Pith. sign in

REVIEW 4 major objections 4 minor 34 references

HyperKD: Distilling Cross-Spectral Knowledge in Masked Autoencoders via Inverse Domain Shift with Spatial-Aware Masking and Specialized Loss

T0 review · 4 major / 4 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read HyperKD transfers a six-band multispectral teacher's mid-layer features to a 218-band hyperspectral student, raising reconstruction quality and downstream task performance.

desk verdict Useful cross-spectral KD recipe for hyperspectral MAEs, but the experiments never isolate the distillation signal from the new loss and masking, so the headline reconstruction gains are under-supported. read the letter →

arxiv 2508.09453 v1 pith:FNIF24JD submitted 2025-08-13 cs.CV cs.LG

Abdul Matin , Tanjim Bin Faruk , Shrideep Pallickara , Sangmi Lee Pallickara This is my paper
classification cs.CVcs.LG
keywords knowledgedistillationmaskedautoencoderhyperspectralimagingvisiontransformerremotesensingfoundationmodelEnMAPGaborfilterspectraldomainadaptation
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

This paper claims that a pretrained multispectral foundation model can be turned into a useful teacher for a hyperspectral student, even though the student has far more spectral channels. HyperKD does this by averaging EnMAP bands into six synthetic bands that match the teacher's six HLS bands, distilling the teacher's layer-8 features into the student using KL divergence, and masking the most structurally salient patches as reconstruction targets. On EnMAP data (218 bands), the student's reconstruction improves from PSNR 24.61 / SSIM 0.55 without distillation to 31.02 / 0.77 with it, and the pretrained encoder also improves crop-type, land-cover, and soil-carbon downstream tasks. The authors argue this 'inverse' knowledge transfer, from a simpler teacher to a more complex student, can make geospatial foundation models usable for hyperspectral imagery despite sparse revisit times.

What carries the argument

The load-bearing object is the synthetic-band alignment of Eq. (4): each HLS band's spectral interval selects overlapping EnMAP bands, and those are averaged to produce a six-band input the teacher can consume. Around this sit the layer-8 feature distillation with a learned FC projection and KLD loss, the patch significance score $S_p = \|G_p\|$ or $\|W_p\|$ from Gabor or wavelet responses that selects which patches to mask, and the composite loss $L_{\text{total}} = \alpha(\lambda_1 L_{\text{MSE}} + \lambda_2 L_{\text{SSIM}}) + \beta L_{\text{KD}}$.

What would settle it

A decisive check: replace the spectral-range averaging in Eq. (4) with six randomly selected EnMAP bands, or with bands that deliberately avoid HLS overlap, and retrain HyperKD under the same budget. If PSNR stays near 31, the alignment is not what carries the gain; if it collapses toward the 24-27 range, the synthetic-band proxy is the active ingredient.

Watch

Extended reading notes

Core claim

HyperKD's central claim is that the spectral gap between a 6-band multispectral teacher and a 218-band hyperspectral student can be bridged by an inverse knowledge-distillation setup built on a masked autoencoder. The teacher is a pretrained ViT-MAE (Prithvi-100M); the student is a same-architecture MAE that accepts 218 EnMAP bands. The paper's recipe has three parts: (1) synthesize six teacher-compatible bands by averaging EnMAP bands whose wavelength ranges overlap each HLS band; (2) distill at transformer layer 8 by aligning features with a learned fully connected layer and Kullback-Leibler divergence; (3) compute per-patch significance with Gabor filters or wavelets and mask the most sig

Load-bearing premise

The entire transfer depends on the six averaged synthetic bands being a faithful stand-in for what the teacher expects; if averaging EnMAP bands over each HLS band loses or skews the signal, the teacher's features have little to teach.

Editorial extensions

If this is right

  • HyperKD raises average reconstruction PSNR from 24.61 to 31.02 and SSIM from 0.55 to 0.77 on the California test set compared with a student trained without distillation (Table I).
  • The gains transfer to unseen regions and extended geography: Test Dataset 2 and the CA+CO+KS set both show HyperKD ahead of baseline KD and the no-KD student (Tables I-II).
  • Pretraining with HyperKD improves downstream tasks: cultivated crop classification Top-1 accuracy rises from 90.80 to 92.91, crop-type mean Top-1 from 67.41 to 73.27, and soil organic carbon MAE falls from 0.045 to 0.038.
  • Layer-8 feature distillation with KLD outperforms L1 and JS losses, and Gabor-guided masking outperforms random masking.
  • The teacher's guidance is strongest in spectral bands that overlap HLS, with smaller but still positive gains elsewhere (Figure 4).

