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

Abstract Recent advances in neural weather forecasting have shown significant potential for accurate short-term predictions. However, adapting such gridded approaches to smaller, topographically complex regions like Switzerland introduces computational challenges, especially at high spatial (1 km) and temporal (10 min) resolution. This study investigates whether graph neural network (GNN) models can provide both skill and computational advantages for short-term weather prediction in complex terrain by training and comparing them against operational numerical weather prediction (NWP) systems. Two GNN formulations are explored: one trained using analysis data from an operational nowcasting model as its target dataset [artificial intelligence (AI)-based version of Integrated Nowcasting through Comprehensive Analysis (INCA) forecasting system (AINCA)], and another trained using a weighted combination of radar, surface station, and NWP forecast data to assess performance in the absence of analysis targets ( -AINCA). Both models are implemented within the Anemoi framework and leverage surface observations together with past and future NWP states, enabling an observation-guided interpolation strategy that enhances short-term accuracy while preserving physical consistency. Results over the test period show that both GNNs consistently outperform operational NWP forecasts when verified against nowcasting analyses across most variables and lead times. Relative to the local nowcasting system, scores against analyses show AI gains beyond 2-h lead time (with early lead disadvantages attributable to warm-start initialization), while verification against held-out stations reveals no systematic degradation at short lead times and frequent improvements 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 findings indicate that high-resolution, observation-guided GNNs can match or exceed the skill of established forecasting systems for short lead times, even when trained without nowcasting analyses. Significance Statement The World Meteorological Organization (WMO) defines nowcasting as forecasting with local detail, by any method, over a period from the present to 6 h ahead, including a detailed description of the present weather. Accurate short-term weather forecasts are critical for protecting lives, infrastructure, and ecosystems, especially in regions with complex terrain where local conditions can change rapidly. This study demonstrates that graph-based deep learning models can deliver high-resolution nowcasts comparable to those produced by an operational numerical weather prediction (NWP) system, but at a fraction of the computational cost. Importantly, we show that such models can achieve competitive performance even without access to specialized nowcasting analyses, relying instead on readily available data such as station observations, satellite, radar, and NWP forecasts. These findings highlight the potential of deep learning to extend the benefits of real-time, high-precision weather forecasting to regions and services where traditional nowcasting infrastructure is unavailable.