Pith. sign in

REVIEW 3 major objections 6 minor 41 references

Parameter-Efficient Fine-Tuning of Multispectral Foundation Models for Hyperspectral Image Classification

T0 review · 3 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read A 0.056%-parameter KronA+ adapter brings a multispectral foundation model within a point or two of full fine-tuning on hyperspectral data.

desk verdict KronA+ is a simple, honest extension of LoRA+/KronA with a useful five-dataset HSI comparison, but test-set hyperparameter selection and single runs leave the headline advantage over KronA unverified. read the letter →

arxiv 2505.15334 v1 pith:V6DLAZ7X submitted 2025-05-21 cs.CV

Bernardin Ligan , Khalide Jbilou , Fahd Kalloubi , Ahmed Ratnani This is my paper
classification cs.CV
keywords remotesensinghyperspectralimageclassificationfoundationmodelfine-tuningparameter-efficientKroneckeradapterKronA+LoRA+SpectralGPT
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 authors set out to show that a hyperspectral classifier can be built by fine-tuning a foundation model pretrained on ordinary multispectral imagery, rather than by training a hyperspectral-specific model from scratch. They compare several parameter-efficient fine-tuning strategies on SpectralGPT across five hyperspectral datasets and introduce KronA+, which assigns different learning rates to the two matrices in a Kronecker-product adapter. Their experiments indicate that KronA+ reaches overall accuracies close to full fine-tuning, for instance 82.76% versus 85.62% on Indian Pines and 98.55% versus 98.97% on Botswana, while updating only about 0.056% of the model's parameters and adding roughly 0.2 MB of storage. The paper also claims that fully fine-tuning SpectralGPT for 50 epochs is competitive with, and sometimes better than, dedicated hyperspectral backbones trained for up to 600 epochs. If true, this makes it much cheaper to apply a general remote-sensing foundation model to hyperspectral classification tasks.

What carries the argument

The central object is KronA+, a Kronecker-product adapter in which the weight update is $\Delta W = s(A \otimes B)$, where $\otimes$ denotes the Kronecker product, a way of forming a large block matrix from two smaller ones. The two factor matrices are trained with learning rates $\eta$ and $\lambda\eta$, with $\lambda > 1$, following the same asymmetric-learning-rate idea that separates LoRA+ from LoRA. This adapter is applied to the query and value projection weights of every attention layer in a frozen SpectralGPT encoder, while the input pipeline reduces hyperspectral cubes to 12 principal-component bands, resizes patches to 32x32, and tokenizes them into 8x8x3 blocks matching SpectralGPT's pretraining format.

What would settle it

Run KronA and KronA+ on the same five datasets with all hyperparameters, including patch size and $\lambda$, chosen on a held-out validation split rather than the test split, and average over several seeds; if KronA+ no longer exceeds KronA, or the margin falls inside seed-to-seed noise, the paper's central improvement claim is not supported.

Watch

Extended reading notes

Core claim

The central claim is that KronA+, a Kronecker-product adapter with update $\Delta W = s(A \otimes B)$ and asymmetric learning rates $\eta$ and $\lambda\eta$, is the most effective parameter-efficient fine-tuning method tested for adapting SpectralGPT to hyperspectral image classification. It outperforms LoRA, LoRA+, LoKr, BitFit, and KronA on all five datasets, reaching overall accuracies within about 0.4 to 3 points of full fine-tuning while using only about 0.05 million trainable parameters, roughly 0.056% of the 85.3 million parameters updated by full fine-tuning, and adding about 0.2 MB of storage. The paper also claims that full fine-tuning of SpectralGPT, a multispectral foundation model, is competitive with dedicated hyperspectral architectures and beats the hyperspectral-specific foundation model HyperSIGMA on Pavia, Houston, and Botswana while using only a quarter of the training epochs.

Load-bearing premise

The load-bearing premise is that single-run accuracies on a fixed test split, with patch sizes (Section 4.3) and $\lambda$ values (Section 5.6.3) selected by looking at that split's accuracy, are unbiased estimates of generalization; if those choices do not transfer to new data, the reported KronA+ gains could shrink or disappear.

