Haixin Sun
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Haixin Sun has a total of 13 co-authors. Here are the top co-authors by number of shared publications:
- Fanlei Meng
The School of Electronic and Information Engineering, Changchun University, Changchun 130022, China
4 shared publications - Jingwen Xu
College of Electronic and Information Engineering, Changchun University, Changchun 130022, China
2 shared publications - Qiuguang Cao
College of Electronic and Information Engineering, Changchun University, Changchun 130022, China
2 shared publications - Dan Shao
The School of Electronic and Information Engineering, Changchun University, Changchun 130022, China
1 shared publication Fanlei Meng
School of Electronic and Information Engineering, Changchun University, Changchun, 130022, China
1 shared publication- Hongyue Cao
The School of Electronic and Information Engineering, Changchun University, Changchun 130022, China
1 shared publication
Latest Publications
- Open Access
- 3 July 2026
- Published by MDPI AG in Remote Sensing
- Vol. 18 (13) , 2173
MS3CHFormer: A Multi-Scale Spatial–Spectral Convolutional Hybrid Transformer for Hyperspectral Image Classification
Deep learning methods that integrate convolutional neural networks (CNNs) and Transformers have achieved remarkable progress in hyperspectral image (HSI) classification. However, existing methods still suffer from insufficient multi-scale spatial–spectral feature modeling, a lack of efficient interaction mechanisms between local and global features, and the inherent high computational complexity and redundant information of Transformers, which limit model performance. To address these issues, a Multi-Scale Spatial–Spectral Convolutional Hybrid Transformer model (MS3CHFormer) is proposed in this article. Specifically, a Multi-Scale Spatial–Spectral Convolution Module (MS3ConvM) is first constructed. Through a multi-branch and multi-receptive-field design, it jointly models spatial and spectral features at different scales, thereby enhancing the representation capability of complex ground objects. Then, a Token-Selective Sparse Transformer Encoder (TSSTE) is designed, which adaptively selects tokens and performs sparse modeling via a Dynamic Correlation-Aware Attention (DCAA) mechanism, effectively reducing computational complexity while suppressing redundant information and further reinforcing key feature representations. Furthermore, a Local–Global Feature Fusion Module (LGFFM) is designed to achieve deep complementary fusion of CNN and Transformer features by mapping them into different representation spaces. Finally, a Detail-Preserving Enhancement Module (DPEM) introduces original detail information through residual connections to compensate for detail loss in high-level semantic representations, thereby enhancing the representation capability of boundaries and fine-grained structures. Experiments and comparative analyses on four public HSI datasets demonstrate that the proposed MS3CHFormer outperforms state-of-the-art methods and achieves superior classification accuracy under limited training samples, exhibiting excellent robustness and generalization ability.
- 3 July 2026
- Published by MDPI AG in Remote Sensing
- Vol. 18 (13) , 2173
Open AccessMS3CHFormer: A Multi-Scale Spatial–Spectral Convolutional Hybrid Transformer for Hyperspectral Image Classification
Deep learning methods that integrate convolutional neural networks (CNNs) and Transformers have achieved remarkable progress in hyperspectral image (HSI) classification. However, existing methods still suffer from insufficient multi-scale spatial–spectral feature modeling, a lack of efficient interaction mechanisms between local and global features, and the inherent high computational complexity and redundant information of Transformers, which limit model performance. To address these issues, a Multi-Scale Spatial–Spectral Convolutional Hybrid Transformer model (MS3CHFormer) is proposed in this article. Specifically, a Multi-Scale Spatial–Spectral Convolution Module (MS3ConvM) is first constructed. Through a multi-branch and multi-receptive-field design, it jointly models spatial and spectral features at different scales, thereby enhancing the representation capability of complex ground objects. Then, a Token-Selective Sparse Transformer Encoder (TSSTE) is designed, which adaptively selects tokens and performs sparse modeling via a Dynamic Correlation-Aware Attention (DCAA) mechanism, effectively reducing computational complexity while suppressing redundant information and further reinforcing key feature representations. Furthermore, a Local–Global Feature Fusion Module (LGFFM) is designed to achieve deep complementary fusion of CNN and Transformer features by mapping them into different representation spaces. Finally, a Detail-Preserving Enhancement Module (DPEM) introduces original detail information through residual connections to compensate for detail loss in high-level semantic representations, thereby enhancing the representation capability of boundaries and fine-grained structures. Experiments and comparative analyses on four public HSI datasets demonstrate that the proposed MS3CHFormer outperforms state-of-the-art methods and achieves superior classification accuracy under limited training samples, exhibiting excellent robustness and generalization ability.