Reading between the lines

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

  • If the gains come mainly from masking hard patches rather than from spectral transfer, the same recipe could improve single-domain MAEs on any high-dimensional imagery without a teacher at all.
  • Because the teacher is frozen and only layer-8 features are used, HyperKD might work with any ViT-MAE teacher, so the method could be ported to newer geospatial foundation models as they appear.
  • The synthetic-band averaging is a crude proxy; a learned spectral mapping, such as a small 1x1 convolution, could replace Eq. (4) and likely improve reconstruction in bands that do not overlap HLS.
  • A direct testable extension is to apply HyperKD with different teacher-student sensor pairs (e.g., Sentinel-2 as teacher, a different hyperspectral sensor as student) to see whether improvements scale with the degree of spectral overlap.
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

4 major / 4 minor

Summary. The paper proposes HyperKD, an inverse knowledge-distillation framework that transfers representations from a multispectral masked-autoencoder teacher (Prithvi-100M) into a hyperspectral MAE student (EnMAP, 218 bands). The method has three components: spectral-range channel alignment that averages EnMAP bands into six synthetic bands (Eq. 4), Gabor/wavelet feature-guided masking, and a composite reconstruction loss combining MSE, SSIM, and a feature-distillation KLD term. The experimental section reports reconstruction PSNR/SSIM on two test datasets and downstream results for land cover classification, crop type identification, and soil organic carbon prediction. The main claim is that HyperKD significantly improves representation learning in MAEs relative to a student trained without KD and to a standard KD baseline.

Significance. If established, the result would be practically valuable: it would show a way to bootstrap hyperspectral MAE pretraining from a widely available multispectral foundation model despite sparse hyperspectral data. The paper addresses an important and timely problem, and the proposed pipeline is coherent in outline. The paper also explicitly identifies several nontrivial implementation details (tile construction, single-timestamp alignment, mid-layer distillation) that could help reproducibility. However, the central attribution of gains to the cross-spectral KD mechanism is not supported by the experiments as currently reported. The reconstruction comparison changes reconstruction loss, masking strategy, and KD loss simultaneously; the downstream improvements are partly negligible or unquantified; and the synthetic band alignment, a load-bearing assumption, is not validated. These issues prevent the paper from making its central claim credible, although they are addressable with additional experiments rather than indicating a fundamentally flawed method.

major comments (4)
  1. [Table I, Section V-C] The central claim that KD improves reconstruction is not supported by a controlled comparison. Row 1 (Student) uses HUBER loss and no KD; rows 4/5 (HyperKD) use MSE+SSIM loss, KLD KD, and Gabor/wavelet masking. These rows differ in reconstruction loss, masking, and KD simultaneously, so the PSNR/SSIM gains (24.61 to 31.02; 0.55 to 0.77) cannot be attributed to the cross-spectral teacher signal. Section V-C credits the specialized loss for the SSIM increase from 0.65 to 0.77, but this contrast compares BaseKD (HUBER, L1, random) with HyperKD (MSE+SSIM, KLD, Gabor), not a matched loss-only comparison. A controlled ablation is required, e.g., the same student with MSE+SSIM and Gabor masking with and without KLD, and with random masking and KLD.
  2. [Section V-B2 / Section V-B] The KD-function ablation does not adequately isolate the KD contribution. It compares L1, KLD, and JS within a fixed BaseKD setup, but there is no 'no KD' row under the same reconstruction loss and masking. Moreover, the section states 'All experiments were performed with a 75% masking ratio and MSE for reconstruction loss', which is inconsistent with Table I where rows 1-2 use HUBER and rows 3-5 use MSE+SSIM; thus the ablation numbers are not directly comparable with the headline results. The ablation also uses only 1000 train / 100 val / 300 test samples and 40 epochs, so transferability to the full setup is unclear.
  3. [Section IV-A3, Eq. (4)] The synthetic band construction is a load-bearing assumption and is not validated. The six synthetic bands are formed by averaging all EnMAP bands whose full wavelength range lies inside each HLS band (Eq. 3-4), but no analysis shows that these averages are a faithful proxy for the teacher's original multispectral input. The teacher is described in Section V-A1 as an 18-channel model (6 spectral bands x 3 timestamps), whereas the alignment produces only 6 channels; the paper does not explain how the temporal dimension is handled. A noisy or biased teacher input would weaken the teacher's layer-8 features and the learned KD signal, yet no experiment quantifies this.
  4. [Section VI, Tables IV-VII] Downstream claims are overstated. Land cover mean Top-1 is 75.42 vs 75.49 (+0.07 p) in Table IV; Table V reports only three classes and no mean; crop type mean improves 67.41 -> 73.27 but on a noisy, imbalanced dataset; SOC MAE changes from 0.045 to 0.038. No error bars, multiple seeds, or significance tests are provided, so 'significantly improves representation learning' is not supported for downstream tasks.
minor comments (4)
  1. [Table III] The JS divergence row reports 'SSIM: 27.08'; this is likely a typo for 0.64. Please correct.
  2. [Section IV-A3, Eq. (3)] The condition actually describes EnMAP bands fully contained in the HLS interval, not merely overlapping bands. Clarify whether partial overlaps are considered, and if not, justify the choice.
  3. [Section V-A1 vs IV-A] The teacher is described as an 18-channel model (6 spectral bands x 3 timestamps), while the alignment procedure produces only 6 synthetic bands. Clarify how the temporal dimension is passed to the teacher and whether the teacher is run on a single timestamp or on replicated/repeated channels.
  4. [Throughout] No random seed, number of repeated runs, or standard deviations are reported. For claims of 'significant' improvements, report at least three seeds and error bars, especially for the downstream tasks where gains are small.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: HyperKD's claims rest on an external pretrained teacher and independent reconstruction/downstream evaluation, not on self-referential fits.

