Abstract:Super-resolution (SR) is a promising cost-effective downscaling methodology for producing high-resolution climate information from coarser counterparts. A particular application is downscaling regional reanalysis outputs (predictand) from the driving global counterparts (predictor). This study conducts an intercomparison of various SR downscaling methods focusing on temperature and using the CERRA reanalysis (5.5 km resolution, produced with a regional atmospheric model driven by ERA5) as example. The method proposed in this work is the Swin transformer and two alternative methods are used as benchmark (fully convolutional U-Net and convolutional and dense DeepESD) as well as the simple bicubic interpolation. We compare two approaches, the standard one using the full domain as input and a more scalable tiling approach, dividing the full domain into tiles that are used as input. The methods are trained to downscale CERRA surface temperature, based on temperature information from the driving ERA5; in addition, the tiling approach includes static orographic information. We show that the tiling approach, which requires spatial transferability, comes at the cost of a lower performance (although it outperforms some full-domain benchmarks), but provides an efficient scalable solution that allows SR reduction on a pan-European scale and is valuable for real-time applications.

What problem does this paper attempt to address?

### Problems the Paper Attempts to Solve This paper aims to address how to effectively enhance the spatial resolution of regional reanalysis data through Super-Resolution (SR) technology. Specifically, the focus of the research is on how to transform low-resolution temperature information from global reanalysis data (such as ERA5) into high-resolution regional reanalysis data (such as CERRA). The paper compares different SR downscaling methods, particularly the Swin Transformer model based on Transformer, and explores the effects of two methods: global input and block input. ### Main Background Issues 1. **Importance of Reanalysis Data**: - Reanalysis datasets are a primary source of information for climate analysis, providing long-term (decades) physically consistent gridded data on an hourly/daily basis. - Global reanalysis data (such as ERA5) typically have relatively coarse spatial resolution (on the order of tens of kilometers), limiting their application in local and regional climate analysis, especially in the study of extreme weather events. 2. **Development of Downscaling Methods**: - Dynamic Downscaling: Uses regional atmospheric models (LAMs) to simulate within a limited area to improve resolution. Although this method has good physical consistency, it is computationally expensive. - Statistical Downscaling: Based on data-driven methods, it improves resolution by learning the statistical relationship between coarse-resolution atmospheric variables and high-resolution local variables. In recent years, deep learning techniques (such as Convolutional Neural Networks, Generative Adversarial Networks, etc.) have made significant progress in statistical downscaling. 3. **Application of Super-Resolution Technology**: - Super-Resolution technology was initially applied in the field of computer vision and later successfully used for downscaling meteorological and climate data. These technologies can recover high-resolution details from low-resolution data, improving the usability and accuracy of the data. ### Research Objectives - **Compare Different SR Downscaling Methods**: The paper compares the Swin Transformer-based model with traditional Convolutional Neural Networks (such as U-Net and DeepESD) and simple bicubic interpolation methods. - **Evaluate Global and Block Methods**: The study examines the effects of two strategies: global input and block input. The global method takes the entire region as input, while the block method divides the region into multiple small blocks for separate processing. - **Improve Computational Efficiency and Scalability**: The block method improves computational efficiency and scalability while maintaining performance, making it suitable for real-time applications in large-scale regions. ### Key Challenges - **Computational Resource Limitations**: The global method requires a large amount of computational resources when processing large-scale regions, limiting its feasibility in practical applications. - **Spatial Generalization Ability**: The block method needs to have spatial generalization ability, i.e., to maintain good performance in unseen regions. - **Maintaining High-Resolution Details**: Both global and block methods need to preserve the details and structural features of high-resolution data while improving resolution. ### Conclusion The paper experimentally verifies the superior performance of the Swin Transformer in super-resolution downscaling tasks, particularly in the downscaling of temperature data. Although the block method is slightly inferior to the global method in terms of performance, its efficient computational characteristics and good generalization ability make it an ideal choice for large-scale regional applications.

Transformer based super-resolution downscaling for regional reanalysis: Full domain vs tiling approaches

CSwT-SR: Conv-Swin Transformer for Blind Remote Sensing Image Super-Resolution with Amplitude-Phase Learning and Structural Detail Alternating Learning

Resolution-Agnostic Transformer-based Climate Downscaling

A two-step downscaling method for high-scale super-resolution of daily temperature — a case study of Wei River Basin, China

DeepSD: Generating High Resolution Climate Change Projections through Single Image Super-Resolution

A High-Resolution Land Surface Temperature Downscaling Method Based on Geographically Weighted Neural Network Regression

A Novel Residual-Pyramid-attention Super Resolution Model for Mesoscale Meteorological Forecasting Spatial Downscaling

HiT-SR: Hierarchical Transformer for Efficient Image Super-Resolution

Repeatable high-resolution statistical downscaling through deep learning

TTSR: A Transformer-Based Topography Neural Network for Digital Elevation Model Super-Resolution

Downscaling Land Surface Temperature Derived from Microwave Observations with the Super-Resolution Reconstruction Method: A Case Study in the CONUS

A Spatial Downscaling Method for Deriving High-Resolution Downward Shortwave Radiation Data Under All-Sky Conditions

Deep Learning for Daily 2‐m Temperature Downscaling

Generative Diffusion-based Downscaling for Climate

Transforming Image Super-Resolution: A ConvFormer-Based Efficient Approach

Configuration and Intercomparison of Deep Learning Neural Models for Statistical Downscaling

Are Deep Learning Methods Suitable for Downscaling Global Climate Projections? Review and Intercomparison of Existing Models

Deep-Learning-Based Gridded Downscaling of Surface Meteorological Variables in Complex Terrain. Part I: Daily Maximum and Minimum 2-m Temperature

Successful Precipitation Downscaling Through an Innovative Transformer-Based Model

Deep-Learning-Based Gridded Downscaling of Surface Meteorological Variables in Complex Terrain. Part II: Daily Precipitation