Pith. sign in

REVIEW 3 major objections 4 minor 1 cited by

SpectralX: Parameter-efficient Domain Generalization for Spectral Remote Sensing Foundation Models

T0 review · 3 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read SpectralX shows optical remote sensing models can master spectral imagery with 11.8M trainable parameters.

desk verdict The architecture story is coherent and worth reading, but the headline numbers do not survive contact with Table V, where DSTC's class-wise IoUs average 59.5 yet the table reports 54.5. read the letter →

arxiv 2508.01731 v1 pith:6B4TTZNW submitted 2025-08-03 cs.CV

Yuxiang Zhang , Wei Li , Mengmeng Zhang , Jiawei Han , Ran Tao , Shunlin Liang This is my paper · ORCID
classification cs.CV
keywords domaingeneralizationparameter-efficientfine-tuningspectralremotesensingmultispectralimagehyperspectralfoundationmodelssemanticsegmentationmixtureofadapters
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper sets out to show that remote sensing foundation models pretrained on ordinary optical RGB imagery can be adapted to multispectral and hyperspectral images without training a spectral foundation model from scratch. It proposes SpectralX, a parameter-efficient fine-tuning framework that keeps the optical backbone frozen and updates only 11.8 million parameters through a two-stage procedure. The first stage learns spectral representations by masked reconstruction; the second stage adds a lightweight adapter and trains for semantic segmentation on limited labeled data. If the claims hold, organizations could reuse existing optical foundation models for spectral Earth observation instead of paying for large-scale spectral pretraining, and those adapted models would generalize to unseen regions and seasons better than current spectral foundation models.

What carries the argument

Three components carry the argument. HyperT (Hyper Tokenizer) replaces the patch-embedding layer: a lightweight CNN downsamples the spectral image, local and global self-attention separate spatial and spectral attributes, spatial-grid and wavelength position embeddings are added, and cross-attention with channel-split queries produces 196 attribute tokens. AoMoA (Attribute-oriented Mixture of Adapter) is inserted into the FFN of selected transformer blocks (the 6th, 12th, 18th, and 24th in a ViT-Large backbone); it uses attribute-specific routers with noisy top-2 selection to combine shared adapter experts, then rescales the output. Are-adapter (Attribute-refined Adapter) runs only in the task stage, generating match maps by querying low-level semantic features with high-level attribute tokens and refining the selected tokens. The two training stages use a masked-reconstruction loss and a cross-entropy segmentation loss, respectively.

What would settle it

Re-run all baselines with the exact protocol used for SpectralX, including the same UperNet head, epochs, learning rate, batch size, augmentation, and repeated seeds, and check whether SpectralX still holds the best mIoU in all eight domain-generalization tasks; if the gaps shrink to noise, the central claim collapses.

Watch

Extended reading notes

Core claim

The central claim is that SpectralX, with 11.8 million trainable parameters, adapts remote sensing foundation models designed for optical modalities to spectral modalities and uses limited labeled data to improve domain generalization. On eight domain-generalization tasks built from WHUOHS, DFC2020, and MTS12, the paper reports that SpectralX achieves the best mIoU in every setting, including regional gaps in hyperspectral data and seasonal gaps in multispectral data. The improvement is attributed to an explicit decomposition of spatial and spectral attributes: a Hyper Tokenizer builds attribute tokens, an Attribute-oriented Mixture of Adapter routes tokens through attribute-specific experts, and an Attribute-refined Adapter lets high-level queries refine low-level semantic features during task training.

Load-bearing premise

The reported margins assume every baseline was fine-tuned under the same segmentation head, optimization schedule, data augmentation, and number of runs as SpectralX, but the paper does not report those training protocols for the baselines.

Editorial extensions

If this is right

  • Optical remote sensing foundation models can be adapted to multispectral and hyperspectral imagery without spectral pretraining, making spectral interpretation substantially cheaper.
  • With 11.8 million trainable parameters, roughly 4% of the 303.4 million used in full fine-tuning, the adaptation is feasible with modest compute and small labeled sets.
  • The reported gains on unseen regions and seasons mean a model trained on one site or season can be deployed to another without retraining on target data.
  • The adapter design transfers across at least two different optical backbones, Scale-MAE and SatMAE++, so it is not tied to a single foundation model.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The paper does not test other multi-dimensional inputs, so applying the same spatial-spectral attribute decomposition to multi-temporal stacks, SAR polarimetry, or fused optical-radar data is an untested extension that the design naturally suggests.
  • At the dataset scales used here, the results hint that a well-designed parameter-efficient adapter on an optical backbone can rival or beat a spectral foundation model pretrained on hundreds of thousands of hyperspectral images; whether that advantage survives when the spectral foundation model is trained on much larger data is not established.
  • Because all eight tasks are regional or seasonal splits of the same three datasets, an immediate next test is whether the gains persist across sensor differences, atmospheric conditions, or continents outside these benchmarks.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 4 minor

