Deep generative model super-resolves spatially correlated multiregional climate data

Area of Science:

• Climate Science
• Machine Learning
• Geophysics

Background:

• Climate model downscaling is essential for long-term climate change projections impacting societal decisions.
• Current super-resolution methods struggle to maintain spatial correlations in climatological data, vital for large-scale systems like transportation infrastructure.

Purpose of the Study:

• To develop a machine learning approach for climate model downscaling that preserves inter-regional spatial correlations.
• To enhance the accuracy of climate change impact assessments for spatially extensive systems.

Main Methods:

• An adversarial network-based machine learning approach was employed for super-resolution (downscaling) of climate simulation outputs.
• A novel Physics Informed Super-Resolution Generative Adversarial Network (PI-SRGAN) was developed, integrating physical information.
• A variant, Precipitation Source Inaccessible SRGAN (PS-SRGAN), was explored using pressure fields for precipitation downscaling.

Main Results:

• The PI-SRGAN successfully reconstructed inter-regional spatial correlations with high magnification (up to 50x) while maintaining pixel-wise statistical consistency.
• Integration of climatologically relevant physical information significantly improved downscaling performance compared to standard methods.
• The PS-SRGAN demonstrated unexpectedly effective downscaling performance for precipitation fields, even when the direct precipitation field was unavailable.

Conclusions:

• Machine learning, specifically adversarial networks, can effectively perform climate model downscaling while preserving critical spatial correlations.
• The PI-SRGAN offers a promising tool for inter-regionally consistent climate change impact assessments.
• The PS-SRGAN highlights the potential of using related meteorological fields for downscaling precipitation, opening new avenues for climate data reconstruction.