Encoding Gridded Atmospheric Data with Classical and Quantum Methods

Abstract

Accurate encoding of spatial meteorological data is critical for applying quantum machine learning (QML) in climatology and atmospheric sciences—important domains for geospatial methods. This study explores quantum encoding techniques adapted to the constraints of limited qubit space, focusing on gridded numerical weather prediction (NWP) outputs related to low-visibility events at multiple Czech airports. Beyond quantum encoding, we investigate the role of dimensionality reduction (PCA, t-SNE, UMAP, Isomap) and its integration with quantum amplitude and angle encoding schemes. We assess their capacity to represent visibility transitions using fidelity-based measures. Results indicate spatial heterogeneity in encoding effectiveness, with no single method dominating across all locations. To better preserve spatial and physical structure, we introduce expert-informed groupwise embeddings applied separately on meteorological clusters, rather than across the entire dataset. This approach improves the physical relevance and continuity of spatial patterns. Results suggest that temporal features (diurnal cycles), complicate fidelity assessment and emphasize the importance of temporal segmentation and data curation. Our findings demonstrate that combining classical dimensionality reduction, quantum encoding, and domain expertise offers a promising path toward effectively representing complex spatial-temporal atmospheric patterns. This work supports future development of quantum-assisted weather forecasting systems.