Summary. The paper proposes SpectralX, a parameter-efficient fine-tuning framework that adapts optical RGB-pretrained Remote Sensing Foundation Models (Scale-MAE, SatMAE++) to multispectral and hyperspectral inputs. The method introduces a Hyper Tokenizer (HyperT) to generate spatial-spectral attribute tokens, an Attribute-oriented Mixture of Adapter (AoMoA) for layer-wise fine-tuning with attribute-specific routing, and an Attribute-refined Adapter (Are-adapter) for task-oriented refinement. A two-stage training scheme is used: masked reconstruction for modality adaptation, followed by semantic segmentation with a UperNet head. Experiments are reported on WHUOHS, DFC2020, and MTS12 under both no-domain-gap and domain-gap settings, and the paper claims state-of-the-art mIoU in eight domain-generalization tasks while fine-tuning only 11.8M parameters.

Significance. If the empirical results were reliable, SpectralX would be a practically valuable contribution: it would show that expensive spectral pretraining can be bypassed by adapting optical RSFMs to spectral modalities with a small number of trainable parameters, while improving cross-domain segmentation. The method is modular and the ablation study gives some credit to each proposed component. However, the central quantitative evidence is undermined by a major arithmetic inconsistency in the main comparison table, by class-wise results that are implausibly close to a baseline with a different backbone, and by the complete absence of baseline training protocols and error bars. These issues prevent the paper from currently supporting its state-of-the-art claims.

major comments (3)
  1. [Table V] In the WHUOHS no-domain-gap comparison, DSTC's 24 class-wise IoUs (71.5, 71.3, 63.7, 37.9, ..., 97.1) sum to 1428.6, giving a mean of 59.5, yet the table reports DSTC's mIoU as 54.5. SpectralX*ScaleMAE's class-wise values sum to 1432.0, consistent with its reported 59.7. Thus, on the numbers actually printed, SpectralX's advantage over DSTC is +0.2 mIoU, not the +5.2 implied by the reported mIoU values. Since this table is a central piece of evidence for the paper's modality-adaptation and state-of-the-art claims, the table must be corrected and the conclusions re-examined before the empirical claims can be accepted.
  2. [Table V] The class-wise IoU pattern between SpectralX*ScaleMAE and DSTC is also problematic: classes C1, C2, C3, C4, C7, C14, C16, and C20 are identical (71.5, 71.3, 63.7, 37.9, 56.3, 57.9, 47.2, 63.0), and most remaining classes differ by only 0.1-0.4 IoU, despite the two methods using different backbones (ViT-L vs. PVTv2-B1) and different training paradigms. Such a near-duplicate pattern is highly implausible for independent methods and suggests a copying or rounding error in table preparation, or a deeper data-integrity problem. The authors need to verify every row in Tables V-X and provide raw per-class results.
  3. [Section IV-C/D] No training protocol is reported for any baseline method or for SpectralX: the paper omits learning rate, batch size, number of epochs, optimizer, data augmentation, input resolution, number of runs, and error bars. All comparisons in Tables V-X are single numbers with no measure of variance. Without this information, the reported mIoU gaps—especially the corrected 0.2-point gap on WHUOHS—cannot be distinguished from noise or from protocol differences, and the claim that SpectralX outperforms state-of-the-art methods is not supported. The authors should report the full protocol and run all methods with multiple seeds.
minor comments (4)
  1. [Eq. (2)] Equation (2) sums over i=0,...,Na, but the adapters are defined as A1,...,ANa; the sum should start at i=1. Shortly after Eq. (2), the symbol C is used without definition in the sentence 'The aggregation features of spectral attributes C are obtained by using similar routing policies...'.
  2. [Eq. (1)] Equation (1) defines Wgate in R^{2 x (r/4) x Na}, yet the text says the first dimension corresponds to the spatial and spectral attributes; the indexing of the routing tensors is not explained and should be clarified.
  3. [Abstract and Section I] The abstract says 'we proposed SpectralX' rather than 'we propose SpectralX'; phrases such as 'more multispectral/hyperspectral data remain lack of the corresponding foundation models' and 'limiting their adaptability on spectral image' are grammatically awkward and should be rewritten.
  4. [Section III-A] The paper describes SpectralX as having three stages, but stage 3 is inference on unseen scenes rather than a training stage; the terminology conflates training phases with deployment and should be clarified.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found: SpectralX is an empirically evaluated PEFT architecture; the only notable issue is an arithmetic inconsistency in Table V, which is a correctness concern rather than circular reasoning.

