A super-resolution network based on dual aggregate transformer for climate downscaling

Area of Science:

• Climate Science
• Deep Learning
• Image Super-Resolution

Background:

• Deep learning models show promise for climate downscaling tasks.
• Existing models struggle with capturing complex details for high-resolution climate data.
• Current methods lack dynamic reassignment of rainfall variable importance.

Purpose of the Study:

• To propose a novel Climate Downscaling Dual Aggregation Transformer (CDDAT) model.
• To enhance the extraction of rainfall features and provide storm microphysical/dynamical information.
• To improve the generation of high-resolution climate data through multivariate fusion.

Main Methods:

• A hybrid model combining a Lightweight CNN Backbone (LCB) with High Preservation Blocks (HPBs).
• A Dual Aggregation Transformer Backbone (DATB) utilizing adaptive self-attention (spatial window and channel).
• Multivariate fusion operation based on a convolutional neural network.

Main Results:

• The CDDAT model effectively extracts rich rainfall features.
• The network achieves high texture restoration for rainfall images.
• State-of-the-art results were obtained in climate downscaling tasks using the NJU-CPOL dataset.

Conclusions:

• The proposed CDDAT model significantly advances climate downscaling capabilities.
• CDDAT offers improved detail and dynamic feature handling for climate data.
• This approach sets a new benchmark for high-resolution climate data generation.