Pith. sign in

REVIEW 5 major objections 5 minor 41 references

Knowledge-guided Complex Diffusion Model for PolSAR Image Classification in Contourlet Domain

T0 review · 5 major / 5 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 A diffusion model that runs on complex-valued PolSAR data in the contourlet domain reports the highest classification accuracy and cleanest edges on three radar datasets.

desk verdict A plausible architecture-level contribution to PolSAR classification whose SOTA claim is undermined by an under-described, internally inconsistent experimental report. read the letter →

arxiv 2507.05666 v1 pith:6SVBLSBX submitted 2025-07-08 cs.CV eess.IV

Junfei Shi , Yu Cheng , Haiyan Jin , Junhuai Li , Zhaolin Xiao , Maoguo Gong , Weisi Lin This is my paper · ORCID
classification cs.CVeess.IV
keywords PolSARimageclassificationcomplex-valueddiffusionmodelcontourlettransformknowledge-guidedcomplexU-Netedgepreservationlandcoverremotesensing
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 tries to establish that a diffusion model can be made to work directly on complex-valued PolSAR data, without discarding phase information, and that it can classify land cover more accurately than current deep-learning baselines while keeping sharp edges. The proposed model, CD-KCDM (called HL-CDM in the experiments), first decomposes the PolSAR image with a complex contourlet transform into low-frequency and high-frequency subbands. A complex U-Net diffusion process learns statistical texture from the low-frequency part, while multiscale, multidirectional high-frequency coefficients are injected as structural knowledge through cross-attention, and an enhancement module fuses both for classification. On the Xi'an, San Francisco, and Flevoland datasets, the method reports the highest overall accuracy, average accuracy, and Kappa among the compared methods, with visibly cleaner region boundaries and more homogeneous regions.

What carries the argument

The load-bearing object is the complex-valued non-subsampled contourlet transform, which decomposes the PolSAR coherency data into one low-pass subband and high-pass subbands at 3 scales and 8 directions. The low-frequency stream is processed by the knowledge-guided complex diffusion model (KCDM): a complex U-Net denoiser trained with complex Gaussian noise, with high-frequency structural knowledge injected at each level through a complex cross-attention module. A separate high-frequency detail enhancement module (CAFE) uses cross-attention between directional subbands and dilation-convolution multi-scale fusion, and the two feature streams are concatenated and classified by a CNN. This mechanism does the work of separating statistical texture from geometric structure so each can be modeled by the appropriate tool.

What would settle it

A replication that retrains SARDDPM and SpectralDiff using their original task-specific input representations plus the same label budget, and reports both OA/Kappa and an edge-pixel F1 score, would settle the claim. The edge-preservation claim is falsified if boundary accuracy is no better than the diffusion baselines; the accuracy claim is falsified if the OA gap over properly tuned baselines shrinks to within run-to-run noise.

Watch

Extended reading notes

Core claim

The central claim is that splitting PolSAR data in the complex contourlet domain and letting a diffusion model handle only the low-frequency component, with high-frequency structure supplied as a cross-attention guide, solves the two problems real-valued diffusion models have on PolSAR data: it preserves phase information because the diffusion process runs on complex Gaussian noise, and it preserves edges because structural detail is provided directly rather than recovered from smoothed statistics. The paper reports that this design outperforms all five baseline approaches on all three datasets, with overall accuracies of 96.55% on Xi'an, 99.48% on San Francisco, and 99.45% on Flevoland, and attributes the gains to joint low-frequency statistical modeling and high-frequency edge guidance.

Load-bearing premise

The comparison assumes all five baselines were adapted fairly to PolSAR classification and trained under a protocol (5% labeled pixels, 16 diffusion steps, 1000 iterations) that does not disadvantage them, especially SARDDPM and SpectralDiff, which were originally designed for despeckling and hyperspectral classification.