full rationale

SpectralX's central claim is that a parameter-efficient two-stage adapter framework improves domain generalization of optical RSFMs on spectral benchmarks. The derivation chain is architectural: HyperT (Section III-B) replaces the patch projection with cross-attention tokenization, AoMoA (Section III-C) is a MoE-style adapter inserted into FFN layers, and the Are-adapter (Section III-D) refines tokens via cross-attention match maps. These components are defined by their own equations (Eqs. 1-6) and are trained with standard MAE reconstruction and cross-entropy losses; no component is defined in terms of the target mIoU, and no fitted parameter is renamed as a prediction. The reported superiority rests on comparisons against external baselines on WHUOHS, DFC2020, and MTS12, not on a self-citation chain or on a quantity that is enforced by construction. The ablation study (Table IV) tunes components on the same benchmarks, but that is ordinary empirical design, not circularity, because the final evaluation is on held-out target domains (Tables VIII-X). Two non-circular concerns belong in correctness review: (1) Table V's DSTC class-wise IoUs average about 59.5, not the reported 54.5, which would shrink SpectralX's w/o-domain-gap margin on WHUOHS; and (2) the paper gives no training protocol for baselines, so reported gaps could partly reflect protocol differences. Neither issue makes the method's derivation equivalent to its inputs, so the circularity score is 0.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The central claim rests on the transferability of pretrained optical RSFMs, the effectiveness of masked reconstruction for modality adaptation, and the fairness of the experimental protocol. None of these is externally verified; the protocol fairness is particularly fragile given missing baseline training details.

free parameters (4)
  • Number of attribute tokens L = 196
    Set equal to the number of image patches for 224x224 input with patch size 16; design choice, not data-fitted.
  • Number of adapters Na = 4
    Chosen by hand for the adapter bank; no sensitivity analysis is provided.
  • Top-K routing K = 2
    Chosen by hand for the router; no sensitivity analysis is provided.
  • AoMoA insertion layers = 6, 12, 18, 24
    Adapters inserted into these four transformer blocks in encoder and decoder; no ablation on layer choice is reported.
assumptions (4)
  • domain assumption Pretrained optical RSFMs (Scale-MAE, SatMAE++) provide reusable representations that can be adapted to spectral modalities through lightweight fine-tuning.
    This is the premise of the entire approach, stated in Section I and III-A.
  • domain assumption Masked reconstruction training in stage1 improves downstream spectral segmentation performance.
    Supported only by the paper's own ablations (Table IV); no external evidence is given.
  • domain assumption The comparison protocol treats all methods equitably (same segmentation head, training budget, and data splits).
    The paper does not document baseline training protocols; the validity of all reported comparisons depends on this.
  • domain assumption Wavelength-based position embeddings can be interpolated to arbitrary spectral band configurations.
    Assumed in Section III-B for datasets with different bands; no analysis of interpolation error is provided.

how reviews work

0 comments
Cite this review

Pith. "Pith review of SpectralX: Parameter-efficient Domain Generalization for Spectral Remote Sensing Foundation Models." pith.science (2026). https://pith.science/paper/6B4TTZNW

@misc{pith2026250801731,
  author       = {Pith},
  title        = {Pith review of: SpectralX: Parameter-efficient Domain Generalization for Spectral Remote Sensing Foundation Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6B4TTZNW}},
  note         = {Machine review of arXiv:2508.01731}
}
read the original abstract

