pith. sign in

arxiv: 2605.03371 · v1 · submitted 2026-05-05 · 💻 cs.CV

SoDa2: Single-Stage Open-Set Domain Adaptation via Decoupled Alignment for Cross-Scene Hyperspectral Image Classification

Yiwen Liu , Minghua Wang , Jing Yao , Xin Zhao , Gemine Vivone This is my paper

Pith reviewed 2026-05-08 01:24 UTC · model grok-4.3

classification 💻 cs.CV
keywords open-set domain adaptationhyperspectral image classificationcross-scene transferdecoupled alignmentsingle-stage trainingGaussian mixture modelspectral-spatial features
0
0 comments X

The pith

SoDa² performs open-set cross-scene hyperspectral classification by decoupling spectral and spatial alignment inside a single training stage.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper claims that mixing spectral and spatial features before alignment creates unnecessary domain shift and that two-stage training inflates cost, so it builds a single-stage method that extracts and aligns the two modalities separately. A contribution-aware extractor first disentangles the spectral sequence from the spatial details, then a decoupled module minimizes Maximum Mean Discrepancy on each modality in isolation. A dual-branch network keeps one set of features aligned and another set unconstrained; a Gaussian mixture model fitted to the squared cosine similarity between the two branches labels pixels as known or unknown without any examples of the unknown categories. If this separation works, models trained on one scene can be transferred to another scene that contains extra classes while keeping both accuracy and training speed high.

Core claim

SoDa² is a single-stage open-set domain adaptation framework that uses a contribution-aware dual-modality extractor to separate spectral and spatial signals, applies independent MMD alignment to each signal type, and employs a dual-branch architecture whose squared-cosine similarity distribution is modeled by a Gaussian mixture model to recognize unknown classes in the target domain without prior examples of those classes.

What carries the argument

The decoupled alignment module, which independently minimizes Maximum Mean Discrepancy on spectral features and on spatial features produced by a contribution-aware dual-modality extractor inside a single-stage dual-branch network.

If this is right

  • Classification accuracy on target scenes containing unknown categories improves over mixed-feature alignment baselines.
  • Training cost drops because the entire pipeline runs in one stage instead of two.
  • Domain-invariant features become finer-grained once spectral and spatial discrepancies are minimized separately.
  • Open-set recognition works without any labeled or unlabeled examples of the unknown classes in the target domain.
  • Model transferability across scenes increases because the alignment does not force unknown classes into the known-class space.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same decoupling pattern could be tried on other paired modalities such as RGB plus depth or multispectral plus LiDAR where features are currently entangled before alignment.
  • If the Gaussian mixture model step proves robust, the approach might replace more expensive open-set detectors that require negative samples or outlier synthesis.
  • Extending the single-stage structure to continual learning across many sequential scenes would test whether the dual-branch memory remains stable over time.
  • The contribution-aware weighting inside the extractor could be inspected on datasets where one modality is known to be noisier than the other.

Load-bearing premise

The dual-modality extractor really separates spectral from spatial information and the Gaussian mixture model on squared cosine similarity can separate known from unknown classes without any examples of the unknown categories.

What would settle it

Run the three groups of HSI datasets through SoDa² and through the strongest prior two-stage open-set methods; if SoDa² does not show higher overall accuracy or higher unknown-class detection rate while using fewer training epochs, the central claim does not hold.

Figures

Figures reproduced from arXiv: 2605.03371 by Gemine Vivone, Jing Yao, Minghua Wang, Xin Zhao, Yiwen Liu.