full rationale

Walking the derivation chain: the teacher (Prithvi, an external pretrained MAE) is fixed; the student is a separate 218-band EnMAP MAE. The synthetic 6-band alignment (Eq. 4) is a preprocessing aggregation, not a quantity that reappears as the prediction target. The KD loss (KLD on layer-8 features) is an auxiliary training signal; the reported PSNR/SSIM are computed on the student's own 218-band reconstructions versus ground-truth EnMAP tiles, so the evaluation is not equivalent to the loss's fitted target. No fitted parameter is renamed a prediction: λ1/λ2/α/β are hyperparameters, and the masking schedule was selected on validation, which is standard optimization, not circularity. The paper contains no load-bearing self-citation chain and no uniqueness theorem imported from the authors' prior work; Prithvi is cited as an external model. The most substantial weakness is experimental, not circular: Table I changes reconstruction loss, masking, and KD simultaneously, and the ablations (Section V-B) only vary KD loss within a fixed BaseKD setup, so the isolated contribution of KD to the headline PSNR/SSIM gains is not cleanly established. That is an ablation/attribution limitation, not a derivation that reduces to its inputs. The unvalidated assumption that averaged EnMAP bands overlapping each HLS band are a faithful teacher-input proxy (Section IV-A3) is a soundness concern. Accordingly, no circular step meeting the quoted-evidence threshold is present; the correct finding is no significant circularity.

Assumptions & free parameters 6 free parameters · 3 assumptions · 0 invented entities

The central claims rest on several hand-chosen or fitted constants (loss weights, filter parameters, layer choice) and on domain assumptions about spectral alignment and feature transferability. None of these are validated against independent benchmarks or ablated cleanly.

free parameters (6)
  • alpha, beta (total loss weights in Eq. 12)
    Weights controlling reconstruction vs. KD loss; values not reported, presumably tuned on validation.
  • lambda1, lambda2 (reconstruction loss weights in Eq. 11)
    Weights for MSE and SSIM terms; values not reported.
  • Gabor filter parameters (lambda, theta, psi, sigma, gamma)
    Parameters for the spatial masking score; chosen by hand, not specified.
  • Wavelet type and scale
    DWT settings for the alternative masking score; not specified.
  • KD layer (8) and KD loss (KLD) = layer 8, KLD
    Selected via ablation on reduced subsets, not predetermined.
  • Start of random masking at epoch 100 = epoch 100
    Used for final Gabor/wavelet models (Table I rows 4,5); tuned detail.
assumptions (3)
  • domain assumption Averaging overlapping EnMAP bands yields a synthetic multispectral band that is semantically equivalent to the corresponding HLS band (Eq. 4).
    Invoked in Section IV-A3; no radiometric validation that the synthesized bands preserve the information Prithvi was trained on.
  • domain assumption Middle-layer (layer 8) features of the teacher are domain-invariant and transferable to the hyperspectral student via a learned linear projection.
    Invoked in Section IV-A4; only supported by an ablation on a 1000-sample subset.
  • domain assumption Patches with high Gabor/wavelet response are the most challenging to reconstruct, and masking them improves knowledge distillation.
    Invoked in Section IV-B; no evidence that these scores correspond to reconstruction difficulty beyond the reported results.