Recent advances in Remote Sensing Foundation Models (RSFMs) have led to significant breakthroughs in the field. While many RSFMs have been pretrained with massive optical imagery, more multispectral/hyperspectral data remain lack of the corresponding foundation models. To leverage the advantages of spectral imagery in earth observation, we explore whether existing RSFMs can be effectively adapted to process diverse spectral modalities without requiring extensive spectral pretraining. In response to this challenge, we proposed SpectralX, an innovative parameter-efficient fine-tuning framework that adapt existing RSFMs as backbone while introducing a two-stage training approach to handle various spectral inputs, thereby significantly improving domain generalization performance. In the first stage, we employ a masked-reconstruction task and design a specialized Hyper Tokenizer (HyperT) to extract attribute tokens from both spatial and spectral dimensions. Simultaneously, we develop an Attribute-oriented Mixture of Adapter (AoMoA) that dynamically aggregates multi-attribute expert knowledge while performing layer-wise fine-tuning. With semantic segmentation as downstream task in the second stage, we insert an Attribute-refined Adapter (Are-adapter) into the first stage framework. By iteratively querying low-level semantic features with high-level representations, the model learns to focus on task-beneficial attributes, enabling customized adjustment of RSFMs. Following this two-phase adaptation process, SpectralX is capable of interpreting spectral imagery from new regions or seasons. The codes will be available from the website: https://github.com/YuxiangZhang-BIT.

Figures

Figures reproduced from arXiv: 2508.01731 by the authors.

Figure 1
Figure 1. Limitations of existing RSFMs and SpectralFMs. RGB, MSIs, and HSIs constitute diverse sources of remote sensing information. However, RSFMs [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. The framework of SpectralX consists of spectral modality adaptation (stage1), task-oriented generalization training (stage2) and unseen scenes [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. The flowchart of HyperT. A lightweight CNN is employed to downsample the spectral image. The downsampled features are then fed into the [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (4 more)
Figure 5
Figure 5. Figure 5: The flowchart of Are-adapter. The attribute token representations [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]
Figure 6
Figure 6. Figure 6: Visualization of predictive segmentation maps without and with domain gap in WHUOHS. [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]
Figure 7
Figure 7. Figure 7: Visualization of predictive segmentation maps without and with domain gap in DFC2020. [PITH_FULL_IMAGE:figures/full_fig_p008_7.png]
Figure 8
Figure 8. Figure 8: Visualization of feature maps extracted by HyperSIGMA and SpectralX on DFC2020. [PITH_FULL_IMAGE:figures/full_fig_p011_8.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. CrossEarth-Gate: Fisher-Guided Adaptive Tuning Engine for Efficient Adaptation of Cross-Domain Remote Sensing Semantic Segmentation

    cs.CV 2025-11 conditional novelty 6.0 of 10

    A Fisher-information-guided dynamic selection over a toolbox of LoRA, adapter, and frequency-adapter modules improves cross-domain remote sensing segmentation over static PEFT methods.

Reference graph

Works this paper leans on

40 extracted references · 28 canonical work pages · cited by 1 Pith paper

  1. [1]

    Deep learning enables satellite-based monitoring of large populations of terrestrial mammals across heterogeneous landscape,

    Z. Wu, C. Zhang, X. Gu, I. Duporge, L. F. Hughey, J. A. Stabach, A. K. Skidmore, J. G. C. Hopcraft, S. J. Lee, P. M. Atkinson, et al. , “Deep learning enables satellite-based monitoring of large populations of terrestrial mammals across heterogeneous landscape,” Nature com- munications, vol. 14, no. 1, p. 3072, 2023

    work page 2023

  2. [2]

    A human-machine collaborative approach measures economic development using satellite imagery,

    D. Ahn, J. Yang, M. Cha, H. Yang, J. Kim, S. Park, S. Han, E. Lee, S. Lee, and S. Park, “A human-machine collaborative approach measures economic development using satellite imagery,” Nature Communica- tions, vol. 14, no. 1, p. 6811, 2023

    work page 2023

  3. [3]

    War city profiles drawn from satellite images,

    Z. Hou, Y . Qu, L. Zhang, J. Liu, F. Wang, Q. Yu, A. Zeng, Z. Chen, Y . Zhao, H. Tang,et al., “War city profiles drawn from satellite images,” Nature Cities, pp. 1–11, 2024. 12

    work page 2024

  4. [4]

