Application of multi-stage convolution and attention mechanism based pyramid-spectral-attention Mamba network in unmixing

HU Jiwei; TAN Yangyang; JIN Qiwen

doi:10.11834/jrs.20255023

Models and Methods | Views : 0 下载量: 205 CSCD: 0

R-PDF
PDF
Export
Share
Collection
Album

Application of multi-stage convolution and attention mechanism based pyramid-spectral-attention Mamba network in unmixing
Vol. 30, Issue 1, Pages: 213-230(2026)
Received：16 January 2025，

Published：07 January 2026
DOI： 10.11834/jrs.20255023
稿件说明：

移动端阅览

胡辑伟，谭扬洋，靳淇文.2026.金字塔—光谱—注意力曼巴网络的高光谱图像解混.遥感学报，30（1）： 213-230 DOI： 10.11834/jrs.20255023.

Hu J W，Tan Y Y and Jin Q W. 2026. Application of multi-stage convolution and attention mechanism based pyramid-spectral-attention Mamba network in unmixing. National Remote Sensing Bulletin， 30（1）：213-230 DOI： 10.11834/jrs.20255023.

摘要

线性解混模型因其高效性（计算简单、可扩展性）、明确的物理意义和易于处理受到广泛关注。在各种线性解混方法中，自编码器解混技术在数据拟合能力和深度特征提取能力方面展现出了显著的优势。然而，这种方法也存在一些局限性。例如，输入数据因含有噪声使得模型在处理过程中的泛化能力下降；面对多尺度特征时候往往有一些冗余问题；为了在保持长距离依赖的同时确保计算复杂度为线性，还需要深入理解空间和光谱的特性。为此，本文提出了一种自注意力模块来对高光谱图像进行去噪处理，并提出了一种新的基于Mamba模型的高光谱图像模型。首先设计了一个空间Mamba块用于提取空间特征；然后在光谱方面，提出了一种光谱Mamba块；最后，将光谱向量分成多个组，挖掘不同光谱组之间的关系，并提取光谱特征。将这3种模块融合进多阶段卷积自编码器网络，形成了多阶段曼巴注意力解混框架PSAMN（Phase-wise Mamba Attention Unmixing Framework）。将本文提出的PSAMN模型在合成高光谱数据集和真实高光谱数据集上进行比较实验。实验结果表明把本文提出的解混模型比现有的算法更具有有效性和竞争力。

Abstract

In recent years

the linear mixture model (LMM) has emerged as an important methodology in hyperspectral image unmixing

attracting considerable attention because of its computational efficiency

conceptual simplicity

and interpretability. LMM provides straightforward computations

scalability

and clear physical insights into the unmixing process

thus making it a good option for researchers and practitioners. Among the various linear unmixing techniques

autoencoder-based approaches have demonstrated notable advantages in data fitting capabilities and deep feature extraction. These methods use neural networks to capture complex patterns in hyperspectral data

thereby enabling more precise and robust unmixing outcomes. However

despite their strengths

these methodologies have limitations. A critical issue is the presence of noise within input data

which can affect the model’s generalizability during processing. Noise introduces uncertainty that propagates through the network

thus degrading performance in real-world applications. Furthermore

redundancy often occurs when addressing multiscale features

which are commonly associated with hyperspectral images

thus complicating the learning process and potentially reducing the overall efficacy of the model. Another major issue is the need to maintain long-range dependencies while ensuring computational efficiency. This balance can be achieved by understanding the spatial and spectral characteristics

which is a complicated task in high-dimensional datasets such as hyperspectral images. These challenges become even more pronounced when incorporating advanced architectures such as Mamba blocks

which aim to handle global interactions and require sophisticated design strategies. To address these issues

this study proposes a novel framework that integrates an attention module specifically tailored for denoising hyperspectral images. Furthermore

we introduce MambaHSI

a new hyperspectral image model based on the Mamba framework. This proposed model incorporates two key innovations: a spatial Mamba block and a spectral Mamba block. The spatial Mamba block is designed to simulate global interactions across the entire image at the pixel level

capturing complex relationships between pixels over extensive spatial extents. Meanwhile

the spectral Mamba block partitions spectral vectors into multiple groups

enabling the exploration of inter-group relationships and the extraction of meaningful spectral features. By grouping spectral vectors

the model reduces redundancy and enhances its capacity to effectively represent different spectral patterns. These components are integrated into a multistage convolutional autoencoder network