Editorial extensions

If this is right

  • On the three tested datasets, the method sets the best reported OA, AA, and Kappa among the compared methods, improving overall accuracy by 0.34 to 6.96 percentage points over the baselines depending on the scene.
  • Because the diffusion process is complex-valued, the method can be applied directly to coherency or covariance matrices without discarding phase, unlike real-valued diffusion baselines.
  • The dual-stream design means statistical texture and structural edges are learned separately, so classification maps should stay clean inside regions and along boundaries even in heterogeneous terrain.
  • The same 3-level, 8-direction contourlet decomposition works across C-band data (Xi'an, Flevoland) and L-band data (San Francisco), suggesting the framework is not tied to one sensor or frequency band.

Reading between the lines

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

  • Beyond the paper: the same low/high-frequency split with a complex diffusion low-frequency stream could be tested on other complex-valued imaging modalities, such as InSAR phase filtering or complex MRI reconstruction, where phase and edge structure matter in a similar way.
  • Beyond the paper: knowledge-guided cross-attention is a general mechanism for conditioning a diffusion model on structural side information, so it could be ablated on natural-image tasks where a high-frequency prior is available.
  • Beyond the paper: a direct test of the edge-preservation claim would quantify boundary accuracy, for example an F1 score on edge pixels, because the reported OA, AA, and Kappa numbers are pixel-wise accuracy metrics that do not by themselves measure edge quality.
  • Beyond the paper: since the comparison used 16 diffusion steps and 1000 training iterations, an interesting follow-up is whether the accuracy gap persists with substantially longer training; if it narrows, part of the contribution may be training efficiency rather than architecture alone.
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

5 major / 5 minor

Summary. The paper proposes CD-KCDM (called HL-CDM in the experiments), a complex-valued diffusion model in the contourlet domain for PolSAR image classification. The method first applies a complex non-subsampled contourlet transform to separate PolSAR data into low- and high-frequency subbands, then trains a knowledge-guided complex U-Net on the low-frequency component while using high-frequency structural features as a conditioning signal via cross-attention. A separate high-frequency enhancement module refines multiscale and multidirectional features, and the two streams are fused for classification. Experiments on three PolSAR datasets report OA/AA/Kappa improvements over five baselines, and ablations on the Xi'an dataset support the individual module contributions.

Significance. If the empirical claims were substantiated, the paper would offer a novel and plausible combination of complex-valued diffusion, contourlet-domain decomposition, and knowledge-guided cross-attention for PolSAR classification, with a credible emphasis on edge preservation and region homogeneity. The motivation is clear and the architecture is coherent. However, the current evidence does not establish state-of-the-art performance: the adaptation of the diffusion baselines is not described, several reported numbers are internally inconsistent, no uncertainty quantification is provided, and the data and code are withheld. These issues are fixable in a revision but are load-bearing for the paper's central claim.

