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

This paper addresses the problem of climate downscaling. Previous research on image super-resolution models has demonstrated the effectiveness of deep learning for downscaling tasks. However, most existing deep learning models for climate downscaling have limited ability to capture the complex details required to generate High-Resolution (HR) image climate data and lack the ability to reassign the importance of different rainfall variables dynamically. To handle these challenges, in this paper, we propose a Climate Downscaling Dual Aggregation Transformer (CDDAT), which can extract rich and high-quality rainfall features and provide additional storm microphysical and dynamical structure information through multivariate fusion. CDDAT is a novel hybrid model consisting of a Lightweight CNN Backbone(LCB) with High Preservation Blocks (HPBs) and a Dual Aggregation Transformer Backbone(DATB) equipped with the adaptive self-attention. Specifically, we first extract high-frequency features employing LCB, which adopts HPBs to dynamically reduce the resolution of rainfall processing features and to extract rainfall image depth features at low cost. Then we utilize the DATB to alternately apply spatial window and channel self-attention for spatial and channel features aggregation. Furthermore, we introduce a multimodal fusion operation based on a convolutional neural network. Finally, we evaluate the CDDAT using the NJU-CPOL dataset. The experimental results demonstrate that the proposed network can perform high texture restoration of rainfall images and achieve state-of-the-art-results in climate downscaling tasks.