Observation-guided Interpolation Using Graph Neural Networks for High-Resolution Nowcasting in Switzerland

Recent advances in neural weather forecasting have shown significant potential for accurate short-term forecasts. However, adapting such gridded approaches to smaller, topographically complex regions like Switzerland introduces computational challenges, especially when aiming for high spatial (1km) and temporal (10 min) resolution. This paper presents a Graph Neural Network (GNN)-based approach for high-resolution nowcasting in Switzerland using the Anemoi framework and observational inputs. The proposed architecture combines surface observations with selected past and future numerical weather prediction (NWP) states, enabling an observation-guided interpolation strategy that enhances short-term accuracy while preserving physical consistency. We evaluate two models, one trained using local nowcasting analyses and one trained without, on multiple surface variables and compare it against operational high-resolution NWP (ICON-CH1) and nowcasting (INCA) baselines. Results over the test period show that both GNNs consistently outperform ICON-CH1 when verified against INCA analyses across most variables and lead times. Relative to the INCA forecast system, scores against INCA analyses show AI gains beyond 2h (with early-lead disadvantages attributable to INCA’s warm start from the analysis), while verification against held-out stations shows no systematic degradation at short lead-times for AI models and frequent outperformance across surface variables. A comprehensive verification procedure, including spatial skill scores for precipitation, pairwise significance testing and event-based evaluation, demonstrates the operational relevance of the approach for mountainous domains. These results indicate that high-resolution, observation-guided GNNs can match or exceed the skill of established forecasting systems for short lead times, including when they are trained without nowcasting analyses.

💡 Research Summary

This paper introduces a graph neural network (GNN)–based framework for high‑resolution (1 km) and high‑frequency (10 min) nowcasting over the mountainous terrain of Switzerland. The authors build on the open‑source Anemoi platform and combine a rich set of observational inputs—151 SwissMetNet surface stations, partner stations for out‑of‑sample verification, radar‑derived precipitation composites, selected MSG satellite infrared and visible channels, and a 90 m digital elevation model—with past and future states from the ICON‑CH1 numerical weather prediction (NWP) model.

Two GNN variants are trained: (i) a “analysis‑driven” model that uses the operational INCA nowcasting analyses as ground truth, and (ii) a “forecast‑only” model that never sees INCA analyses, relying solely on NWP, observations, and topography. In both cases the graph’s nodes correspond to observation sites, radar grid cells, satellite pixels, and NWP grid points; edges encode geographic proximity, azimuth, and elevation differences. Message passing is governed by a data‑dependent attention weight αij, allowing the network to re‑weight connections dynamically according to the prevailing flow regime (e.g., north‑west to south‑east wind shifts). This design enables the model to capture non‑linear, spatially heterogeneous interactions that are especially pronounced in complex terrain.

The training data span August 2023 onward, covering two years of 10‑minute cycles. Evaluation is extensive: (a) point‑wise error metrics (RMSE, MAE, CRPS) against ICON‑CH1 forecasts, (b) the same metrics against INCA analyses, (c) spatial skill scores for precipitation (ETS, FSS), (d) pairwise significance testing (p < 0.05), and (e) event‑based verification for heavy precipitation (>10 mm). Results show that both GNNs consistently outperform ICON‑CH1 across all surface variables and lead times. Compared with INCA, the AI models lag slightly at the very first 0–30 min, reflecting INCA’s warm‑start advantage, but they surpass INCA from 30 min onward, with clear gains beyond the 2‑hour horizon. Importantly, the forecast‑only GNN achieves skill comparable to the analysis‑driven version, demonstrating that high‑quality nowcasts can be produced without dedicated nowcasting analyses—a crucial finding for meteorological services that lack such products.

Verification on held‑out partner stations confirms that the GNNs do not degrade performance at short leads; instead they frequently exceed the operational baseline for temperature, dew point, wind, and precipitation. Spatial analyses reveal that the observation‑guided interpolation strategy yields sharper precipitation fields and better captures localized extremes than the traditional interpolation‑plus‑blending approach used in INCA. Visualizations of the attention weights illustrate physically plausible behavior: edges pointing up‑slope or down‑wind receive higher weights when the flow aligns with those directions.

The paper emphasizes reproducibility: all data preprocessing, model configuration, training, and inference steps are scripted within Anemoi, and the trained models are released publicly. The authors discuss operational relevance, noting that the GNN inference latency fits within the 10‑minute update cycle required for nowcasting, and that the graph structure scales efficiently to the irregular observation network of Switzerland.

In conclusion, the study demonstrates that (1) graph‑based, observation‑guided deep learning can deliver sub‑hourly, 1‑km nowcasts that match or exceed the skill of established NWP (ICON‑CH1) and operational nowcasting (INCA) systems, (2) such performance is attainable even without access to high‑quality nowcasting analyses, and (3) the flexible graph architecture is well suited to complex topographies where traditional convolutional networks struggle. Future work is suggested on extreme‑event learning, multi‑scale graph designs, and real‑time deployment across other mountainous regions.