Kai Yu

Cloud contamination in multi-temporal remote sensing images, particularly due to thick clouds, significantly impacts the accuracy of environmental monitoring tasks. While existing Low-Rank Tensor Completion (LRTC) methods address cloud removal, they struggle with strict multi-temporal correlation requirements and poor generalization due to parameter sensitivity. To overcome these limitations, we propose WLRT-DR, a novel two-stage framework combining Wavelet-domain Low-Rank Tensor Completion with Deep Regularization. First, we introduce discrete wavelet transform (DWT) to decompose multi-temporal imagery into frequency-specific subbands (LL, LH, HL, HH). Empirical analysis reveals distinct low-rank properties in each subband, enabling targeted low-rank processing that better preserves spatial textures and structural continuity compared to unified spatial-domain LRTC. Second, we develop a dual-phase architecture where wavelet-domain LRTC (Stage 1) is synergistically combined with deep network regularization (Stage 2). This hybrid strategy leverages the complementary strengths: wavelet-domain low-rank recovery efficiently reconstructs fundamental spatial features, while pretrained deep networks impose non-linear constraints to enhance spectral fidelity. Comprehensive validation on simulated (Landsat data covering diverse Beijing landscapes) and real-world data demonstrates that WLRT-DR delivers enhanced reconstruction accuracy, robustness to complex spatio-temporal patterns, and improved processing efficiency. The framework provides an effective alternative for cloud reconstruction by integrating traditional tensor completion and deep learning in a novel manner.

Optical satellite imagery has become indispensable for Earth observation, yet thick cloud contamination severely impedes continuous monitoring capabilities. Current temporal-based methodologies present dual limitations: conventional approaches suffer from inefficient processing speeds and poor robustness, while existing deep learning-based solutions exhibit constrained temporal adaptability by relying on limited temporal information from single or fixed-period images. These inherent shortcomings consequently hinder their scalability in large-scale real-world applications. To address this challenge, we propose a Multi-temporal Adaptive Cloud Removal Transformer (MACRT), which uses a temporal aggregation strategy to integrate information from multiple reference images, reduce the constraint of reference image selection, and provide a more practical cloud removal solution. Since neighboring pixels of remote sensing images have strong correlations, the proposed network architecture introduces a novel spatio-spectral multi-head attention (SSMHA) module. This module decomposes the attention computation into two branches: coarse-resolution spatial and spectral, which work together to capture long-range dependencies across both dimensions while maintaining a compact parameter count. Extensive experiments on multispectral datasets (Landsat and Sentinel-2) demonstrate the superior performance of MACRT. Quantitative evaluations on Landsat imagery reveal improvements of 1.7 % in SSIM and 2.1° in SAM over state-of-the-art methods, achieving enhanced structural preservation and spectral accuracy. The source code, pretrained models, and all experimental results will be publicly available at https://github.com/Zhangyu-student/ACRT upon the paper’s acceptance.

Recently, the volume-gradient-based band selection (VGBS) method has attracted more and more attention in the field of band selection. It is a ranking-based unsupervised algorithm which applies the sequential backward selection strategy to successively remove the most abundant band. The key finding of VGBS is that the band redundancy corresponds to the volume gradient matrix with respect to hyperspectral images. However, we have found that VGBS requires to update the gradient matrix after each band removal, which includes the calculation of the matrix inverse, determinant, and multiplication, and thus is time-consuming when the number of bands is large. In this letter, we first find that the norm of the row of the gradient matrix has a one-to-one correspondence to the diagonal element of the covariance matrix of the image. Further, we develop a recursive formula to calculate the inverse of the covariance matrix. The experimental results show the effectiveness of the method, i.e., we can reduce the computational complexity of VGBS with an order of magnitude.