    Landsat-8: Science and product vision for terrestrial global change research,

    D. P. Roy, M. A. Wulder, T. R. Loveland, C. E. Woodcock, R. G. Allen, M. C. Anderson, D. Helder, J. R. Irons, D. M. Johnson, R. Kennedy, et al. , “Landsat-8: Science and product vision for terrestrial global change research,” Remote sensing of Environment , vol. 145, pp. 154– 172, 2014

    work page 2014

  5. [5]

    The enmap spaceborne imaging spectroscopy mission for earth observation,

    L. Guanter, H. Kaufmann, K. Segl, S. Foerster, C. Rogass, S. Chabrillat, T. Kuester, A. Hollstein, G. Rossner, C. Chlebek, et al. , “The enmap spaceborne imaging spectroscopy mission for earth observation,”Remote Sensing, vol. 7, no. 7, pp. 8830–8857, 2015

    work page 2015

  6. [6]

    Whu-ohs: A benchmark dataset for large- scale hersepctral image classification,

    J. Li, X. Huang, and L. Tu, “Whu-ohs: A benchmark dataset for large- scale hersepctral image classification,” International Journal of Applied Earth Observation and Geoinformation , vol. 113, p. 103022, 2022

    work page 2022

  7. [7]

    A simple framework for contrastive learning of visual representations,

    T. Chen, S. Kornblith, M. Norouzi, and G. Hinton, “A simple framework for contrastive learning of visual representations,” in International conference on machine learning . PMLR, 2020, pp. 1597–1607

    2020

  8. [8]

    Masked au- toencoders are scalable vision learners,

    K. He, X. Chen, S. Xie, Y . Li, P. Doll ´ar, and R. Girshick, “Masked au- toencoders are scalable vision learners,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition , 2022, pp. 16 000–16 009

    2022

Show all 40 references
  1. [9]

    Momentum contrast for unsupervised visual representation learning,

    K. He, H. Fan, Y . Wu, S. Xie, and R. Girshick, “Momentum contrast for unsupervised visual representation learning,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2020, pp. 9729–9738

    2020

  2. [10]

    Learning transferable visual models from natural language supervision,

    A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal, G. Sastry, A. Askell, P. Mishkin, J. Clark, et al., “Learning transferable visual models from natural language supervision,” in International conference on machine learning . PMLR, 2021, pp. 8748–8763

    2021

  3. [11]

    Segment anything,

    A. Kirillov, E. Mintun, N. Ravi, H. Mao, C. Rolland, L. Gustafson, T. Xiao, S. Whitehead, A. C. Berg, W.-Y . Lo,et al., “Segment anything,” in Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023, pp. 4015–4026

    2023

  4. [12]

    Emerging properties in self-supervised vision transformers,

    M. Caron, H. Touvron, I. Misra, H. J ´egou, J. Mairal, P. Bojanowski, and A. Joulin, “Emerging properties in self-supervised vision transformers,” in Proceedings of the IEEE/CVF international conference on computer vision, 2021, pp. 9650–9660

    2021

  5. [13]

    Towards geospatial foundation models via continual pretraining,

    M. Mendieta, B. Han, X. Shi, Y . Zhu, and C. Chen, “Towards geospatial foundation models via continual pretraining,” in Proceedings of the IEEE/CVF International Conference on Computer Vision , 2023, pp. 16 806–16 816

    2023

  6. [14]

    Scale-mae: A scale-aware masked autoencoder for multiscale geospatial representation learning,

    C. J. Reed, R. Gupta, S. Li, S. Brockman, C. Funk, B. Clipp, K. Keutzer, S. Candido, M. Uyttendaele, and T. Darrell, “Scale-mae: A scale-aware masked autoencoder for multiscale geospatial representation learning,” in Proceedings of the IEEE/CVF International Conference on Comp...

    2023

  7. [15]

    Cross-scale mae: A tale of multiscale exploitation in remote sensing,

    M. Tang, A. Cozma, K. Georgiou, and H. Qi, “Cross-scale mae: A tale of multiscale exploitation in remote sensing,” Advances in Neural Information Processing Systems , vol. 36, 2024

    2024

  8. [16]

    Satmae: Pre-training transformers for temporal and multi-spectral satellite imagery,

    Y . Cong, S. Khanna, C. Meng, P. Liu, E. Rozi, Y . He, M. Burke, D. Lo- bell, and S. Ermon, “Satmae: Pre-training transformers for temporal and multi-spectral satellite imagery,” Advances in Neural Information Processing Systems, vol. 35, pp. 197–211, 2022