Editorial extensions

If this is right

  • A general multispectral foundation model can be repurposed for hyperspectral classification with a tiny adapter, so hyperspectral-specific pretraining is not a strict requirement for strong accuracy.
  • KronA+ closes most of the gap to full fine-tuning, within about 3 points of overall accuracy on Indian Pines and about 0.4 points on Botswana, while storing roughly 0.2 MB instead of the full checkpoint.
  • Because KronA+ uses only about 0.05 million trainable parameters, fine-tuning can run on a single consumer GPU and many dataset-specific adapters can be kept for the same frozen backbone.
  • Full fine-tuning of SpectralGPT beats the dedicated hyperspectral foundation model on Pavia, Houston, and Botswana with one quarter of the training epochs, suggesting that epoch count and pretraining data breadth matter as much as spectral specificity.
  • KronA+ outperforms the state-of-the-art hyperspectral backbones used in this study on the Pavia and Houston datasets, which would make it a practical choice when labeled hyperspectral data are scarce.

Reading between the lines

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

  • The asymmetric-learning-rate trick that separates KronA+ from KronA would likely also boost LoKr and other structured PEFT adapters, because it is the same mechanism that separates LoRA+ from LoRA; the paper does not test this.
  • With a roughly 0.2 MB adapter, one frozen SpectralGPT backbone could serve many hyperspectral sites or classes by storing one tiny adapter per task, a deployment pattern the paper's storage numbers make plausible but do not demonstrate.
  • The near-unity $\lambda$ values on most hyperspectral datasets, compared with $\lambda \approx 24$ in the text-domain LoRA+ results, suggest that spectral classification needs only a mild learning-rate imbalance; this is an editorial reading, not a claim the paper makes.
  • If the full-fine-tuning-versus-HyperSIGMA result survives across more sensors, it would shift the hyperspectral foundation-model debate from 'pretrain on hyperspectral data' to 'pretrain broadly, adapt cheaply'; this is an inference from the paper's cost-performance discussion.
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 / 6 minor

Summary. The paper proposes an efficient fine-tuning framework that adapts the multispectral foundation model SpectralGPT to hyperspectral image classification. It compares several PEFT methods (LoRA, LoRA+, KronA, LoKr) on five HSI datasets and introduces KronA+, a variant that applies the LoRA+ idea of different learning rates to the two Kronecker factors. The main empirical claims are that KronA+ is the best PEFT method, reaching overall accuracies close to full fine-tuning while updating only about 0.056% of the parameters (~0.05M parameters, ~0.2 MB storage), and that full fine-tuning of SpectralGPT is competitive with dedicated hyperspectral backbones, even outperforming HyperSIGMA on three datasets while using fewer epochs. The paper also includes ablation studies on token size, embedding dimension, and the lambda scaling factor.

Significance. If the empirical claims survive stricter evaluation, the paper would be a useful contribution: it provides a systematic comparison of PEFT methods for adapting a multispectral foundation model to hyperspectral data, and it demonstrates that a simple extension of KronA is parameter-efficient and competitive. The explicit reporting of trainable parameter counts and adapter storage is a strength. However, the central claim that KronA+ is the best PEFT method currently rests on a single-run protocol in which key hyperparameters were selected using test-set accuracies, and on dataset splits whose totals do not match the standard benchmarks. These issues make the claim conditional on an evaluation protocol that has not been shown to generalize to unseen data.