major comments (5)
  1. [IV-B; Tables I-III] The central SOTA claim rests on a comparison in which two diffusion baselines are adapted from other tasks without any description of the conversion. SARDDPM is originally a despeckling model and SpectralDiff is originally a hyperspectral classifier, yet the paper only states in Section IV-B that all methods are trained under the same settings (5% labeled pixels, 16 diffusion steps, 1000 iterations, patch size 16). No adaptation layers, input representations, or stopping criteria are described, and 1000 iterations is very short for diffusion-based training. Because the reported margin over SpectralDiff on Xi'an is only 0.34% OA, undertrained or misadapted baselines could erase the claimed advantage. Please describe the exact adaptation of each baseline and provide evidence, such as training curves or longer training runs, that the baselines are not artificially weak.
  2. [IV-C; Tables I-III] The numerical report contains impossible and swapped values. In Table I, CVCNN is reported with Kappa=93.02 and OA=89.59, which is mathematically impossible because Kappa cannot exceed OA. In addition, the improvement list in the San Francisco subsection (2.88, 0.96, 1.10, 20.21, 2.86) matches the Flevoland table, while the Flevoland subsection list (5.67, 0.55, 0.95, 3.18, 1.99) matches the San Francisco table. These errors make the reported accuracy differences unreliable and must be corrected before the SOTA claim can be assessed.
  3. [IV-B; IV-D; Conclusion] No variance, error bars, or statistical significance tests are reported for any of the accuracy numbers. Since training uses a random 5% labeled subset and the diffusion component is stochastic, single-run differences such as the 0.34% OA advantage on Xi'an cannot be distinguished from noise. The Conclusion's statement that the method provides 'statistically higher accuracy' is therefore unsupported. Please report means and standard deviations over multiple runs, or at least paired significance tests, for all tables.
  4. [III-B; Eqs. (11)-(16)] The diffusion-formula presentation contains inconsistencies that affect reproducibility. Eq. (12) and Eq. (13) use α_t where Eq. (11) has the accumulated product \bar α_t for the marginal noising distribution, and Eq. (16) gives σ_t^2 = 1 − α_{t−1}, which is not the standard DDPM posterior variance and is algebraically wrong as written. The surrounding text also calls μ and σ the variance and mean, whereas μ is the mean and σ is the standard deviation. Please correct these equations and specify the exact reverse-sampling procedure used at inference.
  5. [IV-A; IV-C] The manuscript states that dataset details and implementation code will be released only in the final version 'due to anonymity requirements.' For a paper whose main claim is empirical SOTA performance, withholding the dataset specifications, ground-truth generation, train/test splits, and baseline adaptation code prevents verification and makes the numerical claims non-reproducible. Please provide the necessary details in the manuscript or as a public supplement.
minor comments (5)
  1. [Title/Abstract vs. IV-C] The proposed method is named CD-KCDM in the title, abstract, and Section III, but is called HL-CDM throughout Section IV; please unify the name.
  2. [IV-C-2] The San Francisco discussion mentions misclassification patches 'particularly in Urban and Cropland regions,' but the San Francisco dataset has no Cropland class; this appears to be copied from the Flevoland section.
  3. [III-B; Eq. (10)] Eq. (10) has a notation typo: 'jIm(L)t+1' should be j Im(L_{t+1}), and the sentence calling μ and σ the variance and mean should be corrected.
  4. [Algorithm 1] Step 4 of Algorithm 1 introduces PCA-based dimensionality reduction, which is not mentioned in Section III-D or Eq. (23); please clarify whether PCA is part of the method or an omission.
  5. [IV-B] The phrase 'select layer 1 for low-frequency feature extraction' is ambiguous: it should be stated which diffusion time-step or U-Net layer is used to produce the classification feature F1 and whether this selection is re-tuned per dataset.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the classification results are empirical benchmarks against external datasets and baselines, not derivations reducible to fitted inputs or self-citations.

full rationale

The paper does not present a mathematical derivation whose output is encoded in its inputs; it proposes an architecture (CD-KCDM, renamed HL-CDM in the experiments) and evaluates it on three real PolSAR datasets. The core equations (Eqs. 8-26) define a complex-valued forward and reverse diffusion process, a cross-attention fusion mechanism, and a softmax classifier. None of these equations contains the reported OA, AA, or Kappa values, and those numbers are obtained by training on 5% labeled pixels and evaluating on the remaining data, so they are not fixed by construction. The 'knowledge' guidance is high-frequency contourlet information injected into the low-frequency diffusion stream; this is a feature-fusion design, not a fitted parameter renamed as a prediction. The only author-overlapping citation is reference [40] (SGCN-CNN), which is used as a baseline method in Tables I-III; it is not invoked to justify the proposed modules, to supply a uniqueness theorem, or to rule out alternative designs, so it is not load-bearing. The ablation study and hyperparameter choices are standard model selection on the Xi'an dataset. Concerns about whether SARDDPM and SpectralDiff are fairly adapted from their original tasks and adequately trained under the shared 1000-iteration protocol are legitimate experimental-fairness questions, but they are not circularity: a weak baseline does not make the proposed method's derivation equivalent to its input. Overall, no step reduces to its own input by construction, and no central claim is forced by a self-citation chain. Score 0.

