Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Climate Variable Downscaling with Conditional Normalizing Flows

Christina Winkler, Paula Harder, David Rolnick

TL;DR

The paper tackles the challenge of obtaining local-scale climate information from globally coarse simulations by using conditional normalizing flows to model the conditional distribution $p_{Y|X}$ of high-resolution variables given low-resolution inputs. This approach enables both density estimation and efficient sampling, allowing explicit uncertainty quantification through the conditional distribution. The authors demonstrate the first application of CNFs to climate downscaling on ERA5 Total Column Water data, show competitive performance against baselines, and provide uncertainty maps that highlight regions of higher predictive variability. The work advances probabilistic downscaling by delivering physically consistent, uncertainty-aware high-resolution fields suitable for local impact assessment and risk analysis.

Abstract

Predictions of global climate models typically operate on coarse spatial scales due to the large computational costs of climate simulations. This has led to a considerable interest in methods for statistical downscaling, a similar process to super-resolution in the computer vision context, to provide more local and regional climate information. In this work, we apply conditional normalizing flows to the task of climate variable downscaling. We showcase its successful performance on an ERA5 water content dataset for different upsampling factors. Additionally, we show that the method allows us to assess the predictive uncertainty in terms of standard deviation from the fitted conditional distribution mean.

Climate Variable Downscaling with Conditional Normalizing Flows

TL;DR

The paper tackles the challenge of obtaining local-scale climate information from globally coarse simulations by using conditional normalizing flows to model the conditional distribution of high-resolution variables given low-resolution inputs. This approach enables both density estimation and efficient sampling, allowing explicit uncertainty quantification through the conditional distribution. The authors demonstrate the first application of CNFs to climate downscaling on ERA5 Total Column Water data, show competitive performance against baselines, and provide uncertainty maps that highlight regions of higher predictive variability. The work advances probabilistic downscaling by delivering physically consistent, uncertainty-aware high-resolution fields suitable for local impact assessment and risk analysis.

Abstract

Predictions of global climate models typically operate on coarse spatial scales due to the large computational costs of climate simulations. This has led to a considerable interest in methods for statistical downscaling, a similar process to super-resolution in the computer vision context, to provide more local and regional climate information. In this work, we apply conditional normalizing flows to the task of climate variable downscaling. We showcase its successful performance on an ERA5 water content dataset for different upsampling factors. Additionally, we show that the method allows us to assess the predictive uncertainty in terms of standard deviation from the fitted conditional distribution mean.
Paper Structure (11 sections, 2 equations, 3 figures, 1 table)

This paper contains 11 sections, 2 equations, 3 figures, 1 table.

Table of Contents

  1. Introduction
  2. Contributions
  3. Background
  4. Experiments
  5. ERA5 Hourly Water Content Dataset:
  6. Experimental setup:
  7. Qualitative Evaluation
  8. Quantitative Evaluation
  9. Sample Uncertainty
  10. Conclusion
  11. Acknowledgments

Figures (3)

  • Figure 1: Super resolution results on the ERA5 water content TCW test data for 2 $\times$ upsampling. Samples are taken from the CNF $x_{hr} \sim p(x_{hr} | x_{lr})$ with $\tau=0.8$. Best viewed electronically.
  • Figure 2: Super resolution results on the ERA5 TCW water content test data for 4 $\times$ upsampling. Samples are taken from the CNF $x_{hr} \sim p(x_{hr} | x_{lr})$ with $\tau=0.8$. Best viewed electronically.
  • Figure 3: The top row depicts the ground truth, conditional mean, different high-resolution realizations for one low-resolution image and computed standard deviation from the conditional mean for a 2$\times$ upsampling factor. The bottom row displays the same experiment for an upsampling factor of 4$\times$.