Figure 1
Figure 1. Figure 1: Key Challenges of OSDA in HSI. and spatial information and is widely applied in various fields such as precision agriculture, environmental monitoring, and mineral identification [1–8]. Among these applications, HSI classification serves as a core task of HSI, with the primary objective of assigning a class label to each pixel based on its spectral and spatial characteristics [9–11]. With the rapid advance… view at source ↗
Figure 2
Figure 2. Figure 2: Comparison of Open Set Domain Adaptive Models. view at source ↗
Figure 3
Figure 3. Figure 3: Framework diagram of a single-stage open-set domain adaptation method based on decoupled alignment (SoDa view at source ↗
Figure 4
Figure 4. Figure 4: Framework diagram of contribution-aware dual-modality feature extraction. view at source ↗
Figure 5
Figure 5. Figure 5: Open set recognition module. similarity as the consistency metric. For a target sample x T j , the cosine similarity between its aligned feature f T a,j ∈ F T a and intrinsic feature f T b,j ∈ F T b is computed as Equation (9): sim(x T j ) = f T a,j · f T b,j view at source ↗
Figure 6
Figure 6. Figure 6: Visualization of classification results for the PU-PC task. (a) Ground-truth. (b) OSBP. (c) STA. (d) DAMC. (e) UADAL. (f) ANNA. (g) MTS. (h) view at source ↗
Figure 7
Figure 7. Figure 7: Visualization of classification results for the HU13–HU18 task. (a) Ground-truth. (b) OSBP. (c) STA. (d) DAMC. (e) UADAL. (f) ANNA. (g) MTS. view at source ↗
Figure 8
Figure 8. Figure 8: Visualization of classification results for the ZY–GF task. (a) Ground-truth. (b) OSBP. (c) STA. (d) DAMC. (e) UADAL. (f) ANNA. (g) MTS. (h) view at source ↗
read the original abstract

Cross-scene hyperspectral image (HSI) classification stands as a fundamental research topic in remote sensing, with extensive applications spanning various fields. Owing to the inclusion of unknown categories in the target domain and the existence of domain shift across different scenes, open-set domain adaptation techniques are commonly employed to address cross-scene HSI classification. However, existing open-set cross-scene HSI classification methods still face two critical challenges: (1) domain shift issues arising from the direct alignment of mixed spectral-spatial features; (2) high computational costs caused by two-stage training strategies. To address these issues, this paper proposes a single-stage open-set domain adaptation method with decoupled alignment (SoDa$^2$) for cross-scene HSI classification. A contribution-aware dual-modality feature extraction is customized to disentangle the characteristics from spectral sequence signals and spatial details, selectively and adaptively enhancing discriminative features. The decoupled alignment module minimizes the Maximum Mean Discrepancy to independently reduce the spectral discrepancy and the spatial discrepancy between the source and target domains, extracting more fine-grained domain-invariant features. A cost-effective single-stage dual-branch framework is designed to learn MMD-constrainted aligned features and constraint-free intrinsic features for adaptive distinction between known and unknown classes. This framework employs a Gaussian Mixture Model to model the squared cosine similarity distribution between the two feature types, enabling open-set recognition without prior knowledge of unknown classes. Extensive experiments on three groups of HSI datasets demonstrate that SoDa$^2$ outperforms state-of-the-art methods, achieving superior classification accuracy and model transferability for open-set cross-scene tasks.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The manuscript proposes SoDa², a single-stage open-set domain adaptation method for cross-scene hyperspectral image (HSI) classification. It introduces a contribution-aware dual-modality feature extractor to disentangle spectral sequence signals and spatial details, a decoupled alignment module that applies Maximum Mean Discrepancy (MMD) independently to spectral and spatial features to mitigate domain shift, and a cost-effective dual-branch framework that learns MMD-constrained aligned features alongside constraint-free intrinsic features. Open-set recognition is performed by modeling the squared cosine similarity distribution between these two feature types with a Gaussian Mixture Model (GMM), enabling distinction of unknown classes without any prior examples or knowledge of them. Extensive experiments on three groups of HSI datasets are reported to show superior classification accuracy and transferability over state-of-the-art methods.

Significance. If the central claims hold, the work offers a practical advance for remote sensing by replacing two-stage open-set DA pipelines with a single-stage decoupled approach that separately handles spectral and spatial discrepancies. The dual-branch design for open-set detection via GMM on feature similarities is a targeted contribution to HSI cross-scene tasks, where unknown categories and domain shifts are prevalent. The paper's emphasis on computational efficiency and fine-grained domain-invariant features could be valuable if the separability assumption is validated.