how reviews work

0 comments
Cite this review

Pith. "Pith review of HyperKD: Distilling Cross-Spectral Knowledge in Masked Autoencoders via Inverse Domain Shift with Spatial-Aware Masking and Specialized Loss." pith.science (2026). https://pith.science/paper/FNIF24JD

@misc{pith2026250809453,
  author       = {Pith},
  title        = {Pith review of: HyperKD: Distilling Cross-Spectral Knowledge in Masked Autoencoders via Inverse Domain Shift with Spatial-Aware Masking and Specialized Loss},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/FNIF24JD}},
  note         = {Machine review of arXiv:2508.09453}
}
read the original abstract

The proliferation of foundation models, pretrained on large-scale unlabeled datasets, has emerged as an effective approach in creating adaptable and reusable architectures that can be leveraged for various downstream tasks using satellite observations. However, their direct application to hyperspectral remote sensing remains challenging due to inherent spectral disparities and the scarcity of available observations. In this work, we present HyperKD, a novel knowledge distillation framework that enables transferring learned representations from a teacher model into a student model for effective development of a foundation model on hyperspectral images. Unlike typical knowledge distillation frameworks, which use a complex teacher to guide a simpler student, HyperKD enables an inverse form of knowledge transfer across different types of spectral data, guided by a simpler teacher model. Building upon a Masked Autoencoder, HyperKD distills knowledge from the Prithvi foundational model into a student tailored for EnMAP hyperspectral imagery. HyperKD addresses the inverse domain adaptation problem with spectral gaps by introducing a feature-based strategy that includes spectral range-based channel alignment, spatial feature-guided masking, and an enhanced loss function tailored for hyperspectral images. HyperKD bridges the substantial spectral domain gap, enabling the effective use of pretrained foundation models for geospatial applications. Extensive experiments show that HyperKD significantly improves representation learning in MAEs, leading to enhanced reconstruction fidelity and more robust performance on downstream tasks such as land cover classification, crop type identification, and soil organic carbon prediction, underpinning the potential of knowledge distillation frameworks in remote sensing analytics with hyperspectral imagery.

Figures

Figures reproduced from arXiv: 2508.09453 by the authors.

Figure 1
Figure 1. Architecture of HyperKD: A novel cross-domain knowledge distillation framework for HSI adapting in-middle layer insights from Prithvi’s multispectral [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Impact of Teacher-Student Layer Pairings on Feature Distillation [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Image Reconstruction Quality Comparison: HyperKD vs. Baseline KD vs. Student Model Without KD. HyperKD achieves superior performance [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Comparison of reconstruction accuracy at individual channel levels. [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: Downstream task implementation consisting of the frozen HyperKD [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

34 extracted references · 30 canonical work pages

  1. [1]

    The enmap imaging spectroscopy mission towards operations,

    T. Storch, H.-P. Honold, S. Chabrillat, M. Habermeyer, P. Tucker, M. Brell, A. Ohndorf, K. Wirth, M. Betz, M. Kuchler et al., “The enmap imaging spectroscopy mission towards operations,” Remote Sensing of Environment, vol. 294, p. 113632, 2023

    work page 2023

  2. [2]

    Prisma: The italian hyperspectral mission,

    R. Loizzo, R. Guarini, F. Longo, T. Scopa, R. Formaro, C. Facchinetti, and G. Varacalli, “Prisma: The italian hyperspectral mission,” inIGARSS 2018-2018 IEEE international geoscience and remote sensing sympo- sium. IEEE, 2018, pp. 175–178

    work page 2018

  3. [3]

    Planet imagery and data,

    Planet Labs PBC, “Planet imagery and data,” https://www.planet.com, 2023, accessed: May 16, 2025

    work page 2023

  4. [4]

    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

    work page 2024

  5. [5]

    Remoteclip: A vision language foundation model for remote sensing,

    F. Liu, D. Chen, Z. Guan, X. Zhou, J. Zhu, Q. Ye, L. Fu, and J. Zhou, “Remoteclip: A vision language foundation model for remote sensing,” IEEE Transactions on Geoscience and Remote Sensing , 2024

    work page 2024

  6. [6]

