TSSMamba: A temporal–spectral–spatial state space model for multi-temporal remote sensing cloud removal

Abstract

Cloud cover in optical remote sensing imagery severely limits its effectiveness for downstream applications. Compared with cloud removal methods that rely on Synthetic Aperture Radar (SAR) data or single cloud‐free references, multi-temporal imagery provides richer, more stable and more reliable auxiliary information. However, existing methods often exploit only limited temporal observations and fail to fully capture the deep synergistic relationships among spatial, spectral, and temporal features, leading to reconstruction results that suffer from detail loss and spectral distortion in cloud-contaminated regions. To address this issue, we propose TSSMamba, a temporal–spectral–spatial state space model for multi-temporal cloud removal. The framework jointly models temporal dynamics, spectral responses and spatial structures of cloud-obscured imagery. A dual-stream architecture is designed to separately learn temporal–spectral and temporal–spatial dependencies, while a cross-dimensional fusion module strengthens interactions between the two streams and improves spatial coherence and spectral consistency in the restored images. Experiments on three public cloud removal datasets demonstrate that TSSMamba significantly outperforms state-of-the-art methods in both quantitative metrics and visual quality. On the STGAN Dataset, Sen2_MTC and SEN12MS-CR-TS, TSSMamba achieves PSNR gains of 0.98, 0.59, and 1.11 dB and SSIM improvements of 0.0123, 0.0130, and 0.0115, respectively, over MSC-GAN, thereby confirming its superior cloud removal capability. Code will be made available at: https://github.com/zhangcy23/TSSMamba.