major comments (3)
  1. [Section 4.3 and Section 5.6.3] Hyperparameters are selected using the same test accuracies that are later reported as results. Section 4.3 states that patch sizes were chosen as those "leading to the best performance," and Section 5.6.3 plus Table 14 select lambda values from test-OA curves (Figures 11-14). This is a selection-on-the-test-set procedure, so the reported differences between methods (e.g., Table 9, KronA+ 98.55% vs KronA 98.12% on Botswana) are not unbiased estimates of generalization. Please re-run the evaluation with a validation split for hyperparameter selection (or nested cross-validation), report results on a held-out test set, and show that the KronA+ advantage persists.
  2. [Tables 1 and 2] The dataset split tables do not match the standard benchmark totals. In Table 1, the Indian Pines training and test pixels sum to 10,366, whereas the standard Indian Pines benchmark contains 10,249 labeled pixels. In Table 2, Pavia University sums to 43,923, whereas the standard labeled set contains 42,776. Consequently, the class-wise testing distributions differ from those used in the cited state-of-the-art papers, so the comparisons in Tables 6-10 against published SpectralFormer, MambaHSI, and HyperSIGMA numbers are not directly aligned. Please correct the splits or justify the discrepancy and re-run the comparisons.
  3. [Tables 6-10] The paper reports a single run per configuration with no error bars, variance, or significance tests. Because several method differences are small (e.g., Table 9: KronA 98.12% vs KronA+ 98.55% vs FFT 98.97%; Table 7: LoRA 88.27% vs LoRA+ 90.18%), the claim that KronA+ is the best PEFT method is not statistically supported. Please provide means and standard deviations over multiple seeds (at least 3-5 runs) or a paired test over the same test split to establish that the observed ordering is not due to random seed variation.
minor comments (6)
  1. [Table 6] The KronA and KronA+ OA values are concatenated as "81.6282.76"; a space or separator is missing.
  2. [Section 5.6.1 and Section 5.6.2] The token-size and embedding-dimension ablations do not specify which fine-tuning method (FFT or a PEFT method) is used; please state this for reproducibility.
  3. [Section 3.2.5] The notation O(L(r1r2 + p/r1 · p/r2)) is ambiguous; please define the dimensions and whether the second term means (p/r1)·(p/r2) or p/(r1·r2).
  4. [Various] There are naming inconsistencies, e.g., "Spectralformer" vs "SpectralFormer", "Hypersigma" vs "HyperSIGMA", and "WHU-Hi-LongKou" vs "Whu-HI Longkou"; please standardize throughout.
  5. [Data and Code Availability] The GitHub URL contains a space ("PEFT HSIC"); please provide a properly encoded URL.
  6. [Figures 11-14] Please clarify whether the OA-versus-lambda curves are averaged over multiple runs or come from a single run, and whether the selected lambda values are stable across seeds.

Circularity Check

2 steps flagged · score 2.0 of 10

No derivation-level circularity; KronA+ is an empirical extension whose reported advantage is mildly circular because lambda and patch sizes are chosen on the test set whose OA is then reported.

  1. fitted input called prediction [Section 5.6.3 (Table 14, Figures 11-14) with reported results in Tables 6-10, especially Table 9]
    "Figures 12 and 14 show OA values with varying λ. For Botswana, LoRA+ achieves an OA of 97.56% at λ = 1.15, outperforming simple LoRA (OA = 97.31%), while KronA+ peaks at λ = 1.5, surpassing simple KronA (OA = 98.12%)."

    The KronA+ learning-rate ratio λ is selected by maximizing the test-set OA curves in Figures 11-14, and the same test-set OA values are then reported in Tables 6-10 as evidence that KronA+ is the best PEFT method. On Botswana, KronA+ is evaluated at its chosen λ=1.5 and achieves 98.55% OA versus 98.12% for KronA; this gap is the result of a test-set search for λ, not an unbiased out-of-sample estimate. The claim that KronA+ surpasses KronA is therefore statistically forced by the selection procedure rather than independently predicted.

  2. fitted input called prediction [Section 4.3 (Implementation Details)]
    "In this work, we evaluated patch sizes of 7 × 7, 9 × 9, 11 × 11, 17 × 17, 23 × 23, and 27 × 27 for each dataset and selected the ones leading to the best performance."

    Patch sizes are selected by best performance on the same evaluation sets whose OA is later reported in Tables 6-10, so the absolute OA numbers are optimistically selected. This is a milder issue than the λ selection because it applies to all methods equally, but it is part of the same pattern of tuning on the test set and then presenting the resulting accuracy as evidence of generalization.

full rationale

The paper does not contain a derivation that reduces to its inputs: KronA+ is defined as KronA with LoRA+-style distinct learning rates (Eq. 12 analog applied to Eq. 5), and no uniqueness, ansatz, or self-citation chain is used to force the result. The central comparison is empirical. The main circularity concern is evaluative rather than derivational: λ (Section 5.6.3) and patch sizes (Section 4.3) are chosen by looking at test-set OA, after which Tables 6-10 report those same test accuracies as evidence that KronA+ is the best PEFT method. This selection-on-test-set bias means the KronA+ vs KronA gap, e.g., +1.5% OA on Botswana, is not an unbiased estimate of generalization; however, it does not make the method definitionally equivalent to its input. The paper also makes no load-bearing self-citations and relies on external baselines. Score 2 reflects the mild selection circularity in the reported headline comparison.

