, at station locations, for rain onset, and using oracle, displacement, and amplitude metrics.
the work suggests that for precipitation nowcasting, dense spatial data like radar and satellite are more valuable than scattered ground sensors. this has implications for designing operational forecasting systems, where collecting and integrating point observations may not be worth the effort if radar and nwp are already available. the study is limited to the nordic region and the specific model architecture, but it provides evidence that multimodal fusion should prioritize certain data types.
why it matters: it helps ai practitioners focus on high-impact data sources for weather prediction, avoiding wasted effort on sparse observations.