Multiscale bilateral tensor ring decomposition with gradient fidelity for multitemporal remote sensing thick cloud removal

Abstract

Reconstructing multi-temporal remote sensing images contaminated by thick clouds remains challenging due to the over smoothing and cloud mask sensitivity of single-scale tensor methods. To address this, we propose the Multi-Scale Bilateral Tensor Ring decomposition with Gradient Fidelity (MS-BTR-GF) model. By constructing a 3-scale pyramid (original, 1/2, and 1/4 resolutions), MS-BTR-GF integrates cross-scale consistency constraints to preserve hierarchical spatiotemporal spectral correlations. To counteract low-rank over-smoothing, a localized gradient-domain fidelity operator is applied specifically to cloud zones during the first 50 iterations. The non-convex model is solved using the Proximal Alternating Minimization (PAM) framework, yielding closed-form solutions with quasi-linear complexity. Quantitative evaluations on the SEN12MS-CR-TS dataset (256 × 256 × 13 × 6) demonstrate that MS-BTR-GF consistently outperforms state-of the-art methods. Notably, in the Americas region, it achieves a 15.96% improvement in the Edge Preservation Index (EPI) over the standard BTR baseline. Real-world Landsat-8 experiments further confirm its superior generalization across highly heterogeneous coastlines and mountainous terrains.