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

The central empirical claim rests on a small number of premises: that the contourlet split separates texture from structure in PolSAR data, that complex Gaussian diffusion correctly models the low-frequency coefficients, and that the comparison protocol treats baselines fairly. The paper provides no code, data, or error bars, so these premises cannot be checked from the preprint alone.

free parameters (7)
  • Diffusion reverse steps T = 16
    Set in Section IV-B; controls denoising fidelity and is not varied in any sensitivity analysis.
  • Contourlet decomposition levels L = 3
    Chosen from ablation on Xi'an (Section IV-D.2) and applied to all datasets, so it is effectively fitted to one benchmark.
  • Contourlet directional subbands per level = 3 (8 directions)
    Chosen manually in Section III-A; no sensitivity study is provided.
  • U-Net feature layer = 1
    Selected from ablation on Xi'an (Section IV-D.3); deeper layers lower accuracy, so the choice affects the reported results.
  • Training sample fraction = 5% train / 1% validation
    Set in Section IV-B; the reported state-of-the-art numbers depend on this label budget.
  • Training iterations = 1000
    Set in Section IV-B; may be too short for diffusion baselines, directly affecting the comparison.
  • Noise schedule alpha_t = not specified
    Required for equations (9) and (22) but the schedule values are never given, reducing reproducibility.
assumptions (4)
  • standard math Complex Gaussian reparameterization: the forward noising process of the low-frequency coefficients has the closed form Lt = sqrt(alpha_bar_t) L0 + sqrt(1 - alpha_bar_t) epsilon with i.i.d. real and imaginary parts.
    Used in equations (9)-(14) to justify the diffusion loss in equation (22). This is standard diffusion-model math.
  • domain assumption Low-frequency contourlet coefficients primarily contain statistical/texture information, while high-frequency coefficients contain edge structure.
    Stated in Sections I and III-B; the entire knowledge-guidance design depends on this separation being true for PolSAR data.
  • ad hoc to paper A 3-level, 8-direction non-subsampled complex contourlet transform yields a sufficient multi-resolution representation for all three datasets.
    Level and direction settings are chosen by hand and by ablation on Xi'an (Section III-A and IV-D.2), with no independent justification.
  • ad hoc to paper All baseline methods can be fairly compared under the same training schedule (5% labels, 16 diffusion steps, 1000 iterations).
    Section IV-B states the shared protocol; no evidence shows that this schedule is suitable for SARDDPM or SpectralDiff, so the fairness of the comparison is an unverified premise.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Knowledge-guided Complex Diffusion Model for PolSAR Image Classification in Contourlet Domain." pith.science (2026). https://pith.science/paper/6SVBLSBX

@misc{pith2026250705666,
  author       = {Pith},
  title        = {Pith review of: Knowledge-guided Complex Diffusion Model for PolSAR Image Classification in Contourlet Domain},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6SVBLSBX}},
  note         = {Machine review of arXiv:2507.05666}
}
read the original abstract

Diffusion models have demonstrated exceptional performance across various domains due to their ability to model and generate complicated data distributions. However, when applied to PolSAR data, traditional real-valued diffusion models face challenges in capturing complex-valued phase information.Moreover, these models often struggle to preserve fine structural details. To address these limitations, we leverage the Contourlet transform, which provides rich multiscale and multidirectional representations well-suited for PolSAR imagery. We propose a structural knowledge-guided complex diffusion model for PolSAR image classification in the Contourlet domain. Specifically, the complex Contourlet transform is first applied to decompose the data into low- and high-frequency subbands, enabling the extraction of statistical and boundary features. A knowledge-guided complex diffusion network is then designed to model the statistical properties of the low-frequency components. During the process, structural information from high-frequency coefficients is utilized to guide the diffusion process, improving edge preservation. Furthermore, multiscale and multidirectional high-frequency features are jointly learned to further boost classification accuracy. Experimental results on three real-world PolSAR datasets demonstrate that our approach surpasses state-of-the-art methods, particularly in preserving edge details and maintaining region homogeneity in complex terrain.