    Ringmo: A remote sensing foundation model with masked image modeling,

    X. Sun, P. Wang, W. Lu, Z. Zhu, X. Lu, Q. He, J. Li, X. Rong, Z. Yang, H. Chang et al. , “Ringmo: A remote sensing foundation model with masked image modeling,” IEEE Transactions on Geoscience and Remote Sensing, vol. 61, pp. 1–22, 2022

    work page 2022

  7. [7]

    Foundation models for generalist geospatial artificial intelligence,

    J. Jakubik, S. Roy, C. Phillips, P. Fraccaro, D. Godwin, B. Zadrozny, D. Szwarcman, C. Gomes, G. Nyirjesy, B. Edwards et al., “Foundation models for generalist geospatial artificial intelligence,” CoRR, 2023

    work page 2023

  8. [8]

    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

Show all 34 references
  1. [9]

    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,” 2023. [Online]. Available: https://arxiv.org/abs/2212.14532

    2023 arXiv

  2. [10]

    The harmonized landsat and sentinel-2 surface reflectance data set,

    M. Claverie, J. Ju, J. G. Masek, J. L. Dungan, E. F. Vermote, J.-C. Roger, S. V . Skakun, and C. Justice, “The harmonized landsat and sentinel-2 surface reflectance data set,” Remote sensing of environment , vol. 219, pp. 145–161, 2018

    2018

  3. [11]

    Knowledge distillation: A survey,

    J. Gou, B. Yu, S. J. Maybank, and D. Tao, “Knowledge distillation: A survey,” International Journal of Computer Vision , vol. 129, no. 6, pp. 1789–1819, 2021

    2021

  4. [12]

    Distilling the knowledge in a neural network,

    G. Hinton, O. Vinyals, and J. Dean, “Distilling the knowledge in a neural network,” 2015. [Online]. Available: https://arxiv.org/abs/1503.02531

    2015 arXiv

  5. [13]

    Knowledge distillation with feature self attention,

    S.-G. Park and D.-J. Kang, “Knowledge distillation with feature self attention,” IEEE Access, vol. 11, pp. 34 554–34 562, 2023

    2023

  6. [14]

    Efficient object detection in optical remote sensing imagery via attention-based feature distillation,

    P. Shamsolmoali, J. Chanussot, H. Zhou, and Y . Lu, “Efficient object detection in optical remote sensing imagery via attention-based feature distillation,” IEEE Transactions on Geoscience and Remote Sensing , 2023

    2023

  7. [15]

    Self-supervised learning with adaptive distillation for hyperspectral image classification,

    J. Yue, L. Fang, H. Rahmani, and P. Ghamisi, “Self-supervised learning with adaptive distillation for hyperspectral image classification,” IEEE Transactions on Geoscience and Remote Sensing , vol. 60, pp. 1–13, 2022

    2022

  8. [16]

    A novel knowledge distillation method for self-supervised hyperspectral image classification,

    Q. Chi, G. Lv, G. Zhao, and X. Dong, “A novel knowledge distillation method for self-supervised hyperspectral image classification,” Remote Sensing, vol. 14, no. 18, p. 4523, 2022

    2022

  9. [17]

    Siamohot: A lightweight dual siamese network for onboard hyperspectral object tracking via joint spatial-spectral knowledge distillation,

    C. Sun, X. Wang, Z. Liu, Y . Wan, L. Zhang, and Y . Zhong, “Siamohot: A lightweight dual siamese network for onboard hyperspectral object tracking via joint spatial-spectral knowledge distillation,” IEEE Trans- actions on Geoscience and Remote Sensing , 2023

    2023

  10. [18]

    Random shuffling data for hyperspectral image classification with siamese and knowledge distillation network,

    Z. Yang, Y . Cao, X. Zhou, J. Liu, T. Zhang, and J. Ji, “Random shuffling data for hyperspectral image classification with siamese and knowledge distillation network,” Remote Sensing, vol. 15, no. 16, p. 4078, 2023

    2023

  11. [19]

    Self-supervised learning with adaptive distillation for hyperspectral image classification,

    J. Yue, L. Fang, H. Rahmani, and P. Ghamisi, “Self-supervised learning with adaptive distillation for hyperspectral image classification,” IEEE Transactions on Geoscience and Remote Sensing , vol. 60, pp. 1–13, 2021