- Open Access
- 11 April 2026
- Published by MDPI AG in Algorithms
- Vol. 19 (4) , 300
Spectral–Spatial State Space Model with Hybrid Attention for Hyperspectral Image Classification
Hyperspectral image (HSI) classification requires the extraction of discriminative features from high-dimensional spatial–spectral data. While the Mamba architecture has shown promise in long-sequence modeling with linear complexity, its application to HSI remains constrained by two major hurdles: the unidirectional causal scanning which fails to capture non-causal global dependencies, and the serialization-induced loss of two-dimensional spatial topology and local textures. To overcome these limitations, we propose HAMamba, a novel Hybrid Attention State Space Model. HAMamba facilitates deep representation learning through two core components: a Multi-Scale Dynamic Fusion (MSDF) module and a Hybrid Attention Mamba Encoder (HAME). Specifically, the MSDF module augments spatial perception through parallelized feature extraction and dynamically weighted integration. The HAME synergizes a Bidirectional Sequence Scan Mamba (BSSM) to establish global semantic context and a Spatial–Spectral Gated Attention (SSGA) module to refine local structural details. Comprehensive experiments on four public benchmark datasets demonstrate that the proposed HAMamba significantly outperforms state-of-the-art approaches, achieving a superior balance between classification accuracy and computational efficiency.
- 11 April 2026
- Published by MDPI AG in Algorithms
- Vol. 19 (4) , 300
Open AccessSpectral–Spatial State Space Model with Hybrid Attention for Hyperspectral Image Classification
Hyperspectral image (HSI) classification requires the extraction of discriminative features from high-dimensional spatial–spectral data. While the Mamba architecture has shown promise in long-sequence modeling with linear complexity, its application to HSI remains constrained by two major hurdles: the unidirectional causal scanning which fails to capture non-causal global dependencies, and the serialization-induced loss of two-dimensional spatial topology and local textures. To overcome these limitations, we propose HAMamba, a novel Hybrid Attention State Space Model. HAMamba facilitates deep representation learning through two core components: a Multi-Scale Dynamic Fusion (MSDF) module and a Hybrid Attention Mamba Encoder (HAME). Specifically, the MSDF module augments spatial perception through parallelized feature extraction and dynamically weighted integration. The HAME synergizes a Bidirectional Sequence Scan Mamba (BSSM) to establish global semantic context and a Spatial–Spectral Gated Attention (SSGA) module to refine local structural details. Comprehensive experiments on four public benchmark datasets demonstrate that the proposed HAMamba significantly outperforms state-of-the-art approaches, achieving a superior balance between classification accuracy and computational efficiency.
- Open Access
- 13 March 2026
- Published by MDPI AG in Remote Sensing
- Vol. 18 (6) , 886
A Theory-Guided Transformer for Interpretable Hyperspectral Unmixing
Hyperspectral unmixing (HU) is fundamental for conducting quantitative analyses in remote sensing, yet existing methods face a persistent tradeoff between model performance and physical interpretability. Although deep learning models achieve superior performance, even “gray-box” models that incorporate physical constraints still suffer from an intrinsically opaque decision-making process, which hinders their trustworthiness in critical applications. To address this challenge, this paper introduces a theory-guided unmixing framework aimed at enhancing mechanistic interpretability called the sparse and subspace-attentive transformer unmixing network (SSTU-Net). Unlike heuristic architectures, SSTU-Net is rigorously derived from the first principles of sparse rate reduction (SRR) theory. Its core modules—the multi-head subspace self-attention (MSSA) and the iterative shrinkage-thresholding algorithm (ISTA)—directly implement the essential mathematical steps of information compression and sparsification within the SRR theory, respectively. Extensive experiments on both synthetic and real hyperspectral datasets demonstrate that SSTU-Net achieves competitive performance compared to representative state-of-the-art methods—including advanced autoencoder-based networks (e.g., CyCU-Net and DAAN) and recent transformer-based unmixing architectures (e.g., DeepTrans and MAT-Net)—while strictly adhering to theoretically predicted evolutionary trajectories. More importantly, a series of specifically designed structural interpretability validation experiments mechanistically confirm the theoretically predicted behaviors, such as layer-wise information compression, feature sparsification, and subspace orthogonalization. These results reveal the internal working mechanisms of SSTU-Net, validating the feasibility and significant potential of our principled theory-guided framework for developing high-performance and trustworthy intelligent models in remote sensing.
- 13 March 2026
- Published by MDPI AG in Remote Sensing
- Vol. 18 (6) , 886
Open AccessA Theory-Guided Transformer for Interpretable Hyperspectral Unmixing
Hyperspectral unmixing (HU) is fundamental for conducting quantitative analyses in remote sensing, yet existing methods face a persistent tradeoff between model performance and physical interpretability. Although deep learning models achieve superior performance, even “gray-box” models that incorporate physical constraints still suffer from an intrinsically opaque decision-making process, which hinders their trustworthiness in critical applications. To address this challenge, this paper introduces a theory-guided unmixing framework aimed at enhancing mechanistic interpretability called the sparse and subspace-attentive transformer unmixing network (SSTU-Net). Unlike heuristic architectures, SSTU-Net is rigorously derived from the first principles of sparse rate reduction (SRR) theory. Its core modules—the multi-head subspace self-attention (MSSA) and the iterative shrinkage-thresholding algorithm (ISTA)—directly implement the essential mathematical steps of information compression and sparsification within the SRR theory, respectively. Extensive experiments on both synthetic and real hyperspectral datasets demonstrate that SSTU-Net achieves competitive performance compared to representative state-of-the-art methods—including advanced autoencoder-based networks (e.g., CyCU-Net and DAAN) and recent transformer-based unmixing architectures (e.g., DeepTrans and MAT-Net)—while strictly adhering to theoretically predicted evolutionary trajectories. More importantly, a series of specifically designed structural interpretability validation experiments mechanistically confirm the theoretically predicted behaviors, such as layer-wise information compression, feature sparsification, and subspace orthogonalization. These results reveal the internal working mechanisms of SSTU-Net, validating the feasibility and significant potential of our principled theory-guided framework for developing high-performance and trustworthy intelligent models in remote sensing.