Figures

Figures reproduced from arXiv: 2507.05666 by the authors.

Figure 1
Figure 1. Framework of the proposed Knowledge-guided complex contourlet diffusion model. [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Non-subsampled contourlet decomposition information with multi-scale, multi-directional structural de￾tails. This geometric knowledge helps to enhance edge details and guides the diffusion model in learning refined features at the edges. Consequently, two modules are designed in this subsection: the structural knowledge extraction module and the knowledge-guided complex diffusion model. 1) Structural knowledge extra… view at source ↗
Figure 3
Figure 3. Network structure of the Knowledge-guided complex U-Net. [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Qualitative results comparison on the Xi’an dataset. [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: Qualitative results comparison on the San Francisco dataset. [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: Qualitative results comparison on the Flevoland dataset. [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]
Figure 7
Figure 7. Figure 7: Performance variations with different NSCT decomposition levels on [PITH_FULL_IMAGE:figures/full_fig_p009_7.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

41 extracted references · 34 canonical work pages

  1. [1]

    Environment monitoring of shanghai nanhui intertidal zone with dual-polarimetric sar data based on deep learning,

    G. Liu, B. Liu, G. Zheng, and X. Li, “Environment monitoring of shanghai nanhui intertidal zone with dual-polarimetric sar data based on deep learning,” IEEE Transactions on Geoscience and Remote Sensing , vol. 60, pp. 1–18, 2022

    work page 2022

  2. [2]

    Unsupervised terrain classification preserving polarimetric scattering characteristics,

    J.-S. Lee, M. R. Grunes, E. Pottier, and L. Ferro-Famil, “Unsupervised terrain classification preserving polarimetric scattering characteristics,” IEEE Transactions on Geoscience and Remote Sensing , vol. 42, no. 4, pp. 722–731, 2004

    work page 2004

  3. [3]

    Spam-net: A cnn-based sar target recognition network with pose angle marginalization learning,

    J. Oh, G.-Y . Youm, and M. Kim, “Spam-net: A cnn-based sar target recognition network with pose angle marginalization learning,” IEEE Transactions on Circuits and Systems for Video Technology , vol. 31, no. 2, pp. 701–714, 2020

    work page 2020

  4. [4]

    Observations of cutting practices in agricultural grasslands using polarimetric sar,

    K. V oormansik, T. Jagdhuber, K. Zalite, M. Noorma, and I. Hajnsek, “Observations of cutting practices in agricultural grasslands using polarimetric sar,” IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing , vol. 9, no. 4, pp. 1382–1396, 2015

    work page 2015

  5. [5]

    An improved cloude-pottier decompo- sition using h/α/span and complex wishart classifier for polarimetric sar classification,

    C. Fang, H. Wen, and W. Yirong, “An improved cloude-pottier decompo- sition using h/α/span and complex wishart classifier for polarimetric sar classification,” in 2006 CIE international conference on radar . IEEE, Conference Proceedings, pp. 1–4

    work page 2006

  6. [6]

    Feature motivated polarization scattering matrix decomposition,

    W. L. Cameron and L. K. Leung, “Feature motivated polarization scattering matrix decomposition,” in IEEE International Conference on Radar. IEEE, Conference Proceedings, pp. 549–557

    work page

  7. [7]

    A reflection symmetry approximation of multilook polarimetric sar data and its application to freeman–durden decomposi- tion,

    W. An and M. Lin, “A reflection symmetry approximation of multilook polarimetric sar data and its application to freeman–durden decomposi- tion,” IEEE Transactions on Geoscience and Remote Sensing , vol. 57, no. 6, pp. 3649–3660, 2019

    work page 2019

  8. [8]

    Four- component scattering power decomposition with rotation of coherency matrix,