Assumptions & free parameters 7 free parameters · 4 assumptions · 0 invented entities

No new physical entities. All extra degrees of freedom are adapter matrices and hyperparameters, listed under free_parameters. The central claim is empirical; the axioms record assumptions about transfer, PCA preservation, and test-set hyperparameter generalization.

free parameters (7)
  • lambda_KronA+_per_dataset = 1.02 (Indian Pines), 8 (Pavia), 1.05 (Houston), 1.5 (Botswana), 1.02 (Longkou)
    Selected in Section 5.6.3 by maximizing OA on the test set; the Pavia value 8 shows the ratio is dataset-dependent.
  • lambda_LoRA+_per_dataset = 1.025, 1.3, 1.15, 1.15, 1.08
    Used for the LoRA+ comparison and also selected by test OA, per Figures 11-14 and Table 14.
  • patch_size_per_dataset = 15 (Indian Pines), 9 (Pavia), 9 (Houston), 9 (Botswana), 27 (Longkou)
    Section 4.3 says patch size is chosen based on a detailed study of the effect of different patch sizes on the performance of each dataset, i.e., by test performance.
  • KronA_matrix_shape_A = (384,2) for Indian Pines, Botswana and Longkou; (24,32) for Pavia; (16,48) for Houston
    Section 4.3 reports KronA shapes were set 'after tuning' for each dataset; the shape changes the parameter count and expressiveness.
  • LoKR_hyperparameters = f=8, r=8
    User-defined factor and low-rank decomposition for LoKr, set without ablation in Section 4.3.
  • normalization_scheme_per_dataset = BigEarthNet standardization for Indian Pines, Houston, Pavia, Botswana; min-max for Longkou
    Section 3.1 states standardization was tried first, then min-max was adopted when it gave better results, a data-dependent choice.
  • LoRA_rank_and_scale = r=4, alpha=4 (alpha/r=1)
    Standard hyperparameter choice for LoRA, reported in Section 4.3 and not tuned in this paper.
assumptions (4)
  • domain assumption SpectralGPT pretrained on 12-band multispectral data provides a usable representation for 12-band PCA projections of hyperspectral images.
    The whole pipeline assumes the multispectral pretrained encoder transfers to PCA-reduced HSI data; stated in Section 3.1 without independent evidence.
  • domain assumption PCA fit to the full HSI cube preserves the discriminative spectral content needed for classification.
    Band reduction to 12 components is applied before patch extraction and split creation; the paper does not state whether PCA is fit on training pixels only.
  • domain assumption Test-set-optimal hyperparameters generalize to unseen pixels.
    Patch size, lambda, and matrix shapes are tuned on the test split and then the same split is used for reported accuracies; Sections 4.3 and 5.6.3.
  • domain assumption Public benchmark splits with disjoint train and test pixels give a fair estimate of model performance.
    The authors state this in Section 4.1; it is an evaluation convention rather than a proven property.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Parameter-Efficient Fine-Tuning of Multispectral Foundation Models for Hyperspectral Image Classification." pith.science (2026). https://pith.science/paper/V6DLAZ7X

@misc{pith2026250515334,
  author       = {Pith},
  title        = {Pith review of: Parameter-Efficient Fine-Tuning of Multispectral Foundation Models for Hyperspectral Image Classification},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/V6DLAZ7X}},
  note         = {Machine review of arXiv:2505.15334}
}
read the original abstract

Foundation models have achieved great success across diverse domains, including remote sensing (RS), thanks to their versatility and strong generalization abilities. However, most RS foundation models are designed for multispectral data, while hyperspectral imagery (HSI) - with its hundreds of spectral bands - remains less explored. Fine-tuning such models for downstream tasks is also challenging, often demanding considerable memory and storage. In this paper, we propose an efficient framework to fine-tune SpectralGPT, a multispectral foundation model, for hyperspectral image classification (HSIC). We explore several Parameter-Efficient Fine-Tuning (PEFT) methods, including Low-Rank Adaptation (LoRA), Kronecker-based adaptation (KronA), Low-Rank Kronecker (LoKr), and the recent LoRA+, which uses distinct learning rates for low-rank adapters scaled by a factor lambda. Inspired by LoRA+, we introduce KronA+, which applies a similar mechanism to the Kronecker matrices. We evaluate our approach on five datasets from different sensors, showing competitive performance with state-of-the-art HSI models. Our full fine-tuning (FFT) setup for SpectralGPT even outperforms a dedicated hyperspectral foundation model on some datasets while requiring only a quarter of the training epochs. Under the same number of epochs, KronA+ reaches similar performance with far fewer trainable parameters - just 0.056 percent - and adds only approximately 0.2 megabytes of storage, making it the most effective PEFT method tested.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