    2022

  9. [17]

    Rethinking transformers pre-training for multi-spectral satellite imagery,

    M. Noman, M. Naseer, H. Cholakkal, R. M. Anwer, S. Khan, and F. S. Khan, “Rethinking transformers pre-training for multi-spectral satellite imagery,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2024, pp. 27 811–27 819

    2024

  10. [18]

    Spectralgpt: Spectral remote sensing foun- dation model,

    D. Hong, B. Zhang, X. Li, Y . Li, C. Li, J. Yao, N. Yokoya, H. Li, P. Ghamisi, X. Jia, et al. , “Spectralgpt: Spectral remote sensing foun- dation model,” IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024

    2024

  11. [19]

    Hypersigma: Hyperspectral intelligence comprehension foundation model,

    D. Wang, M. Hu, Y . Jin, Y . Miao, J. Yang, Y . Xu, X. Qin, J. Ma, L. Sun, C. Li, et al. , “Hypersigma: Hyperspectral intelligence comprehension foundation model,” arXiv preprint arXiv:2406.11519 , 2024

    2024 arXiv

  12. [20]

    LoRA: Low-rank adaptation of large language models,

    E. J. Hu, Y . Shen, P. Wallis, Z. Allen-Zhu, Y . Li, S. Wang, L. Wang, and W. Chen, “LoRA: Low-rank adaptation of large language models,” in International Conference on Learning Representations , 2022

    2022

  13. [21]

    Parameter efficient fine-tuning via cross block orchestration for segment anything model,

    Z. Peng, Z. Xu, Z. Zeng, L. Xie, Q. Tian, and W. Shen, “Parameter efficient fine-tuning via cross block orchestration for segment anything model,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2024, pp. 3743–3752

    2024

  14. [22]

    Airs: Adapter in remote sensing for parameter-efficient transfer learning,

    L. Hu, H. Yu, W. Lu, D. Yin, X. Sun, and K. Fu, “Airs: Adapter in remote sensing for parameter-efficient transfer learning,” IEEE Transactions on Geoscience and Remote Sensing , vol. 62, pp. 1–18, 2024

    2024

  15. [23]

    Global land-cover mapping with weak supervision: Outcome of the 2020 ieee grss data fusion contest,

    C. Robinson, K. Malkin, N. Jojic, H. Chen, R. Qin, C. Xiao, M. Schmitt, P. Ghamisi, R. H¨ansch, and N. Yokoya, “Global land-cover mapping with weak supervision: Outcome of the 2020 ieee grss data fusion contest,” IEEE Journal of Selected Topics in Applied Earth Observations an...

    2020

  16. [24]

    Cnn, rnn, or vit? an evaluation of different deep learning architectures for spatio-temporal representation of sentinel time series,

    L. Zhao and S. Ji, “Cnn, rnn, or vit? an evaluation of different deep learning architectures for spatio-temporal representation of sentinel time series,” IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing , vol. 16, pp. 44–56, 2022

    2022

  17. [25]

    Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks,

    Z. Chen, J. Wu, W. Wang, W. Su, G. Chen, S. Xing, M. Zhong, Q. Zhang, X. Zhu, L. Lu, et al. , “Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition ...

    2024

  18. [26]

    Neural plasticity-inspired multimodal foundation model for earth observation,

    Z. Xiong, Y . Wang, F. Zhang, A. J. Stewart, J. Hanna, D. Borth, I. Papoutsis, B. L. Saux, G. Camps-Valls, and X. X. Zhu, “Neural plasticity-inspired multimodal foundation model for earth observation,” arXiv preprint arXiv:2403.15356 , 2024

    2024

  19. [27]

    Geo- bench: Toward foundation models for earth monitoring,

    A. Lacoste, N. Lehmann, P. Rodriguez, E. Sherwin, H. Kerner, B. L¨utjens, J. Irvin, D. Dao, H. Alemohammad, A. Drouin, et al., “Geo- bench: Toward foundation models for earth monitoring,” Advances in Neural Information Processing Systems , vol. 36, 2024

    2024

  20. [28]

    Hyperfree: A channel-adaptive and tuning-free foundation model for hyperspectral remote sensing imagery,

    L. Jingtao, L. yingyi, W. Xinyu, P. Yunning, S. Chen, W. Shaoyu, S. Zhendong, K. Tian, J. Xiao, L. Tangwei, Z. anran, and Z. Yanfei, “Hyperfree: A channel-adaptive and tuning-free foundation model for hyperspectral remote sensing imagery,” in Proceedings of the IEEE/CVF Confer...