    Y . Yamaguchi, A. Sato, W.-M. Boerner, R. Sato, and H. Yamada, “Four- component scattering power decomposition with rotation of coherency matrix,” IEEE Transactions on geoscience and remote sensing , vol. 49, no. 6, pp. 2251–2258, 2011

    work page 2011

Show all 41 references
  1. [9]

    A local competition-based wishart classifier,

    J. Peng, J. Yin, and J. Yang, “A local competition-based wishart classifier,” in 2021 SAR in Big Data Era (BIGSARDATA) . IEEE, Conference Proceedings, pp. 1–4

    2021

  2. [10]

    Change detection in urban areas with high resolution sar images using second kind statistics based g0 distribution,

    Z. Shan, C. Wang, H. Zhang, and F. Wu, “Change detection in urban areas with high resolution sar images using second kind statistics based g0 distribution,” in 2010 IEEE International Geoscience and Remote Sensing Symposium. Ieee, Conference Proceedings, pp. 4600–4603

    2010

  3. [11]

    Segmentation of textured polarimetric sar scenes by likelihood approximation,

    J.-M. Beaulieu and R. Touzi, “Segmentation of textured polarimetric sar scenes by likelihood approximation,” IEEE Transactions on Geoscience and Remote Sensing , vol. 42, no. 10, pp. 2063–2072, 2004

    2004

  4. [12]

    Support vector machine for multifrequency sar polarimetric data classification,

    C. Lardeux, P.-L. Frison, C. Tison, J.-C. Souyris, B. Stoll, B. Fruneau, and J.-P. Rudant, “Support vector machine for multifrequency sar polarimetric data classification,” IEEE Transactions on Geoscience and Remote Sensing, vol. 47, no. 12, pp. 4143–4152, 2009

    2009

  5. [13]

    Fuzzy classification of earth terrain covers using complex polarimetric sar data,

    L. Du and J. Lee, “Fuzzy classification of earth terrain covers using complex polarimetric sar data,” International Journal of Remote Sensing, vol. 17, no. 4, pp. 809–826, 1996

    1996

  6. [14]

    Polsar image classification based on statistical distribution and mrf,

    J. Yin, X. Liu, J. Yang, C.-Y . Chu, and Y .-L. Chang, “Polsar image classification based on statistical distribution and mrf,” Remote Sensing, vol. 12, no. 6, p. 1027, 2020

    2020

  7. [15]

    Polarimetric multipath convolutional neural network for polsar image classification,

    Y . Cui, F. Liu, L. Jiao, Y . Guo, X. Liang, L. Li, S. Yang, and X. Qian, “Polarimetric multipath convolutional neural network for polsar image classification,” IEEE Transactions on Geoscience and Remote Sensing , vol. 60, pp. 1–18, 2021

    2021

  8. [16]

    Spatial feature-based convolutional neural network for polsar image classification,

    R. Shang, J. Wang, L. Jiao, X. Yang, and Y . Li, “Spatial feature-based convolutional neural network for polsar image classification,” Applied Soft Computing, vol. 123, p. 108922, 2022

    2022

  9. [17]

    Mp-resnet: Multipath residual network for the semantic segmentation of high-resolution polsar images,

    L. Ding, K. Zheng, D. Lin, Y . Chen, B. Liu, J. Li, and L. Bruzzone, “Mp-resnet: Multipath residual network for the semantic segmentation of high-resolution polsar images,” IEEE Geoscience and Remote Sensing Letters, vol. 19, pp. 1–5, 2021

    2021

  10. [18]

    C-cnn: Contourlet convolutional neural networks,

    M. Liu, L. Jiao, X. Liu, L. Li, F. Liu, and S. Yang, “C-cnn: Contourlet convolutional neural networks,” IEEE Transactions on Neural Networks and Learning Systems , vol. 32, no. 6, pp. 2636–2649, 2020

    2020

  11. [19]

    Tcspanet: Two-staged contrastive learning and sub-patch attention based network for polsar image classification,