- Open Access
Spatial-Channel Multiscale Transformer Network for Hyperspectral Unmixing
In recent years, deep learning (DL) has been demonstrated remarkable capabilities in hyperspectral unmixing (HU) due to its powerful feature representation ability. Convolutional neural networks (CNNs) are effective in capturing local spatial information, but limited in modeling long-range dependencies. In contrast, transformer architectures extract global contextual features via multi-head self-attention (MHSA) mechanisms. However, most existing transformer-based HU methods focus only on spatial or spectral modeling at a single scale, lacking a unified mechanism to jointly explore spatial and channel-wise dependencies. This limitation is particularly critical for multiscale contextual representation in complex scenes. To address these issues, this article proposes a novel Spatial-Channel Multiscale Transformer Network (SCMT-Net) for HU. Specifically, a compact feature projection (CFP) module is first used to extract shallow discriminative features. Then, a spatial multiscale transformer (SMT) and a channel multiscale transformer (CMT) are sequentially applied to model contextual relations across spatial dimensions and long-range dependencies among spectral channels. In addition, a multiscale multi-head self-attention (MMSA) module is designed to extract rich multiscale global contextual and channel information, enabling a balance between accuracy and efficiency. An efficient feed-forward network (E-FFN) is further introduced to enhance inter-channel information flow and fusion. Experiments conducted on three real hyperspectral datasets (Samson, Jasper and Apex) and one synthetic dataset showed that SCMT-Net consistently outperformed existing approaches in both abundance estimation and endmember extraction, demonstrating superior accuracy and robustness.
Open AccessSpatial-Channel Multiscale Transformer Network for Hyperspectral Unmixing
In recent years, deep learning (DL) has been demonstrated remarkable capabilities in hyperspectral unmixing (HU) due to its powerful feature representation ability. Convolutional neural networks (CNNs) are effective in capturing local spatial information, but limited in modeling long-range dependencies. In contrast, transformer architectures extract global contextual features via multi-head self-attention (MHSA) mechanisms. However, most existing transformer-based HU methods focus only on spatial or spectral modeling at a single scale, lacking a unified mechanism to jointly explore spatial and channel-wise dependencies. This limitation is particularly critical for multiscale contextual representation in complex scenes. To address these issues, this article proposes a novel Spatial-Channel Multiscale Transformer Network (SCMT-Net) for HU. Specifically, a compact feature projection (CFP) module is first used to extract shallow discriminative features. Then, a spatial multiscale transformer (SMT) and a channel multiscale transformer (CMT) are sequentially applied to model contextual relations across spatial dimensions and long-range dependencies among spectral channels. In addition, a multiscale multi-head self-attention (MMSA) module is designed to extract rich multiscale global contextual and channel information, enabling a balance between accuracy and efficiency. An efficient feed-forward network (E-FFN) is further introduced to enhance inter-channel information flow and fusion. Experiments conducted on three real hyperspectral datasets (Samson, Jasper and Apex) and one synthetic dataset showed that SCMT-Net consistently outperformed existing approaches in both abundance estimation and endmember extraction, demonstrating superior accuracy and robustness.
- Open Access
- 1 November 2023
- Published by IOP Publishing in Journal of Physics: Conference Series
- Vol. 2640 (1) , 012006
Lightweight Transformer with Transposed-attention for Hyperspectral Unmixing
Hyperspectral unmixing primarily involves accurately estimating the end members and abundances from mixed pixels in hyperspectral images. In our task, we design a transposed-attention mechanism in the lightweight Transformer to capture the effective information between channels to provide better performance and faster computing speed. We compare the proposed model with those of the CyCU, FCLS, and DHTN methods. The quantitative and qualitative results indicate that our method outperforms the quality of endmember spectral and abundance maps and has a competitive advantage in calculation speed.
- 1 November 2023
- Published by IOP Publishing in Journal of Physics: Conference Series
- Vol. 2640 (1) , 012006
Open AccessLightweight Transformer with Transposed-attention for Hyperspectral Unmixing
Hyperspectral unmixing primarily involves accurately estimating the end members and abundances from mixed pixels in hyperspectral images. In our task, we design a transposed-attention mechanism in the lightweight Transformer to capture the effective information between channels to provide better performance and faster computing speed. We compare the proposed model with those of the CyCU, FCLS, and DHTN methods. The quantitative and qualitative results indicate that our method outperforms the quality of endmember spectral and abundance maps and has a competitive advantage in calculation speed.