major comments (2)
  1. [Dual-branch framework and GMM modeling] The open-set recognition mechanism (described in the abstract and the dual-branch framework section) asserts that the GMM fitted to squared cosine similarities between MMD-aligned and intrinsic features can separate known from unknown classes without calibration samples or prior knowledge. No derivation, bound, or analysis is supplied showing why the similarity statistic must be bimodal or separable for unknowns across HSI scenes; if the distributions overlap or the GMM fit is initialization-sensitive, the reported gains in open-set accuracy would not follow.
  2. [Experiments section] The experimental claims of outperforming SOTA methods on three groups of HSI datasets rest on the assumption that the contribution-aware extractor truly disentangles spectral/spatial cues and that decoupled MMD alignment produces a reliable bimodal signal. Without reported ablations isolating each component, statistical significance tests, or details on data splits and unknown-class handling, it is unclear whether the performance improvements are attributable to the proposed mechanisms or to other factors.
minor comments (2)
  1. [Abstract] The abstract refers to 'three groups of HSI datasets' without naming the specific scenes or datasets, which reduces clarity for readers assessing generality.
  2. [Method] Notation for the dual-modality extractor and the exact form of the squared cosine similarity used in the GMM could be formalized with equations for reproducibility.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive comments. We respond to each major comment below and indicate revisions to strengthen the manuscript.

read point-by-point responses

  1. Referee: The open-set recognition mechanism (described in the abstract and the dual-branch framework section) asserts that the GMM fitted to squared cosine similarities between MMD-aligned and intrinsic features can separate known from unknown classes without calibration samples or prior knowledge. No derivation, bound, or analysis is supplied showing why the similarity statistic must be bimodal or separable for unknowns across HSI scenes; if the distributions overlap or the GMM fit is initialization-sensitive, the reported gains in open-set accuracy would not follow.

    Authors: We agree that the original manuscript provides no theoretical derivation or bound guaranteeing bimodality of the squared cosine similarity distribution. The design is motivated by the expectation that MMD-aligned features emphasize domain-invariant cues while intrinsic features preserve scene-specific details, producing higher similarity for known classes. In revision we will add visualizations of the similarity distributions on all three dataset groups to show observed separation, plus an empirical sensitivity study of GMM initialization (multiple random seeds) demonstrating stable open-set accuracy. These additions supply the requested analysis without claiming a formal proof. revision: partial

  2. Referee: The experimental claims of outperforming SOTA methods on three groups of HSI datasets rest on the assumption that the contribution-aware extractor truly disentangles spectral/spatial cues and that decoupled MMD alignment produces a reliable bimodal signal. Without reported ablations isolating each component, statistical significance tests, or details on data splits and unknown-class handling, it is unclear whether the performance improvements are attributable to the proposed mechanisms or to other factors.

    Authors: We accept that the current Experiments section lacks component-wise ablations, significance testing, and explicit protocol details. We will revise to include: (i) ablations removing the contribution-aware extractor, the decoupled MMD module, and the dual-branch structure individually; (ii) statistical significance tests (e.g., McNemar’s test across repeated runs) against the strongest baselines; (iii) precise data-split descriptions (source/target sample counts, train/val/test ratios, and selection of unknown classes in the target scene); and (iv) confirmation that unknown classes receive no samples or labels during training. These changes will clarify attribution of the reported gains. revision: yes

Circularity Check

0 steps flagged

No significant circularity; method is a proposed architecture on standard components

full rationale

The paper describes a single-stage framework that applies MMD for decoupled spectral-spatial alignment and GMM on squared cosine similarity between aligned and intrinsic features to separate known from unknown classes. No equations, derivations, or performance claims reduce by construction to quantities fitted inside the paper itself; the separation mechanism is presented as an empirical outcome of the dual-branch design rather than a self-defining or tautological prediction. The approach builds on established MMD and GMM techniques without load-bearing self-citations or ansatzes that collapse the central result to its inputs.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only review yields no explicit free parameters, axioms, or invented entities; the method relies on standard MMD and GMM techniques from prior literature.

pith-pipeline@v0.9.0 · 5610 in / 1251 out tokens · 43912 ms · 2026-05-08T01:24:14.093432+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

43 extracted references · 1 canonical work pages

  1. [1]

