计算机科学
人工智能
空间分析
模式识别(心理学)
特征提取
地理信息系统
计算机视觉
遥感
上下文图像分类
高光谱成像
空间语境意识
数据建模
特征(语言学)
支持向量机
图像分割
作者
Yong Wang,Jiaxin Zheng,Yuyang Wang,Pei-Qiu Huang
标识
DOI:10.1109/lgrs.2026.3673195
摘要
Deep learning has been widely applied for hyperspectral image (HSI) few-shot classification tasks. However, 1) existing methods directly process whole spectral bands, ignoring the diversity among bands and thus weakening the spectral representation capability; and 2) they often fail to adequately focus on the central pixel information that represents the target, leading to poor classification performance. To alleviate these two issues, we propose a multi-band spectral information and multi-scale center-guided spatial information learning framework (MBS-MSCS), which includes two parts: multi-band spectral information learning (MBSIL) module and multi-scale center-guided spatial information learning (MSCSIL) module. Specifically, MBSIL employs both intra-band and inter-band attentions to assign adaptive weights to different spectral bands, thereby extracting multi-band spectral features and enhancing the representation ability for different spectral bands. Afterward, MSCSIL makes use of multi-scale cropping and central filling on the multi-band spectral features to fully leverage the central pixel information and output multi-scale center-filling patches. Theses patches are subsequently processed via feature extraction, fusion, and decoding to obtain discriminative spectral-spatial features for classification. Extensive experiments on three datasets demonstrate that MBS-MSCS outperforms state-of-the-art methods.
科研通智能强力驱动
Strongly Powered by AbleSci AI