forming the multistage Mamba attention unmixing framework (PSAMN). PSAMN combines the strengths of spatial and spectral modeling with the power of attention mechanisms

thereby comprehensively addressing the aforementioned challenges. The attention module plays a key role in mitigating the impact of noise by emphasizing relevant features and suppressing irrelevant ones

thereby enhancing the model’s robustness. The multistage architecture ensures that the model can progressively refine its representations

achieving superior accuracy in unmixing tasks. To validate the efficacy of the proposed framework

we conducted extensive comparative experiments on synthetic and real hyperspectral datasets. Results demonstrate that PSAMN outperforms existing state-of-the-art algorithms in terms of unmixing accuracy

robustness

and computational efficiency. On synthetic datasets

PSAMN achieves superior performance metrics

including lower reconstruction errors and higher endmember extraction precision. On real-world datasets

the framework exhibits strong adaptability to varying noise levels and complex spectral–spatial structures

thus being highly competitive in practical applications. In summary

the proposed PSAMN framework is a significant advancement in the field of hyperspectral image unmixing. By addressing key challenges such as noise sensitivity

redundancy in multiscale features

and the need for efficient long-range dependency modeling

PSAMN establishes a new benchmark for future research. Its innovative integration of spatial and spectral Mamba blocks

coupled with an attention-driven denoising module

provides a comprehensive solution to the inherent complexities of hyperspectral data analysis. With the growing application of hyperspectral imaging in remote sensing

environmental monitoring

and medical imaging

frameworks such as PSAMN will play a crucial role in unlocking the full potential of hyperspectral data.

关键词

Keywords

references

Biswas A , Md Abdullah Al N , Imran A , Sejuty A T , Fairooz F , Puppala S and Talukder S . 2023 . Generative adversarial networks for data augmentation //Zheng B, Andrei S, Sarker M K and Gupta K D, eds. Data Driven Approaches on Medical Imaging . Cham : Springer: 159 - 177 [ DOI: 10.1007/978-3-031-47772-0_8 http://dx.doi.org/10.1007/978-3-031-47772-0_8 ]

Cai W , Jiang J and Ouyang S . 2022 . Tensor decomposition and adaptive weighted graph total variation for hyperspectral image denoising . IEEE Geoscience and Remote Sensing Letters , 19 : 5509805 [ DOI: 10.1109/LGRS.2021.3113078 http://dx.doi.org/10.1109/LGRS.2021.3113078 ]

Chen J , Chen Y S , Li W H , Ning G Q , Tong M W and Hilton A . 2021 . Channel and spatial attention-based deep object co-segmentation . Knowledge-Based Systems , 211 : 106550 [ DOI: 10.1016/j.knosys.2020.106550 http://dx.doi.org/10.1016/j.knosys.2020.106550 ]

Chen Y , Lin Z , Zhao B , Wang Y . 2020 . Hyperspectral unmixing based on deep prior of spatial–spectral sparsity . IEEE Transactions on Geoscience and Remote Sensing , 58 ( 7 ): 4758 – 4771 .[ DOI: /10.1109/TGRS.2020.2969899 http://dx.doi.org//10.1109/TGRS.2020.2969899 ]

Ghosh P , Roy S K , Koirala B , Rasti B and Scheunders P . 2022 . Hyperspectral unmixing using transformer network . IEEE Transactions on Geoscience and Remote Sensing , 60 : 5535116 [ DOI: 10.1109/TGRS.2022.3196057 http://dx.doi.org/10.1109/TGRS.2022.3196057 ]

Wang H , Zhao W , Xie W , Qi Q . 2023 . MambaSSM: Linear-time sequence modeling with selective state spaces for hyperspectral unmixing . IEEE Geoscience and Remote Sensing Letters , 20 : 6004505 .[ DOI: 10.1109/LGRS.2023.3234747 http://dx.doi.org/10.1109/LGRS.2023.3234747 ]

He C X , Sun L , Huang W , Zhang J W , Zheng Y H and Jeon B . 2021 . Temporal and spatial low-rank learning network for hyperspectral image mixed denoising . Signal Processing , 184 : 108060 [ DOI: 10.1016/j.sigpro.2021.108060 http://dx.doi.org/10.1016/j.sigpro.2021.108060 ]

Heinz D C and Chang C-I . 2001 . Fully constrained least squares linear spectral mixture analysis method for material quantification in hyperspectral imagery . IEEE Transactions on Geoscience and Remote Sensing , 39 ( 3 ): 529 - 545 [ DOI: 10.1109/36.911111 http://dx.doi.org/10.1109/36.911111 ]