    Hy- perspectral image denoising via nonlocal rank residual modeling,

    Z. Zha, B. Wen, X. Yuan, J. Zhou, and C. Zhu, “Hy- perspectral image denoising via nonlocal rank residual modeling,” inICASSP 2023 - 2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2023, pp. 1–5

    2023

  2. [2]

    Joint contextual representation model-informed interpretable network with dictionary aligning for hy- perspectral and lidar classification,

    W. Dong, T. Yang, J. Qu, T. Zhang, S. Xiao, and Y . Li, “Joint contextual representation model-informed interpretable network with dictionary aligning for hy- perspectral and lidar classification,”IEEE Transactions on Circuits and Systems for Video Technology, vol. 33, no. 11, pp. 6804–6818, 2023

    2023

  3. [3]

    Cross-satellite hierarchical multimodal de- noising for hyperspectral imagery,

    M. Wang, B. Shang, Y . Li, X. Zhao, L. Gao, X. Sun, and L. Ren, “Cross-satellite hierarchical multimodal de- noising for hyperspectral imagery,”ISPRS Journal of Photogrammetry and Remote Sensing, vol. 232, pp. 94– 108, 2026

    2026

  4. [4]

    Nonlocal structured sparsity regulariza- tion modeling for hyperspectral image denoising,

    Z. Zha, B. Wen, X. Yuan, J. Zhang, J. Zhou, Y . Lu, and C. Zhu, “Nonlocal structured sparsity regulariza- tion modeling for hyperspectral image denoising,”IEEE Transactions on Geoscience and Remote Sensing, vol. 61, pp. 1–16, 2023

    2023

  5. [5]

    Dual- interaction spatiotemporal fusion network for agricultural remote sensing imagery,

    M. Wang, B. Shang, J. Yao, R. Zhao, and X. Zhao, “Dual- interaction spatiotemporal fusion network for agricultural remote sensing imagery,”IEEE Journal of Selected Top- ics in Signal Processing, pp. 1–15, 2025

    2025

  6. [6]

    Early detection of pine wilt disease by combining pigment and moisture content indices using uav-based hyperspectral imagery,

    R. Hou, B. Zhang, G. Fang, S. Yang, L. Guo, W. Huang, J. Yao, Q. Jiao, H. Sun, and J. Yan, “Early detection of pine wilt disease by combining pigment and moisture content indices using uav-based hyperspectral imagery,” Remote Sensing, vol. 17, no. 11, p. 1833, 2025

    2025

  7. [7]

    Model- guided coarse-to-fine fusion network for unsupervised hyperspectral image super-resolution,

    J. Li, K. Zheng, W. Liu, Z. Li, H. Yu, and L. Ni, “Model- guided coarse-to-fine fusion network for unsupervised hyperspectral image super-resolution,”IEEE Geoscience and Remote Sensing Letters, vol. 20, pp. 1–5, 2023

    2023

  8. [8]

    Newly established forests dominated global carbon sequestration change induced by land cover conversions,

    D. Peng, B. Zhang, S. Zheng, W. Ju, J. M. Chen, P. Ciais, H. Guo, Y . Pan, L. Yu, Y . Xuet al., “Newly established forests dominated global carbon sequestration change induced by land cover conversions,”Nature Communi- cations, vol. 16, no. 1, p. 6570, 2025

    2025

  9. [9]

    Deep learning for hyperspectral image classification: A survey,

    V . Kumar, R. S. Singh, M. Rambabu, and Y . Dua, “Deep learning for hyperspectral image classification: A survey,”Computer Science Review, vol. 53, p. 100658, 2024

    2024

  10. [10]

    Cross-domain hyperspectral image classification based on bi-directional domain adaptation,

    Y . Zhang, W. Li, W. Jia, M. Zhang, R. Tao, and S. Liang, “Cross-domain hyperspectral image classification based on bi-directional domain adaptation,”IEEE Transactions on Circuits and Systems for Video Technology, vol. 35, no. 12, pp. 12 038–12 051, 2025

    2025

  11. [11]

    Pfs3f: Probabilistic fusion of superpixel-wise and semantic-aware structural features for hyperspectral image classification,

