Diffusion Model-based Probabilistic Downscaling for 180-year East Asian Climate Reconstruction

Fenghua Ling,Zeyu Lu,Jing-Jia Luo,Lei Bai,Swadhin K. Behera,Dachao Jin,Baoxiang Pan,Huidong Jiang,Toshio Yamagata

2024-04-05

Abstract:As our planet is entering into the "global boiling" era, understanding regional climate change becomes imperative. Effective downscaling methods that provide localized insights are crucial for this target. Traditional approaches, including computationally-demanding regional dynamical models or statistical downscaling frameworks, are often susceptible to the influence of downscaling uncertainty. Here, we address these limitations by introducing a diffusion probabilistic downscaling model (DPDM) into the meteorological field. This model can efficiently transform data from 1° to 0.1° resolution. Compared with deterministic downscaling schemes, it not only has more accurate local details, but also can generate a large number of ensemble members based on probability distribution sampling to evaluate the uncertainty of downscaling. Additionally, we apply the model to generate a 180-year dataset of monthly surface variables in East Asia, offering a more detailed perspective for understanding local scale climate change over the past centuries.

Atmospheric and Oceanic Physics,Machine Learning,Geophysics

What problem does this paper attempt to address?

The main objective of this paper is to propose a probabilistic downscaling method based on a diffusion model (Diffusion Probabilistic Downscaling Model, DPDM) to address the limitations of existing downscaling techniques in climate data downscaling, particularly for the needs of regional climate change research.

Specifically, the paper aims to address the following key issues:
1. **Limitations of traditional downscaling methods**: Traditional dynamic downscaling methods (such as regional climate models) can provide high-resolution climate information but come with high computational costs and significant data storage and processing demands. They are also prone to propagating systematic biases from global climate models. Statistical downscaling methods, while less computationally intensive, lack robustness in handling uncertainties.
2. **Improving the accuracy and reliability of downscaling results**: By introducing the concept of probabilistic distribution, DPDM can not only provide more accurate local details but also generate a large number of ensemble members based on probabilistic sampling to assess uncertainties in the downscaling process.
3. **Reconstruction of long-term climate data**: Using DPDM to downscale historical climate data, generating a 180-year dataset of monthly surface variables for East Asia, providing more detailed information for understanding local-scale climate changes over the past centuries.
4. **Assessment of extreme weather events**: By downscaling extreme weather events (such as floods and droughts), evaluating DPDM's capability in addressing these events, and comparing its results with other methods.

In summary, by proposing the innovative DPDM downscaling method, the paper aims to improve existing downscaling techniques, particularly in handling uncertainties, enhancing accuracy, and addressing extreme weather events. Additionally, the paper demonstrates the potential application of DPDM in reconstructing long-term climate data.