Statistical spatial downscaling of significant wave height in a regional sea from the global ERA5 dataset

Global wave datasets typically provide coarse spatial resolution, which is insufficient for detailed regional applications. Recently, neural networks have been used for fast downscaling of environmental fields, while their inherent randomness can induce prediction instability. This study aims to enhance the spatial resolution of significant wave height (SWH) in a regional sea by proposing an ensemble neural network and multivariate linear regression. Model performance is evaluated for both self-variable downscaling from low-resolution SWH and cross-variable downscaling from low-resolution wind fields. SWH and wind data from the global ERA5 dataset in the western Black Sea are used as the low-resolution input, and high-resolution SWH data from a regional model are used as the reference. The results indicate that both models perform well in self-variable SWH downscaling. In cross-variable downscaling, the ensemble model yields reasonable results, while linear regression gives significantly higher errors. The ensemble method effectively reduces the prediction instability of the base neural network, while linear regression can serve as a simple and efficient baseline for spatial downscaling in the self-variable approach. This work provides valuable insights into downscaling methodologies, which are crucial for improving spatial wave predictions and supporting diverse coastal applications. • Ensemble SRResNet gives much stable results in spatial SWH downscaling than SRResNet. • SWH can be downscaled from both low-resolution SWH and wind using ensemble SRResNet. • MLR performs similarly to ensemble SRResNet with SWH input but fails with wind input.