高光谱成像
判别式
模式识别(心理学)
人工智能
卷积(计算机科学)
计算机科学
全光谱成像
空间分析
子空间拓扑
数学
计算机视觉
光谱带
上下文图像分类
光谱特征
块(置换群论)
融合
接头(建筑物)
遥感
支持向量机
空间关系
作者
Mengdi Cheng,Haixin Sun,Jingjing Shao
出处
期刊:
日期:2025-12-05
卷期号:: 362-366
标识
DOI:10.1109/spcnc68200.2025.11406480
摘要
In hyperspectral images (HSIs), both spatial and spectral features play important roles in classification performance; hence pixel-wise classification requires learning discriminative representations from both spatial and spectral domains. In this paper, we propose a dual-branch spatial-spectral Mamba (DBSSMamba) network for HSI classification, in which a Convolution Enhancement Block (CEB) is employed to enhance the features of local patches and to simultaneously generate both local spatial-block features and the central-pixel spectral vector. On this basis, a Spatial Scan Mamba branch (SpaSM) is constructed to perform bidirectional row/column state-space scanning to efficiently model long-range spatial dependencies within each patch, and a Spectral Scan Mamba branch (SpeSM) is designed to conduct forward and backward scanning along the central spectral sequence to capture correlations between different spectral bands. Furthermore, to fully exploit the complementarity between spatial and spectral features, a Token-level SpatialSpectral Fusion (TSSF) module is introduced to perform gated, token-wise fusion of spatial and spectral representations, adaptively suppressing inconsistent spatial-spectral combinations and yielding more discriminative joint representations. Experimental results on the Pavia University and Xuzhou HSI datasets demonstrate that the proposed DBSSMamba achieves superior performance compared with state-of-the-art classification methods.
科研通智能强力驱动
Strongly Powered by AbleSci AI