    2021

  12. [20]

    Multimodal online knowledge distillation framework for land use/cover classification using full or missing modalities,

    X. Liu, F. Jin, S. Wang, J. Rui, X. Zuo, X. Yang, and C. Cheng, “Multimodal online knowledge distillation framework for land use/cover classification using full or missing modalities,” IEEE Transactions on Geoscience and Remote Sensing , 2024

    2024

  13. [21]

    Strm-kd: Semantic topological relation matching knowledge distillation model for smart agriculture apple leaf disease recognition,

    D. Li, W. Zhang, and Y . Liu, “Strm-kd: Semantic topological relation matching knowledge distillation model for smart agriculture apple leaf disease recognition,” Expert Systems with Applications , vol. 263, p. 125824, 2025

    2025

  14. [22]

    A survey on computational spectral reconstruction methods from rgb to hyperspectral imaging,

    J. Zhang, R. Su, Q. Fu, W. Ren, F. Heide, and Y . Nie, “A survey on computational spectral reconstruction methods from rgb to hyperspectral imaging,” Scientific reports, vol. 12, no. 1, p. 11905, 2022

    2022

  15. [23]

    Lambertian reflectance,

    S. J. Koppal, “Lambertian reflectance,” in Computer vision: a reference guide. Springer, 2021, pp. 729–731

    2021

  16. [24]

    On information and sufficiency,

    S. Kullback and R. A. Leibler, “On information and sufficiency,” The annals of mathematical statistics , vol. 22, no. 1, pp. 79–86, 1951

    1951

  17. [25]

    Divergence measures based on the shannon entropy

    J. Lin, “Divergence measures based on the shannon entropy.” IEEE Trans. Inf. Theory , vol. 37, no. 1, pp. 145–151, 1991. [Online]. Available: http://dblp.uni-trier.de/db/journals/tit/tit37.html#Lin91

    1991

  18. [26]

    Theory of communication. part 1: the analysis of information. j inst electr eng-part iii: Radio commun eng 93 (26): 429–441,

    D. Gabor, “Theory of communication. part 1: the analysis of information. j inst electr eng-part iii: Radio commun eng 93 (26): 429–441,” 1946

    1946

  19. [27]

    Meyer, Wavelets and operators: volume 1

    Y . Meyer, Wavelets and operators: volume 1 . Cambridge university press, 1992, no. 37

    1992

  20. [28]

    Gabor filter-based edge detection,

    R. Mehrotra, K. R. Namuduri, and N. Ranganathan, “Gabor filter-based edge detection,” Pattern recognition, vol. 25, no. 12, pp. 1479–1494, 1992

    1992

  21. [29]

    Image quality assess- ment: from error visibility to structural similarity,

    Z. Wang, A. Bovik, H. Sheikh, and E. Simoncelli, “Image quality assess- ment: from error visibility to structural similarity,” IEEE Transactions on Image Processing , vol. 13, no. 4, pp. 600–612, 2004

    2004

  22. [30]

    A perceptually relevant mse-based image quality metric,

    H. L. Tan, Z. Li, Y . H. Tan, S. Rahardja, and C. Yeo, “A perceptually relevant mse-based image quality metric,” IEEE Transactions on Image Processing, vol. 22, no. 11, pp. 4447–4459, 2013

    2013

  23. [31]

    Structural texture similarity metrics for image analysis and retrieval,

    J. Zujovic, T. N. Pappas, and D. L. Neuhoff, “Structural texture similarity metrics for image analysis and retrieval,” IEEE Transactions on Image Processing, vol. 22, no. 7, pp. 2545–2558, 2013

    2013

  24. [32]

    Annual nlcd (national land cover database)—the next generation of land cover mapping,

    U. G. Survey, “Annual nlcd (national land cover database)—the next generation of land cover mapping,” 2025

    2025

  25. [33]

    Examining the characteristics of the cropland data layer in the context of estimating land cover change,

    K. Copenhaver, Y . Hamada, S. Mueller, and J. B. Dunn, “Examining the characteristics of the cropland data layer in the context of estimating land cover change,” ISPRS international journal of geo-information , vol. 10, no. 5, p. 281, 2021

    2021

  26. [34]

    Gridded national soil survey geographic (gnatsgo) database for the conterminous united states,

    S. S. Staff, “Gridded national soil survey geographic (gnatsgo) database for the conterminous united states,” 2019

    2019

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.