Meteorological Applications Науки об атмосфере
<scp>LoomingWinds</scp> : A Terrain‐Guided and Spectral Feature‐Enhanced Network for Near‐Surface Wind Field Downscaling
Аннотация
ABSTRACT Numerical weather prediction (NWP) systems encounter computational constraints in efficiently generating high resolution (HR) wind field distributions, particularly for precision‐demanding applications such as environmental protection and renewable energy planning. Deep learning wind field downscaling provides an efficient alternative to generate refined wind fields at targeted spatial resolution. However, two critical challenges remain unaddressed: (1) how to develop trainable modules that systematically integrate HR terrain data into guidance mechanisms for wind fields downscaling processes, and (2) how to effectively extract spectral features to comprehensively characterize the variability characteristics of wind fields. In this study, we propose LoomingWinds, a terrain‐guided framework leveraging the Global 30 Arc‐Second Elevation (GTOPO30) digital elevation model (DEM) product to achieve HR reconstruction of 10 m wind fields (including both 10 m zonal wind speed and 10 m meridional wind speed) from 5 to 1 km resolution. Specifically, we design a Terrain‐Fusion Attention Module (TFAM) that establishes effective exploration of cross‐scale dependencies between topographic characteristics and wind dynamics, refining near‐surface wind fields characteristics through terrain‐informed modulation. Subsequently, a Fourier feature enhancement module (FFEM) is introduced to recover wind dynamics variation and disentangle spectral‐spatial features by combining global frequency representations with local spatial convolutions. In addition, an enhanced loss function with frequency domain constraints is developed to strengthen the reconstruction of high frequency meteorological signatures in wind field patterns. Comprehensive experimentation substantiates that our proposed LoomingWinds achieves a mean squared error of 0.347 m/s (averaged over the U and V components, i.e., 0.345 m/s for U and 0.348 m/s for V) in the JMA nonhydrostatic model (JMA‐NHM) dataset, corresponding to 3.9% and 3.6% reductions, respectively, compared to the second‐best performing model.