DAV-Diff: A Diffusion Model for High-Quality Remote Sensing Image Reconstruction Based on Multi-Level Attention and v $$ v $$ -Prediction

Abstract

With the advancement of deep learning, diffusion models have gained prominence as a powerful method for high-fidelity image generation by progressively removing noise to restore structural and textural details. In remote sensing image super-resolution (RSISR), which typically utilizes paired high-resolution (HR) and low-resolution (LR) images for supervised training, applying diffusion models directly faces three persistent challenges. First, common preprocessing techniques such as bicubic interpolation lose substantial prior information, resulting in a lack of high-frequency details and weakened spatial structural features. Second, existing denoising networks often exhibit insufficient global modeling capability, performing poorly in capturing long-range spatial relationships and evaluating channel-wise feature importance. Third, traditional noise prediction ($ϵ$-prediction) struggles to restore fine details under high-noise conditions, frequently failing to preserve the complex structures of ground objects in remote sensing scenes. To address these issues, this paper proposes DAV-Diff, a fast-sampling conditional diffusion model for RSISR. DAV-Diff coordinates three task-oriented components to improve conditional prior representation, long-range dependency modeling, and high-noise detail recovery in RSISR. The first is a multi-dilated residual attention block (MD-RAB), which enhances the extraction of prior information from LR images and compensates for detail loss caused by interpolation. The second component is a dual-attention mechanism embedded within the denoising network to strengthen long-range spatial dependency modeling and adaptive channel feature weighting. The third is a residual-domain $v$-prediction strategy that replaces traditional noise prediction, mitigating detail degradation in high-noise environments while preserving structural features of complex ground objects. Experimental results show the effectiveness of DAV-Diff against state-of-the-art methods; ablation studies, computational complexity analysis, and sampling-step sensitivity experiments further quantify the contribution and efficiency of each component.