This title appears in the Scientific Report :
2022
Please use the identifier:
http://hdl.handle.net/2128/31419 in citations.
Statistical Downscaling of Surface Temperature and Precipitation with Deep Neural Networks
Statistical Downscaling of Surface Temperature and Precipitation with Deep Neural Networks
In light of the success of superresolution (SR) applications in computer vision, recent studies have started to develop statistical downscaling methods for meteorological data based on deep neural networks (DNNs). DNNs are attractive, because they are computationally cheap, once they are trained.In...
Saved in:
Personal Name(s): | Gong, Bing (Corresponding author) |
---|---|
Langguth, Michael (Corresponding author) / Ji, Yan / Mozaffari, Amirpasha / Mache, Karim / Schultz, Martin | |
Contributing Institute: |
Jülich Supercomputing Center; JSC |
Imprint: |
2022
|
Conference: | Platform for Advanced Scientific Computing Conference 2022, Basel (Switzerland), 2022-06-27 - 2022-06-30 |
Document Type: |
Conference Presentation |
Research Program: |
Earth System Data Exploration Earth System Data Exploration MAchinE Learning for Scalable meTeoROlogy and cliMate Domain-Specific Simulation & Data Life Cycle Labs (SDLs) and Research Groups |
Link: |
Get full text OpenAccess |
Publikationsportal JuSER |
In light of the success of superresolution (SR) applications in computer vision, recent studies have started to develop statistical downscaling methods for meteorological data based on deep neural networks (DNNs). DNNs are attractive, because they are computationally cheap, once they are trained.In this study, deep neural networks are developed to downscale hourly 2 meter temperature and precipitation over the complex terrain of Central Europe. Our approach is based on advanced generative adversarial networks (GANs) and transformer networks. The merit of this choice is that GANs encourage the generator to preserve the strong spatial variability from the data, while the transformer can capture the temporal dependencies. The experiments are designed to generate high-resolution temperature (0.1°) from low resolution (0.8°), and time-evolving high-resolution precipitation (1 km) from low resolution (4 km/8 km). The DNNs are fed with several relevant static and dynamic predictors and comprehensively evaluated by grid point-level errors, and error metrics for spatial variability and the generated probability distribution. Our results motivate the further development of DNNs that can be potentially leveraged to downscale other challenging Earth system data such as cloud cover or wind in operational workflows. |