Unrolled Low-Rank Tensor Completion for SAR-Guided Cloud Removal in Multispectral Images
TL;DR
This work first formulate cloud removal as a joint optimization problem that preserves spectral correlations via low-rank priors while exploiting SAR information, which provides robust structural and textural guidance for restoring spatial details occluded in the optical images.
Abstract
We address the challenge of cloud removal in multispectral satellite images (MSIs), where clouds obscure critical spatial and spectral information. Cloud removal aims to reconstruct cloud-free MSIs by recovering both spatial structures and spectral signatures from partially occluded observations. While traditional model-based algorithms provide a strong theoretical foundation and interpretability by explicitly incorporating intrinsic priors, their performance and generalization are often limited by handcrafted formulations, limiting robustness across diverse cloud conditions. In contrast, recent deep learning approaches achieve superior reconstruction performance but often overlook the physical properties of MSIs, such as spectral redundancy, interband correlations, and low-dimensional subspace structures, which are essential for accurate MSI reconstruction. In this work, we propose a deep unrolled tensor completion framework with synthetic aperture radar (SAR)-guided detail injection for MSI cloud removal, effectively integrating model- and data-driven approaches to leverage their complementary strengths. Specifically, we first formulate cloud removal as a joint optimization problem that preserves spectral correlations via low-rank priors while exploiting SAR information, which provides robust structural and textural guidance for restoring spatial details occluded in the optical images. We further introduce implicit regularizers to compensate for modeling inaccuracies in model-based priors and solve the resulting optimization problem using an iterative scheme, which is unfolded into a multistage deep network where each stage represents a single iteration. Experimental results on multiple real and synthetic datasets demonstrate that the proposed algorithm achieves superior cloud removal performance with better interpretability compared to state-of-the-art algorithms, while requiring a smaller number of training samples.