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

TL;DR

This work tackles high-resolution nowcasting in Switzerland, a region with complex terrain that challenges traditional methods. It introduces an observation-guided Graph Neural Network within the Anemoi framework to fuse surface observations, radar, satellite data, and NWP states, producing 1 km, 10-minute nowcasts. The study shows that the GNN models outperform the ICON-CH1 baseline across most variables and lead times when verified against INCA analyses and SwissMetNet stations, and can match or exceed INCA performance for lead times beyond about two hours, while offering significant inference speed advantages. The approach demonstrates operational viability in mountainous domains, highlights reproducibility via Anemoi, and points to future work in uncertainty quantification and cross-region transfer.

Abstract

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.

Figures (14)

• Figure 1: Left: all available weather stations from SwissMetNet (red bullets) and partner (black triangles) stations. Number of MeteoSwiss/partner stations available per period of 10 minutes on Aug 1 2023 (right).
• Figure 2: Composite and raster data: rain rate from radar (top left), infrared channel from MSG satellite (top right), topography (bottom).
• Figure 3: Schematic representation of the observation-guided interpolation model used for nowcasting. This is the model used for training; at inference time, we use the last hour of observations and [$t_0, t_0+12$h] from ICON--CH1 to forecast every 10 minutes over the next 12-hours.
• Figure 4: 2-meter temperature in $^{\circ}$C on the 27 Sept. 2023 at midnight for ICON--CH1 on its original triangular mesh (a) next to INCA nowcasting data for the same datetime (b). A triangular mesh is indexed with 1D edge coordinates, whereas a regular grid represents a set of areas of equal size formed by 2D coordinates.
• Figure 5: Average root mean squared error, Pearson correlation, Fraction skill score for rain rate for thresholds 0.1, 0.5, 1 and 3 mm.h$^{-1}$ over 10km spatial windows. Scores are computed versus INCA analysis; ICON--CH1 and INCA forecasts are used as baselines for comparison. Test set aggregate.