文献类型：期刊文献

英文题名：Cascaded Local-Nonlocal Pansharpening with Adaptive Channel-Kernel Convolution and Multi-Scale Large-Kernel Attention

作者：Yin, Junru[1] Huang, Zhiheng[1] Chen, Qiqiang[1] Huang, Wei[1] Sun, Le[2] Wu, Qinggang[1] Hou, Ruixia[3]

第一作者：Yin, Junru

通信作者：Yin, JR[1]

机构：[1]Zhengzhou Univ Light Ind, Coll Comp Sci & Technol, Zhengzhou 450002, Peoples R China;[2]Nanjing Univ Informat Sci & Technol, Sch Comp Sci, Nanjing 210044, Peoples R China;[3]Chinese Acad Forestry, Res Inst Resource Informat Tech, Beijing 100091, Peoples R China

年份：2025

卷号：18

期号：1

外文期刊名：REMOTE SENSING

收录：;WOS:【SCI-EXPANDED(收录号:WOS:001657729800001)】；

基金：This study was supported by the National Natural Science Foundation of China (Grant No. 62375133) and the Henan Province Science and Technology Breakthrough Project (Grant No. 252102211065; Grant No. 252102210131) and the Central Guidance for Local Science and Technology Development Fund Project of Henan (Grant No. Z20251831017) and Funded by Basic Research Program of Jiangsu under (Grant No. BK20250043).

语种：英文

外文关键词：pansharpening; adaptive convolution; large-kernel convolution; attention mechanism

摘要：Highlights What are the main findings? A cascaded local-nonlocal pansharpening network (CLNNet) was developed to progressively refine spatial details and spectral fidelity by stacking PLNF modules. Coupling adaptive channel-kernel convolution with multi-scale large-kernel attention within each PLNF improved texture preservation and spectral fidelity. What are the implications of the main findings? The developed CLNNet could enable the generation of HRMS images for downstream remote sensing applications (e.g., classification and detection). The progressive local-nonlocal fusion strategy offers strong potential for other multi-source image fusion tasks requiring both fine textures and long-range dependencies.Highlights What are the main findings? A cascaded local-nonlocal pansharpening network (CLNNet) was developed to progressively refine spatial details and spectral fidelity by stacking PLNF modules. Coupling adaptive channel-kernel convolution with multi-scale large-kernel attention within each PLNF improved texture preservation and spectral fidelity. What are the implications of the main findings? The developed CLNNet could enable the generation of HRMS images for downstream remote sensing applications (e.g., classification and detection). The progressive local-nonlocal fusion strategy offers strong potential for other multi-source image fusion tasks requiring both fine textures and long-range dependencies.Abstract Pansharpening plays a crucial role in remote sensing applications, as it enables the generation of high-spatial-resolution multispectral images that simultaneously preserve spatial and spectral information. However, most current methods struggle to preserve local textures and exploit spectral correlations across bands while modeling nonlocal information in source images. To address these issues, we propose a cascaded local-nonlocal pansharpening network (CLNNet) that progressively integrates local and nonlocal features through stacked Progressive Local-Nonlocal Fusion (PLNF) modules. This cascaded design allows CLNNet to gradually refine spatial-spectral information. Each PLNF module combines Adaptive Channel-Kernel Convolution (ACKC), which extracts local spatial features using channel-specific convolution kernels, and a Multi-Scale Large-Kernel Attention (MSLKA) module, which leverages multi-scale large-kernel convolutions with varying receptive fields to capture nonlocal information. The attention mechanism in MSLKA enhances spatial-spectral feature representation by integrating information across multiple dimensions. Extensive experiments on the GaoFen-2, QuickBird, and WorldView-3 datasets demonstrate that the proposed method outperforms state-of-the-art methods in quantitative metrics and visual quality.