    Y . Zhang, P. Duan, L. Liang, X. Kang, J. Li, and A. Plaza, “Pfs3f: Probabilistic fusion of superpixel-wise and semantic-aware structural features for hyperspectral image classification,”IEEE Transactions on Circuits and Systems for Video Technology, vol. 35, no. 9, pp. 8723– 8737, 2025

    2025

  12. [12]

    Hyperspectral image classification with convolutional neural network IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING 13 and active learning,

    X. Cao, J. Yao, Z. Xu, and D. Meng, “Hyperspectral image classification with convolutional neural network IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING 13 and active learning,”IEEE Transactions on Geoscience and Remote Sensing, vol. 58, no. 7, pp. 4604–4616, 2020

    2020

  13. [13]

    Hyperspectral image transformer classification networks,

    X. Yang, W. Cao, Y . Lu, and Y . Zhou, “Hyperspectral image transformer classification networks,”IEEE Trans- actions on Geoscience and Remote Sensing, vol. 60, pp. 1–15, 2022

    2022

  14. [14]

    Integration of hyperspectral imaging and autoencoders: Benefits, applications, hyperparameter tunning and chal- lenges,

    G. Jaiswal, R. Rani, H. Mangotra, and A. Sharma, “Integration of hyperspectral imaging and autoencoders: Benefits, applications, hyperparameter tunning and chal- lenges,”Computer Science Review, vol. 50, p. 100584, 2023

    2023

  15. [15]

    Special: zero-shot hyperspectral image classification with clip,

    L. Pang, J. Yao, K. Li, J. Zhou, D. Meng, and X. Cao, “Special: zero-shot hyperspectral image classification with clip,”arXiv preprint arXiv:2501.16222, 2025

    work page arXiv 2025

  16. [16]

    Semi-active convolutional neural networks for hyperspectral image classification,

    J. Yao, X. Cao, D. Hong, X. Wu, D. Meng, J. Chanussot, and Z. Xu, “Semi-active convolutional neural networks for hyperspectral image classification,”IEEE Transac- tions on Geoscience and Remote Sensing, vol. 60, pp. 1–15, 2022

    2022

  17. [17]

    An unsupervised domain adaptation method towards multi-level features and decision boundaries for cross- scene hyperspectral image classification,

    C. Zhao, B. Qin, S. Feng, W. Zhu, L. Zhang, and J. Ren, “An unsupervised domain adaptation method towards multi-level features and decision boundaries for cross- scene hyperspectral image classification,”IEEE Trans- actions on Geoscience and Remote Sensing, vol. 60, pp. 1–16, 2022

    2022

  18. [18]

    Locality robust domain adaptation for cross-scene hyperspectral image classification,

    J. Zhang, W. Li, W. Sun, Y . Zhang, and R. Tao, “Locality robust domain adaptation for cross-scene hyperspectral image classification,”Expert Systems with Applications, vol. 238, p. 121822, 2024

    2024

  19. [19]

    Cross-domain hyperspectral image classifica- tion,

    Z. Jiang, J. Li, S. Xu, Z. Liu, D. Ma, Q. Wang, and Y . Yuan, “Cross-domain hyperspectral image classifica- tion,”Pattern Recognition, p. 111836, 2025

    2025

  20. [20]

    A shift reduction domain generalization network for hyperspectral image cross-domain classification,

    Y . Qi, D. Liu, J. Zhang, and Y . Zhang, “A shift reduction domain generalization network for hyperspectral image cross-domain classification,”IEEE Transactions on Geo- science and Remote Sensing, 2025

    2025

  21. [21]

    Pseudo-label-assisted subdomain adaptation for hyper- spectral image classification,

    Z. Feng, S. Tong, S. Yang, X. Zhang, and L. Jiao, “Pseudo-label-assisted subdomain adaptation for hyper- spectral image classification,”IEEE Transactions on Cir- cuits and Systems for Video Technology, vol. 34, no. 6, pp. 4729–4744, 2024

    2024

  22. [22]

    Domain adaptation in remote sensing image classification: A survey,

    J. Peng, Y . Huang, W. Sun, N. Chen, Y . Ning, and Q. Du, “Domain adaptation in remote sensing image classification: A survey,”IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, vol. 15, pp. 9842–9859, 2022

    2022

  23. [24]

