Coupled Transformer Framework for Thick Cloud Removal of Multitemporal Remote Sensing Images
Abstract
Multitemporal remote sensing images (MTRSIs) provide rich temporal information for Earth observation, but thick clouds severely hinder their practical applications. Supervised methods rely on paired cloudy/cloud-free training data, which limits their flexibility across different MTRSIs. Recently, unsupervised methods have attracted increasing attention because they do not require paired training data. However, their representation capability remains limited, particularly in modeling long-range spatiotemporal dependencies required in challenging scenarios with large cloud coverage. To address this challenge, we propose the first unsupervised coupled transformer framework (C-Transformer) for cloud removal, which offers flexible adaptability and long-range modeling ability. Going beyond a straightforward transplantation of Transformers to raw MTRSIs, the fundamental idea behind C-Transformer is to uncover and model the intrinsic spatiotemporal coherence hidden behind cloud-contaminated observations. Specifically, grounded in coupled tensor decomposition, C-Transformer disentangles an MTRSI into a coupled abundance tensor and uncoupled spectral signatures, and employs an untrained Transformer to implicitly represent the former, thereby capturing long-range dependencies while preserving temporal–spectral variability through spectral-guided deep decoders. Extensive experiments on both simulated and real-world datasets demonstrate that C-Transformer achieves the best average performance among competing methods both qualitatively and quantitatively (an average peak signal-to-noise ratio (PSNR) improvement of 2.6 dB).