SADER: Structure-Aware Diffusion Framework with DEterministic Resampling for Multi-Temporal Remote Sensing Cloud Removal

TL;DR

SADER addresses cloud contamination in multi-temporal remote sensing by integrating structure-aware diffusion with mean-guided denoising, cloud-aware loss, and deterministic resampling. The method introduces MTCDN to fuse temporal cues with a conditional backbone, and employs a cloud-focused loss plus brightness consistency to stabilize training across cloud-heavy regions. A deterministic resampling strategy guided by a frozen MAE structural prior refines predictions without excessive computation. Across two challenging datasets, SADER achieves consistent improvements in PSNR, SSIM, LPIPS, and spectral fidelity, demonstrating stronger spatial reconstruction, structural preservation, and temporal coherence suitable for large-scale EO applications. The work provides a practical, scalable framework for robust cloud removal in multi-temporal remote sensing tasks, with publicly available code.

Abstract

Cloud contamination severely degrades the usability of remote sensing imagery and poses a fundamental challenge for downstream Earth observation tasks. Recently, diffusion-based models have emerged as a dominant paradigm for remote sensing cloud removal due to their strong generative capability and stable optimization. However, existing diffusion-based approaches often suffer from limited sampling efficiency and insufficient exploitation of structural and temporal priors in multi-temporal remote sensing scenarios. In this work, we propose SADER, a structure-aware diffusion framework for multi-temporal remote sensing cloud removal. SADER first develops a scalable Multi-Temporal Conditional Diffusion Network (MTCDN) to fully capture multi-temporal and multimodal correlations via temporal fusion and hybrid attention. Then, a cloud-aware attention loss is introduced to emphasize cloud-dominated regions by accounting for cloud thickness and brightness discrepancies. In addition, a deterministic resampling strategy is designed for continuous diffusion models to iteratively refine samples under fixed sampling steps by replacing outliers through guided correction. Extensive experiments on multiple multi-temporal datasets demonstrate that SADER consistently outperforms state-of-the-art cloud removal methods across all evaluation metrics. The code of SADER is publicly available at https://github.com/zyfzs0/SADER.

Figures (7)

• Figure 1: Multi-temporal conditional diffusion for remote sensing cloud removal. (a) The diffusion mechanism, illustrating the iterative noise injection and sampling process driven by a mean--noise schedule. (b) Prior methods that incorporate multi-temporal information via simple input concatenation before a single predictive model. (c) Latest diffusion method (EMRDM) that integrates multi-temporal cues into the network, while the diffusion mechanism remains fixed. (d) The proposed SADER framework, which incorporates multi-temporal information into both the network and the diffusion mechanism. Here, the diffusion model couples a network for step-wise prediction with a diffusion mechanism that iteratively refines the reconstruction.
• Figure 2: The diffusion workflow on multi-temporal remote sensing data. The first column shows the auxiliary image. Columns 2-6 correspond to the mean-guided forward diffusion process. Columns 7-11 illustrate the mean-guided sampling process for cloud removal.
• Figure 3: Architecture of the MTCDN. Key components include: (a) a U-Net-like diffusion backbone and a conditional network, where the backbone takes the diffusion target $\mathbf{x}_t^{\mathrm{diff}}$, cloudy mean image $\mathbf{x}_t^{\mu}$, and auxiliary data $\mathbf{A}_t$ as input, and the conditional network processes $\mathbf{x}_t^{\mu}$ and $\mathbf{A}_t$; (b) encoder block with downsampling and conditional modulation; (c) Temporal Fusion Block (TFBlock) using self-attention and cross-attention for temporal feature fusion; (d) Hybrid Attention Block (HABlock) combining global and neighborhood attention; (e) notation legends.
• Figure 4: Pipeline of the cloud-aware attention loss. Multi-temporal cloudy observations are first used to construct region-aware masks for separating cloud and cloud-free areas. Here, $\cup$ denotes the temporal union of multi-temporal cloudy images to form the cloud union image. $\complement$ denotes its complement to obtain the cloud-free region. $\mathrm{Norm.}$ normalizes cloud thickness to $[0,1]$ to generate thickness-aware weights, and $\mathrm{Binarize}$ produces a binary mask.
• Figure 5: Visual comparison of the proposed method against representative baseline approaches on two multi-temporal cloud removal benchmarks. For each example, the cloud-free target and multi-temporal cloudy observations ($T\!-\!1$, $T\!-\!2$, $T\!-\!3$) are shown, followed by the corresponding reconstructions produced by different methods. (a)-(b) are from the Sen2_MTC_New dataset, while (c)-(d) are from the SEN12MS-CR-TS(EA) dataset. All results are visualized using the same post-processing and value rescaling protocol as EMRDM liu2025effective.