Multi-Scale Downscaling with Bayesian Convolution Network for ENSO SST Pattern

The downscaling of climate data has always been a hotspot in meteorological research. The traditional method is to adjust numerical climate models to simulate high-resolution situation, which is quite time and resource consuming. Recently, deep learning methods have provided an appreciable new insight and successfully downscale in multiple climate phenomena. However, existing models simply treat downscaling problem as image super-resolution problem and extract the spatial features from high resolution (HR) climate data directly, which ignores the detailed physical varieties and produce low credibility. In this paper, we take El Niño-Southern Oscillation (ENSO) events as example and formulate SST pattern downscaling as a multi-scale pattern fusion problem. Meanwhile, in order to cover that all conditions describing the internal physical structure, we find the optimal probability distributions of the output HR SST pattern. Based on this formalization, we further implement a deep learning model named Multi-scale Bayesian convolution network (MSBCNN). We evaluate our model on the monthly air-sea data from 1870 to 2014 and carry out real-world 2014/2015 ENSO SST pattern downscaling. The results demonstrate the effectiveness and superiority of our formalization, which outperforms than the other traditional approaches and state-of-the-art models.