S2G Deep Interpolation: A Deep Learning Framework for Generating High-Resolution Wind Gust Fields from Sparse Mesonet Observations

Abstract

The growing deployment of surface observational networks (so-called Mesonets) has the potential to transform many sectors, such as agriculture, water resource management, renewable energy assessment, disaster risk reduction, power outage prediction, and high-impact weather monitoring. However, network observations are sparse, whereas these downstream applications typically require gridded data. Traditional interpolation methods, such as Barnes interpolation, do not perform well in complex terrain. Therefore, in this study, we propose a deep learning–based framework, called Sparse-to-Gridded Deep Interpolation (S2G-DI), to generate high-resolution spatially continuous fields from sparse surface observations. We focus on wind gusts, which are challenging to interpolate due to their highly localized and transient nature, but are critical for assessing weather-related hazards. We experimented with three deep learning architectures, employing convolutional neural networks and transformers with increasing complexity, and our sensitivity analysis shows that a UNet architecture with meteorological and topographic inputs significantly outperforms Barnes interpolation, reducing RMSE by over 30% and better preserving fine-scale features. The model remains robust to missing data and generalizes well even with limited training data. Evaluations during extreme wind gust events confirm that the framework captures both spatial structure and intensity more reliably than traditional methods. However, for some of these cases, there is still room for improvement for the S2G-DI framework.