人工智能
模式识别(心理学)
高光谱成像
冗余(工程)
计算机科学
特征向量
正投影
特征提取
特征(语言学)
投影(关系代数)
线性子空间
编码器
卷积神经网络
光谱空间
水准点(测量)
数学
序列(生物学)
像素
自编码
算法
特征学习
矩阵分解
深度学习
降维
反向
计算
计算机视觉
计算复杂性理论
维数之咒
人工神经网络
理论(学习稳定性)
作者
Dan Yang,Jiale Chen,Junsuo Qu,Yanli Feng,Linquan Li,Xiaobo Jia
出处
期刊:Electronics
[Multidisciplinary Digital Publishing Institute]
日期:2026-08-21
卷期号:15 (16): 3757-3757
标识
DOI:10.3390/electronics15163757
摘要
Hyperspectral image classification relies heavily on the effective modeling of spatial–spectral representations. Recent deep learning architectures, including Transformers and state space models (SSMs), have shown promise for HSI classification. However, under few-shot scenarios, they may suffer from optimization instability in early-stage feature reduction, weakened local spatial–spectral correlations after direct sequence flattening, and attenuation of center-pixel spectral information caused by deep spatial aggregation. To mitigate these issues, we propose orthogonal projection spatial–spectral Mamba (O-Mamba), a lightweight architecture for few-shot HSI classification. First, we introduce a task-driven orthogonal projection module (TOPM) for learnable end-to-end spectral dimensionality reduction. In this module, orthogonal parameterization, supervised initialization, and an auxiliary loss jointly improve the stability of the projection process, reducing feature redundancy and mitigating representation collapse. Second, we design a 3D spatial–spectral Mamba encoder that employs 3D Convolutional Neural Networks (CNN) as local tokenizers to preserve local spatial–spectral structures and then uses Mamba to capture long-range sequence dependencies with linear complexity with respect to sequence length. Finally, to alleviate over-smoothing in the target-pixel representation, we propose a decoupled target–context fusion strategy. This mechanism separately preserves the original spectral signature of the center pixel and fuses it with high-level contextual features, which may improve the separability of spectrally similar classes. Extensive experiments on four benchmark datasets show that O-Mamba achieves competitive classification performance under the evaluated few-shot settings, while maintaining a relatively small model size and low computational cost compared with representative CNN, Transformer, and Mamba-based methods.
科研通智能强力驱动
Strongly Powered by AbleSci AI