RDSRNet: A Recursive Decoupling Super-Resolution Network for Meteorological Remote Sensing Data

Zhiren Zhou, Ziqi Zhao, Chunxu Duan, Q P Zhang, Lin Song, Richard Mitchell

IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing · 01 January 2026 · Nanjing University of Information Science & Technology (NUIST), University of Reading

Paper

In one sentence. RDSRNet recursively decouples structure and detail to super-resolve radar and satellite meteorological data, claiming gains across both data types.

Abstract

Deep-learning super-resolution (SR) methods can improve the spatial resolution of meteorological remote sensing data. However, existing studies have two limitations: First, meteorological remote sensing data suffer from imbalanced feature distribution: structural components (clear-sky or weak-echo regions) are spatially more extensive and continuous than detail components (strong convective regions). Under such circumstances, existing methods that directly learn the mapping from LR to HR in a unified feature space inevitably bias the optimization process toward structural components, ultimately leading to insufficient reconstruction quality of detail components; and second, most existing methods are only validated on a single type of meteorological data (e.g., radar or satellite), and their cross-data-type effectiveness remains insufficiently explored. To address these issues, this article proposes a recursive decoupling super-resolution network (RDSRNet). The recursive decoupling algorithm decomposes the original features into structural components and detail components, whose decoupling and reconstruction rely on subproblems of the same form but smaller scale. The SR result is finally obtained through the recursive reconstruction of subproblems at all levels. To better reconstruct the structural and detail components, a structural–detail feature reconstruction module is designed, which uses a dual-branch network for targeted reconstruction of the two components, alleviating the problem of poor reconstruction effect in detail regions caused by imbalanced feature distribution. Finally, the omni self-attention block is employed to complete feature fusion along both spatial and channel dimensions. In addition, a composite interaction connection is designed in the dual-branch network to amplify the differences between the two components, avoiding an average solution while ensuring the integrity of the structural component and improving the reconstruction accuracy of the detail component. Experiments on radar composite reflectivity and Himawari-9 satellite data demonstrate that RDSRNet outperforms the existing advanced methods across peak signal-to-noise ratio, structural similarity index measure, spectral angle mapper, and root-mean-square error metrics, especially achieving remarkable improvements in structural fidelity and detail clarity for convective regions. Information entropy and mean gradient magnitude are employed to analyze the two components from global and local dimensions, revealing that the detail components obtained through recursive decoupling possess richer information than the structural components; ablation studies further validate the performance contributions of our core designs.

Labels

Labels and summary are generated by Claude and reviewed weekly. Wrong? Edit this entry.