    Y . Cui, F. Liu, X. Liu, L. Li, and X. Qian, “Tcspanet: Two-staged contrastive learning and sub-patch attention based network for polsar image classification,” Remote Sensing, vol. 14, no. 10, p. 2451, 2022

    2022

  12. [20]

    Polsar image classification using a hybrid complex- valued network (hybridcvnet),

    M. Q. Alkhatib, “Polsar image classification using a hybrid complex- valued network (hybridcvnet),” IEEE Geoscience and Remote Sensing Letters, 2024

    2024

  13. [21]

    Complex-valued convo- lutional neural network and its application in polarimetric sar image classification,

    Z. Zhang, H. Wang, F. Xu, and Y .-Q. Jin, “Complex-valued convo- lutional neural network and its application in polarimetric sar image classification,” IEEE Transactions on Geoscience and Remote Sensing , vol. 55, no. 12, pp. 7177–7188, 2017

    2017

  14. [22]

    Polsar image classification using attention based shallow to deep convolutional neural network,

    M. Q. Alkhatib, M. Al-Saad, N. Aburaed, M. S. Zitouni, and H. Al- Ahmad, “Polsar image classification using attention based shallow to deep convolutional neural network,” in IGARSS 2023-2023 IEEE International Geoscience and Remote Sensing Symposium . IEEE, Conference Proceedi...

    2023

  15. [23]

    Improved denoising diffusion probabilis- tic models,

    A. Q. Nichol and P. Dhariwal, “Improved denoising diffusion probabilis- tic models,” in International conference on machine learning . PMLR, Conference Proceedings, pp. 8162–8171

  16. [24]

    Diffusion models beat gans on image synthesis,

    P. Dhariwal and A. Nichol, “Diffusion models beat gans on image synthesis,” Advances in neural information processing systems , vol. 34, pp. 8780–8794, 2021

    2021

  17. [25]

    Photorealistic text-to-image diffusion models with deep language understanding,

    C. Saharia, W. Chan, S. Saxena, L. Li, J. Whang, E. L. Denton, K. Ghasemipour, R. Gontijo Lopes, B. Karagol Ayan, and T. Sali- mans, “Photorealistic text-to-image diffusion models with deep language understanding,” Advances in neural information processing systems , vol. 35, p...

    2022

  18. [26]

    Auffusion: Leveraging the power of diffusion and large language models for text-to-audio generation,

    J. Xue, Y . Deng, Y . Gao, and Y . Li, “Auffusion: Leveraging the power of diffusion and large language models for text-to-audio generation,” IEEE/ACM Transactions on Audio, Speech, and Language Processing , 2024

    2024

  19. [27]

    Diffusion-based time series imputa- tion and forecasting with structured state space models,

    J. M. L. Alcaraz and N. Strodthoff, “Diffusion-based time series imputa- tion and forecasting with structured state space models,” arXiv preprint arXiv:2208.09399, 2022

    2022 arXiv

  20. [28]

    Autoregressive denoising diffusion models for multivariate probabilistic time series forecasting,

    K. Rasul, C. Seward, I. Schuster, and R. V ollgraf, “Autoregressive denoising diffusion models for multivariate probabilistic time series forecasting,” in International conference on machine learning . PMLR, Conference Proceedings, pp. 8857–8868

  21. [29]

    Dif-fusion: Toward high color fidelity in infrared and visible image fusion with diffusion models,

    J. Yue, L. Fang, S. Xia, Y . Deng, and J. Ma, “Dif-fusion: Toward high color fidelity in infrared and visible image fusion with diffusion models,” IEEE Transactions on Image Processing, vol. 32, pp. 5705–5720, 2023

    2023

  22. [30]

    Score-based generative classifiers,

    R. S. Zimmermann, L. Schott, Y . Song, B. A. Dunn, and D. A. Klindt, “Score-based generative classifiers,” arXiv preprint arXiv:2110.00473 , 2021

    2021 arXiv

  23. [31]

