11:30 〜 11:45
[MGI27-10] Precipitation super-resolution using diffusion model with d4PDF
キーワード：Downscaling, Diffusion model, AI, dfPDF
Accurate high-resolution weather prediction is important for predicting disastrous extreme events, planning infrastructure, and guiding disaster mitigation efforts. For that purpose, regional weather prediction models have been used to employ dynamical downscaling from coarse-resolution global models, providing detailed distributions of meteorological variables such as precipitation. However, these methods often impose significant computational resources, which is necessary for solving complex physical equations with spatially fine model grids.
In this study, we explore an alternative downscaling approach through a diffusion model (DDPM), a deep-learning technique recently gaining attention for generating high-resolution data from coarse input data. We aim to develop a downscaling method that significantly reduces computational load while keeping adequate accuracy in reproducing local heavy precipitation events. This study focuses on precipitation forecasts over Aomori Prefecture in the northern part of Japan, a region characterized by varied topography and complex weather patterns.
We selected the ensemble climate forecast database known as d4PDF, which provides predictions at a 20 km horizontal resolution over Japan, as the input data. These coarse-resolution data include precipitation and essential meteorological variables such as temperature, pressure, dew point depression, precipitable water, and wind. We hypothesized that such variables are informative for capturing precipitation events even when the coarser-resolution dataset indicates weak or no rainfall. As reference data, we used dynamically downscaled 5.0 km resolution d4PDF data, thereby setting a robust target for supervised training.
A series of experiments revealed that a diffusion model, combined with relevant atmospheric variables, can be an effective and computationally efficient alternative for downscaling to infer high-resolution precipitation. By capturing the underlying initial and boundary conditions that possibly produce intense precipitation, our method substantially improves the representation of heavy rainfall events while mitigating biases. These findings have a great potential for future research to refine machine learning–based downscaling techniques for diverse climate applications.
In this study, we explore an alternative downscaling approach through a diffusion model (DDPM), a deep-learning technique recently gaining attention for generating high-resolution data from coarse input data. We aim to develop a downscaling method that significantly reduces computational load while keeping adequate accuracy in reproducing local heavy precipitation events. This study focuses on precipitation forecasts over Aomori Prefecture in the northern part of Japan, a region characterized by varied topography and complex weather patterns.
We selected the ensemble climate forecast database known as d4PDF, which provides predictions at a 20 km horizontal resolution over Japan, as the input data. These coarse-resolution data include precipitation and essential meteorological variables such as temperature, pressure, dew point depression, precipitable water, and wind. We hypothesized that such variables are informative for capturing precipitation events even when the coarser-resolution dataset indicates weak or no rainfall. As reference data, we used dynamically downscaled 5.0 km resolution d4PDF data, thereby setting a robust target for supervised training.
A series of experiments revealed that a diffusion model, combined with relevant atmospheric variables, can be an effective and computationally efficient alternative for downscaling to infer high-resolution precipitation. By capturing the underlying initial and boundary conditions that possibly produce intense precipitation, our method substantially improves the representation of heavy rainfall events while mitigating biases. These findings have a great potential for future research to refine machine learning–based downscaling techniques for diverse climate applications.