[Paper Review] Deep generative model super-resolves spatially correlated multiregional climate data

Norihiro Oyama, Noriko N. Ishizaki+2 more|arXiv (Cornell University)|Sep 26, 2022

Cryospheric studies and observations4 citations

TL;DR

This paper proposes a physics-informed deep generative model, $π$ SRGAN, that super-resolves coarse global climate model (GCM) outputs to high spatial resolution (up to 50× magnification) while preserving inter-regional spatial correlations and local statistical consistency. By integrating sea-level pressure and topography as auxiliary inputs, the method outperforms conventional statistical downscaling in reconstructing realistic, physically consistent climate fields, particularly for temperature and precipitation.

ABSTRACT

Super-resolving the coarse outputs of global climate simulations, termed downscaling, is crucial in making political and social decisions on systems requiring long-term climate change projections. Existing fast super-resolution techniques, however, have yet to preserve the spatially correlated nature of climatological data, which is particularly important when we address systems with spatial expanse, such as the development of transportation infrastructure. Herein, we show an adversarial network-based machine learning enables us to correctly reconstruct the inter-regional spatial correlations in downscaling with high magnification of up to fifty while maintaining pixel-wise statistical consistency. Direct comparison with the measured meteorological data of temperature and precipitation distributions reveals that integrating climatologically important physical information improves the downscaling performance, which prompts us to call this approach $π$SRGAN (Physics Informed Super-Resolution Generative Adversarial Network). The proposed method has a potential application to the inter-regionally consistent assessment of the climate change impact. Additionally, we present the outcomes of another variant of the deep generative model-based downscaling approach in which the low-resolution precipitation field is substituted with the pressure field, referred to as $ψ$SRGAN (Precipitation Source Inaccessible SRGAN). Remarkably, this method demonstrates unexpectedly good downscaling performance for the precipitation field.

Motivation & Objective

• To address the limitation of existing fast downscaling methods in preserving spatially correlated patterns across distant regions in climate simulations.
• To improve the fidelity of super-resolved climate data for long-term impact assessments in infrastructure, energy, and urban planning.
• To develop a deep generative model that integrates climatologically relevant physical variables (e.g., pressure, topography) to enhance realism and consistency in downscaling.
• To demonstrate that incorporating physical constraints into GAN-based super-resolution improves performance over purely statistical or interpolation-based methods.

Proposed method

• The method employs a generative adversarial network (GAN) architecture adapted from SRGAN, with the generator trained to produce high-resolution climate fields from low-resolution GCM outputs.
• Auxiliary inputs—sea-level pressure and topography—are fed into the generator to inform physical realism and improve spatial correlation reconstruction.
• The discriminator network is trained to distinguish between real high-resolution observations and generated high-resolution outputs, encouraging realistic texture and spatial structure.
• A physics-informed loss function is applied, ensuring that generated outputs align with known physical relationships in climate systems.
• The model is trained using a combination of adversarial loss, perceptual loss (VGG feature matching), and pixel-wise L1 loss to balance realism and statistical accuracy.
• A variant, $ψ$ SRGAN, is also developed to downscale precipitation using pressure fields as input when precipitation data are unavailable.

Experimental results

Research questions

• RQ1 Can a deep generative model preserve inter-regional spatial correlations in multiregional climate downscaling while achieving high magnification (up to 50×)?
• RQ2 How does integrating physical auxiliary data (e.g., sea-level pressure, topography) improve the realism and accuracy of super-resolved climate fields compared to purely statistical methods?
• RQ3 To what extent does the proposed $π$ SRGAN outperform conventional statistical downscaling methods like CDFDM in preserving local statistics and spatial correlation patterns?
• RQ4 Can a GAN-based model achieve high-quality precipitation downscaling even when the precipitation field itself is not available as input, as in the $ψ$ SRGAN variant?

Key findings

• The $π$ SRGAN method successfully reconstructs spatial correlation coefficients across regions with high fidelity, closely matching ground truth distributions from measured meteorological data.
• The model achieves a 50× magnification in super-resolution while maintaining pixel-wise statistical consistency with observed data.
• Integration of physical auxiliary variables (pressure and topography) significantly improves the realism of generated climate fields, especially in capturing large-scale spatial patterns.
• The $π$ SRGAN outperforms conventional CDFDM in preserving spatial correlations, which are often lost in standard statistical downscaling approaches.
• The $ψ$ SRGAN variant, which uses pressure fields instead of precipitation as input, surprisingly achieves strong performance in downscaling precipitation, indicating latent physical relationships can be leveraged effectively.
• The method demonstrates improved physical consistency, such as adherence to expected energy spectra, suggesting it captures fundamental climatological dynamics beyond mere statistical matching.