Feature - Compensation - Based Generative Adversarial Network Approach for Meteorological Downscaling
Qinrui Fan, Hui Deng, Xiaochuan Hu, Wenyi Ge, Xiaojie Li, Jing Hu · 2025
With the in - depth development of global climate change research and the increasing demand for refined meteorological services in society, meteorological downscaling technology has become a research hotspot in multiple disciplines. Existing meteorological downscaling methods, such as optimal interpolation, multi - grid variational method, deep learning methods, dynamic downscaling, and statistical downscaling, have their own advantages and disadvantages. Meanwhile, when image super - resolution algorithms are applied to meteorological downscaling, they often fail to fully consider the characteristics of meteorological data. Therefore, this paper proposes a feature - compensation downscaling method based on generative adversarial networks. This method designs an information injection module to periodically extract the latent features of the original meteorological data and inject them into the main network, suppressing the generation of artifacts. It also uses a compensation feature fusion mechanism to optimize the latent features from both spatial and channel dimensions, improving the accuracy and reliability of the downscaling results.