Super-Resolution of Radar/Raingauge-Analyzed Precipitation Using Gaussian Process Regression with a Steering Kernel

Abstract

Super-resolution estimates a high-resolution image from a low-resolution image and has been used for downscaling and resolution enhancement of observations in meteorology. Super-resolution Gaussian process regression with a steering kernel (SRGP-SK) generates more accurate high-resolution images than super-resolution Gaussian process regression, but it has not yet been applied in meteorology. In this study, we applied SRGP-SK to radar/raingauge-analyzed precipitation for a convective case and a stratiform case, and evaluated the results using the structural similarity index (SSIM) and the radially averaged power spectral density (PSD). SRGP-SK achieved the SSIM comparable to that of bicubic interpolation while reconstructing finer precipitation structures: it reconstructed variations down to a wavelength of 6 km, whereas bicubic interpolation reconstructed variations only down to 8 km. We further compared several kernel functions and found that the kernel optimal for SSIM differed from that optimal for the geometric mean PSD ratio, reflecting the different properties that the two measures quantify. This is the first study to demonstrate the usefulness of SRGP-SK in meteorology, and it represents a step toward super-resolution with physical interpretability.