    Image super-resolution via iterative refinement,

    C. Saharia, J. Ho, W. Chan, T. Salimans, D. J. Fleet, and M. Norouzi, “Image super-resolution via iterative refinement,” IEEE transactions on pattern analysis and machine intelligence, vol. 45, no. 4, pp. 4713–4726, 2022

    2022

  24. [32]

    Come-closer-diffuse-faster: Accelerat- ing conditional diffusion models for inverse problems through stochastic contraction,

    H. Chung, B. Sim, and J. C. Ye, “Come-closer-diffuse-faster: Accelerat- ing conditional diffusion models for inverse problems through stochastic contraction,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , Conference Proceedings, pp. 12 ...

  25. [33]

    Score-based generative neural net- works for large-scale optimal transport,

    M. Daniels, T. Maunu, and P. Hand, “Score-based generative neural net- works for large-scale optimal transport,” Advances in neural information processing systems, vol. 34, pp. 12 955–12 965, 2021

    2021

  26. [34]

    Exploring multi-timestep multi-stage diffusion features for hyperspectral image classification,

    J. Zhou, J. Sheng, P. Ye, J. Fan, T. He, B. Wang, and T. Chen, “Exploring multi-timestep multi-stage diffusion features for hyperspectral image classification,” IEEE Transactions on Geoscience and Remote Sensing , 2024

    2024

  27. [35]

    Spectraldiff: A generative framework for hyperspectral image classification with diffusion models,

    N. Chen, J. Yue, L. Fang, and S. Xia, “Spectraldiff: A generative framework for hyperspectral image classification with diffusion models,” IEEE Transactions on Geoscience and Remote Sensing , vol. 61, pp. 1– 16, 2023

    2023

  28. [36]

    c2n2: Complex-valued contourlet neural network,

    M. Liu, L. Jiao, X. Liu, L. Li, F. Liu, S. Yang, Y . Guo, and P. Chen, “c2n2: Complex-valued contourlet neural network,” IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing , vol. 17, pp. 4478–4491, 2024

    2024

  29. [37]

    Knowledge- aware geometric contourlet semantic learning for hyperspectral image classification,

    X. Geng, L. Li, L. Jiao, X. Liu, F. Liu, and S. Yang, “Knowledge- aware geometric contourlet semantic learning for hyperspectral image classification,” IEEE Transactions on Circuits and Systems for Video Technology, 2024

    2024

  30. [38]

    Moving targets detection and analysis on multi-look polarimetric sar images using pwf method,

    B. Zou, T. Wei, and L. Zhang, “Moving targets detection and analysis on multi-look polarimetric sar images using pwf method,” in IGARSS 2008- 2008 IEEE International Geoscience and Remote Sensing Symposium , vol. 3. IEEE, Conference Proceedings, pp. III–1190–III–1193

    2008

  31. [39]

    Polarimetric multipath convolutional neural network for polsar image JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, MAY 2024 11 classification,

    Y . Cui, F. Liu, L. Jiao, Y . Guo, X. Liang, L. Li, S. Yang, and X. Qian, “Polarimetric multipath convolutional neural network for polsar image JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, MAY 2024 11 classification,” IEEE Transactions on Geoscience and Remote Sensing , vol. ...

    2024

  32. [40]

    Combine superpixel-wise gcn and pixel-wise cnn for polsar image classification,

    H. Jin, T. He, J. Shi, and S. Ji, “Combine superpixel-wise gcn and pixel-wise cnn for polsar image classification,” in IGARSS 2023-2023 IEEE international geoscience and remote sensing symposium . IEEE, Conference Proceedings, pp. 8014–8017

    2023

  33. [41]

    Sar despeckling using a denoising diffusion probabilistic model,

    M. V . Perera, N. G. Nair, W. G. C. Bandara, and V . M. Patel, “Sar despeckling using a denoising diffusion probabilistic model,” IEEE Geoscience and Remote Sensing Letters , vol. 20, pp. 1–5, 2023

    2023

Pith tools

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