    Supervised contrastive learning-based unsupervised do- main adaptation for hyperspectral image classification,

    Z. Li, Q. Xu, L. Ma, Z. Fang, Y . Wang, W. He, and Q. Du, “Supervised contrastive learning-based unsupervised do- main adaptation for hyperspectral image classification,” IEEE Transactions on Geoscience and Remote Sensing, vol. 61, pp. 1–17, 2023

    2023

  24. [25]

    Cross-scene wetland mapping on hyper- spectral remote sensing images using adversarial domain adaptation network,

    Y . Huang, J. Peng, N. Chen, W. Sun, Q. Du, K. Ren, and K. Huang, “Cross-scene wetland mapping on hyper- spectral remote sensing images using adversarial domain adaptation network,”ISPRS Journal of Photogrammetry and Remote Sensing, vol. 203, pp. 37–54, 2023

    2023

  25. [26]

    Mind the gap: Open set domain adaptation via mutual-to-separate framework,

    D. Chang, A. Sain, Z. Ma, Y .-Z. Song, R. Wang, and J. Guo, “Mind the gap: Open set domain adaptation via mutual-to-separate framework,”IEEE Transactions on Circuits and Systems for Video Technology, vol. 34, no. 6, pp. 4159–4174, 2024

    2024

  26. [27]

    Wdan: A weighted discriminative adversarial network with dual classifiers for fine-grained open-set domain adaptation,

    J. Li, L. Yang, Q. Wang, and Q. Hu, “Wdan: A weighted discriminative adversarial network with dual classifiers for fine-grained open-set domain adaptation,” IEEE Transactions on Circuits and Systems for Video Technology, vol. 33, no. 9, pp. 5133–5147, 2023

    2023

  27. [28]

    Enhancing multi-source open-set domain adaptation through nearest neighbor classification with self-supervised vision transformer,

    J. Li, L. Yang, and Q. Hu, “Enhancing multi-source open-set domain adaptation through nearest neighbor classification with self-supervised vision transformer,” IEEE Transactions on Circuits and Systems for Video Technology, vol. 34, no. 4, pp. 2648–2662, 2024

    2024

  28. [29]

    It-ose: Exploring optimal sample size for industrial data aug- mentation,

    M. Sun, R. Zhao, Z. Huang, S. Ding, and J. Liu, “It-ose: Exploring optimal sample size for industrial data aug- mentation,”IEEE Transactions on Industrial Informatics, pp. 1–11, 2026

    2026

  29. [30]

    Open set domain adaptation by backpropagation,

    K. Saito, S. Yamamoto, Y . Ushiku, and T. Harada, “Open set domain adaptation by backpropagation,” in Computer Vision – ECCV 2018, V . Ferrari, M. Hebert, C. Sminchisescu, and Y . Weiss, Eds. Cham: Springer International Publishing, 2018, pp. 156–171

    2018

  30. [31]

    Sepa- rate to adapt: Open set domain adaptation via progressive separation,

    H. Liu, Z. Cao, M. Long, J. Wang, and Q. Yang, “Sepa- rate to adapt: Open set domain adaptation via progressive separation,” in2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019, pp. 2922– 2931

    2019

  31. [32]

    Adjustment and alignment for unbiased open set domain adaptation,

    W. Li, J. Liu, B. Han, and Y . Yuan, “Adjustment and alignment for unbiased open set domain adaptation,” in 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023, pp. 24 110–24 119

    2023

  32. [33]

    Open-set domain adaptation for hyperspec- tral image classification based on weighted generative adversarial networks and dynamic thresholding,

    K. Bi, Z. Li, Y . Chen, Q. Du, L. Ma, Y . Wang, Z. Fang, and M. Qi, “Open-set domain adaptation for hyperspec- tral image classification based on weighted generative adversarial networks and dynamic thresholding,”IEEE Transactions on Geoscience and Remote Sensing, vol. 63, pp. 1–17, 2025

    2025

  33. [34]

    Open-set domain adaptation for scene classification using multi-adversarial learning,