41 extracted references · 31 canonical work pages

  1. [1]

    Goldberg, and S

    Ben Zaken, E., Y. Goldberg, and S. Ravfogel 2022. B it F it: Simple parameter-efficient fine-tuning for transformer-based masked language-models. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers) , pp.\ 1--9

    work page 2022

  2. [2]

    Albrecht, J

    Braham, N.A.A., C.M. Albrecht, J. Mairal, J. Chanussot, Y. Wang, and X.X. Zhu. 2024. Spectralearth: Training hyperspectral foundation models at scale. arXiv:2408.08447, cs.CV\/

    arXiv 2024

  3. [3]

    Fendley, J

    Christie, G., N. Fendley, J. Wilson, and R. Mukherjee 2018. Functional map of the world. In 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition , pp.\ 6172--6180

    work page 2018

  4. [4]

    Khanna, C

    Cong, Y., S. Khanna, C. Meng, P. Liu, E. Rozi, Y. He, M. Burke, D.B. Lobell, and S. Ermon 2022. Sat MAE : Pre-training transformers for temporal and multi-spectral satellite imagery. In Advances in Neural Information Processing Systems

    work page 2022

  5. [5]

    Chang, K

    Devlin, J., M.W. Chang, K. Lee, and K. Toutanova 2019, June. BERT : Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North A merican Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers) , pp.\ 4171--4186

    work page 2019

  6. [6]

    Tahaei, I

    Edalati, A., M. Tahaei, I. Kobyzev, V.P. Nia, J.J. Clark, and M. Rezagholizadeh. 2022. Krona: Parameter efficient tuning with kronecker adapter. arXiv:2212.10650, cs.CL\/

    arXiv 2022

  7. [7]

    Eastwood, C.M

    Green, R.O., M.L. Eastwood, C.M. Sarture, T.G. Chrien, M. Aronsson, B.J. Chippendale, J.A. Faust, B.E. Pavri, C.J. Chovit, M. Solis, M.R. Olah, and O. Williams. 1998. Imaging spectroscopy and the airborne visible/infrared imaging spectrometer (aviris). Remote Sensing of Environment\/ 65\/ (3): 227--248. doi:https://doi.org/10.1016/S0034-4257(98)00064-9

    work page doi:10.1016/s0034-4257(98)00064-9 1998

  8. [8]

    Goel, and C

    Gu, A., K. Goel, and C. Re 2022. Efficiently modeling long sequences with structured state spaces. In International Conference on Learning Representations

    work page 2022

Show all 41 references
  1. [9]

    Guo, X., J. Lao, B. Dang, Y. Zhang, L. Yu, L. Ru, L. Zhong, Z. Huang, K. Wu, D. Hu, H. He, J. Wang, J. Chen, M. Yang, Y. Zhang, and Y. Li. 2023. Skysense: A multi-modal remote sensing foundation model towards universal interpretation for earth observation imagery. 2024 IEEE/CV...

    2023

  2. [10]

    Henderson, F.P

    Harold V. Henderson, F.P. and S.R. Searle. 1983. On the history of the kronecker product. Linear and Multilinear Algebra\/ 14\/ (2): 113--120. doi:10.1080/03081088308817548

    1983 doi

  3. [11]

    Ghosh, and B

    Hayou, S., N. Ghosh, and B. Yu 2024. Lo RA +: Efficient low rank adaptation of large models. In Forty-first International Conference on Machine Learning

    2024

  4. [12]

    He, K., X. Chen, S. Xie, Y. Li, P. Dollár, and R. Girshick. 2022. Masked autoencoders are scalable vision learners. Conference on Computer Vision and Pattern Recognition (CVPR)\/

    2022

  5. [13]

    He, L., J. Li, C. Liu, and S. Li. 2018. Recent advances on spectral–spatial hyperspectral image classification: An overview and new guidelines. IEEE Transactions on Geoscience and Remote Sensing\/ 56\/ (3): 1579--1597

    2018

  6. [14]

    Hong, D., Z. Han, J. Yao, L. Gao, B. Zhang, A. Plaza, and J. Chanussot. 2022. Spectralformer: Rethinking hyperspectral image classification with transformers. IEEE Transactions on Geoscience and Remote Sensing\/ 60: 1--15. doi:10.1109/TGRS.2021.3130716

    2022

  7. [15]

    Zhang, X

    Hong, D., B. Zhang, X. Li, Y. Li, C. Li, J. Yao, N. Yokoya, H. Li, P. Ghamisi, X. Jia, A. Plaza, P. Gamba, J.A. Benediktsson, and J. Chanussot. 2024. Spectralgpt: Spectral remote sensing foundation model. IEEE Transactions on Pattern Analysis and Machine Intelligence\/ 46\/ (8...

