高光谱成像
遥感
变更检测
计算机科学
人工智能
全光谱成像
模式识别(心理学)
光谱特征
目标检测
光谱分析
计算机视觉
特征提取
图像处理
作者
Bo Zhang,Guirong Feng,Yuehua Chen,Bingyi Liu,Shili Xiong,Xiaoqiang Lu
标识
DOI:10.1109/tgrs.2026.3738778
摘要
A core challenge in hyperspectral change detection (HCD) is spectral pseudochange, artifacts from imaging conditions that cause severe overlap between changed and unchanged samples, reducing their separability and hindering accurate detection. Existing methods, including traditional difference-driven pipelines and deep learning models, remain vulnerable to such discrepancies. To address this issue, we use pseudoclass separability as a design perspective to guide spatial–spectral representation enhancement, where changed and unchanged labels are regarded as discriminative surrogate categories defined on bitemporal spectral pairs rather than semantic land-cover categories. We propose spatial–spectral difference learning (SSDL), a spatial–spectral difference enhancement framework whose central component is spectral attention (SpeA) for peak-valley spectral-structure enhancement. SpeA is complemented by spatial differential attention (SDA) for competitive suppression of nonchange responses and position-aware multihead spectral linear self-attention (SpeLA) for spectral correlation modeling with a favorable accuracy-complexity tradeoff. Experimental results on three public HCD datasets demonstrate competitive and consistent performance, particularly in scenarios where spectral pseudochange causes strong changed/unchanged distribution overlap, and the code is available athttps://github.com/WUTCM-Lab/SSDL
科研通智能强力驱动
Strongly Powered by AbleSci AI