Climate Variable Downscaling with Conditional Normalizing Flows

The document introduces the application of Conditional Normalizing Flows (CNFs) to climate variable downscaling, demonstrating its ability to provide accurate predictions and assess predictive uncertainty, particularly in the context of super-resolution for climate data.

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.

Method

The authors applied Conditional Normalizing Flows (CNFs) to the task of climate variable downscaling, leveraging a conditional prior and a bijective mapping to learn a distribution for super-resolution on climate data. They trained the conditioned spatio-temporal flow with a specific architecture and experimental setup, allowing for the computation of likelihood values, efficient sampling, and the assessment of predictive uncertainty. The method was evaluated using an ERA5 water content dataset for different upsampling factors, showcasing successful performance and the ability to assess predictive uncertainty in terms of standard deviation from the fitted conditional distribution mean.

Main Finding

The authors discovered that applying Conditional Normalizing Flows (CNFs) to the task of climate variable downscaling allows for the evaluation of predictive uncertainty and provides physically consistent results. They demonstrated the successful performance of CNFs on an ERA5 water content dataset for different upsampling factors, showcasing the ability to generate multiple high-resolution realizations for one initial condition and assess predictive uncertainty through the computation of standard deviation from the fitted conditional distribution mean. Additionally, the method outperformed bicubic interpolation and showed competitive performance compared to a Generative Adversarial Network (GAN) architecture in terms of Mean Absolute Error (MAE), Root Mean Square Error (RMSE), and Continuous Ranked Probability Score (CRPS) on the held-out ERA5 water content test set.

Conclusion

The conclusion of the research is that the application of conditional normalizing flows (CNFs) to climate variable downscaling has been successfully demonstrated, providing physically consistent results and enabling the assessment of predictive uncertainty. The proposed method offers the advantage of density estimation and efficient sampling, and effectively models the stochasticity inherent in the relationships among fine and coarse spatial scales of climate variables. Additionally, the method allows for the computation of uncertainty maps in terms of standard deviation computed from the distribution mean, showcasing its potential for improving the accuracy of local scale predictions in climate modeling.

Keywords

climate variable downscaling, conditional normalizing flows, super-resolution, probabilistic models, statistical downscaling