    2024

  8. [16]

    Giurgiu, S

    Houlsby, N., A. Giurgiu, S. Jastrzebski, B. Morrone, Q. De Laroussilhe, A. Gesmundo, M. Attariyan, and S. Gelly 2019, 09--15 Jun. Parameter-efficient transfer learning for NLP . In Proceedings of the 36th International Conference on Machine Learning , Volume 97 of Proceedings ...

    2019

  9. [17]

    Wallis, Z

    Hu, E.J., yelong shen, P. Wallis, Z. Allen-Zhu, Y. Li, S. Wang, L. Wang, and W. Chen 2022. Lo RA : Low-rank adaptation of large language models. In International Conference on Learning Representations

    2022

  10. [18]

    Huang, X. and L. Zhang. 2009, January. A comparative study of spatial approaches for urban mapping using hyperspectral rosis images over pavia city, northern italy. Taylor & Francis, Inc.\/ 30\/ (12): 3205–3221

    2009

  11. [19]

    Roy, C.E

    Jakubik, J., S. Roy, C.E. Phillips, P. Fraccaro, D. Godwin, B. Zadrozny, D. Szwarcman, C. Gomes, G. Nyirjesy, B. Edwards, D. Kimura, N. Simumba, L. Chu, S.K. Mukkavilli, D. Lambhate, K. Das, R. Bangalore, D. Oliveira, M. Muszynski, K. Ankur, M. Ramasubramanian, I. Gurung, S. K...

    2023 arXiv

  12. [20]

    Tang, B.C

    Jia, M., L. Tang, B.C. Chen, C. Cardie, S. Belongie, B. Hariharan, and S.N. Lim 2022. Visual prompt tuning. In Computer Vision – ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23–27, 2022, Proceedings, Part XXXIII , pp.\ 709–727

    2022

  13. [21]

    Mintun, N

    Kirillov, A., E. Mintun, N. Ravi, H. Mao, C. Rolland, L. Gustafson, T. Xiao, S. Whitehead, A.C. Berg, W.Y. Lo, P. Dollár, and R. Girshick 2023. Segment anything. In 2023 IEEE/CVF International Conference on Computer Vision (ICCV) , pp.\ 3992--4003

    2023

  14. [22]

    Leng, J., T. Li, G. Bai, Q. Dong, and H. Dong. 2016. Cube-cnn-svm: A novel hyperspectral image classification method. IEEE 28th International Conference on Tools with Artificial Intelligence (ICTAI)\/ : 1027--1034

    2016

  15. [23]

    Li, Y., Y. Luo, L. Zhang, Z. Wang, and B. Du. 2024. Mambahsi: Spatial–spectral mamba for hyperspectral image classification. IEEE Transactions on Geoscience and Remote Sensing\/ 62: 1--16

    2024

  16. [24]

    Loshchilov, I. and F. Hutter 2019. Decoupled weight decay regularization. In International Conference on Learning Representations

    2019

  17. [25]

    Guo, J.R

    Lu, S., J. Guo, J.R. Zimmer-Dauphinee, J.M. Nieusma, X. Wang, P. vanValkenburgh, S.A. Wernke, and Y. Huo. 2025. Vision foundation models in remote sensing: A survey. IEEE Geoscience and Remote Sensing Magazine\/ : 2--27

    2025

  18. [26]

    Jia, G.L

    Marsocci, V., Y. Jia, G.L. Bellier, D. Kerekes, L. Zeng, S. Hafner, S. Gerard, E. Brune, R. Yadav, A. Shibli, H. Fang, Y. Ban, M. Vergauwen, N. Audebert, and A. Nascetti. 2024. Pangaea: A global and inclusive benchmark for geospatial foundation models. arXiv:2412.04204, cs.CV\/

    2024 arXiv

  19. [27]

    Blilie, X

    Mulliqi, N., A. Blilie, X. Ji, K. Szolnoky, H. Olsson, S.E. Boman, M. Titus, G.M. Gonzalez, J.A. Mielcarz, M. Valkonen, et al. 2025. Foundation models--a panacea for artificial intelligence in pathology? arXiv preprint arXiv:2502.21264\/

