文献类型：期刊文献

英文题名：Cross-Scene Diffusion-Enhanced Uncertainty Attention Network for Hyperspectral Image Classification

作者：Wang, Chenglong[1];Guo, Yi[1,2,3];Ye, Minchao[4];Xiong, Fengchao[5]

机构：[1]East China Univ Sci & Technol, Dept Comp Sci & Engn, Shanghai 200237, Peoples R China;[2]Natl Engn Lab Big Data Distribut & Exchange Technol, Business Intelligence & Visualizat Res Ctr, Shanghai 200436, Peoples R China;[3]Shanghai Engn Res Ctr Big Data & Internet Audience, Shanghai 200072, Peoples R China;[4]China Jiliang Univ, Coll Informat Engn, Key Lab Electromagnet Wave Informat Technol & Metr, Hangzhou 310018, Peoples R China;[5]Nanjing Univ Sci & Technol, Sch Comp Sci & Engn, Nanjing 210094, Peoples R China

年份：2026

卷号：64

外文期刊名：IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING

收录：;EI(收录号：20261420426015);WOS:【SCI-EXPANDED(收录号:WOS:001737527100010)】；

基金：No Statement Available

语种：英文

外文关键词：Diffusion models; Attention mechanisms; Noise; Image classification; Adaptation models; Hyperspectral imaging; Uncertainty; Data models; Training; Semantics; Cross-scene attention mechanism; cross-scene hyperspectral image (HIS) classification; data augmentation; domain adaptation diffusion module (DADM)

摘要：Cross-scene hyperspectral image (HSI) classification is widely used for its exceptional ability to address the small-sample-size problem. However, important challenges remain, including domain shift and noise between scenes, the need to align features while accounting for sample-level reliability, and the difficulty of fusing domain-invariant and scene-specific features without degrading discriminability. To enhance generalization and reliability with small sample sizes, we propose a cross-scene diffusion-enhanced uncertainty attention network (CDUAN). CDUAN consists of three main components: 1) a domain adaptation diffusion module (DADM) that utilizes target-guided diffusion to synthesize augmented source samples, establishing distribution-level alignment as the first step; 2) a cross-scene uncertainty-aware mutual attention module (CUMAM) is then proposed to ensure feature-level reliability control during alignment across scenes by integrating a sample-level uncertainty mechanism. This mechanism adjusts attention weights to highlight domain-invariant features and, during fusion, retain as much discriminative information as possible; and 3) a cross-scene fusion attention module (CSFAM) realizes fusion-level control of invariance and discriminability across scenes by combining multiscale contrastive pairing with attention-based fusion. CSFAM maximizes relevant cross-scene feature similarity, suppresses noise, and ensures discriminative representation for accurate classification. Extensive experiments on three cross-scene datasets validate the effectiveness of the proposed CDUAN.