Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Generative Adversarial Models for Extreme Geospatial Downscaling

Guiye Li, Guofeng Cao

TL;DR

This work tackles the challenge of generating high-resolution geospatial climate data from coarse inputs by introducing a latent adversarial generator (LAG) within a conditional GAN framework. By combining progressive GAN training (ProGAN) with a latent Gaussian-like variable, the method can produce multiple plausible high-resolution realizations for extreme downscaling factors up to $64\times$, while explicitly modeling uncertainty and preserving mass. Compared against geostatistical and deep-learning baselines (ATP K, DIP, ESRGAN, PhIRE GAN, EDiffSR), LAG achieves superior pixel- and perceptual-quality metrics (relative MSE, SWD) and provides a robust uncertainty-space via Monte Carlo realizations, demonstrated on wind velocity and solar irradiance datasets. The approach offers practical impact for climate data downscaling and suggests future extensions to spatiotemporal data and Bayesian uncertainty integration.

Abstract

Addressing the challenges of climate change requires accurate and high-resolution mapping of geospatial data, especially climate and weather variables. However, many existing geospatial datasets, such as the gridded outputs of the state-of-the-art numerical climate models (e.g., general circulation models), are only available at very coarse spatial resolutions due to the model complexity and extremely high computational demand. Deep-learning-based methods, particularly generative adversarial networks (GANs) and their variants, have proved effective for refining natural images and have shown great promise in improving geospatial datasets. This paper describes a conditional GAN-based stochastic geospatial downscaling method that can accommodates very high scaling factors. Compared to most existing methods, the method can generate high-resolution accurate climate datasets from very low-resolution inputs. More importantly, the method explicitly considers the uncertainty inherent to the downscaling process that tends to be ignored in existing methods. Given an input, the method can produce a multitude of plausible high-resolution samples instead of one single deterministic result. These samples allow for an empirical exploration and inferences of model uncertainty and robustness. With a case study of gridded climate datasets (wind velocity and solar irradiance), we demonstrate the performances of the framework in downscaling tasks with large scaling factors (up to $64\times$) and highlight the advantages of the framework with a comprehensive comparison with commonly used and most recent downscaling methods, including area-to-point (ATP) kriging, deep image prior (DIP), enhanced super-resolution generative adversarial networks (ESRGAN), physics-informed resolution-enhancing GAN (PhIRE GAN), and an efficient diffusion model for remote sensing image super-resolution (EDiffSR).

Generative Adversarial Models for Extreme Geospatial Downscaling

TL;DR

This work tackles the challenge of generating high-resolution geospatial climate data from coarse inputs by introducing a latent adversarial generator (LAG) within a conditional GAN framework. By combining progressive GAN training (ProGAN) with a latent Gaussian-like variable, the method can produce multiple plausible high-resolution realizations for extreme downscaling factors up to , while explicitly modeling uncertainty and preserving mass. Compared against geostatistical and deep-learning baselines (ATP K, DIP, ESRGAN, PhIRE GAN, EDiffSR), LAG achieves superior pixel- and perceptual-quality metrics (relative MSE, SWD) and provides a robust uncertainty-space via Monte Carlo realizations, demonstrated on wind velocity and solar irradiance datasets. The approach offers practical impact for climate data downscaling and suggests future extensions to spatiotemporal data and Bayesian uncertainty integration.

Abstract

Addressing the challenges of climate change requires accurate and high-resolution mapping of geospatial data, especially climate and weather variables. However, many existing geospatial datasets, such as the gridded outputs of the state-of-the-art numerical climate models (e.g., general circulation models), are only available at very coarse spatial resolutions due to the model complexity and extremely high computational demand. Deep-learning-based methods, particularly generative adversarial networks (GANs) and their variants, have proved effective for refining natural images and have shown great promise in improving geospatial datasets. This paper describes a conditional GAN-based stochastic geospatial downscaling method that can accommodates very high scaling factors. Compared to most existing methods, the method can generate high-resolution accurate climate datasets from very low-resolution inputs. More importantly, the method explicitly considers the uncertainty inherent to the downscaling process that tends to be ignored in existing methods. Given an input, the method can produce a multitude of plausible high-resolution samples instead of one single deterministic result. These samples allow for an empirical exploration and inferences of model uncertainty and robustness. With a case study of gridded climate datasets (wind velocity and solar irradiance), we demonstrate the performances of the framework in downscaling tasks with large scaling factors (up to ) and highlight the advantages of the framework with a comprehensive comparison with commonly used and most recent downscaling methods, including area-to-point (ATP) kriging, deep image prior (DIP), enhanced super-resolution generative adversarial networks (ESRGAN), physics-informed resolution-enhancing GAN (PhIRE GAN), and an efficient diffusion model for remote sensing image super-resolution (EDiffSR).
Paper Structure (24 sections, 3 equations, 16 figures, 3 tables)

This paper contains 24 sections, 3 equations, 16 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Related work
  3. Geostatistical methods
  4. Deep learning methods
  5. GAN-based methods and others
  6. Stochastic downscaling
  7. Extreme downscaling
  8. Methodology
  9. Loss functions
  10. Neural networks structures
  11. Experimental design
  12. Datasets
  13. Training details
  14. Accuracy metrics
  15. Model comparisons
  16. ...and 9 more sections

Figures (16)

  • Figure 1: Structures of the generator and critic in the described LAG-based downscaling framework.
  • Figure 2: $4\times$ SR results generated by different methods for two channels (U and V) of a randomly sampled LR image from the wind velocity test dataset. Zoom in for better observation.
  • Figure 3: $4\times$ SR results generated by different methods for two channels (DNI and DHI) of a randomly sampled LR image from the solar radiance test dataset. Zoom in for better observation.
  • Figure 4: $64\times$ SR results generated by the LAG-based downscaling framework for a randomly sampled LR image from (a) wind velocity and (b) solar irradiance test datasets. For each panel, the leftmost column shows the input LR images at two channels (with image size $8\times8$), the rightmost column shows the corresponding HR ground truth images (with image size $512\times512$), and the columns in between shows results of LAG. Zoom in for better observation.
  • Figure 5: Semivariograms of the $4\times$ SR results generated by different models for a randomly sampled LR image from (a) wind velocity and (b) solar irradiance test datasets. Images with similar spatial patterns should have close lines. Zoom in for better observation.
  • ...and 11 more figures