    2025

  20. [28]

    Naseer, H

    Noman, M., M. Naseer, H. Cholakkal, R.M. Anwer, S.H. Khan, and F.S. Khan. 2024. Rethinking transformers pre-training for multi-spectral satellite imagery. 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)\/ : 27811--27819

    2024

  21. [29]

    Quan, Y., S. Dong, W. Feng, G. Dauphin, G. Zhao, Y. Wang, and M. Xing 2020. Spectral-spatial feature extraction based cnn for hyperspectral image classification. In IGARSS 2020 - 2020 IEEE International Geoscience and Remote Sensing Symposium , pp.\ 485--488

    2020

  22. [30]

    Radford, A., J.W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal, G. Sastry, A. Askell, P. Mishkin, J. Clark, G. Krueger, and I. Sutskever 2021, 18--24 Jul. Learning transferable visual models from natural language supervision. In M. Meila and T. Zhang (Eds.), Proceedings of t...

    2021

  23. [31]

    Narasimhan, T

    Radford, A., K. Narasimhan, T. Salimans, and I. Sutskever. 2018. Improving language understanding by generative pre-training. OpenAI\/

    2018

  24. [32]

    Mommert, and D

    Scheibenreif, L., M. Mommert, and D. Borth 2023. Masked vision transformers for hyperspectral image classification. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW) , pp.\ 2166--2176

    2023

  25. [33]

    Charfuelan, B

    Sumbul, G., M. Charfuelan, B. Demir, and V. Markl 2019, July. Bigearthnet: A large-scale benchmark archive for remote sensing image understanding. In IGARSS 2019 - 2019 IEEE International Geoscience and Remote Sensing Symposium . IEEE

    2019

  26. [34]

    Shazeer, N

    Vaswani, A., N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A.N. Gomez, L. Kaiser, and I. Polosukhin 2017. Attention is all you need. In Proceedings of the 31st International Conference on Neural Information Processing Systems , pp.\ 6000–6010

    2017

  27. [35]

    Wang, D., M. Hu, Y. Jin, Y. Miao, J. Yang, Y. Xu, X. Qin, J. Ma, L. Sun, C. Li, C. Fu, H. Chen, C. Han, N. Yokoya, J. Zhang, M. Xu, L. Liu, L. Zhang, C. Wu, B. Du, D. Tao, and L. Zhang. 2024. Hypersigma: Hyperspectral intelligence comprehension foundation model. arXiv:2406.115...

    2024 arXiv

  28. [36]

    Xin, Y., S. Luo, H. Zhou, J. Du, X. Liu, Y. Fan, Q. Li, and Y. Du. 2024. Parameter-efficient fine-tuning for pre-trained vision models: A survey. arXiv:2402.02242, cs.CV\/

    2024

  29. [37]

    Xiong, Z., Y. Wang, F. Zhang, and X.X. Zhu. 2024. One for all: Toward unified foundation models for earth vision. IGARSS 2024 - 2024 IEEE International Geoscience and Remote Sensing Symposium\/ : 2734--2738

    2024

  30. [38]

    Hsieh, Z

    YEH, S.Y., Y.G. Hsieh, Z. Gao, B.B.W. Yang, G. Oh, and Y. Gong 2024. Navigating text-to-image customization: From ly CORIS fine-tuning to model evaluation. In The Twelfth International Conference on Learning Representations

    2024

  31. [39]

    Zhou, and Z

    Zhang, Y., K. Zhou, and Z. Liu. 2024. Neural prompt search. IEEE Transactions on Pattern Analysis and Machine Intelligence\/ : 1--14. doi:10.1109/TPAMI.2024.3435939

    2024

  32. [40]

    Şakacı, S.A. and O. Urhan 2020. Spectral-spatial classification of hyperspectral imagery with convolutional neural network. In 2020 Innovations in Intelligent Systems and Applications Conference (ASYU) , pp.\ 1--4

    2020

  33. [41]

    write newline

    " write newline "" before.all 'output.state := FUNCTION output.doi doi empty skip "doi:" doi * "" * output if FUNCTION format.archive archivePrefix empty "" archivePrefix ":" * if FUNCTION format.primaryClass primaryClass empty "" " [" primaryClass * "] " * if FUNCTION format....

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.