Huo X , Wang M , Zhang H and Zhang Y . 2018 . Deep autoencoder for feature selection in high-dimensional data . IEEE Access , 6 : 23528 - 23535 [ DOI: 10.1109/ACCESS.2018.2831233 http://dx.doi.org/10.1109/ACCESS.2018.2831233 ]

Qu B , Zheng X T , Qian X M and Lu X Q . 2024 . Research progress on hyperspectral anomaly detection . National Remote Sensing Bulletin , 28 ( 1 ): 42 - 54

屈博 , 郑向涛 , 钱学明 , 卢孝强 . 2024 . 高光谱遥感影像异常目标检测研究进展 . 遥感学报 , 28 ( 1 ): 42 - 54 [ DOI: 10.11834/jrs.20232405 http://dx.doi.org/10.11834/jrs.20232405 ]

Qu Y ， Qi H ， Ayhan B ， et al . Hyperspectral image classification with multi-scale convolutional neural network [C ] // 2024 IEEE International Geoscience and Remote Sensing Symposium (IGARSS) . IEEE ， 2024 : 1 - 4 .

Song H H , Wang G J and Zhang K H . 2014 . Hyperspectral image denoising via low-rank matrix recovery . Remote Sensing Letters , 5 ( 10 ): 872 - 881 [ DOI: 10.1080/2150704X.2014.973997 http://dx.doi.org/10.1080/2150704X.2014.973997 ]

Sun H , Zheng X T , Lu X Q and Wu S Y . 2020 . Spectral–spatial attention network for hyperspectral image classification . IEEE Transactions on Geoscience and Remote Sensing , 58 ( 5 ): 3232 - 3245 [ DOI: 10.1109/TGRS.2019.2951160 http://dx.doi.org/10.1109/TGRS.2019.2951160 ]

Tong Q X , Zhang B and Zhang L F . 2016 . Current progress of hyperspectral remote sensing in China . Journal of Remote Sensing (in Chinese) , 20 ( 5 ): 689 - 707

童庆禧 , 张兵 , 张立福 . 2016 . 中国高光谱遥感的前沿进展 . 遥感学报 , 20 ( 5 ): 689 - 707 [ DOI: 10.11834/jrs.20166264 http://dx.doi.org/10.11834/jrs.20166264 ]

Wang J. , Li J. , & Plaza A . ( 2017 ). A new one-dimensional CNN-based end-to-end hyperspectral image unmixing framework . IEEE .[ DOI: 10.1109/IGARSS.2017.8128225 http://dx.doi.org/10.1109/IGARSS.2017.8128225 ]

Zhang B , Sun X , Gao L , Li Y . 2020 . Denoising of hyperspectral imagery using a low-rank tensor decomposition with total variation regularization . IEEE Transactions on Geoscience and Remote Sensing , 58 ( 7 ): 5110 – 5123 .[ DOI: 10.1109/TGRS.2020.2971555 http://dx.doi.org/10.1109/TGRS.2020.2971555 ]

Zhang S , Li S S , Wei G F , Zhang X N and Gao J W . 2022 . Refined multi-scale feature-oriented object detection of the remote sensing images . National Remote Sensing Bulletin , 26 ( 12 ): 2616 - 2628

张省 , 李山山 , 魏国芳 , 张新耐 , 高建威 . 2022 . 面向精细化多尺度特征的遥感图像目标检测 . 遥感学报 , 26 ( 12 ): 2616 - 2628 [ DOI: 10.11834/jrs.20221801 http://dx.doi.org/10.11834/jrs.20221801 ]

Zhong Y F , Feng R Y and Zhang L P . 2014 . Non-local sparse unmixing for hyperspectral remote sensing imagery . IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing , 7 ( 6 ): 1889 - 1909 [ DOI: 10.1109/JSTARS.2013.2280063 http://dx.doi.org/10.1109/JSTARS.2013.2280063 ]

Zhu F Y , Wang Y , Fan B , Xiang S M , Meng G and Pan C H . 2014 . Spectral unmixing via data-guided sparsity . IEEE Transactions on Image Processing , 23 ( 12 ): 5412 - 5427 [ DOI: 10.1109/TIP.2014.2363423 http://dx.doi.org/10.1109/TIP.2014.2363423 ] .