    J. Zheng, Y . Wen, M. Chen, S. Yuan, W. Li, Y . Zhao, W. Wu, L. Zhang, R. Dong, and H. Fu, “Open-set domain adaptation for scene classification using multi-adversarial learning,”ISPRS Journal of Photogrammetry and Remote Sensing, vol. 208, pp. 245–260, 2024

    2024

  34. [35]

    In- tegrating intrinsic information: a novel open set domain adaptation network for cross-domain fault diagnosis with multiple unknown faults,

    Y . Zhang, H. Zhang, B. Chen, J. Zheng, and H. Pan, “In- tegrating intrinsic information: a novel open set domain adaptation network for cross-domain fault diagnosis with multiple unknown faults,”Knowledge-Based Systems, vol. 299, p. 112100, 2024

    2024

  35. [36]

    An open set domain adaptation network based on adversarial learning for remote sensing image scene classification,

    J. Zhang, J. Liu, L. Shi, B. Pan, and X. Xu, “An open set domain adaptation network based on adversarial learning for remote sensing image scene classification,” inIGARSS 2020-2020 IEEE International Geoscience and Remote Sensing Symposium. IEEE, 2020, pp. 1365– IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING 14 1368

    2020

  36. [37]

    Parallel adversarial domain adaptation for cross-dataset hyperspectral image classification,

    Y . Qie, J. Li, D. Shen, Z. Du, X. Ma, J. Wang, and H. Wang, “Parallel adversarial domain adaptation for cross-dataset hyperspectral image classification,”IEEE Transactions on Geoscience and Remote Sensing, vol. 63, pp. 1–15, 2025

    2025

  37. [38]

    Feature disentanglement based domain adap- tation network for cross-scene coastal wetland hyper- spectral image classification,

    Z. Xin, Z. Li, M. Xu, L. Wang, G. Ren, J. Wang, and Y . Hu, “Feature disentanglement based domain adap- tation network for cross-scene coastal wetland hyper- spectral image classification,”International Journal of Applied Earth Observation and Geoinformation, vol. 129, p. 103850, 2024

    2024

  38. [39]

    Two-branch attention adversarial domain adap- tation network for hyperspectral image classification,

    Y . Huang, J. Peng, W. Sun, N. Chen, Q. Du, Y . Ning, and H. Su, “Two-branch attention adversarial domain adap- tation network for hyperspectral image classification,” IEEE Transactions on Geoscience and Remote Sensing, vol. 60, pp. 1–13, 2022

    2022

  39. [40]

    Kullback–leibler divergence metric learning,

    S. Ji, Z. Zhang, S. Ying, L. Wang, X. Zhao, and Y . Gao, “Kullback–leibler divergence metric learning,” IEEE Transactions on Cybernetics, vol. 52, no. 4, pp. 2047–2058, 2020

    2047

  40. [41]

    Integrating structured biological data by kernel maximum mean discrepancy,

    K. Borgwardt, A. Gretton, M. Rasch, P. Kr ¨oger, B. Sch ¨olkopf, and A. Smola, “Integrating structured biological data by kernel maximum mean discrepancy,” Bioinformatics (Oxford, England), vol. 22, pp. e49–57, 08 2006

    2006

  41. [42]

    Deep coral: Correlation align- ment for deep domain adaptation,

    B. Sun and K. Saenko, “Deep coral: Correlation align- ment for deep domain adaptation,” inEuropean Confer- ence on Computer Vision. Springer, 2016, pp. 443–450

    2016

  42. [43]

    Adversarial network with multiple classifiers for open set domain adaptation,

    T. Shermin, G. Lu, S. W. Teng, M. Murshed, and F. Sohel, “Adversarial network with multiple classifiers for open set domain adaptation,”IEEE Transactions on Multimedia, vol. 23, pp. 2732–2744, 2021

    2021

  43. [44]

    Unknown-aware domain adversarial learning for open-set domain adaptation,

    J. Jang, B. Na, D. H. Shin, M. Ji, K. Song, and I.- C. Moon, “Unknown-aware domain adversarial learning for open-set domain adaptation,”Advances in Neural Information Processing Systems, vol. 35, pp. 16 755– 16 767, 2022. Yiwen Liureceived the B.E. degree in Automation and the M.E. degree in Control Science and Engi- neering, Taiyuan University of Technol...

    2022