Pointwise is Pointless? A Multimodal Ablation Study for Precipitation Nowcasting with Graph Neural Networks

A new study examines whether sparse point observations from ground stations improve precipitation nowcasting when combined with graph neural networks, finding that local gains do not automatically translate into better radar-field forecasts [1]. The work, posted to the arXiv preprint repository on June 16, 2026, was authored by Ophélia Miralles and colleagues [1]. It deploys a multimodal graph neural network system over the Nordic radar domain that predicts rain rate every five minutes up to two hours ahead [1]. The researchers trained the model with varying combinations of radar history, MEPS numerical weather prediction, Netatmo surface observations, MSG satellite channels, stochastic noise, and CRPS-based ensemble losses [1]. The study is structured as an ablation of operationally relevant information sources and training objectives [1]. Configurations were compared using diagnostics on the radar grid, at station locations, for rain onset, and through oracle, displacement, and amplitude scores [1]. MEPS numerical weather prediction stabilised radar-only extrapolation, while Netatmo observations improved local station and onset diagnostics [1]. Satellite predictors reduced some station-level biases but could activate rain too early when used deterministically [1]. Configurations trained with CRPS-based losses delivered the most consistent radar-grid gains, and the combined satellite and CRPS setup achieved the best overall oracle and displacement-amplitude score [1]. The findings do not support the conclusion that point observations are uninformative for nowcasting, but they show that local observational skill and spatially coherent radar-field skill are distinct targets [1]. The practical implication is that sparse observations can provide useful local constraints, though their benefit for radar-like fields depends on the training loss, uncertainty representation, and how observation support is encoded in the model [1]. The paper appeared on arXiv, an open-access repository of electronic preprints that, as of November 2024, receives about 24,000 submissions per month and has surpassed two million articles [6]. The repository is not peer-reviewed; papers are approved for posting after moderation [6]. The study’s abstract page also features arXivLabs integrations, a framework launched in 2020 that allows community collaborators to develop tools such as citation explorers and code finders directly on the site [4][5].