Mesh Interpolation Graph Network for Dynamic and Spatially Irregular Global Weather Forecasting

TL;DR

The paper tackles global weather forecasting under spatially irregular and dynamically distributed station data. It introduces Mesh Interpolation Graph Network (MIGN), which combines a regular HEALPix mesh with a parametric spherical-harmonics location embedding to improve spatial generalization and enable predictions at unobserved locations. Through extensive experiments on NOAA GSOD data, MIGN outperforms 13 baselines across six meteorological variables, showing strong global generalization to unseen stations and robustness in data-sparse regions. The work highlights the importance of addressing irregular data distributions on a spherical Earth and demonstrates a practical, scalable approach for global, station-based weather forecasting with potential broad impacts on agriculture, disaster planning, and climate research.

Abstract

Graph neural networks have shown promising results in weather forecasting, which is critical for human activity such as agriculture planning and extreme weather preparation. However, most studies focus on finite and local areas for training, overlooking the influence of broader areas and limiting their ability to generalize effectively. Thus, in this work, we study global weather forecasting that is irregularly distributed and dynamically varying in practice, requiring the model to generalize to unobserved locations. To address such challenges, we propose a general Mesh Interpolation Graph Network (MIGN) that models the irregular weather station forecasting, consisting of two key designs: (1) learning spatially irregular data with regular mesh interpolation network to align the data; (2) leveraging parametric spherical harmonics location embedding to further enhance spatial generalization ability. Extensive experiments on an up-to-date observation dataset show that MIGN significantly outperforms existing data-driven models. Besides, we show that MIGN has spatial generalization ability, and is capable of generalizing to previous unseen stations.

Figures (16)

• Figure 1: (A). Illustrations of spatially irregular station distribution. (B). The probability density of the station in terms of longitude and latitude. (C). The recorded number of stations in the up-to-date NOAA Global Surface Summary of the Day (GSOD) dataset for each year.
• Figure 2: Framework of the model. MIGN architecture follows an encoder-processor-decoder framework.
• Figure 3: The global MAE distribution of SLP in the generalization experiment testing set
• Figure 4: Comparison of different models in data-scarce regions.
• Figure 5: Comparison of model performance with different mesh hyperparameter settings