Document Type
Paper
Start Date
16-10-2024 9:50 AM
End Date
16-10-2024 11:10 AM
Abstract
Facing the challenges of global climate change, precise and high spatial resolution climate data are crucial and in pressing need for scientific research and analysis. However, most existing datasets are only available with very coarse spatial resolution and demand large-scale resolution enhancement. Meanwhile, climate datasets own much more intricate textures than natural images. Statistical downscaling or super-resolution (SR) with the deep-learning-based generative model might be a promising approach to address these challenges. It is worth noting that a learned Bayesian reconstruction with generative models (L-BRGM) method was proposed recently. The proposed Bayesian deep learning framework employs a single pre-trained generative model to solve different image restoration tasks and is feasible for large-scale image downscaling. In this work, we applied it to the statistical downscaling of geospatial datasets. With a case study of wind velocity and solar irradiance datasets, we exhibit that the L-BRGM method can reconstruct high-resolution images with large scaling factors (64×) on climate datasets through Maximum a-posteriori (MAP) estimation.
DOI
10.5703/1288284317800
Included in
Climate Commons, Geographic Information Sciences Commons, Spatial Science Commons, Statistical Methodology Commons
Statistical Downscaling of Climate Datasets with Deep Generative Model and Bayesian inference
Facing the challenges of global climate change, precise and high spatial resolution climate data are crucial and in pressing need for scientific research and analysis. However, most existing datasets are only available with very coarse spatial resolution and demand large-scale resolution enhancement. Meanwhile, climate datasets own much more intricate textures than natural images. Statistical downscaling or super-resolution (SR) with the deep-learning-based generative model might be a promising approach to address these challenges. It is worth noting that a learned Bayesian reconstruction with generative models (L-BRGM) method was proposed recently. The proposed Bayesian deep learning framework employs a single pre-trained generative model to solve different image restoration tasks and is feasible for large-scale image downscaling. In this work, we applied it to the statistical downscaling of geospatial datasets. With a case study of wind velocity and solar irradiance datasets, we exhibit that the L-BRGM method can reconstruct high-resolution images with large scaling factors (64×) on climate datasets through Maximum a-posteriori (MAP) estimation.