    2025

  21. [29]

    Parameter-efficient transfer learning for nlp,

    N. Houlsby, A. Giurgiu, S. Jastrzebski, B. Morrone, Q. De Laroussilhe, A. Gesmundo, M. Attariyan, and S. Gelly, “Parameter-efficient transfer learning for nlp,” in International conference on machine learning . PMLR, 2019, pp. 2790–2799

    2019

  22. [30]

    Visual prompt tuning,

    M. Jia, L. Tang, B.-C. Chen, C. Cardie, S. Belongie, B. Hariharan, and S.-N. Lim, “Visual prompt tuning,” in European Conference on Computer Vision. Springer, 2022, pp. 709–727

    2022

  23. [31]

    Dtl: Disentangled transfer learning for visual recognition,

    M. Fu, K. Zhu, and J. Wu, “Dtl: Disentangled transfer learning for visual recognition,” in Proceedings of the AAAI Conference on Artificial Intelligence, vol. 38, no. 11, 2024, pp. 12 082–12 090

    2024

  24. [32]

    Parameter-efficient fine-tuning for large models: A comprehensive survey,

    Z. Han, C. Gao, J. Liu, J. Zhang, and S. Q. Zhang, “Parameter-efficient fine-tuning for large models: A comprehensive survey,” arXiv preprint arXiv:2403.14608, 2024

    2024 arXiv

  25. [33]

    Enhancing few-shot out-of-distribution detection with pre-trained model features,

    J. Dong, Y . Yao, W. Jin, H. Zhou, Y . Gao, and Z. Fang, “Enhancing few-shot out-of-distribution detection with pre-trained model features,” IEEE Transactions on Image Processing, vol. 33, pp. 6309–6323, 2024

    2024

  26. [34]

    Hada: Hyper-adaptive parameter-efficient learning for multi-view convnets,

    S. Wang, C. Li, Z. Yan, W. Liang, Y . Yuan, and G. Wang, “Hada: Hyper-adaptive parameter-efficient learning for multi-view convnets,” IEEE Transactions on Image Processing , vol. 34, pp. 85–99, 2025

    2025

  27. [35]

    Parameter efficient self-supervised geospatial domain adaptation,

    L. Scheibenreif, M. Mommert, and D. Borth, “Parameter efficient self-supervised geospatial domain adaptation,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2024, pp. 27 841–27 851

    2024

  28. [36]

    Alps: An auto-labeling and pre-training scheme for remote sensing segmentation with segment anything model,

    S. Zhang, Q. Wang, J. Liu, and H. Xiong, “Alps: An auto-labeling and pre-training scheme for remote sensing segmentation with segment anything model,” IEEE Transactions on Image Processing , pp. 1–1, 2025

    2025

  29. [37]

    Unified perceptual parsing for scene understanding,

    T. Xiao, Y . Liu, B. Zhou, Y . Jiang, and J. Sun, “Unified perceptual parsing for scene understanding,” in Proceedings of the European conference on computer vision (ECCV) , 2018, pp. 418–434

    2018

  30. [38]

    Spatial structure constraints for weakly supervised semantic segmentation,

    T. Chen, Y . Yao, X. Huang, Z. Li, L. Nie, and J. Tang, “Spatial structure constraints for weakly supervised semantic segmentation,” IEEE Transactions on Image Processing, vol. 33, pp. 1136–1148, 2024

    2024

  31. [39]

    Energy- based domain adaptation without intermediate domain dataset for foggy scene segmentation,

    D. Jang, S. Lee, G. Choi, Y . Lee, S. Son, and D.-S. Kim, “Energy- based domain adaptation without intermediate domain dataset for foggy scene segmentation,” IEEE Transactions on Image Processing , vol. 33, pp. 6143–6157, 2024

    2024

  32. [40]

    Dual-stage hyperspectral image classification model with spectral supertoken,

    P. Liu, T. Xu, J. Wang, H. Chen, H. Bai, and J. Li, “Dual-stage hyperspectral image classification model with spectral supertoken,” in European Conference on Computer Vision . Springer, 2024, pp. 368– 386. 13

    2024

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.